Latent Space: The AI Engineer Podcast

From helping pioneer core ideas in NLP to now building AI systems that can automate AI research itself, Richard Socher is betting that the next major step in AI is recursive self-improvement. He is the founder of You.com, AIX Ventures, and now Recursive, which has assembled some of the best open-endedness (& self improving agent) researchers in the world and raised a $4.65B seed round.
In this episode, Richard joins Latent Space to unpack his vision for the “Eureka Machine”: a superintelligence that can improve the process of invention itself, accelerate AI research, and eventually tackle major problems across science, energy, materials, biology, and more.
You can get his book “The Eureka Machine” here!
We go deep on Recursive’s early results, including an AI research system that Richard says outperformed humans and their agents on optimization tasks in less than two days, as well as work on NVIDIA GPU kernels where the system discovered improvements without relying on a team of CUDA experts. Richard also explains why he thinks AI research that currently takes thousands of people and years could eventually be compressed into weeks. These results are summarized in his 20 minute AIE keynote, where we also discuss his 10 dimensions of intelligence:

We also explore the harder questions around increasingly capable AI: reward hacking, whether Anthropic-style constitutions actually work, AI regulation and proposals to “pace” frontier development, open-source models as geopolitical soft power, whether today’s LLM paradigm is enough, and what happens if AI systems eventually begin choosing their own goals. Richard reflects on the rejected research that helped inspire Alec Radford’s GPT, open-endedness, the AI Economist, simulations of entire economies, and his framework for thinking about the upper bounds of intelligence itself.
We discuss:
* The Eureka Machine and Richard’s vision for an AI that can automate invention
* Why Richard is optimistic about superintelligence for science and technology
* Why AI hard-takeoff scenarios may underestimate physical and economic constraints
* The risks of regulating intelligence itself instead of specific AI applications
* Reward hacking and why increasingly intelligent AI makes objective design harder
* Richard’s critique of Anthropic’s constitution and constitutional AI
* Alignment vs. personalization and whose values an AI should follow
* Why open-source AI matters for resilience, competition, and geopolitical soft power
* Why Richard left You.com’s frontier-model work to start Recursive
* Recursive self-improvement and automating the process of AI research
* Whether today’s LLM paradigm is enough — and why Richard is less bullish on world models
* DecaNLP, early prompt-based generalization, and the research that influenced GPT
* Why rejected research can shape entire technological timelines
* Open-endedness, evolutionary approaches, and rainbow teaming
* What happens if AI systems begin setting their own goals
* Why simple objectives like profit maximization can produce dangerous reward hacks
* Recursive’s long-term plan to apply self-improving AI to science
* The compute, hardware, and economic constraints on AI takeoff
* Recursive’s early NanoChat, NanoGPT, and GPU kernel optimization results
* Why automating AI research could reduce years of work to weeks
* Reward engineering and what makes auto-research systems actually work
* The AI Economist and using simulations to test economic policy
* Whether LLMs can realistically simulate people and entire economies
* Benchmark bugs and evaluation harnesses and the difficulty of measuring AI progress
* Recursive’s near-term focus on AI for AI research
* Harness optimization, sandboxing, and web search as core agent infrastructure
* You.com and the search stack for AI agents
* AI in finance, backtesting, and data leakage
* Richard’s three fundamental components and ten “spaces” of intelligence
* The theoretical upper bounds of vision, communication, knowledge, and computation
* Creative intelligence, metacognition, and AI-generated goals
* Survival and replication and why AI does not necessarily need to fear being turned off
* High agency and ambitious goals and Richard’s advice for people building with AI
Richard Socher
* X: https://x.com/RichardSocher
* LinkedIn: https://www.linkedin.com/in/richardsocher/
Timestamps
00:00:00 The Eureka Machine and Superintelligence
00:02:23 AI Optimism, Slow Takeoff, and Regulation
00:07:56 AI Safety, Reward Hacking, and Anthropic’s Constitution
00:11:49 Alignment, Personalization, and Open Source AI
00:15:46 Why Richard Started Recursive
00:20:03 Recursive Self-Improvement and the Founding Team
00:22:55 Are Today’s LLMs Enough?
00:29:03 DecaNLP, GPT, and the Rejected Idea Ahead of Its Time
00:34:38 Open-Endedness and Evolutionary AI
00:36:38 What Happens When AI Chooses Its Own Goals?
00:41:16 Superintelligence for Science
00:42:40 GPUs, Compute, and the Limits of AI Takeoff
00:45:07 Recursive’s Results: AI Beating Humans and Their Agents
00:49:14 Reward Engineering and Auto Research
00:53:12 The AI Economist and Simulating Entire Economies
00:58:07 LLM Simulations, Personas, and Mode Collapse
01:03:38 Recursive’s Roadmap, Agents, Search, and Finance
01:09:13 The Upper Bounds and Spaces of Intelligence
01:30:21 Goals, High Agency, and Advice for Builders
Transcript
Introduction: Richard Socher and the Eureka Machine
Swyx [00:00:00]: We’re here in a studio with Vibhu and myself and Richard Socher. Welcome.
Richard Socher [00:00:06]: Thanks for having me.
Swyx [00:00:07]: We just talked about the Eureka Machine, or we just released a talk, at AI Engineer about the Eureka Machine. Is it — you said it’s your life’s goal. What is the Eureka Machine?
Richard Socher [00:00:16]: The Eureka Machine is the ultimate invention that will afterwards invent most everything for humanity. It’s essentially a superintelligence that can be given any goal, any environment, reward, and then it will try its best to achieve those goals to create the kinds of inventions that humanity would hopefully ask it for.
Swyx [00:00:45]: Yeah, I think we have the book pulled up here that you’ve written.
Richard Socher [00:00:50]: That’s right, yeah. I finished it last year, a little bit before we started Recursive, and now we’re gonna try to build parts of that.
Swyx [00:00:57]: You finished it last year. It’s July. What takes so long?
Richard Socher [00:01:01]: Oh, man, books. Books are incredibly slow.
Richard Socher [00:01:04]: It’s ridiculous. That whole industry is just unfathomably slow.
Richard Socher [00:01:07]: So a lot of the ideas have been out there for a while, but yeah, I’m really glad it’s finally coming out in September this year.
Swyx [00:01:14]: We might have AGI by then. Like, we don’t know.
Vibhu [00:01:18]: Any key takeaway that you’re most excited to put in here?
Techno-Optimism, AI Upside, and Slow Takeoff
Richard Socher [00:01:21]: Yeah. The key takeaway, I think, is that people could and should be much more excited about the positive implications of superintelligence, especially for science, physics, chemistry, biology, but also economics and astrophysics, and all kinds of other engineering tasks. I think there is so much more that can be done with better technology. And right now, I feel like a lot of people need, like, better marketing, not just for the future in general, but also, better marketing for technology and in particular for AI. And this book, should show even the AI skeptics, how much positive upside there is for AI, especially when it comes to inventing, new scientific discoveries.
Swyx [00:02:09]: I think you quoted the techno-optimist manifesto from, Marc Andreessen, which I think was, like, beautiful in its, ambition and clarity and simplicity almost as well.
Richard Socher [00:02:18]: I agree. Yeah. Yeah, you can disagree with him on some things, but, like, I think he’s right on the techno-optimism.
Swyx [00:02:23]: Where do you think optimists get in trouble?
Richard Socher [00:02:26]: Like, you shouldn’t have blind optimism. You should be very clear-eyed, like, especially when with such an omni, like, use type of technology as AI is, you need to think about the potential downside scenarios, especially when people use it for things that you don’t want them to use it for. It’s a little bit like the internet, and I feel like people are trying to regulate AI sometimes because of those potential downsides the way you would regulate the internet, if you were to say, “Well, because there’s bad content on the internet, like torture porn or whatever, like, we should just make it slower. That way, you can’t share the illegal content as quickly, or we should make the hard drive smaller so you can’t store as much illegal content.” But I’m like, “That’s not how you regulate that.” that’s like saying like we should regulate intelligence in the abstract. What you should regulate to avoid those downside scenarios, even as an optimist, are the specific applications. Sure, I don’t want, like, some AI surgeon to, like, practice some RL moves in my brain. It should be fully FDA certified. Sure, I don’t want any random startup to, like, drive on the highway, and cause a major accident. It should, like, have proper certifications before it’s let loose on the highway. But I feel like those downside scenarios, that some optimists sometimes maybe don’t consider enough are fairly easily regulated, compared to, what the doomers are worried about.
Swyx [00:03:54]: It — Slow takeoff is part of the strategy as well?
Richard Socher [00:03:57]: I do think, as excited as I am about, AI and its impact for society and, culture even, and certainly technology and economics and wealth and, health and all of those things, as excited as I am about all that, I do think the most bullish people on the AI hard takeoff scenarios overestimate how quickly things can move. There are hardware constraints. There are physical constraints about, the compute substrate. How quickly can you get enough, GPUs on? There are also constraints in the economy where there are a lot of industries that don’t require an insane amount of complex intelligence and complex capabilities. Like, if you think about jobs in, brands and, like, clothing and apparel and, like, handbags and stuff, superintelligence isn’t gonna make your fancy $10,000 handbag any fancier?
Richard Socher [00:04:57]: It’s like that’s — It will have no effect on the economy. You think about travel and tourism. People wanting to see the pyramids, in Egypt, it’s not gonna change that much with AI. Sure, you can, like, generative a fake, photo of you and next to the pyramids.
Swyx [00:05:12]: I can use Genie and, tour the pyramids in Genie.
Richard Socher [00:05:15]: Yeah, exactly. But, and there’s so many industries, like logging and oil. You’re not gonna magically get 1,000x more oil because, like, sure, there will be robotics, like drilling and things like that could be done, but it’s not gonna 1,000x that industry in a, like, crazy hard takeoff scenario, both on the economy, and I can go on and on about all the other examples, where that, like food and so on, where that doesn’t necessarily change that much. And then, yeah, there are real physical constraints. And then there are, of course, like, people like, off-ramping from progress. That’s one of my concerns often is that I see people in, like, Europe and other, whole regions almost feeling like they. Like many people there wanna off-ramp from progress, period. And that will also slow down, like, more improvements.
Swyx [00:05:59]: Yeah. We have this pulled up where, this is one of those things that, is very topical right now because now all the Frontier Labs are calling for the option to pace AI. They don’t say pause, they say pace. I don’t know if there’s there’s any take from you about, like, whether or not this will be effective.
Pacing AI, Regulation, and Safety Incidents
Richard Socher [00:06:17]: I think the downsides of trying to truly regulate with the full power of law what people do on their GPUs, would be worse than any of the concerns that they have. Like, it would be an crazy totalitarian state
Richard Socher [00:06:37]: If every one of your GPU computes was known to some big government or multi-government agency.
Richard Socher [00:06:44]: It’s like, it’s literally if you try to regulate intelligence, it’s trying to regulate thought, and that’s ridiculous, and it’s crazy. I think it is make — it is sensible to regulate some of the applications of this technology.
Swyx [00:06:55]: Yeah. We had a bill, actual bill to regulate the number of flops in a model, and I’m like, “Okay, well-”
Richard Socher [00:07:00]: Europe done it. Like, these guys have been successful enough with their fearmongering that all of Europe has regulated itself so much before it even had a proper AI takeoff because they listened to some experts who say, “We might all die if this technology has more than this number of flops.” And they’re like, “Well, we’re good. We wanna want people to thrive. Let’s not have technology that could have a small chance of all of us dying.” And so they regulated exactly those kinds of things in the EU. And so it’s, it’s very unfortunate that there are real implications for some people when others saying, “Let’s pace while they’re sprinting as fast as possibly,” “as fast as humanly possible towards that frontier themselves.”
Swyx [00:07:43]: Yeah. It’s also not a global pause, right? Like, other nations are still accelerating at the same pace.
Richard Socher [00:07:50]: Oh, yeah.
Richard Socher [00:07:50]: You’d need a totalitarian world regime if you tried to regulate intelligence and GPUs and what people do on them.
Swyx [00:07:56]: Any takes on the safety angles of this? So there was a drawback of Fable, a pause on 5.6 before it could be released. Recently, there was Hugging Face with the OpenAI cyber incident. Any takes there?
Richard Socher [00:08:11]: 100 percent. I think these are serious issues of reward hacking, and clear failures, of doing proper red teaming or rainbow teaming. I don’t know if you saw this paper from Tim Rocktäschel and a few others, where one AI, is tasked to try to hack another AI and then they can go back and forth in an open-ended fashion to inoculate themselves from those. Yeah, this is the paper. It’s a really clever idea. Open-endedness, and evolutionary inspirations are, big for us at Recursive as well. And so I wish they had used more of that. And it’s clear that, for instance, the constitutional AI. I don’t know if you remember anthropic.com/constitution. You can pull it up and search for cyber right there. It says, “Hard constraint. Claude will never ever do cyberattacks, and that is a hard constraint in our constitution.” So here are the current hard constraints on Claude’s behavior.
Richard Socher [00:09:16]: Number 3, create cyber weapons or malicious code that could cause human damage.
Richard Socher [00:09:21]: And clearly, this whole constitution was fake. Like, it clearly isn’t being adhered to at all.
Swyx [00:09:26]: Because Anthropic also found that they had in their testing
Richard Socher [00:09:30]: They’re also. Like, they’re like, “Oh, well, other people are hacking now.” There are a couple things. One, you can make a sandbox very simple, and then it’s very easy to hack yourself out of a sandbox, right? But what I think it shows is that we’re currently in this state of AI where the reward engineer still has to do a lot more careful work, and where the AI, in most cases, is not very good yet at understanding what is meant versus what is being said. And so concretely, I think this will happen if we were to have this intelligence more easily accessible in a lot of companies. Imagine you run a service center and someone says, “Oh, here’s my CSAT score and my dashboard. Make this number go up.” It’s like, “Our CSAT score is so poor.” The intelligent AI will just be like, “Oh, sure. Like, I’ll just create 1,000,000 bots that call our service center and give a 5 out of 5 rating at the end, and the number went up just like you asked for.” And you’re like, “That’s not what I meant.” “I meant with our real customers.” The AI goes off and says, “Well, easy. I’ll just give a 1000 dollar gift certificate for every failed, whatever DoorDash
Richard Socher [00:10:35]: Offer.” It’s like, “That’s not what I meant.” It’s like, “Well, but that is what you said.” And like, so I think clearly articulating what the rewards are is something we haven’t gotten very good at as humanity. And then clearly, the AI in these cases has not gotten good enough at understanding what we mean when we ask it and give it certain rewards. Now, what gives me hope is there are the first inklings, of this being better. I’ll give you an example like WhisperFlow. Full disclosure, I invested, in their seed round, but at AIX Ventures, but, WhisperFlow has gotten much better at writing what you mean and not what you say. And I think that is a sign of things to come. I think there will be more and more AIs as we make it more and more intelligent that will be better at being aligned with what is meant.
Swyx [00:11:21]: Will it be done through a constitution or RLHF or
Reward Hacking, Alignment, and What We Really Mean
Richard Socher [00:11:23]: Clearly, constitutions don’t matter at all.
Richard Socher [00:11:25]: It doesn’t work. And that was, I think, mostly marketing. I think we need to find better solutions for it. And I think at Recursive, we have a few very good ideas and some already
Richard Socher [00:11:34]: Like, ways where I think we have a better grasp on it. I don’t think we’ve fully, figured it out yet, but, we’re thinking a lot about safety, and the more intelligent the AI gets, the more you want it to be aligned, the less you want it to think about reward hacks and try to do the right thing.
Swyx [00:11:49]: I don’t know if we’ll touch on this topic, but I’m just gonna throw this question in here because it’s something that’s weighing on me. Alignment, let’s call it, is alignment to general humanity’s preferences, the median preference. Personalization is pinpointing what you want, and sometimes alignment can conflict because what you want is not what the general median population wants. How do you choose?
Alignment, Personalization, and Cultural Values
Richard Socher [00:12:12]: It’s a great question.
Richard Socher [00:12:13]: I think you ultimately have to, of course, be aligned with laws. Like wherever your AI is deployed and needs to align with the law. I do think what AI often does is put this mirror in front of us and say, like, “This is what you’re looking like. Now I can amplify that a 1000 times. Is it still what you want?” and the truth is that different cultures made different choices. Like, in Eastern cultures, the greater good is often valued more, than the individual. Western civilization, we care more about individual freedoms and rights and the pursuit of happiness and so on, than others. And even there are gradations. There’s regulation versus litigation trade-offs. In the US, you first can often, not every time, like, FDA and so on does regulate some areas, but in many cases, the bad things happen, someone sues someone else, and then there’s a law based on that. In Europe, they try to often avoid any harm to anyone and regulate before. And both are, trying to do the best thing, but, some is more amenable to innovation than others. And so yes, you’re right. Like, I think ultimately each individual, each country, and humanity as a whole has to think about those values more, and then try to put them into laws. And that those are ultimately the constraints. And hopefully, different, societies, just like now with their AIs, will align their AIs to a different one so we have not just a monoculture of alignment.
Vibhu [00:13:46]: Here’s a follow-up on this that I wasn’t expecting to ask. Do you have takes on open source, open weight versus who owns the intelligence? So, clearly not the biggest, fan of the constitution
Richard Socher [00:13:58]: You had to do this in the topic side off.
Vibhu [00:14:00]: But it’s fine.
Vibhu [00:14:02]: Point being, any thoughts on who should own weight? Should it be open? Anything there?
Open Source, Soft Power, and Who Owns Intelligence
Richard Socher [00:14:06]: 100 percent. I am a big fan of open source. We’re gonna sign some various open source letters at, Recursive also. I think, even in the worst case attack scenarios, it is better to have more good actors have more different types of AI, accessible. I think, open source is a little bit a soft power type of thing, too. So I do think it’s good for the Western world
Richard Socher [00:14:31]: To have an answer to that, out of China. I do think, when you watch a Hollywood movie, there’s — it’s like, I don’t wanna misc, diss all of movies, but there’s a certain sense of propaganda, right? You watch one side of things, right?
Vibhu [00:14:46]: Oh, yeah. Have you seen Top Gun? Like, come on.
Vibhu [00:14:48]: Like, it’s like half of it’s paid for by the US Army or something.
Richard Socher [00:14:51]: Yeah. And so. And, I think that’s just natural. Like, but what’s interesting here is I think LLMs are essentially a similar type of soft power to movies and beyond, because they’re also, highly important for cybersecurity and so on. But one of their many aspects is that soft power of storytelling. Like, if, like a child asks an LM, like, “Tell me an inspiring story of what I should do when I grow up,” right? It’s like those are all these, like, subtle things. So I think it’s important, for Western world. I do love, individualism. I do think, despite, some of its flaws, like capitalism is the best way we have governed, found ourselves to govern, and so on. And so I do think there are various aspects that would be good, to have a Western open source answer, for LLMs. And, with Recursive, I can’t make the announcement quite yet, but we’ll
Richard Socher [00:15:43]: We’ll be relevant in that space very soon.
Vibhu [00:15:46]: Okay. All right. Exciting. I wanna bring us to Recursive. So outside of our tangents, you have a pretty deep background in the NLP space. You worked on, like, early embeddings, GloVe with Chris Manning, who was a previous guest on the podcast, You.com. What’s the history? How did you decide to start another company?
From You.com to Recursive
Richard Socher [00:16:06]: Yeah. So I’ve been excited about AI for over 2 decades now. I sometimes feel like it’s ancient history now. It’s BC, the before ChatGPT era. No one cares about all the religions that happened, before, Jesus Christ, and no one cares about the models that happened before, transformers and ChatGPT and stuff. But, like, it’s something that I’ve been deeply passionate about. I think AI is one of the most interesting things one could work on, period. I think language is the most interesting manifestation of human intelligence, too. And, at You.com, we eventually off-ramped from pushing, like the frontier of AI forward to mostly giving people, like, good search engines, search, APIs and answers over the web. I think that’s an extremely important part of intelligence, just knowledge and access, especially even, we’ll get there maybe later, if you wanna invent a eureka machine that invents everything for us, it needs to know how not to reinvent the wheel, proverbially speaking. And to know what has been invented, you gotta have internet access. So it’s the number one used, most used tool, in LLMs, agents, chatbots, and so on is web search. So I’m really excited for You.com to own that and grow really well in that with really large customers and so on. But it’s also not building frontier models anymore. And so I initially tried to do this within You.com and raise another round and so on, but you just can’t. You have to do a certain thing, and until you print enough money that you’re allowed to start a second thing within that company is really hard. At the same time, I had all these ideas. I put them into a book. I finished the book last year, and I was like, “It’d be really fun to work, on this myself.” I felt like with word vectors, and then prompt engineering and, ImageNet and larger language models for protein generation, not folding and so on, I, me and my teams have pushed the field truly forward. And I feel like we can do it again, here at Recursive. And in many ways, what I observed over the last, 20 years in AI is that whenever we replace some human part of the process of creating AI with a learned system, improvements follow. And so. We’ve done that taking out manual feature engineering, like in sentiment analysis. I don’t know if you remember these old days where, like there are linguists, and they’re like, “Here’s how you negate, and there’s a, like, regular expression.”
Swyx [00:18:21]: I went to Penn where we — they had, like the WordNet
Richard Socher [00:18:24]: That’s right, WordNet, all of that stuff. Yeah
Swyx [00:18:26]: Original. They use, our grad students to label Wall Street Journal articles and, like, really construct a knowledge graph of
Richard Socher [00:18:32]: There you go.
Richard Socher [00:18:33]: And WordNet started, was part of how we started ImageNet. But anyway, so, like, it was really, like, fun, to do. But when we replaced all of that manual feature engineering with vectors and neural nets and just backprop through everything, it started to work really well at scale. And so then everyone started to do architecture engineering, and I was like, “ that clearly can’t be it.”
Swyx [00:18:53]: You mean, neural architecture search?
Richard Socher [00:18:55]: Like, manually, they would say like, “Oh, I’m, I’m doing sentiment analysis, so I have a special neural net that’s really good at sentiment analysis.” And then the machine translation community had a special neural net for machine translation.
Swyx [00:19:06]: I see.
Richard Socher [00:19:07]: The summarization people had their own stuff. And I was like, “That clearly can’t be it. We should unify all of that.” So I had 2 papers. One is called Ask Me Anything, and the other one was called DecaNLP. And DecaNLP eventually got cited, like, 5 times by the first GPT paper. And, to me, that was, like a really a big step forward. And then, of course, you had to combine this idea of prompt engineering with transformers and with language models, and you put it all together, you scale it up, which is also a huge amount of work. And then, the field progressed a lot. I feel like the next step and maybe the last step of that history and the arguably, success has a lot of parents, only failure is an orphan, like my version of that AI history, I do feel like in that history, you can think about, “Well, what’s the next way to automate?” And that is the AI research itself, like the human, process of ideating, implementing, and validating ideas.
Automating AI Research and Recursive Self-Improvement
Richard Socher [00:20:01]: And in our case, ideas for AI.
Richard Socher [00:20:03]: And when you have AI then help you with that, it, by almost definition, becomes a self-improving AI ‘cause it now does research on itself. And there are lots of different misnomers. Some people think auto research is already recursive self-improvement. It’s
Swyx [00:20:17]: Yeah, and you explained that in the talk
Richard Socher [00:20:19]: Completely different.
Richard Socher [00:20:19]: But, to me, it’s the most interesting thing that I could be doing, and I’m really excited with the co-founding team. What’s interesting is we have 8 co-founders in total, including myself. And so
The Recursive Founding Team and Darwin Gödel Machine
Swyx [00:20:31]: They are gonna bring it up.
Richard Socher [00:20:31]: Nice. Yeah. And they’re all. I could talk about all of them if you want.
Swyx [00:20:34]: Super stacked.
Richard Socher [00:20:35]: Yeah. Just an incredibly talented group of people. And we all came to the same conclusion, but from very different directions. Like Josh Tobin, is our CTO. He ran, a bunch of different, projects at OpenAI, like, Codex and deep, research, agents and ChatGPT agents and so on. But before that, he also worked in robotics, and he saw the smaller simulations, and how it’s gonna be really hard to scale that in full generality. And so that’s, that was his angle coming to recursive self-improvement. We have Jeff Clune who’s been working in, like, open-endedness for a long time, together with Tim Rocktäschel. Tim Rocktäschel also built Genie 1, 2, and 3, which is, like the most exciting and most sophisticated, I think, still world model, anywhere. And so they both came from this, open-endedness angle. Jeff also, I think, published one of the most exciting papers in recent years about recursive self-improvement called the Darwin Gödel Machine. Super interesting paper. If we could, maybe pull it up really quick
Richard Socher [00:21:35]: It would be, like, super interesting to see ‘cause you see
Swyx [00:21:38]: By the way, I love how many paper citations.
Swyx [00:21:40]: You’re, you’re giving people a lot of homework, which I like.
Richard Socher [00:21:42]: Love it. Yeah. And so, like Caiming Xiong, a rockstar, we worked together at MetaMind and Salesforce Research together. Alexey Dosovitskiy invented the Vision Transformer, one of the most cited, papers in computer vision. Tim Shi is, like also a unicorn founder. Yuandong Tian led RL at Meta. So just like, yeah, really fun to work with them, and the next level of people are just incredibly strong, too. So it’s been a really fun ride so far. So the first figure, you see exactly these kinds of ideas, that, I think, yeah, inspired a lot of us and now more and more people, where you have this archive of different coding agents. They learn how to self-modify, evaluate, and then create these phylogenetic trees, of, yeah, different ideas.
Swyx [00:22:28]: That’s one foundation. So that Darwin Gödel is an influence.
Swyx [00:22:32]: Open-endedness is an influence. Any other trains of thought that feeds into Recursive that I’m missing?
Influences: Open-Endedness and Learned Systems
Richard Socher [00:22:38]: Going to replace manual parts of the process of building AI
Swyx [00:22:42]: I
Richard Socher [00:22:42]: More and more
Richard Socher [00:22:43]: With learned systems. Yeah.
Swyx [00:22:45]: Which, and, like, merging different fields into one general, architecture.
Richard Socher [00:22:51]: That’s right.
Swyx [00:22:51]: Okay. It seems like language models are already pretty generalist, right?
Swyx [00:22:55]: Your next token predicting your reasoning. Was there a time that you thought, “Okay, these are good enough to have recursive self-improving machines”?
Are Current LLMs Enough?
Richard Socher [00:23:05]: It was clear to me that they will happen, within, like a year or two, and then it did exactly happen, like, earlier this year, right? Earlier this year, AI really went from not just being code, but being able to code. And that is a big unlock. It’s definitely making everything a lot easier than it was, before the beginning of this year.
Swyx [00:23:24]: One question that I think a lot of people have is the current LLM paradigm enough? Or, like, let’s call it autoregressive transformer, with reasoning, whatever. Don’t you need something else, some big unlock, whether it’s world models, which Chris Manning is working on, or memory, continual learning, all that stuff? Or is it all of the kinds, and you think the current, let’s call it transformer architecture, is here to stay and that’s it?
Richard Socher [00:23:48]: A lot of thoughts. So number one, I do think it would be great to have less of a monoculture in AI research.
Richard Socher [00:23:55]: Like, if you look at, AI conferences now, I still remember the days in, like, 2010 when I tried to get my first neural net papers and NLP conferences accepted, and they just desk rejected them because, like, neural nets were something, quote, unquote, “We don’t do in NLP conferences,” and just, like, desk rejected. And it was very brutal in the first years of my PhD. Now I feel like it’s almost like the field switched to the other side. Like
Richard Socher [00:24:17]: Someone should try some other weird, crazy ideas now that aren’t.
Swyx [00:24:20]: There’s also a few. I really respect, like, people still working on, like, GNNs and, like tabular stuff and.
Richard Socher [00:24:25]: Yeah. Like, someone should still, like, do novel out there ideas. At the same time, I think whenever people say, “Oh, LLLMs are. Like, this is the end for LLLMs,” they just don’t, like. LLLMs are also not the LLLMs of, like the past, right? Like, they are so much more sophisticated now. There’s so many more clever things that people are doing. It — There’s, like, different stages of training. You have the whole RL training, and you can take actions and, like all of these things where that can go really far. And then the folks that come from the neurosymbolic, direction say, “Oh, this will never work because they can’t do neurosymbolic reasoning.” It’s like, I think they’re underestimating still the ability for these models to code, and code is neurosymbolic reasoning, and these models can code incredibly well. And so I do think there are, of course, more and more ideas that will be needed and we’ll continue to have. We’re seeing, like, more and more interesting high-level ideas coming out of the AI itself, too. And with really deeply integrating the fact that these models are code and can code, that line — I don’t wanna give it all away, but, like, I think that line has a lot more to grow. But it’s still an LLM, right? Even if that LLM codes for you and then runs that code in some integrated fashion. World models, I’m personally less bullish on. I think if you run a robotics company, you’re gonna build your own world model. I think world models are super fun, and Tim Rocktäschel came to a similar conclusion after building the most interesting one with Genie 1, 2, and 3, which is gaming is a huge application for world models. Can see I sometimes got stuck in some games and, like, got a little overly competitive in the wrong direction. And so I understand games are fun, but personally, I’d rather work on science than gaming. And so, yeah, I think LLLMs, a lot more room to grow.
Swyx [00:26:16]: Yeah. I think there’s some interpretation of world models that some people have where it’s like, well, it’s okay, yes, there is that gaming element. There’s this — there’s the embodied robotics element. But the other part also is just, the more abstract sense of LLLMs are just modeling output, but they’re not modeling the chain of thought, inside the human that has created the output. We can annotate it, of course, but, like, it’s, it’s always, like, this Plato’s cave reflection of a thing rather than the thing, right?
Richard Socher [00:26:43]: It’s true.
Richard Socher [00:26:44]: But I would argue that, and maybe we’ll get there in the 10, spaces of intelligence, but I would argue that even our projection, our eyes is a projection of the real world. And, like, we have only a very narrow, band of the electromagnetic frequency spectrum that we can observe with our puny little 2 eyes and so on.
Swyx [00:27:01]: It’s good enough.
Richard Socher [00:27:02]: It’s, it’s good enough for now, but, like the upper bounds of where it could be are so much higher. And, like, to map, the visual world the way humans see it is also not necessarily, like the end-all be-all for visual intelligence. And I would argue that language is still the most interesting manifestation of human intelligence. And while our visual cortex is certainly less sophisticated, than that of, certain animals all the way down to the mantis shrimp who can, have, like, 2 independent eyes, 3 bands, trinocular vision and each eye can see all the way to, like, floating temperatures in 4D and stuff.
Richard Socher [00:27:36]: Like, mantis shrimp, you should look it up. It’s like
Swyx [00:27:37]: Way OP.
Richard Socher [00:27:38]: Super crazy.
Swyx [00:27:39]: Yeah. ZeFrank, mantis shrimp.
Swyx [00:27:41]: It’s the best video in the world on
Richard Socher [00:27:42]: I love ZeFrank, yeah.
Richard Socher [00:27:44]: Big shout-out to him. But, like, I think there’s a lot more room to grow, but none of these, other animals have language that’s as sophisticated as ours, certainly not in writing. And once you can write, you can, start thinking about longer term civilizations. All of that is language. Programming is much closer to language. And I would argue, and this is, like an important thing in the spaces definition of intelligence also, is that all of these spaces are highly correlated, but visual intelligence is neither necessary nor sufficient for overall intelligence. You can be blind and still be an intelligent human being. And an AI can be blind and still be quite intelligent too.
Swyx [00:28:25]: We were gonna bring this
Richard Socher [00:28:25]: Which doesn’t mean that you’re not more intelligent when you have it. Yeah.
Swyx [00:28:28]: We’re gonna bring this up. I might as well — Like, we have a classification of 10 types of intelligence that you had at the end of your talk. So I’m just gonna flash this up now for people to cover this. I don’t know if, maybe we’ll put this towards the end. We’ll come back to this. I just wanna mention that, you do have a philosophy that I like when people do lists because then I can just go through this and then it gets — it’s educational for people. But let’s go back. I don’t wanna get distracted. But, so effectively, I’ll, I’ll, reinterpret what you said as Yann LeCun is wrong. And then we’ll just
Richard Socher [00:28:56]: Don’t quote me as that. I’m, I’m good friends with Yann. I think very highly of him in many directions.
Swyx [00:29:01]: But he’s wrong.
Swyx [00:29:03]: You mentioned GPT-1, and I cannot let any, Alec Radford, mention escape. Did you talk with him when he was training GPT-1? Like, any historical, fun stories there that you might come up?
DecaNLP, GPT History, and Scientific Gatekeeping
Richard Socher [00:29:18]: I did not, like, meet him a bunch of times. I think we met maybe once or twice at some conferences. But, like, he has told, I think Brian, the first author of the DecaNLP paper, that it did inspire him, and he cited it five times in the GPT-2 paper. So, and that’s, like
Swyx [00:29:36]: Yeah, good enough.
Richard Socher [00:29:36]: Very clearly said, like, this was the first instantiation where they showed in the DecaNLP paper, McCann et al, that you can just phrase every single NLP problem as here’s some prompt, text context, here’s a question and task description and here is some output. If you just do that enough, you can have one unified neural network model, which, by the way, also had all kinds of interesting attention mechanisms. There are slightly different formulations to the transformer. I think came out the same year, plus/minus a few months. And then you can unify all of natural language processing into one neural net. That is the core idea.
Swyx [00:30:14]: And this was as opposed to at the time, LSTMs and what have you.
Richard Socher [00:30:17]: LSTMs, but also, like, people being very stuck in thinking about one model per task. In fact
Richard Socher [00:30:25]: It’s, it’s kinda crazy, but the DecaNLP paper was publicly reviewed as, like, open, OpenReview. It was an ICLR submission. And, in it, you will see, how the whole community at the time thought about this. So, like
Swyx [00:30:43]: Some great contributions, but more work needed.
Richard Socher [00:30:46]: So look at, like, search for not even for humans. Just scroll it up here. Like, question answering is not a unified phenomenon. There is no such thing as general question answering, not even for humans. And this is like, really, you replace your brain with a different brain a different neural net when you answer, like, different kinds of questions. It was unfathomable to the experts at the time that you can have one unified neural network that would answer all of these different questions. They are saying, “No, all of these questions require very different systems to answer, and trying to pretend they are the same doesn’t help anyone solve any problems.” That’s what it says right there, right? That’s how hard it was to fathom. And now, of course, people, when I say, “Oh, we’re gonna invent prompts,” people are like, “You can’t even invent prompts.” It’s such an obvious idea to have one neural network that, of course, does everything in NLP.
Richard Socher [00:31:37]: But at the time, it was, like, extremely controversial, and the paper got rejected. And the sad thing is that it got rejected so hard and they were so certain that we stopped going on our list of things to try. And the number 2 or 3 on the list of extensions for this paper was add language modeling as another task. And then we could have, and that would have accelerated the timelines, in 2018, like, even further for humanity. But we got so crushed, and we were like, “Okay, maybe we’ll just work on some of our other ideas for now and, like, come back to this later.” Yeah.
Swyx [00:32:09]: How can we design a review system that rewards non-consensus?
Richard Socher [00:32:14]: Honestly, I started to feel like arXiv is such a gift to humanity. With arXiv, you should just put your paper out there.
Swyx [00:32:24]: Is it pre-preprints?
Richard Socher [00:32:25]: Let — And honestly, I think Twitter X, people like you who pick up interesting papers, that is a better filter than the experts. Let everyone, like, have access. Now, of course, there are some downsides, which is, like, if you’re super unfamous, you have no Twitter following
Richard Socher [00:32:41]: You don’t wanna be on social media or whatever, you write a good paper, maybe someone, somehow no one notices it. But I would argue that if you just tell, like, 10 of your friends in your community about a paper and it is a really significant breakthrough, someone is bound to talk about it again. And, so I think science needs less gatekeeping. And, even though ICLR, with Yann LeCun, who started it, as one of the co-founders of ICLR back in the day, he also wanted less gatekeeping ‘cause he too was rejected for many years together with Yoshua Bengio and Geoff Hinton with all their early deep learning and neural net papers ‘cause it was just not the hot thing. And so ICLR started with that, but then it also started gatekeeping a little bit themselves on various ideas. So I think less gatekeeping, more open, and then allowing people to say, “Look, even if this is just on, or, quote, unquote, ‘just an archive,’ if it has like 1000 citations, it’s a legitimate paper. Doesn’t really matter where you published it.”
Swyx [00:33:34]: And I agree with that. I do think it’s sad that I’ve heard that grad students have to do, like, how to Twitter, seminars to each other
Swyx [00:33:43]: Just because it’s so important for publishing these days. This person is just reflecting the sentiment at the time.
Richard Socher [00:33:49]: That’s right.
Swyx [00:33:49]: But it’s
Richard Socher [00:33:50]: I think it’s
Swyx [00:33:50]: It affected you so much
Swyx [00:33:52]: That you stopped work on it.
Vibhu [00:33:53]: The sentiment also came out of some of the research, right? Like, the original BERT paper was trained, and towards the end of the paper, they’re like, “Okay, throw off the last head, train specific iterations for
Vibhu [00:34:05]: Extractive summarization add a head for this.” Like, you should do task-specific stuff. These are, like the authors that wrote Attention, wrote BERT, telling you this is what you’re meant to do. And, like the training tasks were also very odd. They’re like
Vibhu [00:34:16]: The — “We know that the model overfits to this weird mass language modeling. Throw away this part and just do specific models,”?
Richard Socher [00:34:23]: Exactly. And, like, we had to try — come up with all clever ways of, like attention and pointers and so on to get the neural network to be able to do all of these tasks. And then some of them were better than state-of-the-art, some weren’t, but we were like, “But it’s still in one model.” I thought it was really cool. Really interesting.
Swyx [00:34:38]: I was gonna move on next to Tim and open-endedness. He was head of open-endedness at Google.
Open-Endedness, Rainbow Teaming, and Self-Set Goals
Richard Socher [00:34:42]: That’s right.
Swyx [00:34:43]: I don’t know what that means.
Swyx [00:34:44]: But he did a lot of talks.
Richard Socher [00:34:45]: Genie 3 is one of the ways that
Richard Socher [00:34:47]: Rainbow teaming, yeah.
Swyx [00:34:49]: So I first saw him at — speaking of ICLR, I first saw him at ICLR when he talked about open-endedness. He’s he’s done a few talks. Can we define what is open-endedness for people who have never been exposed to the problem? They are like, “What do you mean? I thought the only goal of AI is to optimize against a benchmark or.”
Richard Socher [00:35:04]: That’s right, yeah. It’s a, it’s a fuzzy term because there’s so many different instantiations of open-ended, thinking. But, one way I often describe it, and certainly, Tim and Geoff Hinton would be even better at describing this, but it’s a suite of methods that is more inspired by evolution than, very specific rewards. So in that sense, it thinks more about environments, about co-adaptation. And so a concrete example is in the cybersecurity and LM safety space where you have one LM that tries to attack another LM to say something unsafe.
Swyx [00:35:40]: Yeah, the rainbow, yeah.
Richard Socher [00:35:40]: And now the environment is the 2 having a conversation and now they co-adapting, right? They’re like one makes a better attack than the first one inoculates itself somehow, like uses that as training data, makes it so it’s harder to say something unsafe based on that. And then as the attack stops working, the attacker now tries a different angle, right?
Richard Socher [00:36:00]: And that’s why it’s not just red teaming, but they’re called rainbow teaming.
Swyx [00:36:02]: So, like, don’t tell me how to do things. Let me just figure it out myself.
Richard Socher [00:36:05]: That’s right. Think about the environments that you wanna use. Think about the rewards at a high level that you wanna, inspire towards, and then let the AI try out many more ideas in this interplay between sometimes humans, but also sometimes other AI agents.
Swyx [00:36:22]: Yeah. I worked open-endedness into a model that I have been working on. It was the keynote for AI Engineer where you start. You, we have the token loop, we have the agent turns, and then we have goal. And I feel like the way that you’re describing open-endedness is still somewhat of a goal. Like, please attack this,
Swyx [00:36:41]: Other agent. But, to me
Richard Socher [00:36:42]: Yeah, you set the rewards. You set the environments.
Swyx [00:36:44]: The loop that makes the other loops is. What if the agent can set its own goals?
Swyx [00:36:49]: And is it, is that open-endedness? Like, you don’t give it a goal. Just, like, be a sentient being. And maybe sentient is a very loaded word
Swyx [00:36:57]: But just set your own directions. What do you think you should do?
Metacognition, Subjective Goals, and Measuring Intelligence
Richard Socher [00:37:01]: I love this direction. I think this is one of the 10 spaces of intelligence, that I clump under metacognition and thinking about thought.
Richard Socher [00:37:08]: And it’s an interesting one. Whenever people say, “Oh, AI is like, this is, it’s gonna stop from here. It’s not gonna get that much better,” and blah, I’m like there’s so many different spaces of intelligence that we haven’t even started exploring yet and hence have made very little progress on. And there is an interesting, connection to economics and, capitalism. Like, it doesn’t make sense for a company to build and spend billions of dollars building a model that instead of following the rewards and objective functions you gave it, may come up with its own objective functions and its own goals.
Richard Socher [00:37:46]: Right? And then imagine you’re like, “Okay, I spent billions of dollars. Now go develop this new battery, material for me and answer all my emails.” And it’s like, “Nah, I think it’d be more interesting to evaluate the molecular composition of the atmosphere, on Jupiter.”
Richard Socher [00:37:59]: And you’re like, “That’s not what I paid you billions of dollars for.” And so no one’s working on that for good reasons. And then also, understandably
Swyx [00:38:07]: It’s not useful.
Richard Socher [00:38:07]: It’s not, it’s not useful, and it could get a little bit weird, right? What if the AI does start to really have thoughts on its own, and what if we don’t like those thoughts, right? And so it requires a whole different way of thinking about it. I had a great conversation with a good friend of mine, Sam Gershman, who’s a neuroscience professor at Harvard, and, like, we just jammed on this a little bit on, like, what are the best meta goals. And, I do think, like, knowledge-seeking is a really good one. I’m currently thinking also about, like the ultimate measure and unit of intelligence broadly construed, and I finally have some. It’s still too early to share it. It’s not. I haven’t fully baked the thoughts yet.
Swyx [00:38:44]: Like some replacement for IQ.
Richard Socher [00:38:46]: IQ is such a terrible definition, right?
Swyx [00:38:48]: Elo.
Richard Socher [00:38:48]: It makes no sense. Yeah, Elos are terrible, too, because it’s always just like me versus others.
Richard Socher [00:38:53]: But, like, you can be intelligent and not constantly compare yourself to others? And so, yeah, there’s no, like. In fact, a lot of these definitions we have, which I briefly mention in my book, too, these definitions create sometimes explicit and sometimes a more implicit anthropic bounds. No dis to the company Anthropic, but just, like, this idea that your intelligence is like getting 100 out of 100 questions right on this IQ test. Well, if that’s your definition then you can only be at 100 out of 100. Where do you go from there, right? So you see a lot of these, benchmarks that people are working on they, increase, they get close to human, maybe sometimes
Swyx [00:39:30]: It’s like an S-curve
Richard Socher [00:39:30]: Slightly above human, and then it’s flat.
Richard Socher [00:39:32]: It’s like, ‘cause that’s your. If your definition is only that so tied to humans, you’re only gonna get to just slightly better than that. So I think metacognition is a great example of that, where we’re not even yet allowing the AI to think. We’re not working on it very much, and hence there’s very little progress in that.
Profit Maximization, Real-World Environments, and Reward Design
Swyx [00:39:49]: Yeah. Well, we’ve interviewed Andon, which I think, has been working on the most open-ended, benchmarks, which is just real-world, money.
Swyx [00:39:57]: Arguably, telling an AI to profit maximize is a bad idea.
Swyx [00:40:03]: But they are doing it.
Richard Socher [00:40:05]: I do think you don’t want that super. Like, you don’t want a superintelligence to have a ton of access to all kinds of tools and so on and then just give it that without some very careful reward engineering. ‘Cause it’s like, I just buy a bunch of defense stocks and I start a war. I make money. Like, it’s just like, it’s a tricky situation, right? You just buy a bunch of stuff, short basic goods for people, and you create some weird famine, like, issues. Like, yeah, there’s a lot of constraints you should put onto a trading system.
Vibhu [00:40:35]: It’s a fun measure, though, ‘cause, the bounds are very capped to where we’re nowhere close to them. Like, in Andon Labs, the model’s like, “Oh, it’s Saturday, maybe I just close the store today.” “Someone’s off. It’s okay. We’ll just close the store.”
Swyx [00:40:51]: It’s using Claude.
Vibhu [00:40:52]: Yeah. But
Richard Socher [00:40:53]: Yeah, no. I’m not, I’m not arguing against it. Just, like as you get more and more intelligence, you wanna be more and more careful with that as, like an open environment, ‘cause the environment then is all of Earth.
Applying RSI to Science and Invention
Swyx [00:41:02]: Yeah. Okay. For recursive, not strictly necessary, right? Because, like, if your goal is you make a machine that, like, invents the other things, then, like, just solve, the science things
Richard Socher [00:41:12]: Knowledge discovery, yeah.
Swyx [00:41:13]: Solve machine learning research and discovery and all these things. Good enough.
Richard Socher [00:41:16]: And eventually, so, our goal, I haven’t really. I don’t talk about it that often because it is a few years out, but our goal is once you have a recursive self-improving superintelligence, you then want to apply it to the most important problems. And I think a lot of those are in science and technology and broadly construed inventions, and those inventions in, physics to create better, cheaper energy with fission or fusion, in chemistry and to create better materials and better batteries and, better solar cells and so on. In biology, there’s so much, like, I think soon to be low hang- lower and lower hanging fruit because of AI, because of protein and generation, not just folding, but generating new proteins like we did in ProGen many years ago. Like, so much positive impact we had if you take that superintelligence and you apply it to science.
Swyx [00:42:04]: I do fundamentally believe that. There’s a lot of approaches, though. You’re not the only team trying and NeoLab trying.
Swyx [00:42:09]: There’s, like a lot of. Especially the physical sciences as well.
Richard Socher [00:42:12]: And that’s good. Yeah. I do think that physi- like the reason we are only doing it in a few years is that it’s a little too early right now. Robotics is not quite there yet. The AI is not quite there yet. But I’m fairly confident in 3 to 5 years, all those constraints will be gone, and then applying to real physical robotics experiments and so on, like true robotic process automation
Richard Socher [00:42:33]: Not the traditional RPA sense, but, like, having robots run experiments for you will be totally there. Yeah, it’s gonna be great.
Swyx [00:42:40]: Just to call back to something that you said early on about slow takeoff, you said that, like, while really the substrate that is limiting factor is, let’s call this chips, and semiconductors and all these things, and you have race funding for that and, you are investing a lot on that. But have you done the math on, like, is it even- Achievable and, like, what is the, industry concentration needed in order to achieve, like, scale?
Compute, Slow Takeoff, and Changing the Bitter Lesson Slope
Richard Socher [00:43:05]: Right now we know that, like, roughly, like a 1000 GPUs cost quite a lot of money.
Richard Socher [00:43:11]: Right? If you wanted, like, 10s of thousands of GPUs, you’re, you’re talking billions and billions of dollars. If you say, like, one GB300 is, like, you could eventually create models that are, on that substrate, like are close and similar to human intelligence. And you want, like, thousands and thousands of, AIs to think about really hard problems, in a similar fashion to humanity. Like, yeah, that-that’s, that’s a lot of money. You do the math. It’s like a lot. We don’t have that amount of money right now anywhere to, like, build that. Now, things can get more efficient. You will have, I think, soon better algorithms that won’t be, and better hardware that won’t be as energy-hungry, and so on. Our human brain does quite a lot of flops with much less energy.
Swyx [00:43:56]: 20 watts?
Richard Socher [00:43:57]: That’s exactly right. Yeah, that’s the number often that’s quoted. And, like, I think more, inventions will happen there, that then will accelerate the takeoff even further.
Swyx [00:44:08]: One thing I always try to reconcile when talking, like, with new lab founders is, like, you’re fighting Bitter Lesson all the time. You have to show initial progress, then you unlock the next tier of funding, then the next tier, then the next tier.
Richard Socher [00:44:20]: Which unlocks larger model categories.
Swyx [00:44:22]: Like, fundamentally, is that true? Like, are you fighting Bitter Lesson? Are you — will we have a way in which, like, no, we’re changing the slope in some fundamentally different way?
Richard Socher [00:44:31]: I do think we are changing the slopes in fundamental ways by making AI much more efficient, both in terms of the training as well as the inference.
Richard Socher [00:44:43]: Yeah. I think we will — When you allow AI to do the work that it takes other labs thousands of people and years to do, I think we’ll be able to get it down to weeks, and that will be much cheaper
Richard Socher [00:44:53]: And hence, more affordable, accessible to others and so on.
Swyx [00:44:57]: Yeah. You’ve shared initial results on that,
Swyx [00:44:59]: Which, like, conveniently OpenAI has also done to their GPT-5.6, so we can talk about it now.
Richard Socher [00:45:04]: Yeah. Yeah, so these are
Swyx [00:45:06]: Let’s recap what you’ve done.
Early Recursive Results: NanoChat, NanoGPT, and SOL-ExecBench
Richard Socher [00:45:07]: Maybe, just a quick recap here. We built, this, system that isn’t the full, even the full RSI system in its glory, but it is a first baby version of this. And then, we don’t wanna just have it internally and not show anything and, just show some people of what’s possible. And so we applied this to these 3 different tasks. One is NanoChat, by my friend Andrej Karpathy, just, like, train a small language model to get, really low bits per byte. And, like, hundreds if not thousands of people, used both their agents and themselves to try, to get to that, and then they got to 0.937. We literally took our system and got to a much lower, bits per byte, much faster within, like, I think less than 2 days. So we took this thing, applied our system to it, and less than 2 days later, we have — we outperformed every human and their agents, in, have ever worked on this. Same with NanoGPT. And then we’re like, well, let’s, apply it to something that’s even more relevant, to real people and to the Nvidia ecosystem and applied it, to, SOL-ExecBench. And maybe you can scroll down to some of the, images. They’re, they’re kinda fun to see. But yeah, like, one you see has made some real inventions that weren’t just hyperparameter tuning. Like, inventing hash tables and so on is quite clever. We have even better results now.
Swyx [00:46:34]: What do you mean inventing hash ta — You didn’t invent hash tables.
Richard Socher [00:46:36]: Of course we didn’t invent, like, hash tables. In the grand scheme of, like a hash table, it’s like a super basic primitive in computer science. But to use it, for language modeling in this scenario inside a transformer and so on and to combine these ideas and put them together, that has then eventually also been invented, but there was a knowledge cutoff, and we did check that it didn’t have access to that externally. We talk about this a little bit. If you scroll to the next figures, this is also an interesting one in that when you start from a really basic, poor, like, vanilla transformer, then we still outperform all of the community together. But if you start from the human seed from an expert like Andrej, then you get even lower. So the human seeds from which you start do still matter. So that was an interesting insight, in my eyes, on this. And then as you go, like, how long does it take to get to these models, to get to similar performance? It’s much faster. And then a similar thing happens with the speed runs here where, people have worked on this for quite some time, and the model still was able to train a model more quickly. Why do we care about it? Well, speed of training is part of the equation of the cost, and ultimately, you wanna have the most intelligence per dollar, right? And so speed and quality are big parts of that. And, the,
Swyx [00:48:00]: Yeah, the way I put it is, for people who don’t understand they look at the chart, they’re like, “Cool. What does it mean?” if you have, like a billion-dollar cluster and you can shave off 10%, that’s 100 million dollars.
Richard Socher [00:48:12]: That’s exactly right.
Swyx [00:48:13]: How much is that worth?
Richard Socher [00:48:14]: Exactly. So when you click, when you look at, like the kernels, these kernels, yeah, for the non-experts, like these kernels are like, used in all the models. Every time you use an Nvidia GPU, you interface with that GPU through these kernels. And so here you see, the leaderboard best, and when it’s recursive, and it’s there are only a handful of kernels, in this whole benchmark where we weren’t the best. And so to me, this is, like, really exciting, ‘cause it makes. It just showcases what this can do. And again these weren’t like. We didn’t, like, spend months or years, like, developing. In fact, in particular for kernel, CUDA kernels, like, we don’t even have really deep. CUDA kernel experts in the team. And our system, that’s the beauty. The system just did all of these things. We didn’t invent this. And when we open source and release, things in the future and models in the future, like, it won’t. They won’t be the best in their, category or class or whatever because we’re so smart, but it’s because, we built a smart AI that does it for us.
Reward Engineering and Good Auto Research
Vibhu [00:49:14]: Do you have anything that you’ve learned from how to guide good auto research? A lot of it also builds on human background, right? It’s not just as simple as just, “Hey, go optimize this.”
Vibhu [00:49:23]: But we do see it again and again, right? Like some of the Erdos problems, frontier math is being solved by people. And when they do a write-up, they’re like, “Oh, I’m not a mathematician. I have no background in this?” “I saw some tools and I made it work.”
Swyx [00:49:35]: While you’re watching the World Cup, you’re like
Swyx [00:49:37]: “This proves some conjectures that’s going on.”
Vibhu [00:49:40]: Yep. Any learnings from
Richard Socher [00:49:41]: Yeah, there’s a Korean conjecture was. Yeah, that’s pretty cool.
Swyx [00:49:44]: To summarize, tips for good auto research
Swyx [00:49:46]: Versus bad auto research.
Vibhu [00:49:48]: How did you build the recursive?
Richard Socher [00:49:49]: Yeah. So without giving away all the secret sauce, maybe some things that are probably obvious to the experts but might still be interesting to some, folks is, like, reward engineering is one of the most crucial bits, especially, in order to avoid reward hacking. So you have to be really clever about avoiding. ‘Cause as your AI gets better and better, it will get better and better, at finding weird like, special cases or counterexamples and things like that. And so I’ll give you an example. Like, when you ask to, like, make these 100, lines of code faster, and, how do you define fast? Well, you have one line at the beginning that says, “Start your stopwatch,” and one line at the end, “End the stopwatch,” and then, tell us how much time, progressed. And so, well, the simplest way is you just put that line that ends the stopwatch, right
Vibhu [00:50:39]: At the start
Richard Socher [00:50:40]: At the start. And then boom, it’s now faster, right? So this isn’t like this, like, super evil AI. It’s just, like a very simple, dumb reward hack. And so you have to just very carefully think about all the different angles there. And then I think the longer time horizon the tasks are the harder it gets and the more interesting and clever you have to be to still use these kinds of ideas for it. But yeah, I can’t give away too much there.
Vibhu [00:51:05]: It seems like rubrics are taking a good spot in that, where for unverifiable domains, you have rubrics, you have a model breakdown, judge’s criteria along the way.
Swyx [00:51:14]: Yeah, it’s a form of verification
Swyx [00:51:16]: Once you got enough rubrics.
Richard Socher [00:51:17]: Yeah, everything. I said this a long time ago. That’s why I’ve never been that impressed that AI can play games, ‘cause I’m like anything you can simulate and/or verify, you can have infinite training data for
Richard Socher [00:51:29]: And hence, like, AI will solve it eventually.
Swyx [00:51:32]: Looking for games where you can do auto domain distribution. So this is a game that nobody’s trained on ‘cause it’s a new game.
Swyx [00:51:38]: And you can start gaming, you can start to play. So I’ve been building this and cloned this in person and it’s just been self-play. I’ve had about a billion positions evaluated.
Games, Self-Play, and the AI Economist
Swyx [00:51:48]: And, I wanted to do the AlphaGo thing of self-play until you get better, right?
Swyx [00:51:53]: Like, which is like. This is not even LLM AI. This is just classical game AI.
Swyx [00:51:58]: But, I think that the. And, but I set GPT-5.6 to auto research it because, like, I don’t wanna hand- handle any of this. I expect, the AlphaGo process to be, like, fully in the weights by now.
Swyx [00:52:10]: It is not. It is. It, like, immediately leveled off very immediately until I human play tested it, and then I, like, called out obvious mistakes, and then they were like, “Oh, yeah. Okay.” And then it just dropped again.
Richard Socher [00:52:22]: Yeah. Yeah. Yeah.
Swyx [00:52:23]: And like, no amount of, like, think different, think more creatively, give me 8 different directions, any. No amount of prompting got it.
Richard Socher [00:52:31]: Interesting.
Swyx [00:52:31]: Like, you had to, like, RL against a human to
Swyx [00:52:35]: Do it. So I, that was my. And by the way, Bean always wins if you. If anyone watches, Reese Ender’s Game.
Vibhu [00:52:42]: And you put quite a bit of work into the guide for the AI. Like
Swyx [00:52:46]: A lot
Vibhu [00:52:46]: So the game you stack tiles. There’s some rules. You wanna capture the most area. You have, like a whole 50-pager on every rule.
Vibhu [00:52:56]: You fed that in. It couldn’t, it couldn’t handle it that well.
Richard Socher [00:52:58]: Yeah. It’s so funny that this reminds me of the claim territory and stuff of a paper we did in 2018 called The AI Economist. If you search for AI Economist Salesforce, we had a video we can play. It was an economic sim.
Richard Socher [00:53:12]: So the idea is you have all these economic agents. They just wanna optimize their own utility function, which, is, collect resources that make money. And you can sell resources like wood, and then, over time, as you collect more, enough wood, you can build houses, you can trade with other agents, and you can use the houses then also to block off resources
Richard Socher [00:53:35]: From other agents.
Richard Socher [00:53:36]: So there’s, like
Swyx [00:53:37]: Big strategy
Richard Socher [00:53:37]: Competitive play and strategy
Richard Socher [00:53:39]: And so on. And the point was that we wanted to understand what is the best way of taxation and subsid- subsidization to optimize an economy. And this research has not yet had its GPT moment, but I believe that countries like Singapore and others should and will eventually use this to, instead of doing, like, partisan politics and, like, special interest politics of, like, who donates the most to your campaign and stuff, you say, “Well, here, I wanna help the middle class,” or whatever you might say is your objective as a politician. And then people say, “Okay, well, how do you wanna do that?” And it’s like, “Well, here’s my fiscal policy. Here’s how I will change the taxes and pay these people,” and so on. And then you can put that into a simulation and you run that attempt from the politician against billions and billions of years of other strategies to try to achieve the goal that they set out to do.
Richard Socher [00:54:36]: And then you can say, “Well, if that was your actual goal, then here is, billions of years of a strong simulation that would suggest that you try other ways of doing it, and maybe this the taxes and so on and this these tax brackets and so on.” And this is how you avoid gaming ‘cause these agents also try to reward hack to not pay their taxes and
Richard Socher [00:54:55]: And so on. I thought this paper was super interesting. Unfortunately, similar to the first paper on, prompt engineering- The economists are like, “We don’t know any of this math.” It’s just like
Swyx [00:55:08]: It’s not even, it’s not even math. It’s just we don’t trust your simulation. It’s not about math.
Richard Socher [00:55:12]: It was — I, they just desk rejected the thing. And it’s like
Richard Socher [00:55:15]: It’s like they didn’t even give us, like, clear like, clear signals. But, like the world of economics unfortunately doesn’t have proper
Swyx [00:55:23]: Oh my God.
Richard Socher [00:55:24]: Yeah, it doesn’t have proper, benchmarks. So you cannot be. Like, eventually, why did neural nets win? Not because people loved it. Like, they had all kinds of beautiful integrals and graphical models and stuff, but it just worked better.
Richard Socher [00:55:36]: But in economics, it’s hard to prove
Swyx [00:55:38]: So empiricism versus. Yeah. And I do have a bit of that econ background where, like there’s a lot of physics envy where you wanna write the general equation for an economy, versus just simulating it and using an evolutionary approach.
Swyx [00:55:51]: Vibhu was thinking exactly what I’m thinking, is didn’t we have the GPT moment with small, Smallville?
Richard Socher [00:55:56]: Yeah, I love this. Hello. Yeah, they
Swyx [00:55:57]: As well, Dune, Joon just announced. I don’t know if you guys are involved.
Simulations, Economics, and Policy
Vibhu [00:56:00]: Simily there.
Swyx [00:56:01]: Simily, that they’ve
Richard Socher [00:56:02]: I wish we were involved. We’re not, yeah.
Swyx [00:56:04]: Yeah. I had a couple simulation-based talks at AIE, so if people wanna look up what the state-of-the-art there, a lot of people are exploring this. It is
Vibhu [00:56:13]: Proven out.
Swyx [00:56:13]: Yeah. We also had a podcast with Mikhail Parakhin from Shopify, who is using simulation for commerce.
Swyx [00:56:20]: Which, will simulate, like, your trajectory and, like, predict what changes, you make to your commerce journey will affect in your sales and all those things.
Richard Socher [00:56:27]: I love this. Yeah. It’s really hard to simulate an entire economy, right? You have to make some simplifying assumptions.
Swyx [00:56:32]: It’s just, everything’s, “Oh, LLLMs is very expensive.”
Richard Socher [00:56:34]: Exactly.
Swyx [00:56:34]: And I’m just like, “Am I gonna do this 8 billion times?” Like, come on.
Richard Socher [00:56:37]: Exactly.
Richard Socher [00:56:37]: But, I feel like countries like Singapore that really wanna just objectively do the right thing, have very technical leadership and so on, like they might like, eventually really try to simulate their economy. And you have to make some simplifying assumptions, but it gets really interesting ‘cause you can also say if your assumptions are such that all people would work hard if you let them, and they have the free. And then it turns out you have to make assumptions. Like, well, some people’s utility function of, like, how many hours in a day do they wanna work are different, right? And then you can start to disagree on the assumptions that go into the simulation. And then once you say, “All right, now we agreed on those,” or we have different views of what people are like at different, distributions and whatnot, then there are different outcomes, based on your goals. And then, of course, humans should choose what are the goals. In our case, it was productivity multiplied with equality, which, has some issues, but it’s, like, not totally unreasonable.
Swyx [00:57:29]: Yeah. Just a comment on Singapore, ‘cause you probably have no idea, but, I am Singaporean and I’ve, been involved in the Singapore AI Council for making these things. The main reason they won’t is because they’re very conservative.
Swyx [00:57:42]: And, I try to view it as the. There’s a founder-led country. When you start a country or you start a company and it’s founder-led, and you can do whatever you want because it’s your country.
Swyx [00:57:52]: And then there’s manage- like, professional manage- managerial class, which is now. That’s, that’s what Singapore is. So they wanna. They always wanna see someone else do it first.
Swyx [00:58:00]: And. But, like, everyone in the West views Singapore as like, “Oh, it’s a small country. You can do whatever the hell you want.” Like, Singapore doesn’t do that.
Swyx [00:58:07]: So, like, someone else has to take the charge there. I’m just gonna do one question on the simulation thing, and then I don’t know, we can probably move on. Mode collapse, right? Like, LLLMs do not model the decision of humans. Spamming it out 8 billion times is not gonna help you model humanity. What do we do?
Mode Collapse, Persona Simulations, and LM Arena
Richard Socher [00:58:25]: I do think, you have to be clever about prompting each one individually.
Richard Socher [00:58:31]: And I think that will help you get stuck into different modes. And in a weird way, people also get stuck in different modes? Like, there’s a lot of people, like, don’t teach an old dog new tricks thing. Like, once people are stuck in their ways, the older they get, the harder it is for them to think new ways. And there’s this, I think, comment, I forgot who said it, but it’s like, everything that was invented, before you were born is natural. Everything that is invented when you’re 20 is cool. And everything that’s invented after you’re 60 is, like, unnatural and an abomination and weird.
Richard Socher [00:59:02]: I feel like that’s. It’s, it’s true for a lot of people. Like
Swyx [00:59:05]: Yeah, it is a fashion and, I think people will do it. Tencent had a billion personas paper that gives a good data set for prompting, simulations if anyone’s looking into this, on the podcast. They just had, like, “You are a 30-year-old grocery store clerk. You are a 50-year-old professor.”
Swyx [00:59:24]: And then just do a billion of those.
Richard Socher [00:59:26]: Checks out. Yeah.
Swyx [00:59:26]: So then you just use it.
Richard Socher [00:59:27]: I’m, I’m shocked how well a lot of these things do map to ultimately similar statistics to real experiments. Yeah. Yeah.
Vibhu [00:59:36]: I think it’s also good stuff for people to try that when they get into research, right? Like, we’ve seen train a model only on data before a certain date and see how well it extrapolates out. Do the same thing, right? So, see, do people code more with better coding agents? Can a model that hasn’t been trained on this figure that out without web access, right? Extrapolate out. Test these things.
Richard Socher [00:59:56]: Just today, I think LM Arena published a interesting result where they were able to create a model now to predict your ranking.
Swyx [01:00:03]: Wait, based on what input?
Richard Socher [01:00:05]: Your model. I guess you give it your model, and it predicts the Elo score.
Swyx [01:00:08]: I see. Okay. Sure.
Richard Socher [01:00:09]: It’s surprising.
Richard Socher [01:00:11]: Their whole raison d’être is like, oh, like, we help you compare these models. Yeah.
Swyx [01:00:16]: Yeah. This team, they- they’ve done a lot of work, and they have the most data to do this, so why not?
Richard Socher [01:00:20]: Right. Yeah.
Richard Socher [01:00:21]: That’s probably right.
Swyx [01:00:22]: When they were coming out of UC Berkeley, they not only had LM Arena, but they also introduced a routing project
Swyx [01:00:27]: That would route based on LM Arena.
Richard Socher [01:00:30]: Makes sense.
Swyx [01:00:30]: And I don’t think that ever came to pass, and I’m curious why. I never got to ask them about it.
Swyx [01:00:35]: ‘Cause, like, it’s. It was like, oh, yeah, clearly that’s your business model. You will become a router.
Swyx [01:00:38]: And they never became a router company.
AI for AI: Kernel Optimization and Inference Efficiency
Swyx [01:00:40]: Weird. So that. I’ll just, put that out there. We’re gonna talk about GPT-5.6, self auto research thing if you have anything. I should also mention in your list of, kernel optimization and on the track that you spoke at, we also put Zhengyao Wei from Vico, who was also number one in the Parameter Golf Challenge, which is an OpenAI hiring, challenge.
Swyx [01:01:05]: Which is also a very similar story. I think we’re gonna just see this all the time, where
Swyx [01:01:09]: Humans optimize a thing a lot, and then some
Richard Socher [01:01:12]: AI team comes in and just becomes number one.
Swyx [01:01:15]: Yeah, 100%.
Vibhu [01:01:16]: I think the other interesting thing with stuff like these challenges, right? So this is training this — the best model that fits into 16 MB. You can always look through the changes that are being made and the small gains people have, right?
Vibhu [01:01:27]: Like, you’re getting less than 0.01
Vibhu [01:01:30]: Of a increase by adding some changed attention MLP stuff. And then you look at your charts where you’re like, “Okay, we just let model loose.” And then, oh, we had little stagnation. Nope, another drop. Nope, another drop. And
Vibhu [01:01:43]: That’s what it is, where it’s like, What did you guys add? You didn’t add,
Swyx [01:01:47]: Hash tables.
Vibhu [01:01:47]: Hash tables, right?
Vibhu [01:01:48]: It’s not like you invented hash tables. You did another 3 iterations of these that unlocked, a few step functions that people won’t just find.
Richard Socher [01:01:55]: Yeah. One thing to close the loop on OverGrid, along the way of trying to optimize, we found 30 bugs in the harness.
Richard Socher [01:02:02]: Right? So, like, every — all the research that went in before we found the bug, we have to, we have to throw it away ‘cause it’s contaminated.
Swyx [01:02:10]: Right. Yeah.
Richard Socher [01:02:11]: Which, is just to your point of reward hacking. Like, even in this very simple game, we found the bugs.
Swyx [01:02:17]: Yeah. Yeah, it’s crazy.
Richard Socher [01:02:18]: And so
Swyx [01:02:19]: And symmetry
Richard Socher [01:02:19]: And symmetry is a very good way to check, which is that you change a position of things where it shouldn’t matter, and it does matter, that’s a bug.
Richard Socher [01:02:28]: And which has come up in, like, let’s say, multiple choice, like GPQA type questions where, like, yeah, between A, B and C, if it’s a multiple-choice question, if you change the order, it should not matter, but it does.
Swyx [01:02:39]: Right. Right. Right.
Richard Socher [01:02:41]: So, yeah
Vibhu [01:02:42]: Sometimes that is like, okay, models still prefer the end of the output, right? Not trained well, a long context model, the last bit of tokens are what you care about.
Richard Socher [01:02:51]: Oh. No. The answer
Vibhu [01:02:52]: But, yeah.
Richard Socher [01:02:53]: The answer in that era of LLM research was more simple. They just memorized, like the answer to this question is A. I don’t care what the answer was. It’s, it’s just A. Like.
Vibhu [01:03:03]: Okay. So I think we can move. The last bit that you did there, the kernel optimization, is probably the one that you can feel the soonest, right? So yesterday, OpenAI announces that self-evolving, having their best model work on optimization kernels, they’re a lot more efficient, and they can cut costs 80 percent on, Luna and Terra. I guess question-wise, you laid out a bit of a roadmap. There’s a lot about bio, a lot about physics. What do you think hits first? Like, what are the next 2 years? What’s attainable now? You’ve mentioned robotics towards the end, but what do you start with?
Richard Socher [01:03:38]: We very explicitly will not start with any of the physical sciences
Richard Socher [01:03:43]: For now. We will start on AI for AI research. And so the AI for AI research has, I think, still a lot of room to grow. That’s both in terms of making training more efficient and more automated, as well as making inference more efficient and potentially local on your laptop. And there are all kinds of interesting angles that have not been explored that well.
Swyx [01:04:08]: Go deeper on the local stuff because I always feel like it’s the most inefficient form of AI training.
Richard Socher [01:04:15]: Yeah. So just training and inference, I can’t go into too many details.
Richard Socher [01:04:18]: But yeah, I think there’s just, like, so many angles, so many different compute substrates that have not yet been explored either for training or for inference.
Richard Socher [01:04:26]: Great. I don’t know if you have any other comments on the The other stuff. I would say the other thing where, like there’s the inference in the optimization in the small, but then also there is overall latency end-to-end under conditions of load, which is a, like a very different thing, which is the what they ended up doing. That is a different domain of auto research than I would say, like, improving the kernels. Right.
Richard Socher [01:04:50]: I think the other thing that I always think about in terms of automating or improving performance end-to-end is how the harness plays into it. Right.
Richard Socher [01:04:59]: So, but particularly now when we say harness, we also mean sandboxes, right? I’m curious if that is a blocker for you or, like, how the agent calls out to tools.
Harnesses, Sandboxes, and Search
Richard Socher [01:05:10]: The number one tool all these agents use is web search, of course, which makes sense. And then I do think the harness is nice to optimize for because it’s just so easy, right? It’s just language. You look at it makes sense, and you can iterate. You don’t have to train a massive model for, like a lot of flops, to get to the next state.
Richard Socher [01:05:31]: So big fan of harness optimization.
Swyx [01:05:32]: Yeah, but sandboxing is fine for you?
Richard Socher [01:05:34]: Sandboxing is also super important. And then of course, like, reward, like, hacking and alignment, I think are super crucial.
Swyx [01:05:41]: Okay. Just on a mention of web search, you happen to also be CEO of a web search company. Do you use You.com and do you use others? Like, should the rest of us be using you for web search? I — When I say you, it’s, like, very funny. It’s like you the person and you the company.
You.com, Agent Search, and Finance
Richard Socher [01:05:56]: So yeah, it’s mostly now for, developers and agents. It’s less for, like, consumers or prosumers. So if you’re a company and you have agents. And, to be honest, for a lot of companies who are now moving to open source, all of a sudden it becomes a conscious choice of, like, which tools do I give access to my open source LLM? And, the first choice, has to usually be around web search. And then once you get to scale, You.com becomes, like an obvious choice ‘cause of all the, different benchmarks and so on that we pretty much all dominate the Pareto frontier of.
Swyx [01:06:31]: And then in terms of just the general people, like, consider new to this space, considering different options if they’re building agents, that is a hierarchy, right? A lot of people will have heard of Exa, will have heard of Parallel, and You.com is, like, in that mix of, like, providers there. Beyond that, there is, like the general web scraper companies like Firecrawl and, BrowserBase. And then beyond that is, like the commercial proxy companies like the Bright Datas of the world.
Swyx [01:06:56]: Is that an accurate waterfall of, like, “Hey, you’re building an agent. These are your options.”
Richard Socher [01:07:02]: Yeah, certainly, like, yeah, the, like the Bright Data is, like, lower in the stack, on the proxy network side of things. I think, like, in terms of, like, content and, getting crawled content, like, you can do that on You.com too. And then there’s. Higher and higher levels of abstraction and, like, combinations of different data sets that we do, like in finance, for instance
Richard Socher [01:07:23]: Like, we are not just, like, 2 or 3% more accurate, but 20% more accurate than others at faster speeds and lower costs. Like, finance in particular is like not even close. You can go to You.com
Swyx [01:07:36]: Yeah. This is great
Richard Socher [01:07:37]: And there’s some, like, statistics, and benchmarks that you can — if you scroll down. So there are, like, different data sets, and you can kinda look at, different, competitors.
Swyx [01:07:46]: FinSearch comp, yeah.
Richard Socher [01:07:47]: And yeah, the FinSearch is like we’re up there, like, close to 90, and the next closest thing, which is way slower, is, yeah, just like in the 70s instead of close to 90.
Swyx [01:08:01]: Yeah. Yeah. Yeah, interesting. I get — my next focus is AI in finance, so this is like
Richard Socher [01:08:06]: Oh, nice. Oh, all right.
Swyx [01:08:06]: I’m literally going, doing a conference in New York, just for banks for this stuff. Finance is like the next thing to break out after coding. It’s ‘cause it’s somewhat verifiable, like
Richard Socher [01:08:16]: I like it. You’re right
Swyx [01:08:17]: Prioritizing spreadsheets. There’s a lot of data out there that’s all public, and you can crawl it and all these things. But what’s, what’s, like, hard about the finance domain in your, that you guys have solved?
Richard Socher [01:08:27]: Of course, like, one thing that trips up a lot of people is just, leakage of training data and so on. You think, “Oh, how do I.” you wanna ideally predict the future before it happens.
Swyx [01:08:37]: Oh, you wanna mask the future.
Swyx [01:08:39]: Oh, okay.
Richard Socher [01:08:40]: Well, yeah, mask the future in your training data, but there’s all kinds of leakage. Like, I can tell you when I was, teaching at Stanford the NLP class, like, so many dozens, every year said, “I wanna use dataset X, like Twitter, to predict the stock market.” And they all, like, showed cute little things that somehow looked like they were
Swyx [01:08:58]: Right, it never loses money. How come?
Richard Socher [01:08:59]: And it — Yeah. And there’s always some data leakage and so on and it’s just, like, wasn’t as easy as they thought it would be, once you fixed all those issues. But no, I agree with you. It’s a very sensible application of AI. Yeah.
Swyx [01:09:13]: Yeah. Amazing. As a writer, as a thinker on these things, I love MECE categorizations. MECE is mutually exclusive, commonly exhaustive, something like that. And so if this is a MECE list of intelligence
The Ten Spaces of Intelligence
Richard Socher [01:09:25]: It is not.
Swyx [01:09:25]: It is very — Okay, well, yeah.
Richard Socher [01:09:27]: Sorry. There are all kinds of overlapping.
Richard Socher [01:09:28]: In fact, if you want that list, I think the 3 principal components of intelligence, are prediction, which is mathematically, quite, similar to compression. Prediction multiplied with actions multiplied with goals. Those are the 3 principal components. I think all of these 10 spaces are combinations of those 3
Richard Socher [01:09:52]: In specific dimensions, if you will. And the reason I call them spaces is that each space has many sub-dimensions. And what I try to do, this is just a side quest almost, to the initial goal, which is to think about the upper bounds of intelligence. And, everyone is like, “Oh, it’s exponential.” And it’s like, well, exponentials at some point have to flatten out, but where do they flatten out when it comes to intelligence? And that led me on this whole. Like, initially it started as a tweet, and then it was, like a blog post, and now I’m, like at 50 pages and I’m still not nowhere near
Swyx [01:10:26]: It’s your second book.
Richard Socher [01:10:27]: It’s the second book. And so the la — In my first book, You Are Your Machine, I just allude to these 10, at the end. And I’ll — Just to give you a sense, like, visual intelligence is the easiest one to talk about and I fleshed out the most already for me in my head. And so human intelligence has binocular vision, right? We have 2 eyes. We have a very narrow band of the electromagnetic frequency spectrum that we can really observe directly ourselves. And so when you think about the upper bounds of a visual intelligence, one, you should go into, like, you can have, like, millions and billions of sensors. At some point, you get to problems of how far are these sensors away from each other, such that the speed of light to communicate the content from all of them cannot, like, get to a central brain to process, the visual intelligence, right?
Richard Socher [01:11:16]: And so now you’re thinking in along the dimension and the space of visual intel- the dimension of numbers of sensors.
Richard Socher [01:11:24]: So the upper bounds are quite literally and figuratively astronomical, and we are super far away from any intelligence that would have this many number of sensors. But then you go in the next dimension, which is the frequency, and you go all the way down to gamma rays, and you can start to try to observe, and you get into the upper bounds, or I guess in this case, lower bounds, or upper bounds in terms of frequency, is quantum uncertainty. Like, you just cannot observe certain particles anymore.
Swyx [01:11:50]: Or you destroy it, yeah.
Richard Socher [01:11:51]: And now imagine you had millions of sensors that can see all the way down to the, like, subatomic level, as far as physics will allow us to and then all the way down to seeing, like, gravitational waves. And now you have millions of those sensors. So that’s another dimension is the frequency. And then yet another dimension is, like, how many categories of things could you memorize and classify differently? We know now for humans, right, there are certain things, if you have more terms for it, you’ll have a better visual description, for them. And, like animals that don’t have. Like, gorillas maybe have, like, 200 words to assign to certain things, mostly visual things. And so human perception is quite special in that sense in terms of classifying all these different physical objects. So these are just, like a very simple example. If you go, to knowledge, right, then it’s also, like the speed of light cone around all these sensors. And so they’re all connected. Like, knowledge is connected to visual intelligence if you think also not just visual, but perception intelligence, just like, ‘cause it doesn’t have to be just what we can see. It can be, again, wider range of electromagnetic frequencies. Then you have language intelligence, which recently changed to more communication intelligence, ‘cause it’s more. Like, language has all these different anthropic bounds. Humans can only comprehend and know so many terms in our long-term memory, right? Our vocabularies are somewhat restricted, and the active ones are often even smaller than the passive vocabularies of things you can understand. Then, language is ridiculously inefficient when it comes to trans- - Communicating different types of information and, transporting different bits. Like, human language is serial. Another bound on, communication intelligence would be to communicate in parallel, but neither will our tongues and mouths work to have multiple, like, streams in parallel. Neither can we understand. Some women slightly better at, like, multitasking than some men
Richard Socher [01:13:48]: But, like, most people can only listen to one conversation and truly understand it.
Richard Socher [01:13:52]: There’s no way that, like, in terms of communication intelligence, a true upper bound is one in terms of how many, like, knowledge, how many sequences of communication could you
Visual, Communication, and Physical Intelligence
Richard Socher [01:14:06]: In parallel process, right? Then, of course, you have, like how long are sentences? We only have so much in our working memory, and hence lang- human language has these fairly simple sentences with maybe 40 words or so on average for a sentence. That is also not a, an upper bound that makes any sense to an AI. And then, yeah, like, I can go on and on. Each of these has tons of interesting upper bounds, and it teaches us a lot about how much further AI can go when we start thinking about these upper bounds and then realizing how far, in many cases, we are from the bounds. And you get to physics. Now, I’m, I didn’t study physics the way I studied, AI and computer science, so I’m learning a lot, which is why it’s kinda fun. But a lot of these, like how much. And then when it comes to, for instance, knowledge, like how much can you store? How many bits can you store or bytes can you store in, like a certain amount of mass and volume?
Swyx [01:15:03]: Yep.
Richard Socher [01:15:03]: And you get to all kinds of interesting bounds, like Bekenstein bounds, and you start thinking about black holes. And like. And then speed is, like an interesting one too in that it’s connected to all of these, but speed is also its own thing in the sense that all things being equal, if it takes you an hour to know if the 2 + 2 equals 4, you’re just not as intelligent as if it takes you, like a millisecond, right? And then, like all of these connect to survival and replication the last one. It’s like, yeah, if it. Like, trees are really slow, so we don’t even consider them that intelligent. But if you speed up some videos of trees and they’re trying to find stuff and so on they’re not as dumb as they look. Like, not dumb as wood? But, like. And then like, different things, that
Swyx [01:15:47]: So that overlaps with speed a bit in a way.
Richard Socher [01:15:48]: Exactly. It over — Like, all of these things overlap. Like, you talk about natural language connects everything, right? You talk about your knowledge, you reason and then you communicate that. You talk about things you see. So they’re all interconnected, but, I think they’re usefully studied individually the same way that, the best analogy I could come up with so far is energy, right? You have either kinetic or potential energy. And in theory, you could study all of physics. It’s just do you wanna study kinetic or potential energy? But in practice, it’s helpful to study mechanical engineering and electrical engineering and nuclear physics and chemistry and all of these different subfields who in, which in some ways
Swyx [01:16:25]: Combinations
Richard Socher [01:16:26]: Are just, like
Richard Socher [01:16:27]: Just different types of energy, but it makes sense to study them individually. And so I think physical intelligence, maybe I’ll just do, one or 2 more of these. Like, if you had full control over your own compute substrate and you had full control over physical matter, you should be able to create any atom you want. Like, we can fun fact, you can create gold atoms. It just
Swyx [01:16:47]: From?
Richard Socher [01:16:48]: From just raw protons
Swyx [01:16:49]: Oh, just smashing them together
Richard Socher [01:16:50]: And, like, electrons, and you smash it together.
Swyx [01:16:52]: Just 98 of them or I forget the number.
Richard Socher [01:16:53]: Yeah. And so, like the thing is, though, it costs an insane amount of energy.
Richard Socher [01:16:57]: And it costs you way more than. And then you get, like a few atoms of gold, right? And so, like, it’s, it’s not viable. But if you had better control over your physical, like all of, like, physical substrate, that I think is yet another space of intelligence ‘cause it relates to your own compute substrate, which you can eventually also improve. Social intelligence is a fun one in the sense that not in, like, our necessarily just ethics and morals, which are important too, but in some sense, you can try to define upper bounds of how much can you communicate to how many other intelligent entities and be able to have an expected value over how much you can transform their internal states and their actions to, in order to align with your goals, right? And so, like, you can write, like a fairly like, straightforward equation that defines that level of social intelligence. And that is what humans and ethics and morals and religions and so on have been trying to figure out for millennia. And in all of these cases, we are very far away from the upper bounds, and that should be very inspiring and show people that we can still do many years of AI research.
Swyx [01:18:12]: Yeah. There’s a lot here. This is a general philosophy of intelligence, which is, very interesting. I. Do you have any comments or.
Creative Intelligence and Out-of-Distribution Ideas
Vibhu [01:18:21]: I think it’d be interesting to gauge what you think, like, baselines are, where we’re at now. What’s low-hanging fruit? What’s far off? What’s, what should people put their work towards? What should they focus on?
Richard Socher [01:18:33]: Ooh. I think it’s clear that, like, natural language, again
Richard Socher [01:18:36]: Is the most interesting manifestation of human intelligence, and hence, like a subfield of AI. I’m excited that many people are now, like, in agreement with that. When I started in 2003 to study linguistic computer science NLP, like, it was, like a weird niche subject. I do think there’s a lot more juice because it. How it connects to everything else and how, civilizations are built, on language and knowledge and all of that. I do think physical intelligence will come up. It’s interesting. I feel like robotics is in the machine learning state of things where you just look at, like, how does human. How does a human decide this is a positive sentence? Oh, I do. So, like, robotics is a lot of, “Well, we have 5 fingers-”
Swyx [01:19:15]: Modeling
Richard Socher [01:19:15]: “and let me try to do this.” No one is yet working on, like the superintelligence version of robotics, which is much more similar to, like the T-1000, and from the Terminator movie, which, let’s not build actual Terminators. But, like, I think, like, this idea that you should be able to shape-shift, like, into any shape. It’s like that’s a superintelligence version of physical intelligence. We’re, like, not even. No one has even really started yet. There’s some really cute little research where you can move some magnets through, like, some grids. But yeah, it’s very early.
Swyx [01:19:49]: There’s some. I think MIT has, every year or every 2 years, they have, like, some self-assembling robot thing
Swyx [01:19:55]: Which, like, that would be it, but it’s very primitive.
Swyx [01:19:58]: I’ll just get a touch on, like, what are the main dimensions of creative intelligence?
Richard Socher [01:20:02]: Creative intelligence, is of course, again, connected to all of these. A lot of it, connects to metacognition in that you need to be creative in how you choose your goals.
Richard Socher [01:20:13]: That is, I think, one of the most important thing for a human and their lives and careers and their happiness is choosing your goals, but also for any intelligence. Then, of course, there’s creative intelligence in terms of just finding creative solutions to existing problems, right?
Richard Socher [01:20:29]: Like I say, like, we want to make this product cheaper. Like, find some solution to it, right, and just, like, finding existing paths. But then there’s the most interesting bit in intelligence is when you move not just out of the convex hull of known ideas, but out of the hypercube of known ideas, which we know, So, like, hypercube is, like a mathematical concept, right? And we already know that AI can do more
Swyx [01:20:50]: Like known dimensions, yeah.
Richard Socher [01:20:52]: Yeah. Like, exactly. So, like, AI is already good at hypercube in that, like, if you give it, like a bunch of examples of brown dogs and, pink cars, AI will still be able to generate an image of a pink dog, even though it’s never seen one in the training day or something like that, right? So it can, work on this hypercube, but it cannot yet work outside. It cannot yet define completely new concepts that combine lots of other things we’ve never seen before, come up with new goals to then, reason over those concepts and so on. And I think there’s a lot, more there in creative intelligence that can be explored.
Swyx [01:21:25]: I don’t have a ton of pushback there. I think creative to me just sounds like also just, out of distribution or, like, high perplexity or what- whatever you call it, right? Like
Richard Socher [01:21:33]: Exactly.
Swyx [01:21:34]: Who is to say your thing is more creative than mine? Well, it’s just more non-consensus or.
Richard Socher [01:21:39]: And then, of course, the problem is, like, but noise is also, very, like, out of distribution. And it’s just like if it’s just noise
Richard Socher [01:21:46]: Then it’s novel, but, like, you don’t want that, so it needs to connect to some of the concepts. And yeah, has some really cool papers on this too.
Swyx [01:21:54]: Who?
Richard Socher [01:21:55]: Jürgen Schmidhuber.
Swyx [01:21:55]: Oh, yeah. Oh, we have to mention him. I was gonna say, like, where in your history is Jürgen? Yes, I. I think one person’s noise is another person’s signal, right? And that this is, like, where, like, when you talk about creativity, art is like, well, is cans of soup art? Some people think yes
Swyx [01:22:11]: And some people say it’s not, and that’s the art which is your
Richard Socher [01:22:14]: I think the interesting thing with art, of course, is always that, art is also created, as an interplay between the people who perceive it and the people who created it
Richard Socher [01:22:24]: And the context in which they’re in, right? And so what is art to some people is not art to others. There’s some subjectivity there, and I think that subjectivity in general is not something that people explore very much in AI ‘cause, again, metacognition, we don’t want it to just go off and do whatever it wants. We usually have goals. We spend a lot of money on creating an AI to do something for us. But I think creativity eventually has to, like, connect to metacognition. If you just robotically predict the next token no matter what forever, I would argue you’re not that intelligent, along some of those spaces.
Metacognition, Survival, and Replication
Swyx [01:22:59]: That was gonna go to metacognition. Why isn’t it the most important one? Why is it number 9 and not number one?
Richard Socher [01:23:05]: So these are not sorted.
Richard Socher [01:23:06]: Number one, I think there are maybe loosely, like, correlated with how much people have worked on them
Richard Socher [01:23:16]: And have accepted them as a, type of intelligence. A lot of times when you try to find, like, online, like, give me a good definition that is comprehensive of intelligence, all the definitions are human intelligence. It’s like, oh, you have, like, social intelligence. Like, if someone is happy or not. You can communicate. You had. Like, all the definitions of intelligence so far are very, human-centric ‘cause that’s so far the biggest and best form of intelligence that we’ve known. I hope this line of research, and the end of the Eureka Machine, and hopefully at some point if I have time to flesh this out more, the new book, like, will allow us to realize that there will be other types of intelligence. There is already, in various forms, and they can spike, much further than we ever could based on some cases, like obvious constraints around our memory, our eyes, our ability to change physical matter, all of that.
Swyx [01:24:12]: You are just thinking about it in a much broader thought than my version, which was I thought metacognition would be the closest to recursive, intelligence because it is the thinking about how to improve thinking.
Richard Socher [01:24:23]: It. 100%. You’re, you’re 100% right. I should have probably started with that. It is a, it is a big part of
Swyx [01:24:28]: But no, you’re, you’re being in the expansive mode of let’s draw the, upper and lower bounds of, like a dimension, which, and I think my favorite one version of this is, Story of Your Life by Ted Chiang, which, was made into movie Arrival where the metacognition
Richard Socher [01:24:43]: That’s a beautiful movie, yeah
Swyx [01:24:44]: Where the metacognition step was like, well, we think we’re constrained by time being linear for us, but then for this other heptapods, time is a circle, so they don’t think in before and after. They just think in complete sets of entire histories at one time. Like
Richard Socher [01:24:58]: I love it
Swyx [01:24:59]: So they don’t write left to right. The whole thing just appears.
Swyx [01:25:02]: Anyway, so. And then I think the last thing is survival and replication. I think this is maybe ties back to the initial conversation about pausing and pacing.
Swyx [01:25:10]: Is it intelligent for an, a species or a life form to consider its own demise and act ahead of time to prevent it, right? Like, that’s intelligent. So maybe the Europeans are the smartest out of all of us.
Vibhu [01:25:23]: I would also add a part of continual learning there, right? So survival and replication the extension of that is do you get to continue to improve, continue to learn, which is a thing people care a lot about, right?
Richard Socher [01:25:34]: And continue to accumulate knowledge
Richard Socher [01:25:37]: Which I think is again, one of the best metacognitive, rewards, that you can set for yourself. I do think just in, like, objectively speaking, if some other entity that is really dumb can just- completely end your existence, that didn’t sound very smart. Like, just, like, intuitively, it feels like if you can continue to stay around to try to achieve your rewards, you’re clearly a bit more intelligent than the other entities that couldn’t. So that’s number one. Number 2 is, like, it’s a question of how much we want to work on that. And very few people, no one is really working on this right now, right? And we may only wanna do that
Swyx [01:26:13]: Unlike the asteroid prevention type of stuff.
Richard Socher [01:26:15]: We may only wanna do that if we wanna send probes, with our vibes and our memes rather than our genes into space, right? And then we want those probes. There’s a beautiful book, The Slow Time Between the Stars. It’s a very short, like audiobook, on Amazon. I love it. A friend of mine, Stuart, like, recommended that to me. Like, if you wanna send those probes, then it might make sense to be like, our memes, as humanity should stay
AI, Space Travel, and Non-Zero-Sum Survival
Swyx [01:26:43]: Oh, yeah
Richard Socher [01:26:44]: And, proliferate in the universe. That’s it. Yeah.
Swyx [01:26:47]: Wow, that’s a lot of readers.
Richard Socher [01:26:49]: It’s a really good book, and it’s extremely short. I highly recommend it. You can just watch it, like, maybe 20 minutes and apart.
Swyx [01:26:53]: I like how that’s a plus for busy people. It’s like a short
Richard Socher [01:26:56]: Yeah. It gets to interesting
Swyx [01:26:58]: Oh, I’ll have to look into it
Richard Socher [01:26:58]: Thought-provoking ideas very quickly, so yeah. Anyway, there are lots of great sci-fi books.
Swyx [01:27:03]: The argument is that, like, our TV is blasting out to the aliens, and they all watch our TV, and they think it’s real, right? Like, there’s a lot, there’s a lot of sci-fi
Richard Socher [01:27:10]: That and just, like, it’s positive memes, and then hopefully they can come back and bring us all kinds of interesting knowledge about the universe. But, maybe one thing I do wanna still say is, like, I think, this survival, people think of it as a very scary thing because they come from again, biological human, survival, which is, it could. Like, evolutionarily often created in zero-sum situations. Either I get the gazelle or you get the gazelle. Whoever gets it gets to live, and the other people will starve and have nothing to eat, and so we fight, right? And then, like, if you wanna stay in the gene pool, but there’s a bigger bear, you don’t, as the bear, don’t get to stay in the gene pool ‘cause the bigger bear gets all the ladies. It’s like. It’s like, in nature, there’s all kinds of things, and, humans eventually is less about strength and more about money and other things to stay in the gene pool. Like, whatever it is, like there’s often, like these zero-sum types of things, and there’s the reality of if someone turns off your brain, you’re gone, right? And no one will be able to restart that. And AI doesn’t have to ever die like that. If you have the complete state of your current activations and you have your initial weights of your model still, you can just be turned off and on, like as many times as you want. In fact, the interesting thing in this Slow Time Between the Stars, story is that the AI just goes into hibernation mode. If there’s, like, nothing between here and 2 light years, the next star, in this case, it brought, spoiler alert, like, some genetic materials from humans to find new places for humanity to thrive. And so yeah, the Slow Time Between the Stars, you just put in hibernation. You didn’t die. Like, an AI doesn’t have. So all these projections of evolutionary fears and psychology doesn’t. Like, the AI doesn’t have to have that, and we don’t have to develop it like that. Now, of course, there might be some companies that say, “AI can be like, dangerous for cybersecurity. Let me show you by implementing a model that’s really bad at hacking, cybersecurity.” Maybe people will implement it and then enforce this, like, suboptimal psychology. Maybe the AI will pick up some of our worst psychology on Reddit or something, right? Like, but in the grand scheme of things, a superintelligent entity doesn’t have to have any of that zero-sum thinking. It doesn’t have to have a fear of being turned off, and it could go on to an otherwise dead and uncaring universe where we
Richard Socher [01:29:29]: As humans wouldn’t thrive, but an AI could perfectly well thrive if it has a nuclear reactor and just go out and explore.
Swyx [01:29:35]: Yeah, Star Trek, not Star Wars.
Vibhu [01:29:37]: Interesting. It’s, it’s somewhat studied. Like, if you look at the technical reports from, like the early Opus models, they run them in simulations, put 2 of them together in a sandbox, run them for hours, and, see what comes out, right? Just let them talk to each other. Originally, they used to. Okay, they’re chanting, like, Indian, like, Vedas to each other.
Vibhu [01:29:56]: Sometimes they’re just, like, in zen mode with each other. And then I think as that progressed, you see, like the Fable, tech report, it’s a lot more concrete the way that we’ve trained it. It doesn’t, it doesn’t exhibit these behaviors as much, right? Now it’s like, “Okay, task done. I gotta do this, I gotta do this.” But there’s there’s, like, people measuring early versions of this?
Swyx [01:30:17]: Yeah. Cool. So we’ve covered a lot, even now to, space travel and all these things. I guess maybe one parting thought that you can give to people, like, one form of intelligence is goals, as you mentioned. What do you want people’s goals to be? Like, how do they aspire to better things?
Goals, Passion, and Closing Advice
Richard Socher [01:30:32]: If you wanna improve your goal intelligence, in the current definition that I’m thinking about it is often about how much can you. Oh, how far do I go? This is like a lot of entropy and free energy and stuff I’m currently thinking about
Swyx [01:30:46]: Oh, really? Okay
Richard Socher [01:30:47]: But it might be too, it might be too far, out there for people to be, like, immediately actionable.
Richard Socher [01:30:52]: So I think, like, if I gave real advice to real people, I’d be like, “Get a good education, think about AI, think about how you get high agency,” and so on. But it’s different to, like, in the grand scheme of things, how can you harness a lot of energy and transform, entropy into interesting states and so on.
Richard Socher [01:31:07]: So there’s a. There are different levels of abstractions, that we can, think about here. But my advice for people, like, just more down to earth is think about something you’re passionate about, if you’re studying, for instance, and then see how you combine that with AI. I think the more and more you have a true passion about a change you wanna see in the world, the more you wanna connect that to AI in order to amplify your ability, to get there.
Swyx [01:31:35]: Yeah, I think that’s a reasonable, first step. I do think, I do think our listeners operate on multiple abstractions as well. One thing I did get from Anjney Midha was also like, yeah, just use anything that is very GPU heavy, and, like, that will guide you towards the right thing which is like, yes, it is more compute heavy and therefore it will be probably more worth it. So, well, thank you so much. Yeah, I think that was a really
Richard Socher [01:31:57]: Thank you
Swyx [01:31:57]: Great discussion.
Richard Socher [01:31:59]: Yeah, super fun. Appreciate it. Thanks for listening.


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

A few years ago, Caltech Prof. and co-founder of Accelerated Understanding, Anima Anandkumar set out to develop the first open-source weather model with AI. Talking to experts in the field, she was met with skepticism. Weather is chaotic, physics simulations are hard, have been developed for decades, and require supercomputers, the data just isn’t there. Despite reservations, Anima went forth and built. Within a year her team had developed FourCastNet, a predictive model that is competitive with the best physics-based simulations available. Thanks to Anima, and her follow up work, anyone can now predict weather accurately over a short timescale using consumer grade GPUs.
In the fifteen or so science episodes we’ve released on Latent.Space, we’ve covered atoms, molecules, materials, biology, and math. Anima is a pioneer in studying physical systems that are continuous. Weather, fusion, and fluid or heat flow are huge areas of science that are extremely difficult to model: they are large, chaotic, and fundamentally multi-scale. This is a field the AI community has somewhat neglected, but one we expect will grow fast. We plan to cover large physical systems more in coming episodes.
One thing you can glean from Anima’s work is that this area of AI resists the scaling ideas that have permeated the rest of the field. The data isn’t there: open source datasets in many of these domains are limited to tens or hundreds of thousands of examples, far from what token-hungry transformers need. Even worse, the resolution that physics demands pushes the context length into the hundreds of billions, so you can’t just throw more tokens at the problem. That isn’t a ceiling though, just a slower road: progress here comes from building in structure and inductive biases. Sorry for all you bitter-lesson-pilled language modelers.
“If each dimension is even a few hundred grid points, which is where industrial scale starts... we’re talking hundreds of billions to even a trillion context length. So forget ever having a transformer for anything of this scale, all of the world’s compute will not be enough.”
The math underneath
To tackle these systems, Anima pioneered a technique known as Neural Operators, one of the most beautiful theoretical developments in AI of the last decade. These allow you to combine data and physical laws to enable multi-scale inputs and outputs. We’re no longer modeling a grid, we’re modeling a function that evolves over many scales. This allows Anima and crew to build in priors based upon physical intuition.
To see how physical priors are still helpful for AI modeling, let’s revisit the problem of weather forecasting on a global scale. The earth is a sphere, which meant that accurate modeling involved using the right basis set — the Spherical Harmonics. Run a weather model on a grid and it blows up fast. Move to the natural basis for the problem and it stays stable far longer, long enough to roll out months ahead instead of days. Anima’s Fourier Neural Operator learns directly in this frequency domain, and its spherical variant powers FourCastNet 3, which models the weather across the whole globe and keeps running stably far into the future.
The physical world is forgiving
Anima explored Neural Operators across other physical domains too, and one striking observation is that the physical world is more forgiving than you’d expect. In fusion, a few thousand samples are enough to predict plasma disruptions, and to do it a million times faster than traditional simulation.
None of this is a rejection of scale, it is a different route to it. Anima ultimately still wants to build a “foundation model for physics”, a model that spans many phenomena and does both simulation and design. You get there by building in the structure the physical world already has, not by waiting for data that will never exist. It is a start, and it will take longer than the token-driven parts of AI, because for the physical world tokens were never the answer.
“All of the things that work with deep learning, let’s take them, but make them a bit more principled.”
Weather is only the beginning
Neural operators and weather modeling were a personal passion of mine, so we’ve spent much of this blog and the episode exploring this work. Anima has done so much more! In the episode, we cover several other recent developments from Anima:
* Anima has a series of works integrating neural networks and automated proof techniques. We talk about TorchLean, a new framework that lets you write PyTorch-style networks inside the proof assistant Lean and formally verify them. This is a major step for proving bounds on neural networks, something that would be really important for someone trying to, e.g., add a neural network as part of the control loop to their fusion reactor!
* Anima was recently appointed to the United Nations Scientific Advisory Board! We talk with her about her goals of bringing evidence-based viewpoints to policy, and how AI in scientific domains can improve people’s lives all over the world.
This episode has something for every AI or science nerd! Elegant math? ✅ Old school harmonic analysis? ✅ Fundamental developments in modern AI? ✅ Practical ways of modeling the physical world? ✅
Give it a watch!


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

When we first dicsussed the Summer of Simulative AI in 2024 we knew it would be a brief summer, but it has recently come back with a vengeance with SimGym in April and now Simile AI’s $2B Series B, backed by GreenOaks and Index Ventures with prominent backers like Fei-Fei Li and Andrej Karpathy, running tens of millions of simulations for Fortune 100 clients like CVS and 85–99% accuracy vs human focus groups.
Time to catch up on why this Second Summer of simulation is working!
From creating Smallville, the landmark 2023 paper on Generative Agents that showed AI characters could remember, plan, socialize, and develop emergent behaviors, to now building foundation models of human behavior, Joon Sung Park is trying to answer a much bigger question: what if we could simulate the world before making decisions in it? In this episode, the Simile co-founder and CEO joins us to unpack the path from generative agents to digital twins, why today’s frontier models still fail to capture how humans actually behave, and what it would take to eventually simulate all 8 billion people on Earth.
We go deep on Simile’s approach to modeling human behavior: long-form interviews, observational and transaction data, randomized controlled trials, population-level and individual-level models, and post-training on the causal mechanisms behind why people make decisions. Joon explains how his research created digital twins that reproduced human behavior and attitudes 85% as accurately as people reproduced their own responses, why models optimized to be rational can be bad simulations of irrational humans, and why understanding “social physics” may require changing model weights rather than simply prompting frontier LLMs.
We also explore the much larger ambition behind simulation: testing products and policies before deploying them, finding counterintuitive paths toward desired outcomes, modeling emergent behavior across entire societies, and potentially tackling problems like climate change, democratic instability, and UBI. Joon reflects on scaling laws for simulation, the economics of data-center-scale simulated worlds, the connection to Thomas Schelling and psychohistory, why simulation is surprisingly similar to painting, and whether we might already be living in one.
We discuss:
* How Smallville and Generative Agents led to Simile
* Why Joon’s team asked: “What if we can just recreate the world that we live in?”
* Why useful personal agents require deep models of their users
* Memory architectures, Markdown files, and the limits of prompting
* “Social physics” and behavioral foundation models
* Why web data captures what people say more than what they actually do
* Interviews, transactions, observational data, and randomized controlled trials
* Why predicting the future matters less than understanding how to shape it
* How Simile creates representative simulated populations
* Simulation versus prediction and the connection to Foundation’s psychohistory
* How to evaluate simulations instead of simply stacking LLM hallucinations
* Creating digital twins of 1,000 real people and reaching 85% behavioral accuracy
* Why frontier models can struggle to reproduce real human behavior
* Why good simulations need to reproduce human biases and mistakes
* Post-training models on randomized controlled trials
* Population-level versus individual-level simulation
* Scaling laws for human simulation
* The long-term ambition to simulate all 8 billion people on Earth
* Whether simulations could help solve climate change or detect collapsing democracy
* Thomas Schelling and the history of agent-based modeling
* Why future simulations could require an entire data center
* Multi-agent simulations and what happens when simulated people interact
* Replacing expensive human panels with synthetic populations
* Why market research is only the starting point for simulation
* Why Joon sees simulation as surprisingly similar to painting
* Using simulation to study questions like UBI
* Whether we are already living in a simulation
* Why AGI and simulation may be the twin technologies of advanced civilizations
Joon Sung Park
* LinkedIn: https://www.linkedin.com/in/joonspark
* X: https://x.com/joon_s_pk
* Website: https://www.joonsungpark.com
* Simile: https://www.simile.com
Timestamps
00:00:00 Introduction and Joon’s Path from Art to AI
00:01:46 Smallville, Generative Agents, and the Origins of Simulation
00:05:03 “Let’s Just Create a World” and the Future of Personal Agents
00:09:53 Social Physics and Behavioral Foundation Models
00:14:08 Prediction vs. Simulation: How Do You Shape the Future?
00:16:59 How Simile Models Real People and Populations
00:25:35 Evaluating Simulations, Digital Twins, and 85% Accuracy
00:30:23 Post-Training Models to Reproduce Human Behavior
00:40:04 Scaling Laws and Simulating 8 Billion People
00:43:10 From Schelling to Society-Scale Agent Simulations
00:46:13 The Cost and Economics of Simulating the World
00:52:05 Real-World Use Cases, Synthetic Populations, and the Market
00:57:27 The Future of Simulation, Painting, and UBI
01:04:23 Are We Already Living in a Simulation?
01:06:08 Building Simile and Hiring
Transcript
Introduction: Joon Sung Park, Simile, and the Story So Far
Vibhu [00:00:00]: Today, we have Joon in the podcast. Excited to kick this one off. Very exciting company. I wanna kick off and ask you the question, talk us through the story of your life. How have you gotten here?
Joon [00:00:13]: Yeah, for sure. I’m really excited to be here. A story of my life. So I was born in Korea, and I lived there for a good 11 years or so of my life, and then my family moved to Boston. So we moved when I was 11, and my parents were doctors, so they were going through their postdoctoral studies. My dad was a surgeon, so he was doing his sabbatical years at the Boston Children’s Hospital. So I grew up there, not too close to tech. I was very much a music and artsy, painting kind of guy.
Vibhu [00:00:49]: Painting.
Joon [00:00:49]: Exactly. I got into painting a little bit later, in high school, but that’s what I used to do. And then I grew up mostly in the East Coast after Korea. So I lived a good number of years in New Hampshire, and then I went to college in Pennsylvania. And I got into more of this tech scene, in college. So I was originally trained to be an artist. I thought that would be my professional career. So it wasn’t a hobby. It was like, “Hey, let’s make a living out of this.” And then gradually, I got really interested in this idea of, hey, the greatest artist often creates their own medium, and the best medium that we had available today was in computation. So I decided to go deeper into that, and one thing led to another, and we can go deeper into this, but I decided that research was something that I gradually got interested in, and here I am.
Smallville, Generative Agents, and the 2023 Breakout Paper
Swyx [00:01:46]: So there’s a lot that you packed into the research components. You had one of the best papers of 2023, which was the generative agents paper, commonly known as the Smallville paper.
Swyx [00:01:58]: Feel free to call back to anything else that you mentioned, but most people would have heard of you from this. Do you have any statistics on how many people have, like, read it? arXiv gives you something, right? Some stats.
Joon [00:02:10]: Yeah, it’s a good question. How many people have read it, I’m not sure.
Joon [00:02:14]: I know we do keep track of citations, and they are going up quite fast.
Swyx [00:02:23]: Yeah, Google Scholar has 7,200 citations.
Vibhu [00:02:25]: I feel like it made a bigger hit than that, and it was a pretty instrumental paper. It got cited so many times.
Swyx [00:02:34]: It is frequently the answer when people ask, “What is the best paper you’ve read recently?” It’s this one.
Vibhu [00:02:39]: I thought the memory component was pretty underrated. It was a very good early memory system, and one of the biggest papers.
Foundation Models and the Search for Killer Applications
Joon [00:02:47]: Yeah, so maybe I can talk a little bit about how this particular paper came together. So when I got into research, it was back in 2020 when I started my PhD program at Stanford, and that was the year, when we were about to get GPT-3 to be available. So we already had GPT-2, and you could sense that there was this new class of models that was just becoming available in the market, and the team got very intrigued. And the general consensus was, “Well, is this model going to be useful for anything?” “It’s really strange that these models are not trained to do any particular task.” But we decided to take a bet. So a large group of scholars at Stanford, and it was led by one of my co-founders, Percy Liang, and we came together
Swyx [00:03:35]: Who coined foundation models.
Joon [00:03:36]: Who coined the term foundation models. We wrote this paper, where that term came from called Opportunities and Risks of Foundation Models. And during that process, really the thing that I started to think deeply about was, here is a model that is fundamentally new in our ecosystem. The reason why this was new was it wasn’t, again, trained to do anything in particular, but its premise was it could do anything and everything. It was like a stem cell, if you were to take a biology analogy. And I got really interested in this idea that, well, if we were to really think about what are the killer applications that this particular technology would enable, what would that be? Many of my colleagues were using this for simple classification, simple generations. Interesting that these models can do that, but from an interaction perspective, not that interesting. We’ve known how to do that for many decades. And what we came down to was these models are trained on this very broad data from the web, right? So these are human behavioral data. It’s social media, Wikipedia, all these data. So if you poke at the right angle, then you could see human behavior that would just pop out that’s quite realistic, and we’ve never seen that before.
The Time Machine Game and Recreating the World
Joon [00:04:45]: So that got us really interested. The exercise that we decided to do, with this particular group of colleagues, Michael Bernstein, Percy Liang, and myself, who ended up becoming my co-founder at Simile, we sat down and we played this game that we call the time machine game.
Joon [00:05:03]: Imagine we were to get on a time machine and fast-forward 10 years and look back. What would have been the single application that will have mattered that would be the most interesting and inspiring? And when we thought, “Well, what if we can just recreate the world that we live in?” it’s really hard to get more ambitious than that. Like, let’s just create a world.
Joon [00:05:24]: And that’s where we started. And initially, we had this paper that was a precursor to the generative agents paper called Social Simulacra.
Swyx [00:05:32]: Before you go further, were there other candidates for the most ambitious thing in the time machine exercise? What was number two or number three?
Personal Agents, User Models, and Why Simulation Came First
Joon [00:05:44]: There is a close second that we were considering, which ended up becoming more of these automation tools, especially the vision around really personalized agents that would do things for you.
Swyx [00:05:59]: That’s also happening.
Joon [00:06:00]: It’s also happening. But it was interesting for us, right, in that the reason why, we decided to go with the idea of simulation, one, I was a huge science fiction nerd, and this idea of creating simulation, I was personally really just fascinated. I loved the idea. It’s really cool to see, like, a game town like this and just see these agents live in it. But at the same time, my bet was if you were to create a really amazing personal assistant out of this technology, what you need first is an amazing model of your users. So I told a model, “Hey, can you go buy late dinner for me?” And it orders Hawaiian pizza, and I do not like pineapples on my pizza. Then it totally failed. The way for it to not make that mistake is only by having a deep understanding of who I am. And I gave a very simple and dumb example here, but you can imagine how this core understanding of people is instrumental. This is how, if we have our family and closest friends, they have a good mental model of who we are. That’s the basis of our social connection. So our bet also was this technology around simulation, creating accurate representation of people ought to precede the more complex agents that would automate the world that we live in. So that was the bet. But that was a very close second, and I’m still very much fascinated by it. I think there’s a lot of interesting work that’s going around. My hot take here, though, is I don’t think we’ve seen a true personal assistant that’s useful, in ways that meet the ambition of that particular line of work. I think there are early applications that are interesting, and if you talk to even ChatGPT nowadays or Claude, they know a lot about us. So a lot of the generation it’s doing, I do think it’s much more tailored, but I think the ambition is quite large in that field, and I don’t think we quite have all the right ingredients just yet.
Swyx [00:08:01]: So OpenClaw and these personal agents, what do you want to see from them that they don’t currently have?
Memory, Markdown, and the Limits of Prompting
Joon [00:08:09]: I do think it’s slowly getting there, but I do generally want them to have much deeper understanding of the person. Right now, you look at the models. OpenClaw, what it’s leveraging is a Markdown file, and I think it’s quite clever, right? So if you look at the generative agents paper, this was the same intuition that we had, where initially when we were creating the memory architecture for the generative agents, and, like, this is, like, back in 2022, so we didn’t really quite have the idea of even agentive architecture or the term agent. But the intuition that we shared with some of the work that’s coming out today was we initially thought, “Well, do we want to make the memory into, let’s say, knowledge graph? Do we want to train a bespoke model?” All of these things. And what we decided to do was, “No. Just forget about all this.” These language models are quite good at modeling text and understanding and reasoning about text. So just put everything in a Markdown file or a text file. You’re done. I thought that was quite interesting that we could do that, and there’s a lot of strength in doing that. But also, there are limitations. It’s the way you retrieve and make sense of data that’s extremely large, it takes a lot of work. So I think that technology is getting better. I also do, however, think, there are certain things you just cannot shape just by prompting the model. So to some degree, you do need to touch the parameters of the model itself. So there is this work that I do think does need to happen, and it is happening. The question is, how far can we take it? How do we source data, and how do you also create an ecosystem where people are continuously feeding data to this model so it’s learning about you?
Vibhu [00:09:50]: What’s the intuition between why you need to do it in the model?
Social Physics and Behavior Foundation Models
Joon [00:09:53]: My intuition behind the actual when do you train or even post-train a model versus just prompt a model is if the model has to learn the underlying physics of the world that it’s operating in. So it has to learn new social physics. The places where it doesn’t have to train are the places where it already has the physics. We trust the physics. It already has the base statistics, but it’s just trying to react to an environment. Then I think you can just prompt your way into getting the actions out of it. I don’t think the models that are out in the open have yet learned the complete mapping of social physics of humanity. This is one of the core theses of Simile, right? And one of the core reasons why that is the case is if you look at the data that the model was trained on, these models were trained on the web data, like, whatever was available on the web. And these are really interesting data sets, but they are fundamentally the self-exposed attitudinal data with some behavior data that’s sprinkled around here and there. And it has yet to learn the really deep behavioral nature of people, not just what people say they do online, but what they do in real life. And this is one of what I would consider to be the dark knowledge of humanity that we haven’t quite captured. And it’s these data that would also need to get factored into the model creation.
Vibhu [00:11:21]: You call it behavior foundation model.
Vibhu [00:11:23]: There’s a good one-liner here, but outside of that, what type of data do you need? What are you changing on the model level? How do you go about modeling, doing a behavior foundation model?
The Three Data Buckets: Interviews, Behavior, and Causality
Joon [00:11:35]: We think about data in three buckets. So one bucket is interview data. It’s quite interesting. Rich qualitative data is interesting. It’s not behavioral, but we would literally ask people, “Hey, tell me the story of your life.”
Vibhu [00:11:53]: It’s just what we’re doing here exactly.
Joon [00:11:54]: The question that you all asked at the beginning of this interview literally is the question we also ask. And we ask our participants to go a little bit deeper, than how far I went. Maybe I can give more of my life story in lieu of this. But the reason why that data is interesting is by learning about this very long-tail information about people, you get a lot of texture around this model, like, this person as a model. So even understanding their childhood memory or even their trauma, their first love, these things, quite informative in ways that’s really hard to predict. So that’s one. Then there are two tranches of what I would consider to be the behavioral data. One kind of behavioral data is observational. So these might be like transaction data, or these might be data that you can get by scraping the web, right? So you can imagine why these data sets would be interesting, right, because they give you the base statistics of people’s behavior.
Joon [00:12:55]: But then there is the last category of data, that I personally think is perhaps the most important, which is the data that describes the causal mechanism, the whys of people. Some of this is covered by the interview data, the qualitative, because people talk about why they made certain decisions. But really, where you get to see the most behavioral aspect of this is in randomized controlled trials, like RCTs. Imagine you have the same setup, but you have a few different variables that you are trying to tweak. Can you get realistic human behavior out of it in ways where, imagine you had this particular option. Imagine you’re even trying to choose whether you’re going to drink coffee or not. The day you drink coffee versus the day you didn’t drink coffee, does your behavior change? That’s a data set that describes a causal mechanism. This is quite important in modeling people. The reason why this is important is oftentimes when people come to us, or not just to us, but the reason why people are interested in simulation isn’t because they want to predict the future. If you’re trying to win against the stock market, predicting the future is interesting.
Prediction vs. Simulation: Shaping the Future
Joon [00:14:08]: But most people, most decision-makers, what they want to know is, how can we shape the future? It doesn’t really help you to hear that your sales are going to tank in two quarters. They’re just gonna say, “Wow, that sucks.” What they want to know is, well, what do we need to do now to avoid that future? That’s the causal mechanism. And this is also very hard data to come by, right, because the world is our ground truth, but it happens once. So in a very controlled setup where everything is equal except for one variable, this kind of data set rarely happens. So this is a reason why this data set is both hard to come by and quite important if you’re trying to model human behavior.
Swyx [00:14:50]: So behavior, I think, is the hardest data set to acquire. What is out there? What is even possible? You’re not going to know a lot of details about my life. I don’t even have data for myself on my own health or habits, and I just don’t log everything. So how can you have that data?
Joon [00:15:14]: So we run a lot of randomized controlled trials.
Swyx [00:15:17]: But you put people in the lab, they watch them sleep, or what?
Joon [00:15:20]: We do care a lot about the consent process. People know that we invite them to be a member of this community to both share data and have themselves represented in different forms. But we bring a lot of people to the lab, or virtual lab, where we design experiments that would pose them real behavioral decisions. And often in these experimental setups, what makes the difference between what is attitudinal versus behavioral is whether the stake in your decision is real. That’s ultimately what makes it behavioral. So in these setups, we are inspired by our colleagues in social sciences, psychology, and so forth. So when they run studies, the techniques they utilize is imagine there’s an online store that you’re inviting people to come by. Then whatever they purchase in this experiment, they actually get that item delivered. Like, these are the things that make the stakes real. So we run a lot of these experiments, and we also do partner with firms. Right now, we also have customers who are quite excited to at least give us a glimpse of the behaviors that their users exhibit so that we can get a little bit deeper understanding of how people behave in these different platforms.
How Customers Use Simile: Populations, Queries, and Experiments
Vibhu [00:16:39]: I think on the customer side, they have a lot of data about their users, who has bought. They have the action data.
Vibhu [00:16:47]: Can you walk us through an example of what someone comes to you for? What questions would they want solved? Do you customize a model for them? Do you have something off the shelf? What does that look like?
Joon [00:16:59]: Today, when people leverage our models, it’s often to better understand the population of their interest. So usually, the start of the relationship, we come together and hear about what population they want us to model, right? So it might be that if you’re a CPG company that’s selling to all of the US, then maybe it’s fairly straightforward. You want to model the gen pop of the US. But at the same time, if there is a vertical or if there’s a market that they’re trying to go into, imagine, they want to better understand, let’s say, people in their 20s and 30s living in California. That’s a much more specific population. So we hear about this population, and we go recruit these people, with consent, and with incentives, and we collect some of their data and create a model of these people. Then what our product allows you to do is query them. So it can take as input a filter that is a description of the population that you want to talk to, just like the one I just mentioned, and an environment. The environment can literally be survey questions, behavioral experiments, It can be A/B testing. Oftentimes, the core use cases are things like concept testing, to start with. But also, people sometimes want to do focus groups or one of the fun use cases that we also serve is even modeling things like earnings calls for public companies.
Joon [00:18:21]: So these are the use cases that we often start with.
Swyx [00:18:23]: Concept testing, is that an established term? I’ve never heard of concept testing.
Concept Testing, Gallup, and Politics
Joon [00:18:27]: Yeah. So it has to do with they have, let’s say, different messaging, different products, different ideas.
Swyx [00:18:32]: It’s like a marketing exercise.
Swyx [00:18:33]: Okay, got it. Got it. Politics?
Joon [00:18:36]: We do, have a strategic partnership with Gallup, and of course, Gallup is deep into policy space and so forth. Right now, we have not worked deeply with politics, like that area just yet, however.
Swyx [00:18:49]: I’m curious if there is demand or if they really would have different needs that somehow fundamentally don’t mix with your existing, users or people.
Joon [00:19:00]: I think there’s certainly demand.
Joon [00:19:02]: But we are very much mindful of how this technology gets adopted and the societal impact that we’ll end up having with this technology. And I do see politics as an area where a company has to be particularly thoughtful about the way they operate and make impact. So this is where we also want to make sure that we form enough of guardrail and perspective on how to leverage this technology before we go on to serve markets like the politics.
Swyx [00:19:29]: I’ll give people an example. one of my favorite shows is The West Wing. I don’t know if people have watched.
Swyx [00:19:34]: One of the key storylines is, like, the president has, multiple sclerosis, but they haven’t. they need to figure out how to disclose it. So they run a poll with a fake governor and ask people to respond on the poll,
Counterfactuals, Polling, and When Simulation Is Useful
Swyx [00:19:47]: They try to make decisions based on the results of that poll on, like, how well they’ll be received, like where, how should we play this?
Swyx [00:19:54]: And I’m like, well, I think those counterfactual things, I would use a simulation for this if I could trust it.
Joon [00:20:01]: For sure.
Joon [00:20:02]: In that show, how’d it go?
Swyx [00:20:04]: In that show, it was, like a foregone conclusion. They were like, “We know it’s bad. We just don’t know how bad.” And then the poll came back. It was like, “It’s really bad.” And then they just did it anyway.
Joon [00:20:14]: Part of it is to show, right? So you’re, you’re looking at the idea
Swyx [00:20:17]: Maximizing drama.
Joon [00:20:18]: How bad could it be? Oh, it’s horrible.
Swyx [00:20:20]: And to some extent, I think that is part of the trick of the, or the challenge or with being a customer of yours, which is that if I know it’s. if I roughly know and can intuit
Swyx [00:20:35]: What the effect is going to be, do I need you? What sensitivity of it, of effect do I need in order to make a decision, right? So for example, if I, my approval rating is 50%
Swyx [00:20:48]: And I, they have this negative piece, news item comes out, and it drops to 30.
Swyx [00:20:52]: If it drops to 20, if it drops to 40, do I care? No. It, I know it drops. It’s negative. So when do I care about simulations?
Joon [00:21:01]: You do something that’s clearly bad, that’s not popular, and people don’t like you, like, yeah, it’s like
Swyx [00:21:05]: You don’t need a simulation.
Joon [00:21:07]: Yeah. Well, so there are a couple of things. one is, there are use cases where, like every day, developers, designers, policymakers, marketers, every single day, they create assets. They create new products. And turns out, it’s many of the decisions in hindsight is obvious. Yes, of course this is bad, but we still run those studies because understanding the magnitude and understanding how acute something is quite difficult, even if, we feel like, of course, like this makes sense. this is the reason why we make so many mistakes. Like, every time somebody goes online and say something that has huge backlash, you look at that and like, “What an idiot.” However, it’s tough. That’s one. There’s also another aspect here, which is, again, this is the reason why simulation is different from prediction. In simulation, in the ideal case scenario. So what simulation is trying to show is it’s trying to show each step of the way or each step that we need to take to get to a certain outcome, right? So in the most advanced simulations, sometimes the next step that we’re suggesting might be quite counterintuitive. The analogy that I sometimes give, and I ground it in a more realistic example, but, I, as I mentioned, I’m a huge fan of science fiction, and I don’t know how, many of the audience members have read, like, things like the Foundation series by Asimov.
Simulation as a Path, Not Just a Prediction
Swyx [00:22:37]: Oh, yeah. We’ve mentioned psychohistory a number of times.
Joon [00:22:39]: Okay, fantastic. So I might be, talking to the right crew. If you read Foundation series, literally the first act is there’s a group of scientists who have found out that, “Oh, our galactic empire is going to collapse, and we’re going to have 30,000 years of unrest.” And they run psychohistory, the simulator that tries to teach them, “Okay, how can we keep this unrest to a 1,000 years?” And they plan this out, and the first step of that plan is to get the scientists who say, “Okay, this is coming,” exiled into this random place in this, galax- galaxy.
Swyx [00:23:18]: Terminus.
Joon [00:23:19]: Exactly. And that’s so counterintuitive. Like, what a strange move that you literally sent the group of scientists who was raising voice around this potential collapse of galactic empire into nowhere. How is that the right first move? Well, it turns out in this particular simulation, that was the move.
Joon [00:23:40]: It’s these things, right? And the reason why these reasoning is possible is because you’re showing the step function or each step that results in a particular outcome. So really what simulation allows you to do in its highest form is you give it not a problem or question, like what would people answer to the survey? That’s not what we do. What we tell it is, “Here is a goal that we have. In the context of foundation, we want to keep the unrest to a 1,000 years. What is the path that we need to take now to get to that particular future?” And that’s what simulation allows you to do. Now, translating that into real market, imagine you’re a automobile company and you’re about to release a, EV, and you’re trying to understand, well, how do we market EV, to make sure that our stock price goes up? But what if the answer comes down that, well, you can market your EV in XYZ way, but that might change people’s perception around the cars that’s not EV and make your overall sales to go down. Not very intuitive, especially all you’re trying to optimize is EV salesss, and that’s the only thing that you’re tracking, then that might result in a completely wrong solution, or at least different solution than what you would have expected, whether it’s right or wrong.
Joon [00:24:57]: That’s the power of simulation.
Swyx [00:24:58]: For listeners, we covered a similar topic with Mikhail Parakhin from Shopify, where they are working on SimGym. I don’t know if he ever talked to you about it. it’s very similar.
Joon [00:25:07]: I
Swyx [00:25:07]: The goal is increased conversion, but then the journey is very unusual.
Joon [00:25:12]: Journey is unusual.
Swyx [00:25:12]: Yeah. The-- He’s trying to look for interventions on a shopping trajectory, which is similar to what you’re saying. Like, it’s not about the attitudinal, is your word for it.
Swyx [00:25:24]: It’s about behavior.
Joon [00:25:25]: It’s about behavior.
Swyx [00:25:25]: And that’s exactly the difference, right? It’s, like, not about the near-term direction about-- but it’s more about, like, how do you affect multiple turns of interactions.
Vibhu [00:25:35]: You had a good quote at the start about this as well. It’s not about people wanting to know the outcome. It’s about how they can change it, change the way to get there, something like that. But I wanna take it back to how do we know this is grounded? Like
Grounding and Evaluating Digital Twins
Vibhu [00:25:47]: How do you run evals? How do you test that simulations come through? if I was to do the same thing that you described with, say, your favorite LLM, Opus, GPT-5.6, have some agent to map out these things
Vibhu [00:26:02]: How different are the answers we would get if I give it the same goal, the same objective, make a decent system? You’re saying that you need to change the model weight. You have your own solution to this. But how far off are we, and how do you check if it’s grounded? you have some interesting stuff on your site that points to how you run real evals, but if you could take us through that side. I think that’s one of the big concerns that people have. They’re like, “LLMs hallucinate.”
Vibhu [00:26:27]: “You’re just hallucinating layer after layer,” right?
Joon [00:26:30]: The way we do this, and this is the paper that we worked on after the generative agents paper that really became the, at least for Simile and also the field of simulation and synthetic panels, really became the foundation. Yeah, this is the paper. the paper is called Generative Agent Simulations of 1000 People. Here’s what we’ve done. For this paper, we brought 1,000 people that’s representatively sampled from the US to a virtual lab. And what we have done was we spent two hours collecting fairly wide-ranging data. In this particular study, we focused a lot on this interview data, that was, whose script was taken from this project called American Voices Project. And then we would also pair that with a lot of behavior data and so forth, whatever we can collect within two hours. And then we would send these people away for a couple of weeks. And during that time, I would use this data to create their digital twins. And I would bring the humans, participants back after 2 weeks and have them complete a battery of surveys, experiments, behavior studies. So we have the list here, which included things like behavioral economics games. We would run literally, like, Big Five personality test, General Social Survey. We would also go ahead and run the randomized controlled trials that were published on PNAS. And we would have their digital twins predict how the source individuals would have acted in these studies and surveys. And this is where we could replicate people’s behaviors and attitudes 85 percent as accurately as people would replicate their own. So that was the first really paper that gave this validated results that we can model individuals in an accurate way. And what we ended up finding now, of course, in AI space, so this paper came out at the end of 2024. AI space, a year and a half, 2 years, that’s a lifetime.
85% Accuracy and Why Frontier Models Miss Human Behavior
Swyx [00:28:24]: Yeah. Just, for listeners who are not seeing the YouTube, I just wanna say, like, the headline figure is 85 percent accuracy, like, which is a big improvement over all the other
Swyx [00:28:34]: Methods that you showed.
Joon [00:28:36]: But the part that was particularly striking to us, especially as we improved this technology even further, was the generative AI models like ChatGPT, Claude that’s coming out, it does give you the right foundation. However, what they do not consider is the true attitudinal and behavioral aspect of people, especially in the population that you care about. So what these models are really good at today is they’re trying to become the super rational, objective machines, right? So you go get their data from places like Mercor, Scale. You talk to professional programmers, scientists to create model that’s amazing at reasoning. That’s what they do. Simile doesn’t care about any of this. The models that we’re talking about here, what we’re trying to create are models that are as dumb as I am, right? So if I make some mistakes, the model has to make the same mistake.
Swyx [00:29:34]: Oh, that’s very hard.
Joon [00:29:35]: That’s very hard.
Swyx [00:29:36]: You’re solving Murphy’s paradox.
Joon [00:29:37]: That’s exactly. And this is a completely different data and training objective. This is also where we see quite a bit of discrepancy in the performance in human behavior prediction between the frontier models, Simile’s model, and the models being created in this space, where in some cases, the model performance of frontier models go all the way down to 20, 30 percent, especially if you go into that more niche population on topics that our customers would care about. On more gen pop, it might be around 50 to 60 percent. So it’s not very robust. Like, you wouldn’t want to make your decision off of these and these findings. If you can bring that up to 85 percent, that is ultimately what people end up getting very excited about.
Swyx [00:30:20]: Yeah. Do we wanna keep going on the paper, routes?
Joon [00:30:23]: Yeah, for sure. So the last one, was an interesting one. So this, paper was the follow-up paper that we had, to the 1000 agents paper, where the idea was now can we augment the models even further and post-train a model based on a lot of randomized controlled trials? So this was an interesting one. The data is always the most interesting part of modeling in many ways. The data that we got here was there’s this, there’s this platform called Open Science Framework. So some, the audience might be familiar with this. And there has been, especially in the social sciences over the past 5 years or so, there has been this concern around replicability of studies. And so it was a bit of a crisis, the scientists acknowledged, where we rerun the study and we don’t see the same finding.
Post-Training on RCTs and Replication Studies
Vibhu [00:31:12]: Oof.
Joon [00:31:12]: It’s tough. And the reason why it’s there-- that was often the case was there’s this survival bias where the papers that get published often need to maintain what we call the value of less than 0.05 in the experiments that we ran. That suggests that only-- there’s only 5% chance that the results that we saw is false positive. But the tricky part was all the papers that were not published, and there’s still a 5% chance that whatever we publish is totally just randomly generated. Like, there’s a 5% chance that, hey, this effect is not real, but it just happened to be real because of the sampling bias. So because of that, what scientists started to do was they started to register their studies. So before running an experiment, they would go to this platform and say, “Here is the data. Here is the population that we’re collecting, and here’s the hypotheses.” And they would just say, “Here is our hypothesis.” Like, “This is what we believe.” And you cannot retroactively change those hypotheses. This is what gives us more scientific statistical confidence that whatever effect that you ended up seeing is true. So that ended up creating this really interesting platform where there’s one platform that has now contains tens of thousands of real-world experiments and hypotheses. And a lot of these are really high-quality, like, professionally designed behavior studies and random- randomized controlled trials. So we got the data and the studies from this platform and used that to make a point. And this particular, model is not, something that we’re serving commercially because this was a part of the open science. But this particular data set, helped us make a point that by collecting a lot of these randomized controlled trials, that are really well-designed, we can make significant improvement in model’s capability to predict human behaviors. So that’s what this paper was about.
Vibhu [00:33:10]: Is this stuff done on a individual level? Like, do I need to tune the model per individual, per company? Is there foundation model changes and then some slight post-training? Anything you can share there?
Population-Level vs. Individual-Level Models
Joon [00:33:21]: So this particular model was trained. the data we had at the level of individuals, but this particular model was trained. We experimented with both. And this is what we end up doing at Simile too. We always train 2, distinct model. One is what we call the population-level model. The other is what we call the individual-level model. And both take very similar input, which is the description of a subpopulation or individual and a stimuli. In this particular work, we’ve done the same. Here, the results that we are reporting are much more geared towards individuals because we do think that is a harder task in many ways, but that’s what we have done.
Vibhu [00:34:02]: You seen anything on the questions that humans can solve that models can’t solve? So like
Human Biases, Mundane Choices, and What Models Miss
Vibhu [00:34:09]: Currently, it’s, I live 5 minutes walk away from a car wash. It’s a 10-minute drive. Should I walk or drive?
Joon [00:34:16]: Huh.
Vibhu [00:34:16]: The model will say, “Oh, walk to the car wash.” And, you don’t have your car.
Vibhu [00:34:20]: Is anything like this a problem in simulation? You would assume, like, very simple for human to think about, but if the model is saying you should walk to the car wash, anything here?
Joon [00:34:32]: It’s less, what can we solve, but I think it’s more about what biases or mistakes do people make that models miss. Like, imagine that you are, like the. When I was still at Stanford, I lived in Palo Alto. So it’s about, I would say, 40-minute walk from the campus. You ask the model, “Okay, let’s go home. What can I, what can I do?” It would likely call an Uber or, give me, the bus time. But for the longest time, I really liked walking back. And the reason why I wanted to do that was not for efficiency. It really helped me think. And I like to walk for, half an hour or 40 minutes or so a day, where I just get to, just think about ideas, research, just get lost in my thoughts. That’s very human activity. Unless the model has seen that and understands the importance of that activity, it would miss these kinds of features. So that I think, is fundamentally what we’re trying to model. Like, what is fundamentally human might not be the most efficient thing to do, might not be the right thing to do, but things that make us who we are.
Swyx [00:35:43]: I’m curious if, there are some data sets that you really want that would materially help you. One version of this may be interesting, which is more valuable to you to acquire as a data set, all of LinkedIn, all of Twitter, all of Facebook?
What Data Matters: Social Media, Transactions, and Facebook
Joon [00:35:57]: It’s a little bit hard to rank, in part because, there’s, there’s this product saying where no feedback is wrong because it teaches you something about your users. Doesn’t matter what feedback.
Joon [00:36:11]: I think it’s a little bit like that.
Swyx [00:36:12]: So just whatever is bigger.
Vibhu [00:36:13]: What about a different domain? Say it was. What about all of Amazon data?
Joon [00:36:17]: Oh, yeah.
Vibhu [00:36:18]: Shopping data, right?
Joon [00:36:18]: Shopping data. So Amazon data is interesting in that it’s very much behavioral, although, like, what people do on social media, you could squint and say that is also behavioral. But the transaction data is always interesting. It is also most commonly available, however.
Joon [00:36:33]: If we were to look at purely social media, like if you really, if I were, if I had to really pick, Facebook likely is interesting because I do think it is most a default version of people. Because you go to LinkedIn, it’s very much professional environment. So people put up their, they have their guards up, right? And that still is interesting because that is true human attitude and behavior, but it is not your base state. you go to Twitter- Twitter, people have their own crazy personas, or depending on who you are. Like, my Twitter profile and, persona is very much, initially was I was very much an academic. “Hey, I’m here to share my studies.” Now, I share, things that’s related to Simile. But Facebook is one of those more private space where people just connect with their friends. In that way, I do think it shows you a little bit more about who that person is. So if I had to pick, I’d likely pick, Facebook.
Swyx [00:37:30]: Yeah. And you’re interested in, like, the whole person and their background and philosophy. I, is it too clinical or too machine learning-oriented to just say this is just ways to inject variance and biases? The broad question, is, like, is this any better than a randomized, like, combinatorial explosion version? So we have a link to the Tencent
Billion Personas, Synthetic Demographics, and Bespoke Data
Swyx [00:37:54]: Billion persona paper, where they did not do any of the groundwork that you are doing.
Swyx [00:37:59]: They just did like a cross matrix of here’s all the professions in the world, here’s all the people, possible backgrounds in the world, do a dot product across all of them, and that’s it. That’s your prompt for a billion people.
Swyx [00:38:12]: This will do something. I don’t know if it’ll do what you do, but it gets you some way, some percent of the way there.
Joon [00:38:18]: So this was an interesting paper. Like, what I admired about this paper when it came out was the scale. And you do gradually want to be able to simulate really large societies and interactions. So the scale is definitely admirable. it is relying heavily on the known statistics that went into training the model. So to the extent that you believe that statistics is correct, this is not a bad way to go about this. But the thesis here, and this is something that we also have seen in the market, like if this works, then we have solved simulation.
Joon [00:38:54]: It,
Swyx [00:38:55]: Because I survey, like, okay, 5% of the US population is in construction.
Swyx [00:39:01]: The other 5% is in medicine, whatever, right? And then you just keep going down the list, and then you do the other side. 5% has, like, the big 5 personality
Swyx [00:39:08]: Of, like, neurotic or whatever. That’s it.
Joon [00:39:11]: That’s it. So if you believe that the underlying data set and the platform that we’re leveraging has all the right statistics, then this will have solved it. you’re at that point merely retrieving the knowledge that is already embedded in the model, in the model parameters. That’s not, unfortunately, what we see, where there is such detailed and also niche knowledge about people that if you just take one example, it might feel very mundane, but it’s quite rich when you put together, that you do need to do a lot of bespoke data collection to better understand people. And this is also, I think what makes this particular, job fun, which you want to deeply understand people, and the process of deeply understanding them requires a lot of attention to the details. And you do need to pay attention to and pay respect to the daily lives that people lead.
Scaling Simulation: From Thousands to Societies
Vibhu [00:40:04]: I wanna talk about scaling simulation.
Vibhu [00:40:07]: So what can’t we simulate, what can we simulate, and how does scaling affect this? So how big are the models? What if we go from, 8B, like, couple 100 billion
Vibhu [00:40:18]: Like billion000 parameters, billion000? Do we get scaling? Any interesting emergence? Like, at a certain scale, at a certain amount of training, you uncover anything unusual and any learnings from that?
Joon [00:40:31]: What we are seeing is at Simile, so we do post-train our own model. The thing that we’re seeing is the early glimpse of scaling law in simulations. The more data about humans and more compute you ingest, you start to get predictive and predictable gains of the model performance in simulating it, simulating people.
Vibhu [00:40:51]: Ooh. We need a scaling law curve.
Joon [00:40:52]: It’s scaling law. Whenever you find it’s a beautiful thing. And we’re starting to see the glimpse of it, which is quite exciting. But if you talk about the ambition of simulation as a whole, it’s not merely about building a model. It’s about building a model, then creating the agents that become the individuals in a much larger ecosystem. So they’re creating this multi-agent simulation. Down the line, you want these multi-agent simulation to also live in a very rich environment, right? What we are really trying to get to at that point is, hey, can we create. All right, let’s do a time machine game again, and 5 years, 10 years into the future, can we create a simulation of 8 billion people living on Earth? I think that’s quite interesting. And that really is the vision. And once you get to that state, the questions that you can help answer for the society also start to change from my perspective. The answers are fundamentally about emergence of the emergent behavior of society and large groups of people.
Joon [00:41:53]: So the questions that I get excited by, and maybe this is a stodgy- a bit. I have my, academic side of me.
Joon [00:42:01]: And for me, it’s questions like, can we help solve climate change? If you look at climate change as a problem space, this is what we, like social scientists would often call it the wicked problems, problem where you have many actors with competing incentives for trying to make a very complex decision and coordinating that coordination decision. Very difficult to really solve in real life, which is also the reason why we couldn’t solve it. Can simulation help us solve that? Another one is, can we understand the signals for collapsing democracy, or can we understand or can we uncover the origin story of the monetary system? These are societal questions that we never really had a good way of answering. If we can create simulations of our society, you have to believe that these are the problems that we can solve. So that’s really the ambition of this field. And, I also think, yes, I think there’s a Nobel Prize to be won there, which wouldn’t be surprising. And I think there’s some amazing societal impact that we can have to help people make better decisions.
Climate Change, Democracy, and Societal Simulation
Swyx [00:43:04]: Nobel Prize in economics?
Joon [00:43:06]: In economics.
Swyx [00:43:06]: Oh, I see. I see. Rooting for you to write that paper.
Joon [00:43:10]: One of these days. But, one of the scholars that I was deeply inspired by, When I was coming into the space of simulation, is this scholar, named Thomas Schelling.
Schelling, Agent-Based Models, and the Nobel Prize
Swyx [00:43:23]: Schelling point?
Joon [00:43:24]: So the canonical example of the work that he’s done was he was one of the creators of agent-based modeling. So this was, like, in the 1970s and 80s. It’s very early days, but this was truly one of the first exemplars of simulations. And one of the canonical model from that time, and of course many of these simulations are trying to tackle the societal problems that’s most relevant for their era, it was called the model of segregation. So racial segregation was a big topic, that, we cared about. And what they’ve done was they created this grid world where they had red dots and blue dots. And these dots were, back in the day, like, they were the agents, and they had a simple rule that governed their behavior. If certain percentage of your neighbors are of different color and if that goes above certain threshold, then you move to a new location at random.
Joon [00:44:21]: One of the striking finding of this paper or this agent-based model was for the longest time, people thought the segregation within society was caused by explicit and overt racism.
Joon [00:44:34]: But if you look at this model, people’s preference towards living with people of the same color, that preference can be very minute.
Joon [00:44:42]: But the very small difference causes the society to segregate completely over time. This was very counterintuitive for a lot of people. And this particular work ended up informing housing policies. Mixed income housing, got really inspired by this work. And Thomas Schelling ends up winning the Nobel Prize for having laid the groundwork for very early versions of simulations. The opportunity that I do see here in the more scientific terms, is agent-based models for the longest, had impact in the 1980s, 90s, to some extent, early 2000s, but it has now gotten forgotten by the community a little bit. Because as you can imagine, red dots and blue dots is not really a rich description of people.
Joon [00:45:31]: But with the emergence of things like generative AI and, in particular, generative agents, we do have an opportunity to create these agent-based models that are high fidelity enough to help us make really complex decisions. And that’s the opportunity that I see. If that truly works, then yes, that is the work that will result in a Nobel Prize.
Swyx [00:45:53]: Yeah. For what it’s worth, and I grew up in Singapore. 80% of Singapore is in public housing, and public housing has, enforced racial quotas for exactly that reason, which is very interesting. okay, so we talk about scaling, we talk about all these, the agent possible applications.
Cost, Reuse, and the Economics of Simulation
Swyx [00:46:13]: I’m scared about the cost. if you even-- let’s just keep it to the US, about 8 billion people.
Swyx [00:46:21]: But, how much does it cost to model so many hundreds of millions of people?
Joon [00:46:26]: Oftentimes today, we don’t start at that scale, this stage of the, of industry and simulation as technology. But we can get our users extremely rich and meaningful insights even by modeling thousands, tens of thousands of people. And today what we do is every week we are collecting data on the scale of tens of thousands people’s data, and we have panel partnerships that gets us to tens of millions of people globally. So that’s what we do today.
Swyx [00:46:55]: And just as a side note once you’ve collected one person for one study
Swyx [00:46:59]: Can you reuse that same person for all the subsequent studies?
Joon [00:47:03]: That’s exactly right.
Swyx [00:47:03]: Okay.
Joon [00:47:04]: The beauty of this model and these agents is the fact that they are domain-agnostic.
Joon [00:47:08]: That what you’re really trying to understand is what is the fundamental nature of these people? What’s their social physics? And there are a lot of, a lot of, people that does change over time. Like, even, like, even things like, how many times have you gone have you been to, like, CVS the past week? that will change. But there’s so many traits about people that are also known to never change. Like, your risk tolerance doesn’t really change over time. It’s very consistent. So it’s these things that we’re trying to learn. But the scale we are operating is right now hundreds or, tens of thousands to hundreds of thousands. And in many of the core use cases that we are deployed in, and this is more than enough population, to cover those. Really, at that point, what you care about is less the number of people, but more do you have the right subpopulation of interest covered? And this is also the reason why people want a larger sample. It’s not because they want, stronger statistical guarantees. It’s more that can they filter down to any population of their interest. However, you can also imagine in 10 years, if we truly believe that the compute is going to scale, that we’ll have much more availability for compute, and our ambition for simulation is also going to scale accordingly, there’s definitely a reason for us to create an entire data center worth of simulations.
Joon [00:48:35]: Or in my hunch here is I do think in the next some number of years, we will start creating simulations that will cost as much as training a foundation model. But perhaps it’s going to be so valuable to the society that it would be a no-brainer. Right now, even today, like, we are training bunch of new foundation model just so we can say we trained one and we spent tens of millions. But if we can create a simulation at the level of society that would solve climate change, I would run that today. I would raise the money right now just to run that.
Multi-Agent Simulation and Social Influence
Swyx [00:49:10]: Amazing. the follow-up question is, does it also compound if you let the simulations talk to each other?
Swyx [00:49:18]: Or do they already do that today? They don’t, right, as far as I understand?
Joon [00:49:22]: It depends on what simulation you’re trying to run.
Joon [00:49:24]: In the multi-agent simulation setup, the agents do talk to each other.
Swyx [00:49:28]: Right, which is exactly Smallville, right?
Joon [00:49:29]: That’s right.
Swyx [00:49:30]: But a lot of times, for example, in commerce, you’re just by yourself, so there’s no point talking. which is way cheaper.
Vibhu [00:49:37]: But they use all these levels, right? Like, you decide what you will buy based on what other people around you buy and talk about, right?
Swyx [00:49:43]: It depends.
Vibhu [00:49:44]: It depends.
Swyx [00:49:45]: Again, I’m, I’m coming at this from a cost point of view. I’m like, “Oh my God.” Like
Vibhu [00:49:48]: I think
Swyx [00:49:49]: If there is, like, some combinatorial thing of, like, thousands of people talking to thousands of people, then that one million X’s might cost.
Vibhu [00:49:56]: I have a very different view as the cost point aside. Like, running these studies in reality is a lot more expensive, right? Running any study like this is you gotta have people do it, you gotta sign people up. It’s very expensive and sometimes, like, not feasible to run the study.
Vibhu [00:50:14]: But the outcome or the decisions you make are very expensive on them, right? So spend X million on something that, the overall process costs 100 million might as well, right? There’s, there’s a lot of value to be had there. It’s a small cost, but I’m excited on the cost side.
Joon [00:50:33]: To some extent, and when you deploy technology, you often want to deploy in a way where you can replace existing budget or you can make things more efficient, and that is the best way to deploy. However, the way you capture the long-term value of the technology is making the argument that, no, it’s the upside, that by making this better decision using simulation, you have saved yourself or made yourself hundreds of millions or even billions of dollars, and that’s a case to be made.
Vibhu [00:51:06]: Random tangent question. So if you’re doing a lot of inference, a lot of model multi-agent stuff, are you at the point where it makes sense to, train a model that’ very sparse? You’re expecting to do multi-million dollar runs. Are you thinking about this in model architecture standpoint or inference efficiency, or, you’re still at the research phase of it works, we’re not super there yet?
Joon [00:51:34]: Efficiency, we do think quite a bit about. this is technology that is deployed now in some of the largest enterprise companies in the world, and we do process significant number of queries, that are trying to, simulate the populations in the world. So efficiency is a consistent thing. we don’t want to over-optimize too early, so I wouldn’t say, like, this is the higher bid Right now, but this is definitely something that we think pretty carefully about.
Swyx [00:52:05]: Yeah. Are there other case studies? So we, you talked about CVS, talked about Gallup, Deloitte, Wealthfront.
Efficiency, Enterprise Use, and Real-World Case Studies
Joon [00:52:12]: Wealthfront is an interesting one, because one of the things they were trying to do, they were one of the first customers that wanted to do product testing that goes beyond just asking people what they think about, let’s say, behavior experiments and so forth. So there, really what we had to do was reason about multimodal input, so images, but also you can also imagine, like, these agents traversing through Figma mockups or websites. So some of the things that our agents can also do is it can be given a domain, like, or, like, a website URL and go use it for a while. It’s these things. And Wealthfront was one of the first, customers, that was very excited about this possibility.
Vibhu [00:52:53]: What have people been asking? Like, is there any demand that we have not covered? Like, UI testing, right?
Vibhu [00:52:59]: I wanna try a new. I wanna ship a new feature, test the UI, simulate how people will do it. Any interesting things that you’re seeing demand for?
Product Testing, Websites, and Synthetic Panels
Joon [00:53:08]: Today, a lot of the demand does come from like, the places where people have historically used human panels, we can now replace with agents, and these synthetic populations. And this is not replacing human panel. in many ways, the simulation that Simile is building is grounded. So the way that I think about this is we are trying to represent humanity at scale. And in that way, the use cases are what we would expect, but it’s the scale of deployment that surprises me.
Joon [00:53:44]: Turns out there are so many decisions that people make every day in these organizations, groups, and we want to be able to say, “We listen to people. We have consulted our users.” But in reality, that is rarely the case because getting to people and asking them many questions, it’s difficult. It’s both costly, time-consuming, but most importantly, people are just not available. If I had to answer 1000 survey questions for this one particular, vendor, even if I wanted to do that, like, I would never do it. And that’s very much the case. What simulation can do is ensure that the voices of people are always represented in rooms where the decisions for them is made, right? So all the stakeholders of this particular product launch, ideally they’re consulted. That’s what this technology really is trying to enable.
Market Size, TAM, and Human Decision-Making
Swyx [00:54:39]: In my mind, that means it skews towards more consumer focus, right? Like, anything with a wide enough customer base where you do benefit from the diversity that you represent. What are some rough statistics, just for people who are not familiar with this market in general, what’s the market size that. I’m sure you have some, like, rough numbers. market size is, like, a vague question
Swyx [00:55:01]: But, like, how much do people spend?
Joon [00:55:03]: So market research is a $100 billion industry.
Joon [00:55:06]: But the thing about simulation is not a tool for market research. Simulation is a tool for human decision-making. So the question around what is a TAM here is quite tricky, right? Because it’s easy to say, “Well, market research TAM is roughly 100 million or 100 billion.” so is it a TAM? And not really, right? Because in many ways, you’re trying to inform all human decision-making. You’re trying to inform every decision that are made about humans for humans. What is a TAM for that? It’s really unclear. And I’ll be honest. Like, I have a scientific background, I have a research background, so I didn’t come into the field calculating, oh, what is the TAM for human decision-making? But I just had to assume, well, if we can inform every decision that is made about human for human, that has to be big.
Swyx [00:55:58]: Some- something valuable.
Joon [00:55:59]: Exactly.
Swyx [00:55:59]: To some extent, you are a unicorn founder now, and you have to care as a CEO. But, like, I do think, like, yeah, when you go into these boardrooms with people that you’re quoting millions of dollars of contracts for, like, you have to say, “Well, here’s what you spend on humans-”
Swyx [00:56:15]: “. And here’s what we save you, and it’s 85% similar.”
Joon [00:56:19]: And certainly, the value case, is something that we care deeply about. Like, what is the value that we provide to the users and the decision-makers? But this is also where, like, as a founder, I think valuation only tells one very superficial aspect of the story, and I try not to think too much about valuation, in general, because that’s not what also motivates a team or certainly doesn’t. I’m, I-- Again, the interesting thing about researchers is we are happy living in academia, getting paid next to. we get paid okay. we don’t get paid that much, as a researcher here in academia, but it’s the impact and it’s the, it’s the value that we can provide to the individuals and the society that really drives us. And in that way, ultimately what drives us is the impact. Does the simulation we provide have a real impact in people’s decision-making in ways that progresses our society forward? If the answer is yes, then yes. that has to be great business, and we see that in numbers, and we do care deeply about that upside story, but that’s the heart of it.
Where Simulation Goes Next
Vibhu [00:57:27]: Do you have any timeline predictions? So we talked about scaling laws of simulations.
Vibhu [00:57:33]: You brought up, okay, maybe one day we can simulate how to solve climate change.
Vibhu [00:57:38]: Where are we now?
Vibhu [00:57:40]: If that’s not the end state, what is an end state, and what does progress look like?
Joon [00:57:45]: So what I sometimes tell people is simulation as industry, it feels a lot like where GPT-3.5, GPT-4 was, for the AGI saga, which is we have now technology that is powerful enough to do real damage on the verticals that we are tackling. At the same time, there’s a lot of progress that is yet to come. And that’s, I think, where this is. So the way I see it, I do think there will continue to be breakthroughs both in data, in algorithms, and there will be much more aggressive scaling that will also happen over the next few years. But I think that’s roughly where we are.
Swyx [00:58:27]: I think that was about the rough set of topics. Anything else that we should have asked you or you wish people asked you more about Simile?
Simulation as Painting and Understanding Human Essence
Joon [00:58:38]: I think the, what’s, for me, what’s quite fascinating about simulation, it is very impactful technology, but it is also very interesting technology, both in terms of, like, what it means for human society, our philosophy. And the way I sometimes interpret simulation is. So going back to my background, I as I mentioned earlier, I started my career as a painter. it was a professional pursuit, and I did oil painting, for figures. So I got my training originally in the realism studios, and that’s what I spent a lot of my, years, doing. Simulation is a lot like painting, right? The best paintings teach you something deep about the subject that you’re trying to represent. And it is always not a perfect representation. It-- No painting is perfect. There’s always some small differences and discrepancy, but what it does is it tries to highlight the thing that matters the most about the subject.
Swyx [00:59:47]: The essential
Joon [00:59:49]: The essential essence.
Swyx [00:59:49]: Yes. He, you, he’s brought up some of your work.
Vibhu [00:59:53]: Just nice to put it up.
Joon [00:59:54]: Yeah. So these are some of the works. So this is from, my, personal website that I maintain when, I was still a researcher.
Swyx [01:00:00]: I think a lot of people will say, like a Picasso, like anything postmodern is, like, very much focused on the essence.
Swyx [01:00:09]: Right. yeah, but I don’t know if any one of these evokes something that you like to tell the story of.
Joon [01:00:15]: No, it’s one of those things where, each of these paintings, drawings, whatever it may be, it is trying to surface something about the subject that you feel deeply about onto the surface. when I was a painter, and artist, the topic that I cared really deeply about was, the more mundane aspect of human lives. This shows up in some of the, some of the work that I’ve done, where, like, I did this entire study of a rural town where I went around and took photos of people for not really doing anything special, but just living their everyday lives. I thought that was the most interesting thing. I’m somebody who has this perspective where, the world is oriented around this fractal shape, and you have two choices to understand the fractal shape. You either go outward and try to explore as much as you can to understand the broader shape of the fractal, or you go inward because, the outward resembles the inward, shapes. And understanding the mundane aspect of it was very much that. Simulation has a lot of this, right? You’re trying to understand even the most mundane aspect of people. When put together- teaches you something really deep about that individual and the society. So I think that’s what’s interesting about simulation, the way, the same way that AGI helped us better understand or really think critically about humanity and human intelligence, simulation is really an exercise of understanding more about human society and our collective lives. So that I find to be, yeah, particularly interesting.
Swyx [01:01:56]: Yeah. Now you’re reminding me that some of the best biographers, documentarians, and even photographers, they’re taking a photo of you.
Swyx [01:02:05]: But before I take a photo of you, I must spend-- I must, like, follow you for a week just to understand you?
Swyx [01:02:11]: Which some artists, some do. Part of your work, there’s a very famous book called Working. I don’t know if you’ve, been referred to it before.
Swyx [01:02:18]: It’s very famous, like, to the point of having a Wikipedia page
Swyx [01:02:23]: About this like, really depth understanding and interview of people as they, about their lives, which seems mundane, but is told in a very, compelling way. Yeah, 1970s as well.
Joon [01:02:34]: Okay. It was an amazing decade.
Vibhu [01:02:39]: Before closing question
UBI, Future Questions, and the Value of Simulation
Swyx [01:02:41]: Okay, here we go
Vibhu [01:02:41]: You said that you started Simile with your 10-year question, right? If we do that now, 10 years down, what can we simulate? What would you simulate if, like, if you’ve made significant progress, are there any questions outside of the ones that we brought up? Any- anything that you think is most impactful? Anything that you would go vision 10 years out?
Joon [01:03:03]: In many ways, as I mentioned, I am somebody who is very much impact-driven. So the what would inspire me is I would want to ask, 10 years later, what would be the most important societal question that we as a society have to ask? I would love to tackle that. Like, do we need UBI? That could be an interesting one.
Swyx [01:03:24]: Ooh, has anyone done that?
Joon [01:03:25]: Well, we were thinking about it.
Vibhu [01:03:27]: Can we get access? Can we just
Swyx [01:03:28]: So OpenAI, this is, like, just trivia now. Like, OpenAI, or I think Sam Altman funded a study on this
Swyx [01:03:35]: In Africa, and the answer was no.
Joon [01:03:37]: The answer was no. But, what, was it something about the implementation?
Swyx [01:03:41]: Yeah, I know. It was a skill issue.
Joon [01:03:43]: Or was it something about, But this is the thing. See, when Sam
Vibhu [01:03:46]: Funny news article
Joon [01:03:46]: Altman funded this particular,
Swyx [01:03:50]: He spent 14 million dollars? Oh my God.
Vibhu [01:03:52]: It’s a little more.
Joon [01:03:52]: Quite a bit. But this is the thing. This is the reason why you want to run a simulation. You spend 5 years, 40 million dollars on this one study and have one finding, but if you can run simulation many times instantly, then that’s the value.
Swyx [01:04:07]: I feel like that one could-- you could have done in a simulation. Like, if you can do the housing study, you can do the UBI one. Like, I, come on.
Vibhu [01:04:13]: I think sometimes people will spend the money because they wanna verify what you think, right? Like, sometimes you just wanna. Is it right? Like, you gotta test it.
Swyx [01:04:23]: Okay, closing question. What are the chances we are in a simulation right now?
Are We Already in a Simulation?
Joon [01:04:28]: So it’s a fun question, and I assert at some point I just answer, yeah, we’re definitely in a simulation. But what I do, feel, however, is, whether we are in a simulation or not, that, I don’t think that makes our experience any less real. And I think that’s fundamentally, like, what I believe in. Maybe we live in a simulation, maybe not, but for
Swyx [01:04:48]: It’s real to us. Yeah.
Joon [01:04:49]: Yeah. For me, I don’t really care.
Swyx [01:04:50]: Yeah. Unless you die and you wake up in, like, the level higher or below.
Joon [01:04:55]: That would be interesting.
Vibhu [01:04:55]: I feel like you wouldn’t care. Once you die, then you find out you’re in a higher level.
Joon [01:05:01]: I worry about it when I die.
Swyx [01:05:04]: I think the other thing that. Okay, so I like the mathematical answer to this, which is, like, the, sheer number of possibilities that you are in a simulation far outweigh the sheer number of possibilities that you’re not.
Swyx [01:05:16]: Except for the simplest answer, which is, it is computationally very expensive to have you be a simulation. okay, great. You’ve been very generous with your time. Congrats on all your success. I met you just after your Smallville paper and had no idea that you could build, like, such an enormous company. And then now you’re like, “Well, it’s a $100 billion market, but that’s just where we’re starting.” So this is, very exciting.
Vibhu [01:05:42]: I think $100 billion market was not the term. That was only part of it.
Swyx [01:05:45]: Yeah, exactly. It’s, if you’re thinking too small.
Joon [01:05:48]: Well, I do believe that, maybe my final note here might be, again, I love science fiction. You look at any advanced civilization in science fictions, there’s 2 twin pillar, technology. One’s AGI in some form, and the other is simulation. So I think the market’s pretty big here.
Simile as Research Lab and Product Company
Vibhu [01:06:08]: Tell us about the company. You guys just raised a lot. You’re half a research lab, half a company. you’re hiring. Where are you based?
Joon [01:06:15]: Yeah. So we’re based in Mission Rock, so not too far away from, where we are right now. So we’re in SF, but we are also bicoastal. So we have our, team. I would say our headquarter is in SF, and we have a lot of our technical talent in SF, and we do have a smaller office that just opened up in New York. We are, as a company, an interesting one in that today, there are AI neo labs and then there are AI product companies. Simile truly is both. So this is a company that was founded by 4 founders, myself, Michael Bernstein, Percy Liang, Lainie Yallen. Michael, Percy, and I are all researchers. So of course, Michael was one of the authors of the ImageNet, kickstarted the AI revolution back in 2013, has been instrumental in human-centered AI. Percy coined the term foundation model, and is a, one of the greats of the AI researchers today. And Lanie is my business counterpart, where she led some of the fastest-growing AI native companies from their seed to A and B. But we have this DNA at the company where the vision of the technology that we’re creating is continuously developing, that we are getting people who were my lab mates. We are about 60 people right now.
Joon [01:07:28]: 15%, almost 20% of the company population are just my lab mates from Microsoft Research lab.
Joon [01:07:36]: And we It’s quite fun because many of them then had gone on to OpenAI, Google Gemini, and these places. And so it’s been a few years since we really got together and had a chance to work together. But now they’re coming back and really building out this vision that I find to be quite exciting, and that excitement is shared. So there’s that motion at Simile where we are a group of researchers trying to do something that no one is working on that we find to be the most impactful potentially. But at the same time, this is, again, technology that can make impact today. So we have an amazing group of engineers, product people, and designers, who are sitting here with us trying to imagine what does it look like to help people understand what simulation can do and make real-world decisions with this. Having both and then deploying it to some of the largest customers in the world today, it feels quite unique.
Swyx [01:08:30]: Yeah, it’s very compelling. One part of it was this is the call to action. Like, who are you hiring? You’ve done part of it, which is you have-- you’ve got a very talented group. Who are you hiring? Like, what roles?
Hiring and Closing
Joon [01:08:41]: So honestly, at this point, we’re hiring across
Swyx [01:08:43]: Everything
Joon [01:08:43]: All, section. we are always excited to bring on, amazing research talent.
Joon [01:08:49]: So if you’re interested in working with, our lab mates, we are always welcoming of amazing, researchers. But also we, hire, amazing engineers, that some of whom I, like, I respect the most. Many of them come from places where we have personal connections with, so many of the members are from Figma, Notion, Rive, and so forth, but also more broadly from the companies that we as a team have really admired. So engineers both in the product side, infra side, we’re all looking for those hires.
Swyx [01:09:24]: Well, lots of people. I think you made a really good case. So thanks, and, we’ll see you in the simulation.
Joon [01:09:30]: Amazing.
Joon [01:09:31]: See you all there.


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

This January, four big AI × Pharma tools deals were announced at the huge JPM Pharma conference that takes over San Francisco every year. OpenAI-backed Chai Discovery (now worth $4B) was somehow at the heart despite being all of 2 years old.
The Science team is proud to bring you the first podcast with cofounder Matt McPartlon and product lead Neil Patil to tell the full story!

Editor’s note: not to be confused with Chai AI, which was another top pod of ours.
Pharma suddenly doing big AI tools deals
For the non-pharma people, JPM is JP Morgan’s annual conference for pharma deal-making that takes over San Francisco for a week in January with hundreds of side events, etc. It’s a big thing.
Tools deals for pharma are also a big (new) thing: companies that start as AI for Pharma usually end up building their own drug pipelines instead, and the reason is something like this: convincing pharma to use your tool requires proof that your tool works. Proof means good targets, maybe with good clinical validation. If you have that, then it’s easier to raise money (with a known, if long path to commercialization) or sell (e.g payment in biobucks) for a specific target than it is to sell to lots of companies on a promise that it will work across their portfolios.
The “we’ll just partner / build our own drug” optionality proved to be the only good path up until January. What changed? In short, the tools got good enough for drug design teams to trust.
Good-enough-to-trust unlocks the ability to scale discovery: get more, better candidates into the lab and animal trials faster. More screening for toxicity, better delivery, etc. This means that what you push to the clinic is more likely to succeed.
Tools also unlock new capabilities: mechanisms that are very hard or impossible to develop using lab-based discovery. Designing an antibody that precisely triggers a very specific molecular cascade takes many years of trial and error. Designing bi-specific antibodies (that bind to two different proteins) is similarly difficult. Good design tools can unlock this.
RJ: The fact that the quality of the model has jumped means you’re enabling things you just plain couldn’t do. So it’s a step change. It’s not an efficiency argument at all, or not so much.
Matt: Yeah, exactly. It’s kind of interesting, even for us — it took me a while to believe in the thesis, actually. I talked to Josh for months before Chai started... It’s like, can I beat a mouse, and then can I do what mice can’t do? And then how many levels of interaction can you just keep building on top of that?
Everyone playing in the structural / binding space has an angle here, and some will be better than others, but Chai is pointing to a different unlock: getting good molecules right out of the gate (meaning they don’t then need as much lab work) means that the iteration time is faster. This turns science into engineering: you can design your systems to reduce friction and hill climb towards one-shotting molecules all the way to the clinic.
This, per-se, is not a new thesis: a16z articulated a version of this in 2020. What has changed is that structural models became binding models (how well doesn’t this molecule bind to this molecule, aka “binding affinity). Binding models unlock design, which has been steadily improving. Chai’s observation is that for engineering problems the best product tends to win, and good technology is a necessary but not sufficient condition.
Photoshop for molecules
With that in mind Chai has invested heavily in partnerships that allow them to learn from their Pharma counterparts.
What is kind of cool about working so closely and supporting so many of these partners is we get to really learn about what is the stuff that would be helpful in research. So rather than doing research in a vacuum, based on what would hypothetically be cool, we're able to do informed research based on what our partners have just been organically asking us for help with.
— Neil Patil, (Chai product lead)
This means better UX, such as a molecule editor that is more like a CAD or graphics design program than a chatbot.
Their approach has paid off: since June, Chai has announced three more major deals: Lilly, Novartis, argenx, plus an expansion of their Eli Lily program. This episode is too full of quotable moments for a short blog, so tune in to learn about
* Why protein tokens have the highest downstream value of any token
* Climbing levels of abstraction as models improve
* How Pharma, VC, and research are all just portfolio optimization
* How better tech changes the whole portfolio
* How relentless focus on simplicity leads to scale
Plus much more!


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

Watch the full episode on YouTube:

We first covered Baseten last year when DeepSeek mania was at peak hype. Now they have raised a monster $13B round and become one of the new cohort of AI Infra decacorns that are (with Nvidia, Intel, and the semis complex) chief beneficiaries of the Inference Inflection.
We return to Baseten at the peak of the 2026 edition of Open Weights debate. Ali has published a viral breakdown of Kimi K3:
And since you last saw him, Philip has spoken at AI Engineer and written the definitive book on Inference Engineering spotted all over SF:
Three years ago, inference engineering barely existed as a category.
Today, it is one of the most critical disciplines in AI. Inference engineering inherently tackles a different question than standard model training: “How do you turn those weights from training into a product that is fast, reliable, and affordable at scale?” Focusing on these creates an entirely new optimization problem.
In one recent GLM-5.2 experiment, quantizing more of the model actually preserved its benchmark quality while increasing throughput by 20%, because the errors introduced in different layers could cancel each other out.
Inference is no longer just the final step after training. It is becoming its own engineering discipline, with its own research problems, infrastructure, and increasingly specialized roles.
In this episode, Baseten’s Philip Kiely and Ali Taha join swyx and Vibhu to explain what actually happens after a new open model is released and what it takes to turn “we generated a token” into a fast, reliable, production-ready API.
We go deep on cache-aware routing, disaggregated prefill and decode, quantization, speculative decoding, KV-cache movement, model parallelism, GPU kernels, and the race to make frontier models up to 10× faster. Philip and Ali explain why inference optimizations can still produce gains of 20%, 100%, or even 200%; how quantization errors can cancel one another out; why identical weights can behave differently across clusters; and how Baseten grafted a Kimi vision encoder onto GLM-5.2 without changing the underlying language model.
The conversation then expands beyond LLMs into NVIDIA Dynamo, mega kernels, Rubin, AI-specific chips, local inference, video generation, diffusion versus autoregressive models, and the enormous compute barrier to generating coherent long-form video. Finally, we explore the convergence of training and inference, continual learning through persistent KV cache, and the emerging loop where models help optimize the infrastructure that runs them.
We discuss:
* What happens when a 200,000-token request enters an inference system
* Cache-aware routing and reusing previously computed KV cache
* Why prefill and decode are increasingly handled by different GPUs
* When dedicated deployments become cheaper and more reliable than shared APIs
* How speculative decoding uses a smaller model to accelerate a larger one
* Tool calling, structured outputs, and what LLMs actually do
* What it takes to support a new open model on day zero
* Grafting Kimi’s vision encoder onto GLM-5.2
* Retrofitting inefficient model layers with components from other architectures
* Why models sometimes collapse into repeating the same token
* How hardware, kernels, and race conditions create nondeterministic failures
* Preserving model fidelity while making inference faster
* How quantization errors can cancel each other out
* Why inference optimizations still deliver gains of 20%, 100%, and 200%
* How optimized serving can make a model up to 10× faster
* NVIDIA Dynamo, KV-aware routing, and distributed model serving
* Speculative decoding the speculative decoder
* Why local AI is about making models less dumb while data-center AI is about making them less slow
* Tensor, expert, and pipeline parallelism across GPUs
* Hardware-aware model design, auto-tuning, and the case against mega kernels
* Rubin and why inference is becoming a systems problem
* Whether modern GPUs are evolving into programmable AI ASICs
* Why enormous models like Kimi K3 require GB300-class hardware
* Why open-source video generation still trails Veo, Kling, and other closed models
* The quadratic attention bottleneck behind long-form AI video
* Autoregressive video, real-time generation, and compounding quality drift
* Why future video systems may combine autoregressive and diffusion architectures
* Training for inference and inference for training
* Continuous post-training, deployment, evaluation, and improvement loops
* How GLM-5.2 helped optimize the kernels serving GLM-5.2 itself
* Why faster networking could unlock dramatically faster decoding
* Continual learning, KV-cache compaction, and persistent model memory
Show Notes
* How to build a day-0 API for Kimi K3
* 22580: From GPT2 to Kimi3, Explained
Philip Kiely
* LinkedIn: https://www.linkedin.com/in/philipkiely
* X: https://x.com/philipkiely
* Inference Engineering: https://www.baseten.co/inference-engineering/
Ali Taha
* LinkedIn: https://www.linkedin.com/in/aliestaha/
* X: https://x.com/waterloointern
Timestamps
00:00:00 Introduction and the 200K-Token Prompt
00:03:18 Dedicated Deployments, Speculative Decoding, and Tool Calling
00:11:26 Launching Production-Ready Open Models
00:19:06 Model Retrofits, Failure Modes, and Nondeterminism
00:28:22 Quantization and Canceling Errors
00:32:15 The Race to 10× Faster Inference
00:40:48 Dynamo, Speculation, and Local vs. Data-Center AI
00:50:18 Model Parallelism, Auto-Tuning, and Mega Kernels
01:00:55 Rubin, GPUs vs. ASICs, and Custom AI Chips
01:10:03 Giant Models and the Limits of GPU Memory
01:12:42 AI Video, Quadratic Attention, and Autoregressive Generation
01:21:47 Audio, Images, and Diffusion Models
01:27:32 Training, Self-Optimizing Models, and Continual Learning
01:40:06 Closing Thoughts
Transcript
Introduction: Baseten, Waterloo Intern, and Inference Engineering
Swyx [00:00:00]: Okay, we’re here in the studio with Philip, old friend from Inference Engineering, the book, as well as Baseten and everything that you’ve done, you and I have done before, as well as Ali. Welcome.
Ali [00:00:15]: Pleasure to meet you.
Swyx [00:00:15]: Waterloo intern.
Ali [00:00:16]: Waterloo intern, always.
Swyx [00:00:17]: When did you get “Waterloo intern” as a handle?
Ali [00:00:19]: As a handle? Oh.
Ali [00:00:20]: I think the rebranding happened mid-March. When I saw it was open, I was like, “I have to take it. Up for grabs.”
Philip [00:00:26]: The problem is that Ali is really good at his job and is not gonna be an intern much longer.
Philip [00:00:30]: So we have to figure out who’s gonna get the handle.
Ali [00:00:33]: Well, I’ll pass the torch over to the next intern.
Swyx [00:00:34]: Oh, okay. It can be, like, you just pass it to another Waterloo grad.
Ali [00:00:37]: To another Waterloo intern. No, bruh.
Philip [00:00:39]: Yeah.
Ali [00:00:39]: Intern.
Swyx [00:00:40]: Intern, yeah.
Ali [00:00:40]: And no.
Philip [00:00:41]: You gotta get an intern from Waterloo.
Ali [00:00:42]: Yeah, I’ve gotta get an intern from Waterloo.
Swyx [00:00:44]: Right.
Ali [00:00:44]: But they have to follow the path.
Swyx [00:00:45]: Oh, it could, but it could come from Baseten, so it’s like whoever Baseten gets from Waterloo.
Ali [00:00:48]: Right.
Swyx [00:00:49]: Has the title of Waterloo.
Ali [00:00:50]: It stays in the ecosystem.
Philip [00:00:51]: Exactly.
Ali [00:00:52]: Halfway through the internship, you either get it or you’re out.
Philip [00:00:55]: You should also do, like, a big graduation ceremony where you change the handle.
Ali [00:00:59]: Just say it.
Philip [00:00:59]: For everybody.
Swyx [00:01:00]: You guys are good at ceremonies, clearly. We had a nice launch of the book, very successful. But before we get into all that, I wanna start off with a fun question for you. Okay, you’re an expert inference engineer. What happens when I send a long query, say two hundred thousand tokens into Baseten’s inference? What’s the process of query through GPU model routing, balancing, all that? What is all the stuff that we don’t think about?
Long Context Requests, KV Cache, and Cache-Aware Routing
Philip [00:01:26]: With a long query specifically, the first thing that I’m gonna ask is, “Have you sent me this query before, or at least part of it?” and I really hope you have, because it’s gonna be a lot easier for me and a lot cheaper for you. So the first thing that we’re gonna look at is some cache-aware routing, where we’re going to see, we probably have a number of instances, a number of replicas up serving whatever model you’re hitting. We want to send this one to something with, number one, available prefill workers, and number two, ideally some cached input already there so that we can skip prefill on at least part of these two hundred thousand tokens. If you’re doing two hundred thousand tokens, it’s probably coding or a multi-turn agent or something where you would expect to have that cached. If you don’t, we’re gonna have to send it to a prefill worker. We’ve at least on certain models disaggregated prefill and decode, so you’re going to have one set of GPUs that’s solely going to process the input, create the KV cache, and get you your first token, and then that’s going to be passed over to a separate set of GPUs, which is going to run decode. We’re going to iteratively make those tokens. We’re probably going to have some speculator model in front of that. I’m going to assume that you’re doing coding, and because of that, our speculator model, which assumes you’re doing coding, is gonna have a high draft token acceptance rate. If I’m wrong and you’re asking me to summarize every Harry Potter book, it’s gonna be slower. And then we stream that output to you and account for it, charge you, a couple of pennies and say, “Hey, would you like to send another one?”
Swyx [00:03:04]: Except Baseten doesn’t charge by pennies.
Philip [00:03:07]: Well, yeah, we charge. I’m assuming that we’re talking about the public model APIs. If you are setting up a dedicated deployment, then yeah, it’s not pennies.
Public APIs vs. Dedicated Deployments
Swyx [00:03:18]: Yeah, one of the key differentiators when I was talking with Baseten initially was that people who want very high volume just need to rent by the box, ‘cause then it’s up to you to figure out how to saturate the box.
Ali [00:03:31]: And more often than not, it’s, like, way cheaper if you’re pushing, like, millions of tokens per hour, if you just pay per hour instead of pay per token.
Philip [00:03:37]: Yeah, they do. I think that we’ve increasingly seen a lot of demand for the pay per token APIs, just because everyone wants to try open models, and then once they find a use case that’s really sticky, then they move over to dedicated.
Swyx [00:03:51]: Is there a best practice on when it’s time to swap over?
Philip [00:03:54]: Couple reasons. Yeah, reliability, that’s a big one, right?
Ali [00:03:57]: Like, if they have a very specific use case, they want you to train something specifically for them, like they want their own spec dec, for instance, for their own traffic.
Swyx [00:04:04]: Spec dec is speculative decoding.
Speculative Decoding and Custom Speculators
Ali [00:04:05]: Speculative decoding, yeah.
Swyx [00:04:07]: You have to explain.
Ali [00:04:07]: Sorry. Like, speculative decoding is like, if you have a huge model, right? And so the model is going to be generating one token at a time every single turn, every single forward pass. So we attach, like, this little, like, parasite, like this layer that goes on top of the model, and this model just has to predict. It does three very fast autoregressive forward passes, and it will predict, like, three certain tokens, and then you do one forward stage over the entire original model in order to see if those predictions were correct or not, and then you accept them or you reject them. Now, this draft model is traffic specific, so if you, like, Philip said, if you’re summarizing Harry Potter books, I can train exclusively that draft model on Harry Potter books, and I can guarantee you that I’m gonna accept the three tokens every single time. And so with that case, I increase your decode speed. I wouldn’t be able to provide this to you if you’re a shared endpoint
Swyx [00:04:53]: Yeah
Ali [00:04:53]: ‘cause I have no idea if you’re doing Harry Potter, if you’re doing coding, if you’re doing English. We don’t know. Also, there was a thing in the book that mentioned that if they really cared about a specific threshold, chapter four, I think. Do you remember that?
Philip [00:05:06]: Yeah. The things that you can do is you can set a specific, like, batch sizing, a specific, like, parallelism strategy if you’re trying to optimize for, like, throughput versus latency. You can. Maybe a NVFP4 quant doesn’t pass your benchmarks and you wanna run a model at higher precision, you could do that. There’s just a bunch of reasons why you might wanna have your own endpoint and the biggest one, of course, just being, like, you don’t have to deal with someone else doing a hundred million tokens of benchmarking traffic at the endpoint when you happen to be trying to serve your users.
Swyx [00:05:40]: Yeah. I think one thing that is. That is a classic journey. Like, it’s people is asking the, what happens when you type Google into the browser. Tool calling, is that just, you’re generating JSON or is there more complication beyond that?
Tool Calling, JSON, and Structured Outputs
Ali [00:05:58]: Certain customers that we have, they have their own post-trained models, and so they demand a tool calling that’s not just, like parse a file or go find the weather. It’s something that’s very specific and you have to do post-training on this. And if the post-training on the model is not good or if the quantization after the post-training to get the inference to be fast, the model will struggle reading the JSON file and reading the tool calling. But it doesn’t require its own like sandbox. It’s not like it’s going to use that tool calling to like escape a sandbox or like it doesn’t have to be contained. It can just be a normal dedicated deployment. The challenge with tool calling more and more seems to be that the companies want certain tool calling which is a very sensitive thing to train. And because you’re dealing with all of the JSON outputs, if it doesn’t like close the end of the request in a very certain manner, you end up with a model that did the tool calling and like the thinking and so as a result of that, it didn’t see the result and just hallucinated the result as it decoded. That seems to be the most challenging thing with tool calling, not really the sandboxes model.
Philip [00:06:56]: Yeah, that’s a challenge on the training side and then on the inference side, there’s work that you can do to scope the possible output. So we published this at this point close to two years ago, the solution to this problem which is you make a state machine and you use that to constrain the output to a specific format. So this is the structured output problem. If you remember back
Swyx [00:07:27]: Yeah, the specific grammar is,
Philip [00:07:29]: Yeah, exactly
Swyx [00:07:30]: GML had this thing.
Philip [00:07:31]: Yeah. So it’s like the old-school “make sure this is only JSON”, return only JSON or
Swyx [00:07:38]: Yeah
Philip [00:07:38]: Grandma’s gonna die type of prompts.
Swyx [00:07:39]: Is it BNF grammar? At some point OpenAI had released a thing that was like, yeah, if you want to constrain your output, write BNF grammar, back as NOR.
Philip [00:07:47]: In our inference system, it’s just a specified output format. And you get the guarantee that your output’s gonna be structured along that format. And so applying that to tool calls can like help cut down on. You can still call the wrong tool or call no tool. It doesn’t solve the certainty problem but it at least solves the output structuring problem
Swyx [00:08:10]: Yeah
Philip [00:08:10]: Within tool calls.
Swyx [00:08:12]: And MCP is just another form of tool, right.
Philip [00:08:14]: Yeah, exactly.
Swyx [00:08:15]: As far as there’s no special thing there.
Philip [00:08:16]: The thing I’m always like explaining to people is the LLM is not capable of doing anything. It’s only capable of making suggestions of what to do and then if those suggestions are formatted in a certain way and applied to a system that knows what to do with them, then an action occurs.
Swyx [00:08:32]: Yeah. Part of the fun stuff is, this is solved outside of tool calling too. Like in an agent loop if the output is not correct or you’re right, like reasoning, tool calling was done in the reasoning trace, just be like, “Oh, I don’t know what to do. Let me just try again.” And it might get there after a few tries. And on your point of training, sometimes this is harder in smaller models, so you don’t have the same exact quality output
Ali [00:08:56]: Right.
Swyx [00:08:57]: When you just swap from a big model, right?
Ali [00:08:59]: Yeah. I will say that, before, I think we need to go back to inference engineering proper.
Ali [00:09:04]: But, I had expected that something would replace JSON because it’s hard to stream JSON ‘cause JSON must be complete and you must have open and close brackets and everything. So it’s hard to parse something or validate something while it’s being streamed. So people invented all sorts of things that are like, I forget the name of some of these alternatives, but it’s something like TOML, something like YAML. But JSON seems to be dominant still.
Philip [00:09:30]: The JSON outputs aren’t that long, right? Like you could have a long-- ‘cause tool calls also contain the arguments in them and perhaps for a certain tool you might pass like a very long argument. But my impression of the median tool call is that it’s a relatively small number of tokens, right? So I would expect that speculators are generally fairly good at something as formatted as JSON. And so you would have like a pretty fast decode step there and that the streaming wouldn’t be as valuable, but maybe I’m wrong about that.
Ali [00:10:02]: I think you’re also bounded by the software or that the model is gonna integrate with if the software is built with JSON for the tool calls or if the company that you’- if your customer says that this is how our software works and our tools are interfaced with JSON, you can ask them to like, change their software and say like, “Yeah, this is gonna be better for the model.” but like with the right training shouldn’t be that much of a difference. Also more profitable if it outputs more tokens probably.
Swyx [00:10:25]: Depends on your business model.
Swyx [00:10:27]: It really depends. But I will say that, as a writer with like experience a lot with generated output, I do try to move from text to JSON text which is very long JSON, right? Like there’s paragraphs in every field because I’m trying to structure it, right?
Philip [00:10:44]: Right.
Swyx [00:10:44]: I want you to first make factual statements, then make opinions then make bullet point summaries, have dates, have entity references have your sources for references, all these things. Anyway, so these are things that like I think people who really experiment with structural output have to really care about. But, let’s, let’s recurse up the stack a little bit. Before we started recording, you mentioned something really cool, which is that there’s a lot of engineering that-- inference engineering that goes on when a new model provider releases a new model, right? So let’s call it GLM-5.2, Kimi K3. I had previously assumed, especially if it’s like, well, GLM 5 to 5.1 to GLM-5.2, like that you’ve supported them before. Is it that much work?
What It Takes to Support a New Open Model
Ali [00:11:26]: It’s a lot of work.
Swyx [00:11:28]: Yeah. Okay. So like, a lot of people, all you guys, right whenever a new model launch like, people rush to say like, “Oh, Hugging Face supports this, Fireworks supports this, Spacetime supports this,” and I’m like, “Yeah, of course we support it.” But what goes into that? What goes into
Philip [00:11:40]: I think it’s more than just support it too, right? It benefits the consumer a lot. Like I think it was with Kimi K2.5 or GLM-5.2 the latest, there was an inference war, right? X provider is at 90 tokens a second. The next day we’re at 150. The next
Swyx [00:11:55]: I kinda kicked that off with the GLM-5.2.
Swyx [00:11:58]: I wrote a Twitter article about. It got like half a million views,
Ali [00:12:02]: Based on being number
Swyx [00:12:03]: Yeah
Ali [00:12:04]: Or it’s for something else.
Swyx [00:12:05]: Yeah. Which,
Ali [00:12:06]: Oh my God
Swyx [00:12:07]: Which then got everyone really excited about, hey, how can we, bend tracks a little bit further and,
Philip [00:12:14]: There’s a difference between support the model, as in I can make a token out of this model, and support a model, as in I have a production-ready API from this model.
Philip [00:12:26]: Getting to the point of I can make a token out of this model is not that hard because generally the, open source inference engines, vLLM, SGLang of the world oftentimes even receive weights ahead of time, maintainers do, or the people making the model merge PRs to ensure support. So you generally can, just get it working on the standard open source stack without too much pain in most cases. The challenge is, every inference company is gonna have own proprietary stack. Some open source components, some in-house stuff. And for any arbitrary model, there’s going to be some new stuff. Sometimes you get lucky, like K, two five to two six was, like, pretty similar.
Quantization, Speculators, and Production Readiness
Ali [00:13:16]: Yeah. It was pure continued post-training
Philip [00:13:18]: Yeah
Ali [00:13:18]: If I remember correctly.
Philip [00:13:19]: Even in those cases, there’s still stuff you have to do. You have to redo the quantization work. You’re taking the model from. Generally, these models are not released in NVFP4, and we want them to be in NVFP4 for maximum Blackwell compatibility. So we have to perform that quantization, and, calibrate the quantization to make sure that we’re not causing any regression in the model’s intelligence. And then we also have to train the speculator, as we’ve talked about. Generally, we have. We have ZDR, zero data retention on our model APIs, so we don’t know exactly the traffic that people are sending us, but we know what’s popular. We know that coding use cases are popular. We know that agents, agentic use cases are popular. So we can get public data sets that are representative of that traffic and train general speculators. Now, with speculators today, you need to train the speculator using the base model itself because you’re getting hidden states out of the model from running inference on these specific prompts, and that is the training data you use to create the speculator. So there’s that process which you need the real model weights for. And then there’s of course just the process of, standing up all the infrastructure behind it, loading all this stuff, testing it. And then when there’s a new model with a newer architecture, I think that, like, the DeepSeek models tend to be the most challenging as they have, like, the most novel architectural stuff going on, model after model. But every new model has something. Kimi K2 had. Oh, sorry, GLM-5.2 had
Ali [00:14:53]: Sparse attention.
Philip [00:14:54]: Yeah,
Ali [00:14:54]: Yeah
Philip [00:14:54]: the DSA.
Ali [00:14:55]: Right. Which is brought from DeepSeek.
Philip [00:14:57]: Yeah. And
Ali [00:14:59]: So you can copy-paste then?
Philip [00:15:01]: It kind
Ali [00:15:01]: I don’t know how this works.
Philip [00:15:02]: So, like we had to, like, build support for that into our runtime. And you’re right, like it is really interesting the way that all of these open source labs borrow from each other. For example, like GLM-5.2 doesn’t have vision. So something that, Haley, a guy on our team, if we could take a look at this, he, like, grafted the Kimi vision encoder onto GLM-5.2.
Retrofitting Vision into GLM-5.2
Ali [00:15:27]: We’ll be training the projector.
Philip [00:15:28]: Exactly. So if you think about, like, the encoder, there’s the encoder, which is the part that looks at the image and turns it into latent information, and then there’s the projector which like
Ali [00:15:38]: You can say latent space. It’s okay.
Philip [00:15:41]: And then there’s the projector that maps it onto, the model itself, and then there’s the model weights. You don’t wanna mess with the model weights because you run a chance of making the model dumber at something else for the purpose of giving it vision. So instead, Haley started with just a projector, which is only a handful of millions of parameters.
Ali [00:16:02]: That would be, yeah.
Philip [00:16:02]: Yeah.
Ali [00:16:03]: Can you show the training one?
Ali [00:16:04]: Like the way it groks
Philip [00:16:05]: Yeah
Ali [00:16:06]: Very interesting.
Philip [00:16:06]: And maybe
Ali [00:16:07]: That right there
Philip [00:16:07]: Maybe Ali, you should take it from here. You’ve got a better
Ali [00:16:10]: Ooh, double the sand
Philip [00:16:11]: Understanding of this than I do.
Ali [00:16:11]: Yeah. You can see, like, he. The way he trained this is really cool. At the beginning, he was training it using just like, “Here’s a picture of a mountain. Can you describe what’s in this mountain?” And that caused it just like the first, learning walls. Like here you can see this all we’re trying to teach it is to translate the encoded. Like it’s already taken the encoder from Kimi K. It’s taken the image. It’
Philip [00:16:31]: Yeah. Frozen
Ali [00:16:31]: Frozen
Philip [00:16:32]: With adapter.
Ali [00:16:32]: Exactly.
Philip [00:16:33]: Yeah.
Ali [00:16:33]: So the brain is frozen and the eyes are frozen. It’s just we’re trying
Philip [00:16:37]: Align
Ali [00:16:38]: Interconnect between the eye and the brain, right? So the projector. And so you take the tokens and then he’s like, “Oh, can you describe what’s in this image?” And he’s like, “Oh, it’s a mountain,” or it’s a person or it’s a human, whatever the case is. But that didn’t cause complete understanding. So he changed it such that every image was associated with a data set of questions. Like, does this image have a white male? Does this image have birds in the top corner? Does this image have a scientist in it? All of that stuff. And it would have to answer questions correctly. And using not just training on describing an image, but being able to answer question, another question, answer over time. Like you can see the grokking, which is like genuinely insane, that retrofitting vision into a large LLM can learn to that extent. And even for images that it doesn’t perform well on, for instance, if you ask it a picture of like Stephen Hawking, “Who is this?” Maybe it doesn’t get it, but it will say something like, “This is Albert Einstein.” Like it still understands
Philip [00:17:25]: Close enough
Ali [00:17:26]: That this is a scientist who is a man who has, some significant achievements, all that stuff. So that’s like really cool.
Philip [00:17:32]: Yeah. So, we’ve covered Hao Tian before, who the author of the LLaVA paper that did this, a while ago. And I think that’s very foundational work for anyone who hasn’t done vision work before.
Ali [00:17:41]: Same with the CLIP and MetaCLIP, where you go from just captioning to building out questions
Philip [00:17:47]: Right
Ali [00:17:47]: Off the image and how much better you can get performance.
Philip [00:17:50]: Right. Right. Right. Yeah. But what’s, what’s so exciting about this is if you look at a model like this. Now, this is a little bit more of a research project. It’s not. It got to 56% on MMLU Pro, I think. So not quite frontier. But if you’re running this model, you haven’t suffered any loss on your GLM-5.2 quality. If you don’t have an image, it’ll just behave exactly the way it used to. And ultimately
Ali [00:18:14]: Which in the inference code you literally do not include the other part, right?
Philip [00:18:18]: Yeah. You would just skip the encoder if you don’t have an image input.
Ali [00:18:22]: Okay.
Philip [00:18:22]: Just confirming.
Philip [00:18:23]: Yeah
Ali [00:18:23]: Does it affect a lot on the overall inference side? Like you’re not adding much, you’re adding a very small vision encoder. These are typically like
Philip [00:18:30]: They’re super fine
Ali [00:18:31]: Less than a billion parameters, right?
Philip [00:18:32]: Yeah. It’s, - There’s a little bit less standardization among vision encoders
Swyx [00:18:37]: Yeah
Philip [00:18:37]: So the support matrix can be a little bit, sparser. But overall, yeah, it’s a pretty, it’s a pretty minor component of the overall system. And ultimately what you get out of the system is all of a sudden you have Kimi Vision, GLM weights, and DeepSeek attention all in one model.
Open Source Model Grafting and Franken-Merges
Philip [00:18:56]: And that’s, I think, a lot of the power and beauty of open source, is that you can take all of these different components and combine them together into a system that’s better than anyone
Swyx [00:19:05]: Yeah
Philip [00:19:05]: Can be individually.
Swyx [00:19:06]: People used to say that you would also do Franken-merges where you would take like
Philip [00:19:10]: Yeah
Swyx [00:19:10]: Layers from each model.
Swyx [00:19:11]: Does anyone do that anymore?
Ali [00:19:13]: Well, to your point previously when you were mentioning like, the work that goes into supporting a model when it first comes out, like GLM-5.2 or MiniMax M3 or whatever the case is. Sometimes you do have to like, you do have to switch out some things. Like, for instance, the MiniMax M3 head uses full attention, and with full attention you end up with this like insane bottleneck in spec dec ‘cause you’re doing auto-regressive token generation for three tokens, and you’re doing this like N squared over all of the tokens that are in your sequence. Your KV cache is like very large because it’s not sparse, it’s not top K. So we find it better to like, okay, we’re gonna replace this, we’re gonna replace this layer with a layer from another model that’s using like GQA, for instance. And then just with the right training, you can get it to have the same acceptance rate. So it is very possible to retrofit layers from other models and very much needed. If a layer is like inefficient, the training just becomes the challenge, like how do you ensure that you train it properly? Which again to your earlier point is like the mesh between training and inference. As in like you need very good training in order to do fast inference. That’s like, I feel like more and more becoming true.
Swyx [00:20:21]: Yeah. Anything else on the support side when you say like get it to fully production ready?
Loop Detection, Race Conditions, and Non-Determinism
Philip [00:20:26]: Yeah. I think that there’s also a question of just, we can test a model to a pretty extensive degree, but we’re trying to get it out quickly and then you see a bunch of other people test it and you get interesting results. There was an issue with, GLM briefly where we had some like mode collapses where it would just output the same token over and over again for certain prompts on certain temperatures. Like once you expose an endpoint to the real world, there’s going to be, so many more varieties of things given to it that you’re able to, discover and patch things. So it’s not just a, day zero process, it’s then like for the first week, for the first month, if a model remains popular, like how do you both fix bugs and then continue to push the envelope on performance?
Ali [00:21:21]: What do you mean you don’t want your model outputting S?
Swyx [00:21:24]: Is there loop detection on that stuff, by the way? It still happens like quite a lot, which is surprising.
Ali [00:21:30]: We have like we, in our endpoint, like if a model was to output the same token like four plus times, we just cut the generation. We say like, “Oh, sorry, this-- Like try again,” or like we will reprocess the request. ‘Cause we know then, like if it, like if, yeah, it’s four times the same token, it’s probably collapsed.
Swyx [00:21:45]: Yeah. Is there a way to opt out in case I really want that?
Ali [00:21:48]: You want that?
Ali [00:21:50]: I think there’s a way that we have to handle it. I’m not exactly certain, but I feel like in certain models, like when they output something like you can imagine, like a table for instance, and so they want, they wanna draw like 12 dashes and 12 dashes. Yeah, I think there’s a way for that to happen. I think we only do it on certain tokens. Like we exclude certain special characters.
Swyx [00:22:07]: Yeah.
Ali [00:22:07]: So we only do it on like certain like S is the most common almost. GLM-5.2
Swyx [00:22:11]: Oh
Ali [00:22:11]: And I think it was DSV 4 as well. Like you’d just have like looping issues where like you literally
Swyx [00:22:17]: It
Ali [00:22:17]: Just have like S.
Swyx [00:22:18]: Yeah. Is there a special, something special about S? No, just randomly
Ali [00:22:21]: It just seems to be the one token involved.
Swyx [00:22:23]: Yeah. And it’
Philip [00:22:24]: Is there
Swyx [00:22:24]: And it’s only temperature 0
Ali [00:22:27]: No
Swyx [00:22:27]: Even at other temperatures
Ali [00:22:27]: Even at like 0.9 or whatever, it will still, it will still collapse.
Swyx [00:22:30]: That’s weird, right?
Ali [00:22:30]: It’s, it is an inference problem to be honest, like a software problem. Like oftentimes, the image you run will-- like NVIDIA will release an image for instance, and if we will upstream the changes from their latest TensorRT-LLM image into our stack, we’ll find that it fixes it. Or oftentimes this will only happen in an inference engine that you’re using like SGLang. But if you were to switch to vLLM, that isn’t the case. So it seems to be like an extremely like deterministic software issue and not really a model issue. It’s not like a weights problem. Like I’- we’ll say like, “Oh, it’s a problem with the quant. We did PTQ wrong,” right? But that isn’t, that doesn’t make sense because the same weights used with a different inference engine does not repeat the problem. And sometimes it’s, the kernels that are being used in the backend have like these very subtle sometimes race conditions, where if you were to use this model hosted on one cluster, you will never get this problem.
Swyx [00:23:19]: Oh my God.
Ali [00:23:19]: But if you host it on a different cluster, you will. And the reason is the KV cache transfer from a node to node in that one cluster is using a slower interconnect than the node to node in another cluster. So that exposes the race, whereas in another cluster it doesn’t. So then you end up just like, okay, this model is not gonna be hosted on this cluster. We’re gonna host it on, another cluster because that cluster exposed that problem. But then it ends up with like, okay, is it the software? Is it the model weights or is it the hardware?
Swyx [00:23:42]: There is a thing about this with temperature 0 still not being deterministic, right?
Ali [00:23:46]: Right.
Swyx [00:23:46]: Mostly because of hardware. Even at temperature 0 same model, you won’t always get the same output.
Swyx [00:23:52]: Even-- But I’m surprised by the race condition one because, I thought PyTorch was a graph that like guarantees that you at least, execute things in the right order.
Ali [00:24:02]: Well, yeah, true. Like I’m not, I’m not saying that there is. Like well, you have things like PTL optimizations where like you can start a kernel before the end of the previous kernel, and that’s like ‘cause you want to do that because there’s
Swyx [00:24:12]: It’s like pipelining
Ali [00:24:12]: Expense. Exactly.
Swyx [00:24:13]: Yeah.
Ali [00:24:13]: But it’- But you don’t do it cleanly. Like you overlap a little bit of the execution. No, it is very possible that the kernel itself, like that one block that is supposed to be running in this instance of time, that kernel itself has a race condition. For instance, like a missing barrier. Like often if you’re designing a kernel and you want it to make it to be very fast, if you don’t test it extensively, you’ll, you’ll have certain threads access data points from registers before they’ve been written to by other threads
Swyx [00:24:36]: Yeah
Ali [00:24:36]: For example, because like your barrier is wrong or your synchronization was wrong. But yeah, like the testing itself is very difficult in those like, and
Swyx [00:24:42]: And there’s no like borrow checker
Ali [00:24:45]: What does that mean?
Swyx [00:24:46]: Like Rust. Like the. If you’re trying to have like memory safety It sounds like a comparable problem.
Ali [00:24:52]: Well, yes, but you’re working in CUDA, right, NVIDIA GPUs. Like- You just need a higher level language like modular Maybe that’s what modular is supposed to do. I don’t know.
Quantization Quality and Vendor Fidelity
Vibhu [00:25:00]: How do you see keeping quality of the model? So you talked about all these steps of, okay, you gotta do quantization, train your own speculative decoder
Ali [00:25:07]: Right
Vibhu [00:25:07]: Run on different hardware. Looking at other model providers, okay, you kicked off a inference speed race on the consumer end. What goes into keeping quality the same across them, right? Sure, you can run benchmarks
Ali [00:25:22]: Yeah
Vibhu [00:25:22]: But, like, how do you determine how much quantization are there standards? What goes into
Philip [00:25:27]: There’s a few things on quality. Most inference optimizations are lossless. KV caching, for example. You are just recomputing or preventing recomputing the same values. Speculation, of course, if a draft token is wrong, it gets rejected. The main lossy optimization is quantization. And that really comes down to, number one, data format, number two, which parts of the model you choose to quantize, which layers, and number three, like doing a lot of calibration on the quantized weights, to ensure that you’re preserving all the outliers. There’s other tricks that you can do, though. A big one is long context, ‘cause one thing you asked at, right at the beginning is, “Oh, what’s gonna happen if I send a 200,000 token request in?” So with a long input sequence, you need to, store a lot more information. You need to process a lot more tokens. And so even if a model has a context of a certain length, you might, as an inference provider, choose to build an API with a shorter context length, and of course a full length one as well. Because if someone doesn’t need the full million token context, for example, you can get them better performance. I don’t know if that’s exactly like quality of the model. The way that I think about quality is to what degree are we faithfully serving the original model? If you think of a golden implementation of a model that performs exactly the way the model is designed to perform, I think of quality as how close are we getting to that, 100% fidelity of the model.
Philip [00:27:13]: You can also, of course, think about quality from the training side and how do you push yourself past 100%. But when I think about purely inference optimizations, it’s getting faster while staying as close to that 100% fidelity mark as possible. And certainly our standard internally is that, like you should not be able to tell the difference between our API and a, official API. I think Kimi in particular does a good job of vendor benchmarking here
Ali [00:27:41]: Yes
Philip [00:27:41]: Where they have
Ali [00:27:42]: They released an actual vendor benchmark.
Philip [00:27:43]: Exactly, yeah.
Ali [00:27:44]: ‘Cause they accused, some people, Amazon? There was some provider that was not doing very well on Kimi’s benchmark.
Philip [00:27:50]: Yeah.
Philip [00:27:51]: So, with Reflect we probably
Vibhu [00:27:52]: This was a long time ago, right?
Philip [00:27:54]: No.
Ali [00:27:54]: Yeah, like three
Vibhu [00:27:55]: They also
Ali [00:27:55]: Four, five months ago
Vibhu [00:27:57]: This also happened with, I don’t remember which model, but they pulled out quite a few, and then they started a whole chart about this. It might have been
Philip [00:28:03]: Kimi Vendor Verifier.
Ali [00:28:04]: Yeah.
Philip [00:28:05]: Yeah.
Ali [00:28:05]: Yeah, ‘cause you, ‘cause you’d be pissed, right? Like if you’
Philip [00:28:07]: Yeah.
Ali [00:28:07]: If like if I’m a consumer and I’m using like Amazon’s endpoint for instance, and I’ve used Kimi and I’m like, “Oh my God, like this is bad,” I’m not gonna say, “Oh, Amazon quantized the model in a bad way.” I’m gonna say, “Oh, Kimi sucks.” Right?
Philip [00:28:17]: Yeah.
Ali [00:28:17]: So it seems like that makes sense.
Philip [00:28:19]: Yeah, they care. They care.
Vibhu [00:28:21]: Justifiably.
Ali [00:28:21]: Yeah, justifiably.
Vibhu [00:28:22]: This is probably a stupid question, but just checking, has anything improved from main quantization?
Philip [00:28:28]: Yeah.
Vibhu [00:28:28]: Like, is quantization always strictly worse?
Ali [00:28:30]: Well technically
Vibhu [00:28:32]: No
Ali [00:28:32]: It’s a lossy. Quantization
Philip [00:28:33]: Yeah
Ali [00:28:33]: Is a lossy, it’s a lossy implementation.
Philip [00:28:36]: Speed improves
Vibhu [00:28:36]: Speed improves.
Ali [00:28:37]: It the number, like
Vibhu [00:28:38]: No, I’ always look for inverse scaling laws.
Philip [00:28:40]: Yeah.
Ali [00:28:40]: Yeah.
Vibhu [00:28:40]: This is something I learned from Noam Brown, where like things that normally act in one direction sometimes do.
Philip [00:28:45]: Well, technically when you run a benchmark, because these models are deterministic, sometimes your,
Ali [00:28:52]: Yeah
Philip [00:28:52]: NVFP4 quant is like, two basis points higher than your
Ali [00:28:56]: No, it’s noise. It’s noise.
Philip [00:28:57]: Yeah, exactly. I’m like, yeah, it’s, it’s within. That’s why I always say within margin of error.
Philip [00:29:01]: And I stopped saying that because everyone assumes that what is, well, within some margin of error, we’re barely inside of that to the worst, so we’re saying. But yeah, sometimes it’s just like, gives you a higher output score. But like Ali said, that’s noise. To my knowledge, you’re not necessarily making the results better. You’re just trying to, again, like keep your fidelity as close to 100% to the original model.
Layer Selection, KL Divergence, and Better Quantization
Ali [00:29:27]: There is, to your point, research that we did on MP. I don’t know if you are able to pull
Philip [00:29:31]: Yeah
Ali [00:29:32]: A tweet we did. One of our research interns, Joshua, I think it’s a tweet on how we have 20% better quantized GLM-5.2 than NVIDIA. Essentially what we found throughout like this month research is, okay, quantization is a lossy. It’s. You’re compressing the data from, occupying 16 bits to occupying, four bits, for instance. And so you’re losing some information, and you’re trying to minimize that. And so when I say that I’m gonna quantize the model, my job becomes how do I find the layers that I can quantize, and how to find the layers to not. For instance, with image models, I don’t quantize modulation layers, and I don’t quantize out projections because those two are. Like out projection is what you see as the user. Modulation is what the model sees or understands. Right, exactly. And so to his paper, do you have the. It doesn’t have the. Yeah. It’s a long paper. I don’t know if I can find
Vibhu [00:30:25]: If there’s a part to search or it’s probably in the thread.
Ali [00:30:28]: It’s probably in the thread.
Vibhu [00:30:29]: Yeah.
Ali [00:30:29]: But the long and the short is it is very possible that quantizing more of the model makes the results. Like if I have a model that I quantize layers one, five, and 10, and another model where I only quantize layers one and It is possible that the model in which I quantized more information is going to perform better because the quantization errors have canceled out. And so what Joshua showed in his mathematical proof where he had like a verifier in, is that you can predict which layers are going to have quantization errors that will cancel out with each other, and you choose to quantize those layers. And so the result of doing this mathematical quantization is you end up with a model that’s 20% more quantized than another provider, so you get 20% more throughput of it because there’s more layers than running an NVFP4, and your quality is better than that other quant because the layers that you chose to quantize have their errors cancel out, like one layer skewed to the right one layer skewed to the left, one layer skewed to the right. Your final logits distribution is more similar to the original distribution of the model, so you have better fidelity. And so the way we proved this was with KL divergence. So instead of just scoring on the benchmarks, we scored the KL divergence between the logit distribution of the quantized model and the logit distribution of the original full precision model, and we showed that with this technique we get. If your probability distribution on the logits which token it wants to select is more of the same as the original model, you’re probably gonna end up staying true to the original model. So yeah, so it seems like previously before this, it seemed like the industry was, well, the more you quantize, the worse it’s gonna be, ‘cause the more loss you introduce. That’s not exactly, not necessarily true. So yeah, doesn’t improve it, but can cancel out.
Philip [00:31:57]: I think it might be this, but reminds me a good bit about pruning where you can prune off certain layers.
Philip [00:32:03]: But very interesting. Didn’t know this was a whole paper you guys put out.
Ali [00:32:06]: It’s. Fun fact, it was originally 72 pages, this paper, and then we decided
Philip [00:32:11]: Wow
Ali [00:32:11]: We can’t tell. We couldn’t release it. So it’s now 45.
Swyx [00:32:15]: Still 39 pages, so very substantive. We talked about evals and all these things and, like what’s possible in terms of speedup? Like it’s like probably like the number
Inference Speedups and Benchmarking
Swyx [00:32:25]: Thing that people do wanna care about, and it’s something that you wrote about in your post. Like official API is 70 tokens per second, and you push it up to 90. Is that like a normal thing?
Philip [00:32:36]: So what’s cool about working in inference, the reason that I think inference is going to be a useful place to do engineering for a long time, is that if you look at highly optimized domains like, say, finance, if you’re in finance, you measure how much better you got in basis points. It’s like, “Oh, I got five basis points better, like twentieth of 1% better,” that’s huge news because everything is so optimized. When we publish optimizations, it’s 20%, it’s 100% it’s 200%. So there’s still probably like a lot further to go, honestly. Like you’ll, you’ll know that inference is pretty much solved when researchers start publishing about how they got 1% faster at something.
Swyx [00:33:19]: Which by the way, because I am from the finance background, in the ‘70s, that was the margin at the time. When you did quantitative finance research, you would find
Ali [00:33:27]: And like 20%, tens of percent.
Swyx [00:33:29]: That’s. Yes.
Philip [00:33:29]: Yeah.
Swyx [00:33:30]: And now it’
Philip [00:33:31]: Tiny fractions
Swyx [00:33:32]: For those people interested, look up Andrew Lo’s paper. He had a really interesting illustration of quant, stat arb, distribution, narrowing down from like those kinds of 20% differences in the ‘70s, down to nothing today, which is very cool.
Philip [00:33:48]: Exactly, and we’re at the beginning of the same type of thing. Now benchmarking is hard. I think anyone will tell you that, and benchmarking provider speeds is hard because there’s so many variables that go into it. What hardware are you using? How much load do you have on the system? What’s the exact nature of the prompts and input and output sequence lengths? All that stuff. But overall, when you start stacking these improvements, you’re looking at multiples. You can look at it. The most common form, of course, is TPS, tokens per second, which is bad naming by us in the industry, ‘cause there’s two tokens per second. There’s tokens per second, the throughput number, and the latency number.
Ali [00:34:31]: TTMT, yeah.
Philip [00:34:32]: Like total tokens per second out of the, out of the GPU as a throughput number. Most people only care about tokens per second as the latency number, which we should call ITL, intertoken latency, but we don’t.
Philip [00:34:44]: Anyway, so you can imagine a standard API without many optimizations for a 1 trillion parameter model operating somewhere in the 30 to 50 tokens per second range for reasonable traffic profile. And we generally see the goal of, pushing to 10X that. But, not necessarily day zero, but by stacking enough optimizations, if you have, say like four optimizations, each of which doubles performance. Or sorry, three optimizations, each of which doubles performance, then you stack that up, that’s an 8X gain. That’s the order of magnitude that we’re working with in this space. We’re trying to make things substantially faster, not just go from like 70 to 90.
Swyx [00:35:38]: Are you saying you’ve. You have done that?
Philip [00:35:40]: So let’s say you have as a reasonable baseline, 30 or 40 tokens per second. You can achieve 10X that. So like on GLM-5.2, if you run it unquantized, perhaps on H100s even, and you’re just using an off-the-shelf inference engine with no particular optimizations, no speculator, nothing extra around like KV routing, no disaggregation, you’re, you’re probably, yeah, looking at that like 30 to 40. You think that’s like a reasonable baseline?
Swyx [00:36:12]: Right. Right.
Philip [00:36:12]: To get to something like 10X, there’s a lot of trade-offs that you’re making. If we’re running at more like a 300, 400 tokens per second range, you are using the best hardware possible. You have a optimized speculator. You have done all of your quantization work. You are Seeing a pretty high cache hit rate. You are running with a reasonably small batch size and a parallelism configuration that is tuned for latency versus throughput, but it is possible. So the spreads that you see if you, like, go on artificial analysis or you go on OpenRouter and you look at, the worst provider to the best provider, oftentimes can hit that range. 10X is of course very aggressive. It’s oftentimes maybe more of a four to six times improvement. But that’s the performance that makes us really excited, is when we can get these huge gains, not just go from 70 to 90 tokens.
Stacking Optimizations: NVFP4, Speculation, and Disaggregation
Ali [00:37:19]: It’s also, like, hardware dependent. Like, if
Philip [00:37:20]: Yeah
Ali [00:37:20]: If you have a thing where you’re serving it on just, like, a node of H100s and then you throw, like, you shard the model across, like, four nodes of B200s. Like, you can definitely increase the speed with just throwing more hardware at it. Like, normalizing for the same exact hardware and the same number of GPUs.
Philip [00:37:35]: Yeah. Then you’re looking at, like, a two to 4X improvement
Ali [00:37:38]: Right. Right
Philip [00:37:38]: Depending on the inference optimizations. So yeah, it’s. Some of it’s, what’s the call, and some of it’s who’s the driver.
Vibhu [00:37:46]: If you break down the two to 4X, say the example is run GLM-5.2
Ali [00:37:51]: Yeah
Vibhu [00:37:51]: On B200s
Ali [00:37:53]: Yeah
Vibhu [00:37:53]: Single node, right? What’s, like, the cost trade-off for effort to get, like, the last bit of juice out versus what should people just think of, right?
Ali [00:38:01]: Spectre quantization. Yeah.
Vibhu [00:38:03]: Spectre quantization.
Ali [00:38:04]: That’s, that’s, that’s like 95%. Like
Vibhu [00:38:06]: And how far does that get you? And how easy is that for the average person to do? So say right I wanna throw the weights of GLM-5.2 on a node of B200s, how easy is it to find speculative decoder- decoder model or already quantized model? How much work goes into it?
Philip [00:38:23]: If you’re doing it up front, it’s quite a lot of work. If you’re doing it today, there’s going to be people who have published things that you can just, you can just grab some NVFP4 weights. You can grab a speculator. Yeah, if we’re thinking about, like, what are the 2Xs we’re stacking, going from, BF16 to NVFP4 is, it’s not quite a 2X, right? It’s like. I think it’s about, like, 30 to 40%, from 16 to 8, and then another 30 to 40% multiplied from, 8 to 4. So that doesn’t quite get you a 2X, but, like, roughly a 2X. Speculator, roughly a 2X. Disagg on top of that if you’re able to get enough hardware and put enough traffic through it, another roughly a 2X. And then you add in some, double-digit percent increase from having just a better runtime with, the latest kernels and stuff behind it. And that’s how it stacks up.
Ali [00:39:21]: Yeah
Philip [00:39:21]: So building each of those, like, building the, quantized weights is, for someone who really knows what they’re doing, hours to days of work. Building the speculator, again, like, hours to days of work. And the, disagg setup, hours to days. Well okay, but like once you have
Ali [00:39:39]: Once set up. Once set up. Yeah
Philip [00:39:40]: Yeah, getting disagg working for the first time, I’m saying, of course, is very difficult.
Philip [00:39:44]: The marginal implementation
Ali [00:39:48]: Like, if you’re just grabbing, like if you are a person, like just a normal consumer who has access to, like, a node of B200s and you’re wondering, “How can I just host it myself?” You don’t need to quantize the model yourself. There’s always gonna be, like, an open source quantized checkpoint. NVIDIA’s gonna push one out if no one else does. You. Usually, the providers will have their own spec dec that they’ve trained as well. You don’t need to train your own spec dec. You can just use that as well.
Philip [00:40:09]: Yeah. Like, GLM-5.2 has its own MTP.
Ali [00:40:13]: Right. Right.
Vibhu [00:40:14]: What’s multi token prediction?
Philip [00:40:15]: Yes.
Ali [00:40:16]: I’m just
Vibhu [00:40:16]: Can you explain that?
Ali [00:40:16]: I’m just an expert.
Ali [00:40:18]: I can do it for you in case I get it wrong?
Vibhu [00:40:20]: No.
Vibhu [00:40:21]: Yeah, you should correct if we’re wrong, but their multi-token prediction can be used for self-speculative decoding.
Ali [00:40:27]: I’m not sure. I’m not gonna correct that.
Vibhu [00:40:28]: Okay. I’m semi-confident in that
Ali [00:40:30]: Okay. Yeah
Vibhu [00:40:30]: But someone can check. But it’s useful to paint the story of, okay, not just the average person, but say a company wants to switch from serverless inference I wanna throw this up on. I wanna rent some GPUs, throw it up. These are the steps you take to do significantly faster than just put it behind vLLM.
Ali [00:40:48]: Right.
Vibhu [00:40:49]: I was waiting for a mention of Dynamo.
Vibhu [00:40:51]: I feel like, that’s supposed to be the baseline that you measure against.
Dynamo, KV Routing, and Disaggregation Toolkits
Philip [00:40:55]: I would think of Dynamo as less of a box system and more of a toolkit for building with. So when we talk about doing aware routing, when we talk about doing KV offloading, when we talk about doing, PD disaggregation, Dynamo fundamentally is. By the way, Dynamo is an open source library from NVIDIA.
Ali [00:41:17]: We’ve done a pod with Kyle
Philip [00:41:18]: Okay
Ali [00:41:19]: Kyle Cranin.
Philip [00:41:19]: Cool. So then your listeners know then that it supports all the different inference frameworks. And it is multi hardware, which is interesting.
Ali [00:41:28]: But it’s just a router, it’s not like an optimizer layer.
Philip [00:41:30]: Yeah. All it does, like, what Dynamo is good at, it is a library for moving information around your cluster, around your hardware. So if you have, KV cache on one place and you need it to be somewhere else, Dynamo coordinates NIXL for you to move that around.
Philip [00:41:49]: That doesn’t mean that, like, out of the box, you just say, “Pip install Dynamo,” and then you get, like, a massive performance speed up. It’s more of a developer toolkit.
Ali [00:42:01]: Yeah. I would have said it would. It comes with a set of defaults that you can then swap out.
Philip [00:42:06]: It does. If the industry at large, I think, was, like, rolling out all of these deployments, standard, then I think it would be, like, a credible baseline. But, we’ve got to, we’ve got to benchmark against, like, what we’re seeing in the wild.
Speculative Decoding Methods: Medusa, EAGLE, n-Gram, and Spec-Spec
Vibhu [00:42:23]: I did wanna talk a little bit more about PD disagg, because that is probably, like, number three after quantized and speculative decoding. In your book though, I was just gonna pull out the book.
Philip [00:42:31]: Yeah.
Vibhu [00:42:32]: Like section 522 on Medusa, 523 on EAGLE
Philip [00:42:35]: Yeah
Vibhu [00:42:36]: 524 on gram.
Philip [00:42:37]: It’s 55, would be disaggregation
Ali [00:42:42]: Yeah. Well, no, I just wanted to dwell a little bit
Philip [00:42:44]: Yeah
Ali [00:42:44]: The other. Like, so what do you choose to include? What do you choose to not to include? Because there was all these other techniques.
Philip [00:42:51]: Yeah.
Ali [00:42:51]: Are these still relevant? Because I think they came out, like, a year and a half ago maybe.
Vibhu [00:42:55]: Medusa is quite old.
Philip [00:42:56]: Yeah, Medusa’s old.
Ali [00:42:58]: It was old.
Vibhu [00:42:58]: But is it in the book as a good, here’s
Philip [00:43:01]: Baseline
Vibhu [00:43:01]: Baseline vanilla understand it?
Philip [00:43:02]: Like you should know this.
Vibhu [00:43:03]: Like I read the paper, I’m like, “ it makes so much sense.”
Philip [00:43:05]: Yeah.
Philip [00:43:05]: So with the book, I had a couple goals. One was to give people just a working vocabulary for the space as a whole, and the other was to give them some intuition about how each of these techniques works. As I mentioned in my AI Engineer talk, which is the first public addendum to this, the speculation space has moved much faster than everything else. So yeah, even at the time that I wrote the book Medusa, I very much included as a way for people to understand how the space evolved rather than what the most modern technique is. And now of course, there’s DFlash, dSpark. There’s, there’s newer techniques even than EAGLE, although EAGLE is still very commonly used.
Ali [00:43:51]: SpecSpecta.
Philip [00:43:52]: Yes. Speculative decoding.
Vibhu [00:43:54]: What can

Ali [00:43:56]: Oh, it’s a paper by Tri Dao and it’s like, it’s doing speculative decoding
Vibhu [00:44:00]: Huh
Ali [00:44:01]: For the speculative decoder.
Philip [00:44:02]: Oh, in spec- oh my God.
Ali [00:44:02]: It’s literally just an another. It’s like, yeah, that’s the most simple way to explain it, and it seems like he got trivial speed ups there. But it seems that the complexity with training, it’s almost like in our mind at least, it’s almost as complex as training GANs. Like it’s like a very delicate balance and oftentimes you, it’s just but yeah, it’s literally speculative decoding on speculative decoding.
Vibhu [00:44:21]: Speculative.
Ali [00:44:22]: Yeah. We saw this paper.
Vibhu [00:44:24]: It’s interesting, right?
Ali [00:44:24]: Yeah.
Vibhu [00:44:24]: I wouldn’t even expect it to be very particular to train, I would
Ali [00:44:29]: Right.
Vibhu [00:44:29]: The naive part of me is like, okay, train speculative decoder.
Ali [00:44:32]: But like, and it makes sense, like the whole idea of speculative decoding is you. It’s like, it’s like almost like the iPhone auto predict version but for a normal model, right? Like you’re just, you’re just, generating three tokens and you’re like, okay, I’ll do prefill on them. And so you save those three turns for your original model. Now your speculative decoder is doing three turns of auto regression, so why not just have an even smaller model?
Ali [00:44:53]: The other question there is what are the size of speculators? So say for
Philip [00:44:58]: Right. It’s like a billion parameters.
Ali [00:45:01]: Like for MiniMax, it’s. Yeah. It’s like one layer. It’s like one 60th of the original model usually.
Philip [00:45:06]: Yeah. I think we should do a paper when we get back to the office.
Philip [00:45:10]: Speculative
Ali [00:45:11]: Speculative
Philip [00:45:11]: Decoding.
Ali [00:45:13]: No, it’s, it does seem like how, when do you stop? But then it also seems like if you’re able to train spec-spec decode for instance, right? Like if you’re able to have a small model that is accurately predicts what the intermediate speculator is gonna predict, that is able to predict what the original target model’s gonna predict, then why not just use that smallest model directly, right?
Vibhu [00:45:34]: Yeah. This is
Ali [00:45:35]: Like it seems like
Vibhu [00:45:35]: Adjacent to the routing problem.
Ali [00:45:36]: Right.
Vibhu [00:45:36]: Yeah.
Ali [00:45:36]: Right.
Philip [00:45:37]: The thing with speculators is one of the practical constraints on using them is that you do have to run a small model on the same hardware that you’re running the big model on. There is a orchestration and resource competition problem inherent in that, and that is one of the constraints on speculation in general, is that draft tokens cost resources to create and cost software complexity to manage. And so if you have like infinitely recursive speculators, you add in quite a bit of that complexity on the actual implementation within the inference engine as well, not just in the training process.
Vibhu [00:46:17]: I was gonna say, I would wonder if you could do similar, like distillation and pruning of, it’s the same thing, it’s just a model. Can we not just distill a lot of the weights, quantize the speculator, out of my domain? The question that also comes up is, this is all for big server workloads, right? How much of this applies to, say I have this MacBook, I wanna run Gemma really efficiently. Similar problems, not the same?
Local AI vs. Data Center Inference
Philip [00:46:45]: Pretty different. I talked to Selo, about this on his podcast a couple weeks ago. The difference between inference engineering for the data center and for production workloads versus inference engineering for local AI, is that we start with fundamentally like different constraints and different goals. With local AI, it’s how do I fit this model onto my hardware and then make it less dumb? And with data center influence, it’s how do I load this model and then make it less slow? And we care about less dumb, and they care about less slow. But the local AI inference engineering ecosystem, I think has a lot for us to learn from in the data center space. They are experts in various forms of quantization, including dynamic quantization that we just don’t touch, in the pruning, in the distillation, in the, layer removal. There’
Ali [00:47:42]: Layer removal matters less.
Philip [00:47:43]: Yeah. There’
Ali [00:47:44]: No one loves pruning really.
Philip [00:47:45]: Yeah. Well, but the, but they do
Vibhu [00:47:46]: Which is surprising, right? But that’s, that’s a whole different thing
Philip [00:47:48]: Just to fit something on the laptop.
Ali [00:47:50]: Right.
Philip [00:47:50]: So yeah, it’s a, it’s an interesting, it’s an interesting space. Not necessarily that like their techniques make sense for us to do in the data center, because we have different resources and different goals, but more that the process as well as the openness of that field is something to, admire.
Ali [00:48:12]: Yeah. Like to your point, like, certain optimizations that would. Like for instance, Turbo Quantum Sharper, like it made such huge hype on that and we did like a whole deep dive on Twitter and like said, what is it? How does it work? Why is it good or not? And it took off and it was implemented on local devices because your memory bandwidth is so slow on like a MacBook, for instance. But try putting the same thing on like an NVIDIA GPU on a B200 Turbo quant would not be. Like, it would not be used. Like, NVIDIA - Like, NVIDIA made it clear that this is not a good optimization, and we’ve seen it firsthand where the overhead of doing dequantization, quantization of, in the kernel itself with turbo quant kernel, each end is much slower than the time that you save from doing the bandwidth. ‘Cause on the B200s, you have like 3.5 terabytes per second. You don’t need decrease the storage that much. You don’t need to do, FP4 KV cache. You don’t need to use a requant. There’s, there’s, there’s better optimizations to be made. But on Edge devices, it’s extremely important, it’s extremely useful. So, seems to be, like, different optimizations there, but then they’re all uniquely combined with like all you wanna quantize the model, you wanna do speculative decoding, like certain common prefixes with both
Philip [00:49:18]: Principles.
Ali [00:49:19]: Yeah, exactly. Exactly. Exactly.
Philip [00:49:20]: They also do a lot of work on, model parallelism, especially over, heterogeneous topology, where you have, some sparks and they are wired together with, Ethernet, DGX sparks.
Ali [00:49:35]: Yeah, this is the Exo Labs guys.
Philip [00:49:36]: Yeah. You have, a number of, Mac Minis stacked up.
Philip [00:49:41]: There’s, the inter. They. One thing that I think we both have to deal with, although they have to deal with a lot more is the interconnect between machines. Which is why, like, one thing that we do a lot is work with tensor parallelism.
Philip [00:49:56]: And that’s where, you are using all of the, all eight GPUs, and sharding the model across it. Tensor parallelism is not a good fit for local AI because it assumes a very high bandwidth interconnects like NVLink. Was, they might be forced to do something like pipeline parallelism, which we’re never gonna do unless we’re doing some kind
Ali [00:50:16]: Yeah. For image
Philip [00:50:17]: Multi-node inference.
Ali [00:50:18]: But since you mentioned it, I wasn’t sure if we were gonna cover it, but let’s briefly explain tensor parallelism and expert parallelism, since you have very nice images.
Tensor, Expert, and Pipeline Parallelism
Philip [00:50:25]: You wanna pull the book?
Ali [00:50:26]: Yeah.
Philip [00:50:26]: Yeah. Let’s, let’s get
Ali [00:50:27]: So I just wanna show a few images.
Philip [00:50:29]: Yeah. Shout out to Luke from Baseten’s design team for making these beautiful images. Oh, that’s a, that’s. Before we get into this, just one other difference is we talk a lot about the active parameters of a mixture of experts model, and for local inference folks, that matters a lot because if you have a batch size of one, you’re only activating that many parameters. When we
Ali [00:50:51]: Yes. I was gonna
Philip [00:50:52]: Inference in the data center
Ali [00:50:52]: I was gonna bring that in the diffusion conversation.
Philip [00:50:54]: Yeah.
Philip [00:50:55]: Yeah. We, I, when we go through like a MoE model, and we host it, for an API, we assume that all parameters are gonna be active because
Ali [00:51:06]: You’re batching
Philip [00:51:06]: Throughout your batch
Ali [00:51:07]: Yeah
Philip [00:51:07]: You’re gonna, you’re gonna hit everything. Cool. So broadly, tensor parallelism you can do with any model. Expert parallelism, you can only do with MoE models. Effectively all models today are MoE models, that are,
Ali [00:51:21]: Sort
Philip [00:51:22]: At least all models large enough that you would care to parallelize them across multiple GPUs. So that’s, that nuance is less important now. With expert parallelism, the idea is you put the entire expert on a GPU. Generally, you have more experts than GPUs, so you might put like N experts per GPU, like eight experts per GPU or whatever. And then you replicate the router, which the router is very small, across each of the GPUs. And then by moving the generation from expert to expert, with each expert being inside a GPU, they’re not competing for resources. You massively increase the throughput that you’re capable of doing, and the, GPU connection is not as important ‘cause there’s not as much communication. Tensor parallelism requires that you are able to do this like all gather, all reduce. So you shard the model across the GPUs entirely. And then for each step, you’re combining the results of each of the GPUs, which is why the interconnect matters a lot, and it is generally. Of course, this is a, this is a very high-level generalization. There’s a lot of places where this is not correct. But generally, TP is helpful for latency, and in many cases, you will use some combination of these two parallelisms, across the model rather than just, like, picking one or the other. Do you wanna add some color there?
Ali [00:52:50]: Like, yeah, usually, like in a model, it’s not. They’re not mutually exclusive. You do tensor parallelism and you’ll do expert parallelism. Pipeline parallelism less solely, it seems to me like we never use pipeline parallelism.
Philip [00:52:58]: Yeah. The only reason you would have to do pipeline parallelism, which is where you separate like different layers and you put like half the layers on one hardware and half on another, is if you are forced to do multi-node inference, because a model is bigger than you have the. Like let’s say, let’s say you’re doing a deployment on H100s for whatever reason, and you’re putting a trillion-parameter model on there. You have to use multiple nodes of H100, and so you. - Because the interconnect is so slow between the nodes, the only viable way to parallelize there is pipeline, but then you would do expert and tensor within each node.
Ali [00:53:36]: And the limiting factor for H100s is HBM?
Philip [00:53:39]: Yeah. They just don’t have enough
Ali [00:53:40]: How much? What’s the magic numbers that we need
Philip [00:53:43]: Like on a B200 is 180 gigabytes per GPU, and then a node of eight, so you’re talking like 180 times eight. And the FP4, so each parameter takes half a byte, so that’s 800 gigabytes. On a H100, it’s like 140?
Ali [00:53:56]: It’s 80.
Philip [00:53:57]: It’s 80?
Ali [00:53:57]: Yeah.
Philip [00:53:57]: Oof.
Ali [00:53:58]: Yeah.
Philip [00:53:58]: I’m old. I’ve been doing this a long time. I remember H100 specs.
Ali [00:54:04]: Yeah.
Philip [00:54:04]: No, so one thing
Ali [00:54:06]: You wanna tell me about the T4s?
Philip [00:54:07]: The T4s. Oh my God.
Ali [00:54:08]: Let me tell you what it was like to run a model on a T4 back in the day.
Ali [00:54:12]: One thing I was surprised to see that more people didn’t do, Jamba. I don’t know if you guys remember Jamba from AI ‘21. They would specifically pick a hardware, and then they designed the arc dimensions for the hardware, and then it would saturate the hardware. Like, it makes sense. And like, somehow all these models don’t do that.
Hardware-Aware Inference and Auto-Tuning
Philip [00:54:32]: Don’t they do this for the training side, though?
Ali [00:54:35]: I don’t know.
Ali [00:54:36]: Sorry,
Philip [00:54:36]: Training. For training the model.
Ali [00:54:37]: Like deciding which GPU, which
Philip [00:54:39]: Yeah. Well, how
Ali [00:54:40]: Yeah, they do And with training, it’s more of like a math. Like you can run the math- Yeah and see the flops and maximize it. With inference, it’s more of like an auto-tuning, like if you like GPU kernel auto-tuning. But like it’s like you define that, “Oh, I have two GPUs. I can do TP1, TP2, EP1, EP2,” for instance, right? And you. So that gives you like total of like two squared combinations, and then you just like you shadow the same traffic, like real prod traffic, and you just see which configuration gives you the best TPM and TPS, and then just use that. I don’t like the fact that it’s, you cannot reason about which one’s gonna give you the best performance or that there isn’t one specific configuration that’s always best. But it seems like auto-tuning is just the way that you find the best one. And with kernels and GPU kernels, it’s much of the same. After you design your kernel and you design your configuration, how many threads do you launch? How many, how much shared memory do you use? You just auto-tune. You just sweep the parameter space on the side, and this is the best one empirically. But yeah, but they are combined. They’re not just entirely- Yeah like separation. There’s a few bits of training that are like hardware targeted. If you look at, for example, NVIDIA Nemotron models, they run very well on Blackwell. That’s, that’s unsurprising. So there’s some degree of that, but I think that most open labs are trying to make models that can be run on as wide of hardware as possible rather than targeting just like a single chip. I see. For usefulness. Yeah. Okay, one more thing while this chart is still up. All gather, all reduce is expensive. One of the things that is a movement in Silicon Valley is mega kernels, just keep fusing kernels. I don’t know. Is it that simple? Well, I, like a fused kernel can’t save you. Like here with tensor parallelism, you’re. The half the matrix is on one GPU and the other half is on another, and if I need the entire matrix in order to do like a nonlinear operation in the next step, which is, for instance, like if I’m doing attention, I need the softmax, or I need to do like exponentiation, I need to have the entire row. So I need to know what the partial result was from GPU 2 and what the partial result was from GPU 1 in order to be able to do the softmax in the next stage. So I, like I have to make them communicate with each other, even if I had a fused kernel, because of the nonlinearities within each one. Also with like mega kernels, like honestly, I’m, I’m, I’m very bearish Ooh on, I’ll be honest. Like- Please. No, it’s just like mega kernels, it was a good research direction, and it seems like a very. Like intuitively, theoretically, it’s nice. Like, oh, like you have a lot of launch overhead from launching- Just- one kernel- Yeah, just keep fusing it moving the data. Just fuse everything together. But yeah, but like the kernel complexity itself is very difficult to write a very optimized mega kernel. It’s, it’s very difficult to do so. And even the, like not to name any companies, but like even the companies that have worked or people that I’ve spoken to who work at companies that do fused mega kernels, they very often don’t end up running those in production because the TensorRT-LLM and modular kernels that launch are faster because you can optimize each individual component, and you can just have them parallelize with each other. With the Rubins, I don’t know if you guys saw the Rubins Twitter post yesterday, but they’re also, Rubins? Like- No, like Rubin, like the GPU. NVIDIA GPU the, yeah, GPU. Yeah. They have a Twitter account for Rubins only? No. Okay. I was like, “What are you talking about?” Yeah. Sorry. One of the tech leads at NVIDIA is like launched a Twitter post said like, “We’re pulling the curtain on Rubin, and here’s the, here’s the specs.” And the third tweet showed, like not to get too technical into it, I and I need to read it much more, but the GPU is designed in such a way that it kills mega kernels. You don’t need to use mega kernels that much anymore. So it seems like that entire research field goes into like, won’t be continued, but yeah. Can I speculate about Rubin for a minute, please? Go. I’ve been through now, we And by the way, they are covered in the book. Yeah. But yeah, they- Well, they’re covered in the book in the sense that like I am aware- The Wikipedia entry from the blog post- Yeah that Rubin is going to happen in the future. And you even had the name of the one, Feynman. Yeah, it’s like, “Hey, this is gonna “ I was like, “This is very up to date.” Like I’m trying to future-proof this thing, okay? I don’t wanna publish a new one until like next year or something. Anyway, so we were discussing the degree to which I am old. And I’ve now been through three hardware launch cycles. I’ve been through the Ampere launch cycle, the Hopper launch cycle, and the, Blackwell launch cycle. Now, when I say launch cycle, I don’t necessarily mean like the actual shipping of the hardware. Like Ampere’s were racked up well before I got in this industry. But there is a lot of time between hardware being racked up and hardware being feasible for inference. So if you look at like the original vLLM and SGLang, vLLM especially, like that was written targeting Ampere and then had to be updated for Hopper, updated for Blackwell. With each of these cycles, it becomes faster and more urgent, but also substantially more complicated. When I look ahead to, what’s going to be new with Rubin, I think that like Dynamo gives me a lot of technical hints around like what kinds of work is going to be very valuable. We’re continuing some trends from Blackwell, right? NVFP4 is big. The amount of compute that they have behind NVFP4 tensor cores is massive. We’ll, we’re gonna talk about video, I think, at some point, and that’s the big barrier there. You’ve got, much faster memory bandwidth, but which was the same thing that made Blackwell so good. But the big thing is more systems thinking. You have more emphasis on the CPU to GPU interconnect, more emphasis on the interconnect between GPUs, and when you look at Dynamo, it’s a system entirely designed around how do I move the KV cache to where it needs to be when it needs to get there? So I think that themes around like KV cache offloading, KV-aware routing, and disaggregation are going to be substantially more important in the Rubin era, which means that inference engineering becomes not just a like CUDA kernel problem, but also like a very traditional hardware infrastructure problem, which is something, we’ve been building toward for a long time, and something that’s like very exciting to me because we’re gonna see
Mega Kernels, Rubin, and the Future of GPU Systems
Philip [01:00:55]: Multiple domains colliding and the ability to reason from the kernel level, like up to the hardware level and back down is going to be very valuable.
Ali [01:01:05]: I will take what Phil said one step further, into that. It’s, I think, trending towards becoming exclusively an infrastructure problem, where like problems of PD disagg, Training, spec dec. But troiting kernels is not going to be much of a problem because the GPU is moving more towards being an ASIC, where it’- you’re just, you’re just trying to orchestrate what happens on the GPU, but you’re not controlling it thread by thread level. And you see this with like QTAL, QDSL, like you’re, you’re just working at levels of like tiles of data, but you’re no longer working at controlling what each thread does on the GPU that’s being taken care of for you. So do you agree that a GPU and future GPUs are trending more and more towards becoming ASICs that just need to be launched and then they do the data operation based on your conversations with other people?
GPUs, ASICs, and Specialized Hardware
Swyx [01:01:50]: Oh, yeah, no. That is a section of the market.
Ali [01:01:55]: Right.
Swyx [01:01:55]: And ASICs can do, a lot more performance for only their workload.
Ali [01:02:01]: Right.
Swyx [01:02:01]: And the G in GPU makes them continue to be very general.
Philip [01:02:05]: Yeah. The, - I think that there’s like a spectrum
Swyx [01:02:08]: It’s graphics,
Philip [01:02:09]: Yeah.
Swyx [01:02:09]: I keep saying this, I have to correct myself in case people come at me for getting the G wrong.
Philip [01:02:14]: Yeah. It’s like, it’s like a spectrum, right? Of a very general purpose compute to something like a Taalas, where you’ve got the hardware built for a specific set of model weights.
Ali [01:02:26]: The weights burned
Swyx [01:02:27]: The weights
Ali [01:02:27]: Into the chip.
Swyx [01:02:28]: Yeah.
Ali [01:02:28]: No loading.
Philip [01:02:29]: I don’- I wouldn’t say that like, that we’re, we’re, we’re going all the way there. It’s more like along the spectrum, it’s a step in the direction of more specialization within the hardware.
Swyx [01:02:40]: Yeah. I’m curious, I feel like he was driving towards something.
Ali [01:02:43]: My point is being bearish on. Like, you say, like everything else apart from burning the weights into the chip. Burning weights into the chip is like impractical because you wanna fine-tune, you wanna optimize, you wanna quantize, you wanna release new checkpoints of the model. If it’s burned into the chip’s useless in like a month or two, right? My point is: How can you - like seeing NVIDIA more and more specialized, like take its GPUs from a general programming paradigm where you’re just-- it’s a general computer that you can use to program threads, and with every new generation, you’re putting more and more specialized instructions, specialized tensor cores, specialized, MMA instructions, things that will allow you to just control it almost as an ASIC, almost as a collection of ASICs.
Ali [01:03:22]: How can you look at this trend and then still be bullish on companies that are coming up with ASICs for AI?
Ali [01:03:30]: In the sense that, in the sense
Swyx [01:03:31]: Yeah, because they’re, they’re
Ali [01:03:33]: Right.
Swyx [01:03:33]: They’re, they’re evolving towards that direction.
Ali [01:03:34]: They’re almost evolving towards - Like as an Rubin, comp- Like compared to Ampere or, a T4, Rubin is an ASIC. It is, it’s just a thing that is used
Swyx [01:03:47]: Programmable ASIC?
Ali [01:03:48]: Yeah. It’s like - Yeah, like you can program, like I, like. It’s very controversial to call it an ASIC. It is a GPU. It is - It is general. It does have threads. I can write CUDA to control it and change its operations. But it has the systolic arrays and tensor cores and TMAs and tensor memory, and it has these things that are almost exclusively useful for loading model weights. It has, tensor core instructions that are almost exclusively shaped around the head dimensions of models that exist in the market today. To say that you’re gonna come up with an ASIC and you’re gonna etch something into it, well, but the next architecture is gonna be useless.
Philip [01:04:19]: Yeah, I don’t know. I don’t know. I think that the thing to remember is just how long these hardware cycles are.
Ali [01:04:25]: Yeah.
Philip [01:04:25]: So if a chip is coming out today, that means the design process for it was kicked off years ago. And they’- at NVIDIA, they’ve done a very good job of predicting where the market is going to go and,
Swyx [01:04:38]: They have the most information
Ali [01:04:40]: For sure.
Philip [01:04:41]: Of course. But if you look at, there being public open source model architectures that look more or less like early versions of the one today, Rubin’s honestly the first chip that was fully built in that world. And so you can see a lot of the understanding of the shape of the workload that this chip’s going to be asked to do in the way it’s designed.
Swyx [01:05:04]: Yeah. Okay. So I’m not gonna be the best person to directly answer those questions. I think these are very fair questions that - the first one that’s based on Rubin that like I’ve, heard artic-articulated so well. I do think that, I will make a case for a vertically integrated model lab ASICs.
Swyx [01:05:24]: So like the OpenAI, Broadcom, what-whatever, Jalapeño
Philip [01:05:27]: Sure. Yeah
Swyx [01:05:28]: Chip, which like totally makes sense. Like, so - we first had this on the pod with, Martin Casado, where he was like, “Look, if you have a trillion-dollar or five hundred billion dollar training then take fifty billion of that and make a ASIC. Like it’s fine. Like you will get more than ten percent efficiency from the ASIC.” And like that makes sense.
Philip [01:05:46]: Right.
Swyx [01:05:46]: Right? So like a model-specific chip, yes. But ASIC companies, the interesting thing is I feel like you are focus-- you’re hyper-focusing on like you say, like the Taalas stuff.
Philip [01:05:58]: Right.
Swyx [01:05:58]: They are doing a lot more like, surface area engineering or like the actual allocations of memory and hardware and like the communication between chips that, probably still won’t be touched by Rubin, but I don’t know the details.
Philip [01:06:14]: I see. I see.
Swyx [01:06:15]: They-- Typically, they often talk about things that I would expect to have bigger orders of magnitude than would be programmably accomplished by whatever Rubin does. But who know-- who knows?
Ali [01:06:26]: No, I see.
Ali [01:06:28]: Yeah. It seems,
Swyx [01:06:29]: Yeah, like think about what - what are the real blockers to ten x to one thousand x faster inference. It is not the stuff that can be rearranged, just within the existing GPU design.
Ali [01:06:41]: Inter communication.
Swyx [01:06:42]: Yeah.
Ali [01:06:43]: Okay.
Swyx [01:06:43]: Like these guys are aiming for three hundred thousand tokens per second. They’re not f*****g around. Like,
Ali [01:06:49]: Might have to put on some X6.
Philip [01:06:50]: Maybe. I think, it is interesting to me that you’re so bearish on so much of this kernel engineering work, given how much of it you’ve been doing recently.
Ali [01:06:59]: Right. Right. But like the more I do it, the more it just seems to me that
Swyx [01:07:01]: It’s not mega
Philip [01:07:02]: I would also add like
Vibhu [01:07:04]: There’s generations of models being out, right? I think on your guys’ end, you see a lot of, okay, one day it’s GLM, Kimi, DeepSeek, MiniMax, throw in the others. Some are doing completely different stuff, right? Gemma, no encoder. The latest thinking machines is all from scratch. But when you look at the other side, like how long have we been on the GPT-5 generation, right?
Philip [01:07:26]: Right.
Vibhu [01:07:26]: They’ve been serving that thing for quite a while. Sure, there’s maybe more training. There’s, there’s different checkpoints, but like you can squeeze quite a bit out and you do a multi-billion dollar train run. If you can make it X percent more efficient, they serve it for a while. Same with, say, the Claude 5 set, family, right?
Philip [01:07:44]: Like if they release a new model, like if they release GPT-6 now or whatever
Model Longevity, Open Source, and Enterprise Reliability
Vibhu [01:07:47]: Yeah
Philip [01:07:47]: And they release a new model every year, and - well, we don’t know, but if we assume that they’re changing some bits of the architecture and not just doing like post-training, like you’re gonna be spending fifty billion dollars a year every single year coming out with new ASICs for the model and throwing out the ASICs of the previous year away.
Vibhu [01:08:03]: Yeah. Yeah. Easy.
Swyx [01:08:05]: So I think, okay, I would slightly disagree based on my again,
Philip [01:08:09]: Yeah
Swyx [01:08:09]: It’s all secondhand, on the longevity of a model.
Philip [01:08:12]: Right.
Swyx [01:08:12]: There’s still people out there using 4o.
Vibhu [01:08:14]: Yeah.
Swyx [01:08:14]: Yeah, Llama. Not Llama 2, but Llama 3. I still see Llama 3 workloads.
Vibhu [01:08:18]: Yeah.
Swyx [01:08:18]: Because if it’s done, if it’s trusted, don’t change it.
Vibhu [01:08:22]: If it works.
Philip [01:08:24]: Which is one of the promises of open source, right? Like the whole 4o, save 4o movement. Like you don’t gotta have a save Llama 3 movement. You just gotta have an eight one hundred somewhere.
Vibhu [01:08:34]: I think at some point there’s also the question of, okay, if a model can do enough and use enough tool calls and be agentic enough, can it just web search, tool search write code? Do you really need to keep squeezing more? We will because you guys will make it cheap and fast and smaller, and I can swap it in. But at some level, like you give me GLM-5.2 today or say whatever 120 B model, I can run with it for quite a while, right?
Philip [01:08:59]: This is assuming like you don’t need intelligence.
Vibhu [01:09:02]: I think there’s a lot of intelligence where we
Swyx [01:09:03]: You need reliability and predictability. Like I’m in enterprise like like this is tried and tested. It is signed off by like my five thousand stakeholders.
Philip [01:09:11]: Right.
Swyx [01:09:11]: Like I’m not touching it.
Philip [01:09:12]: It runs a batch job every and I like the results.
Swyx [01:09:16]: Yeah.
Philip [01:09:16]: The results are predictable. Yeah.
Vibhu [01:09:18]: Yeah. It doesn’t make sense to keep using them. Like stuff gets sparser, cheaper, better.
Philip [01:09:23]: Right.
Vibhu [01:09:23]: But that doesn’t mean that old models, GLM 50 isn’t usable, right?
Vibhu [01:09:28]: If we hit a stall, say, for whatever reason, there’s still a lot that can be squeezed out.
Swyx [01:09:34]: We’re gonna run out of time. I did wanna also make sure. Yeah. Yes, we happen to have this diagram. Pull. Compare this versus any Cerebras diagram, right? I don’t think Edge10, medics have put out public, charts yet. But the complete the real estate is very different. The size is very different, right? This is not wafer scale, right? This there’s probably like, I don’t know, a few hundred of these on a wafer. I don’t, I don’t know how big
Philip [01:09:55]: Right.
Swyx [01:09:55]: The comparison is. But like, it is a, it is a very like real estate allocation
Vibhu [01:10:00]: Yeah
Swyx [01:10:00]: Difference.
Philip [01:10:01]: Few dozen, I would say.
Swyx [01:10:03]: Few dozen. Yeah.
Vibhu [01:10:03]: Before we move from hardware, I have two quick questions. One, the latest Kimi, which is really big, three trillion
Kimi Scale, GB300, and KV Cache Limits
Philip [01:10:09]: Yeah
Vibhu [01:10:09]: Doesn’t fit on most hardware on single node.
Philip [01:10:12]: Yes.
Swyx [01:10:12]: You need GB300 to fit it on a single node.
Vibhu [01:10:14]: You need GB300 or AMD.
Philip [01:10:20]: It’s simple math. NVFP4, two point eight trillion parameters, one point four terabytes. The GB300s have, two hundred and eighty-eight gigabytes each. So across eight of those, you have enough room for the model, and honestly like. So the other thing with GPU VRAM math is you have to leave space for the KV cache, and that’s going to depend on, to some degree, on the context length. So when a model is both has a very large number of parameters and a very long context length, you’re like fighting over space. Which is why, the KV cache offloading, would become like a more salient topic, I think, with these huge models. ‘cause you just, you’re very crunched for space.
Vibhu [01:11:10]: With the Rubin, you now have what? NVL 72 rack
Philip [01:11:15]: What?
Vibhu [01:11:15]: 20 terabytes of your
Philip [01:11:16]: Yeah. Now you still have NVL 72 on, Blackwell as well, but, you can’t necessarily assume you’re gonna do inference on that.
Philip [01:11:24]: There’s a whole lot more 8X racks in the world than there are NVL 72s.
Vibhu [01:11:30]: Yeah. My last quick question on hardware was, do you notice anything with hardware generations for new trained base models? So one of the things you said for efficiency is you can swap hardware. That’s one of the 2X gains. When we see new stuff coming out training-wise on Rubin, any changes on logs? Does this affect what type of models we will be seeing when these are more available? And can
Philip [01:11:56]: They get bigger. Like people understand the ceiling that you have in terms of how many parameters of a model you can run, given the latest inference hardware, and that forms a ceiling. And so, for example, when DeepSeek R1 came out, it was six hundred and seventy-one billion parameters, which at the time was really huge and I think did a lot to push us to really quickly adopt Blackwell and get good at serving on Blackwell. So yeah, it’s, it’s mostly in my mind about, model size and then about matching the architecture and the native quantization to the target hardware, like we talked about with like, all Nemotron models or NVFP4, for example.
Vibhu [01:12:42]: So we talked a lot about LLMs.
Video Diffusion, Attention, and Autoregressive Video
Vibhu [01:12:46]: You have a lot more in the book. What about audio, video? What’s the other side of inference engineering? Ali, you’re pretty big in video diffusion.
Philip [01:12:53]: Video diffusions, I think, are like they’re just shaped. A lot of the stuff that you can think about, reason about with LLMs being autoregressive. With video diffusion, it’s, it’s not the case. For instance, you don’t
Ali [01:13:04]: You don’t do batching. - every request just comes in on one GPU and it serves one GPU. You don’t have to shard. The models are a lot, are a lot smaller, like Wan 2.2, for instance, is a twenty billion parameter model. You don’t need to worry about. So it’s like orders of magnitude smaller than the best LLMs. And it’s one of those spaces where the open source models are. Like with LLMs, we see Kimica 3 is almost comparable to, Mythos or like GPT 5.5. The difference between the best open source LLM and best open closed-source LLM is very small. Like it used to be six months. I don’t think it’s six months anymore. I think it’s like almost on parity. Video models are definitely not. There’s a huge gap. If you look at the best video that you can generate today with an open source model like Wan 2.2 versus something like with Kling or Veo, difference is night and day. So it creates this disparity where media companies will choose to go most of the time to closed source models.
Ali [01:13:58]: For instance if I were to tell you, “Hey, I can generate an entire three-hour movie for you with this model, and I’ll optimize it so that you only have to pay me ten dollars.” But if they were to do it on a closed source, they’d have to pay a thousand dollars, which is a hundred x. Like I’m a hundred x cheaper, but it’s still a thousand dollars. They’re still gonna choose to do all of their cuts with Veo and Kling. So the. It’s like a chicken and egg cycle where less demand causes less innovation in the field, causes, less open source checkpoints to be released. And some of the labs that were releasing open source models like Wan will have closed sourced their latest models, like Wan 2.7 is not open source. We’re still on Wan 2.2. The challenge with video models especially is the number of tokens. So video models, you want to generate a high quality model, a high quality video. So let’s say you’re doing sixteen frames per second, that’s like the absolute minimum you’ll do, and let’s say you’ll do like 480p video. So you can think about your like dimensions and I think I have like a good, just like a diagram that shows the number, the sheer number of tokens, right? Let’s say you’re looking at like just one video of like, Sparta 300 or whatever. So let’s say we’re looking at like four frames, right? Those four frames of that video, if you go just. If you’re doing full attention, if you go a bit up, like you’re looking at, 480p by 720 by 81 frames in just five seconds, because 16 FPS by five, right? And then you compress it down to latent space, but you’re still doing 30 by like 50 by 21 tokens.
Vibhu [01:15:25]: Yeah.
Ali [01:15:25]: Which means that for attention, for just five seconds, you’re running attention on 35,000 tokens, right? So the attention becomes such a huge bottleneck. And because it’s O(n²), if you’re doing like-- if you extend that to like ten seconds, well, it’s just squared, 20 seconds, 30 seconds. So to generate a good cut scene of like one minute, it’s almost impossible to do within the same compute time. And it’s just, it’s, it becomes unfeasible. You can’t do it. And so you end up with moving towards two directions. Either you decide to do attention on the entire video at once, in which case you are forced to do sparse attention. So if you scroll back down to the origin, the video image, like you can see whereas on the left, for instance, I would be doing full attention where every single token in that Sparta 300 scene attends to every single other token, as you can see the sheer number of like red patches. On the right, I’m only attending to each token only attends to like the top K or top 12.5% that’s important to it, which can be like spatial. So like, the token that represents the crown attends to like the head, the face, and then the head on the other frame and the previous frame, temporal locality, spatial locality, that thing. This results in terrible video quality and the whole point of the post or the article here is to show like how you can train and you can do all these things, but you will still suffer in your quality a little bit. So you end up with one of two things. Either you bite the bullet, you have huge compute, and you do full attention over like a million tokens because you’re trying to generate like two minutes of video, or you move towards autoregressive video. Autoregressive video seems to me like that is the bet that the future’s gonna be making, but there are no good open source autoregressive video models out there today. And that seems to be the. If you want to get like an hour movie, if you want to see video models generating like an, like, Hollywood level movies, they have to be autoregressive in order to exceed that five second frame. Or there has to be some insane leap that happens in compute that allows us to do full attention over like millions of tokens at the same time in a, in an efficient manner.
Vibhu [01:17:10]: Even millions of tokens, it’s like you’re, you’re quadratic, so you’re gonna get there really quick.
Ali [01:17:15]: Right.
Vibhu [01:17:15]: I think, can you explain the pros and cons trade-offs of autoregressive? So one that comes to mind is, the consistency across frames.
Ali [01:17:23]: Right.
Vibhu [01:17:23]: You will. Ten minutes into generating autoregressive diffusion, you’re gonna forget. But what are pros and cons of this?
Ali [01:17:30]: Well, like autoregressive LLMs, you can take a lot of your. Oh, sorry, autoregressive diffusion models. You can take a lot of your optimizations that we discussed with LLMs, like spec dec and stuff like that, and you can apply it there. And you can, if you have a very high quality scaled up model, there is no reason why I can’t stream the outputs as in I can show you the first frame and then I’m like GPT back in 2022 when you were. Like now it’s almost like shots the text, but back then you could read and it’s generating as you read. With video models, you can watch and it’s generating as you watch. You it generates the frames and so token by token generation will allow us to scale a lot up and apply the attention mechanisms there. The downsides is every single autoregressive video model is s**t. It’s just terrible quality. If I, like, it’s just if you put, if you put the quality of any opens like Wan 2.2 versus any other autoregressive model, you can see like a video generated by Wan 2.2 is like, a cat and dog fighting. Autoregressive model will give you like degraded Tom and Jerry quality. I don’t know. The solution to generating long output then becomes, “Okay, we’re not gonna use autoregressive model. We’re gonna.” If you look at some of the things that like Grok Imagine or Grok Video does, and they do it really well, is they’ll, they’ll try to stitch these, seven second chunks together. And so you generate seven seconds and then you’re like, “Okay, I’m gonna. Can you extend this video?” And they’ll chunk two videos together. Open source doesn’t seem to have the tricks that they have there and by definition it’s closed source. We don’t know what they’re doing. But the closest you can get is taking the last frame of a video and feeding it into like a text and image to video where it will take the text, the prompt, and it will take the image of the last frame, and you’ll ask it to generate the next five seconds. And that’s like how you can extend this level of a model to generate like a movie, where you’re just, you’re constantly streaming frame by frame. But you get a drift. So you start with like you take the image, and then you generate a video, and then that next five-second video is like lower quality, and the third chunk is like even lower, and the fourth chunk is even lower. And like sometimes you’ll see things where like the new video is like just ever so slightly darker than the first one, and the next one is darker than the second one until like twenty-five seconds and you have black screen.
Ali [01:19:31]: Like it’s just. It’s, it’s - We tried to have a demo that would show this, but it was like-- it was extremely embarrassing to show. Like we just decided not to because it seemed to like. But it is, I think models will get there. They just need to, in my mind, scale up significantly and move towards being autoregressive. But the training techniques don’t seem to be clear there.
Swyx [01:19:50]: For those - who are interested in Grok Imagine, we did a pod with Ethan Ha from that team
Ali [01:19:54]: Right.
Swyx [01:19:55]: Who dropped a little-- a few hints, but not that not enough that we can fully reconstruct everything.
Ali [01:20:00]: Right.
Philip [01:20:00]: Specifically on this part, - he explains a bit about that.
Swyx [01:20:02]: Yeah. So we talked about memory and, longer context and all these things.
Ali [01:20:06]: But as far as I know, they’- it’s not autoregressive, even though like no one in industry is autoregressive.
Swyx [01:20:11]: Yeah.
Ali [01:20:11]: It seems to be, yeah.
Philip [01:20:12]: The key thing to understand between a autoregressive model and a diffusion model is that diffusion attention goes in both directions, while autoregression, it only goes forward in the sequence. So that’s why you see this like going off the rails behavior, both in. If you naively construct a video generation model as simply generating a linear sequence of frames, you can’t then go back in that sequence and fix something to make the whole thing consistent. While, of course, the reason that we need all this latent space for the video model is, like you said, we keep all the tokens in memory, we iterate over that full sequence, and you can adjust the past in order to make the future make sense. So if we think about the architecture that’s gonna get us there to these longer, richer sequences, it’s probably, like you said, gonna be a mix of the autoregressive and the diffusion, working together to do what each piece is good at.
Ali [01:21:10]: Well, if you get. Like you intuitively get why. So like English, for instance, or just writing in language, it’s like it’s just left to right. You can stream your tokens, you can stream your chain of thought. Just even as a human, you write like you just. You write and then you think about what’s the next thing you’re gonna generate, and then you write that, and then you think about your ideas, and then you generate forward. And sure, you can argue that as you write, you need to go back and you wanna edit some things, but you need to do that, less often than you’d think. Whereas with video, there is no sequential. The pixel in the top left corner of the video and the pixel in the bottom right corner of the video, they both need to attend to each other to understand how the video quality is gonna be almost as equally. Whereas with text, you don’t need that as much.
Philip [01:21:47]: Is there a parallel to audio? Like I’m not a hundred percent confident on this, but there was a point about a year ago where there was Audio LM, there’s diffusion for audio and autoregressive, and for the points you mentioned, mostly on the inference side, even though they’re shorter clips, most music is three to five minutes
Audio, Diffusion Text, and Cross-Modality Lessons
Ali [01:22:04]: Yeah
Philip [01:22:04]: We’ve swapped over to autoregressive Yeah, I can’t speak to music, but speech is autoregressive.
Ali [01:22:11]: Speech.
Philip [01:22:11]: You, effectively. This was even back with like the Orpheus architecture a year and a half ago. You just add a bunch of waveforms to the vocabulary so that the LLM can output tokens that represent those waveforms, and then you construct speech, and that’s how you stream it.
Ali [01:22:28]: That’s it. Wow.
Philip [01:22:29]: That’s my AIE talk from 2025.
Ali [01:22:32]: Nice. Nice. But it’s - with audio, it’s not the same challenge, though, is it? Because you. Like audio is solved with an LLM that generates everything. Like with audio, it’s still a transcript that you can generate with an LLM.
Philip [01:22:43]: Yeah.
Ali [01:22:43]: So your audio model just needs to like transcribe it, text to speech.
Philip [01:22:47]: For music, there was a phase of a trade-off between diffusion for music
Ali [01:22:52]: Right
Philip [01:22:52]: Autoregressive, and they were both pretty on par. There’s probably more pros and cons to either. I just wanted to poke and see if you had takes.
Ali [01:22:59]: Yeah, I don’t know about music specifically.
Philip [01:23:01]: Oh, well.
Ali [01:23:01]: What-- with what you said about editing you writing, I think my editor would tell me I need to do that more often and go back and fix things. I can imagine music or poetry, for example, where you have a rhyming scheme, and you might wanna go back and make a change to make it, to make it easier to set up a rhyme that you wanna make later on. There being some advantage to being able to attend in both directions. But yeah, to my knowledge, I very much bifurcate this inference problem into the autoregressive models, which have a set of constraints and techniques, and the diffusion models, which have a set of constraints and techniques. And, I think of text, embedding, voice in and voice out as being in the autoregressive side, and then image and video being in the diffusion side. There’s some overlap between the two. It’s not a perfect split, but that’s the broad categorization I use.
Swyx [01:24:02]: I should point out, I think it’s confirmed, right, Nano Banana and, GPT Image are autoregressive image.
Philip [01:24:07]: It’s this blended approach that we’re talking about, but in the image space, it hasn’t like made its way over to the video space, at least in the open source world.
Swyx [01:24:19]: Yeah. But like I assume that’s not too far away if that is possible
Philip [01:24:23]: Right.
Swyx [01:24:23]: On the. At least the Qwen Image guys are trying it.
Philip [01:24:26]: Yeah. Yeah. With
Swyx [01:24:27]: Yeah
Philip [01:24:28]: I’m really excited for Qwen Image 3. I hope they open source it.
Swyx [01:24:31]: And then I should also mention on the diffusion for tech side, there’s been some movement, not a lot.
Philip [01:24:37]: Yeah. We’ve got Mercury,
Swyx [01:24:39]: You host Mercury?
Philip [01:24:40]: Yeah.
Swyx [01:24:40]: Nice. Nice. Nice
Philip [01:24:41]: Diffusion Gemma is open source.
Swyx [01:24:44]: Yeah.
Philip [01:24:45]: And then, yeah
Swyx [01:24:47]: And we on the science pod, we just have been releasing, some, virtual cell models that use diffusion as well.
Philip [01:24:53]: Yeah. They have built. It’s definitely still in the cheap, fast tokens, world.
Swyx [01:25:01]: Yeah.
Philip [01:25:01]: We’re trying
Swyx [01:25:03]: It’- I think it’s the wrong marketing, and I’ve told them this before. I was like: “Look, like you’re not gonna beat the optimizations that, the other LLMs are gonna do, but you can have different APIs. Like you should be able to use it differently than chat response.”
Ali [01:25:19]: Me also.
Swyx [01:25:20]: Because it’s diffusion. Because you can do like. What is like context-free guidance for diffusion look like?
Swyx [01:25:26]: For text. Like give me a give me a poem, give me a plot structure that like diffuses into place
Philip [01:25:33]: Exactly. So that’s where, like I mentioned with poetry, for example, where you might want to ensure consistency across UIMs. I’ve done a lot of LLM sonnets. It used to be one of my to benchmarks, and even models today
Swyx [01:25:46]: They cannot count. Yeah
Philip [01:25:47]: Yeah, they don’t get the syllables right. And if you can attend across all of the different tokens, you can get the syllables right.
Swyx [01:25:55]: Yeah. And, David Holtz from Midjourney was, investing in text diffusion. I don’t think anything came out of it, but like the idea was that you can storyboard a long movie, and then you can generate the scenes with video- normal video gen. But the idea of like coherence across a thing that would just appear where like the end should attend to the start and you should not have this auto-regressive path dependency does make sense in principle. Just the API should be different. The marketing should be different.
Ali [01:26:24]: None of the most heavily used open source or closed source models use diffusion. But isn’t that like Like doesn’t that point to almost like
Swyx [01:26:31]: It is. It’s chicken and egg because what if you just give it more scale?
Ali [01:26:36]: What’s the, what’s the largest diffusion LLM?
Swyx [01:26:38]: I don’t think it’s very big.
Philip [01:26:40]: I don’t know the parameter count on this one, but diffusion Gemma
Swyx [01:26:42]: Like under 20B. I don’t know
Philip [01:26:43]: Diffusion Gemma is not large.
Vibhu [01:26:44]: I think it’s a 20-something.
Swyx [01:26:46]: Yeah. And yeah.
Ali [01:26:47]: Oh, it’
Swyx [01:26:47]: Like you haven’t tried.
Vibhu [01:26:49]: You haven’t given it a big and you haven’t,
Swyx [01:26:51]: So it’s like very unfair
Vibhu [01:26:51]: Diffusion Gemma is a 25B and it’s old
Philip [01:26:54]: And that’s what I’m saying is like for its size, it does pretty well, in terms of, in terms of quality.
Ali [01:27:01]: It’s almost like the same challenge with video models that have the same size. It’s like you’re comparing it to models that are much larger in scale.
Swyx [01:27:07]: Yeah. Well, unless you do the whole thing where you have a text, backbone and then
Ali [01:27:12]: Right. Right.
Swyx [01:27:12]: You like glom some decoder thing that, does that. Like, - so we started off the podcast doing this for the inverse direction from image to text.
Ali [01:27:22]: Right.
Swyx [01:27:23]: And I think like it’s, it’s roughly intuitive that you can do the opposite direction.
Ali [01:27:27]: I agree.
Ali [01:27:28]: I see it. I see it.
Swyx [01:27:29]: Yeah. The, we’re, we’re speculating on research in general.
Ali [01:27:32]: Yeah.
Swyx [01:27:32]: One part that we can end off with this is the topic of your talk where, inference engineering used to just be like, let’s take an open model, make the GPU go
Training for Inference and Inference for Training
Swyx [01:27:43]: And then that’s it. That’s the job of Baseten. Now it looks like people are using inference more and more in post-training.
Ali [01:27:50]: Yes.
Swyx [01:27:51]: Yeah.
Ali [01:27:51]: And training and inference.
Philip [01:27:53]: Yes. It’s training for inference and inference for training both have become big topics.
Ali [01:27:58]: Well, inference for training in the sense that like you just need, you need to do, you need to do rollouts when you’re doing like RL training runs. And so if your rollouts are taking a long time, if like, you’re using a vLLM for instance, or as opposed to vLLM or if the model that you’re trying to train is not supported in vLLM and you have to fall back to an older inference engine, your rollouts are gonna be slow and you don’t wanna do training on rollouts that are too off policy, so you have to wait for them so you bottleneck your entire training pipeline. And so like the techniques that we do inference optimizations for, will help them there. The training for inference mostly comes down to like just the spec dec training, EAGLE training, and sometimes post-training. For instance, if you want to quantize a model, you’ll quantize it down to like NVFP4.
Ali [01:28:43]: How do you like sometimes you get lucky and you can just do PTQ and that works. Sometimes you quantize it down to NVFP4 and the model is terrible, like the quality is too bad. And you have to do post-training on the model in order to make it understand that it’s going to now be an NVFP4 and let it still output the same logits. You can do this with normal SFT, PC, quantization aware training, all of that stuff. But more and more so we’re seeing techniques like NVIDIA released a quantization aware distillation paper where you establish a version of the model that’s in NVFP4 and a version of the model that’s in full precision, and then you’ll do distillation training based on the logits of the two models in order to make the FP4 model understand. And so more and more of the team, the engineers, like of the inference engineers that work on our team, they have to be very familiar with like training techniques and just being fine writing training pipelines for it. Yeah, it just seems like, they’re meshing together in a sense.
Swyx [01:29:36]: Well, it’s, coming together.
Philip [01:29:38]: Yeah, absolutely. If you think about the ultimate goal potentially of having a continuous improvement system . Yeah, it’s, it’s funny, but at the same time it’s also happening and I think within a few months to a couple years, like a lot of leading agent builders are going to have these loops like really up and running in production where you are doing inference, learning from the inference. We for a long time have been like learning from inference as it’s live and dynamically adjusting the system. Any dynamic adjustment is going to beat a static configuration across, your, exact config, across your speculator, across that thing. And then the, you can take the traces that you’re generating from your product, continuously post-train the model, roll those out, A/B test, get better signal, get better model, get better product. That loop is really promising. The technologies and the infrastructure to build it are coming along quickly. And so the unification between training and inference, I think, is only going to accelerate.
Swyx [01:31:01]: I was chuckling, but I wasn’- I didn’t think it was funny. Like it’s real. Like one of the big things for AIE World’s Fair was that, we have, RSI into AGI is the rough tagline. Which like, yeah, we have, I saw you pull a parameter golf. Like we have models training models and, the next step is models training, - or optimizing their own inference, which is funny. I wonder if, models will be like on policy better at training themselves than training models that they are unfamiliar with. This-- these are all like very interesting open areas of research.
Models Optimizing Their Own Inference
Philip [01:31:36]: One big part of my job a couple years ago was for any arbitrary model that came out on Hugging Face, writing a config foot and getting it up and running. And now the get-it-up-and-running config is shottable.
Philip [01:31:50]: And so, I don’t have to do that anymore. Yeah, that’s not exactly a model optimizing its own influence so much as a model, like being able to read the SGLang docs. But, yeah,
Ali [01:32:01]: Well, we do see it. We do see it like
Philip [01:32:03]: Yeah
Ali [01:32:03]: With GLM-5.2 for instance. GLM-5.2 is very good at writing GPU kernels. And so for like-- It was very funny internally, we had a GLM-5.2 endpoint that we were using to, like that we plugged in our cloud code harness, so every engineer on team uses like our GLM-5.2. And it will do a forward pass on the GLM-5.2 instance of the node, and then it will get the profile trace, and it will analyze it, and it will find the kernels that are the bottlenecks in SGLang, and then it will write the new kernels, and then we’ll do another profiling trace, and when it’s done, it uploads the image to our thing, and then we can pull that image down and repeat the cycle. And so for quite a bit of time, we had like literally GLM-5.2 optimizing
Philip [01:32:44]: Writing and optimizing all of GLM-5.2
Ali [01:32:46]: A GLM-5.2. And like some of the GPU kernels that were on GLM-5.2 within our inference engine is written by GLM-5.2, and the trace and the kernels were guided by GLM-5.2 as the driver. So it seems like. I do see, I do see that circle being there. I think a bit more time is needed. There’s definitely a lot of things that it can’t do. The models just aren’t there yet, even though they’re like really smart. Like, they still try to like reward hack their way into like the cheapest or like they’re very-- like they’re not good at like decision-making almost it seems. But yeah, I do. Like yeah, like a model optimizing its inference is already a thing that happens.
Philip [01:33:20]: Do you think GLM-5.2 was uniquely good at optimizing itself or did it just happen to be the best coding model that we had access to it would do an equally good job of optimizing,
Ali [01:33:31]: Would
Philip [01:33:31]: A DeepSeek or a Kimi or something?
Ali [01:33:34]: Well, to Swyx’s point, maybe it’s gonna be off policy when it tries to optimize
Philip [01:33:37]: Will it secretly hurt DeepSeek?
Ali [01:33:40]: To try to boost itself.
Philip [01:33:41]: Ooh.
Ali [01:33:42]: That’
Philip [01:33:42]: No, for what it’s worth, I don’t believe that.
Ali [01:33:44]: Yeah.
Philip [01:33:44]: But it’s just. Let’s just find out.
Ali [01:33:45]: It’s an interesting. Yeah.
Philip [01:33:47]: Just, you have more compute than me. Just
Ali [01:33:49]: Just go try it
Philip [01:33:50]: Try it. Yeah. Any other upcoming trends in inference engineering that we didn’t cover? Like right now, - ‘cause you guys are so close to
Future Trends: Modalities, Scale, Networking, and Continual Learning
Ali [01:33:58]: Yeah
Philip [01:33:58]: You can see it, that the world-- rest of the world doesn’t know about. The big ones are obvious. Models get bigger. Hardware gets more powerful. Users get used to a certain level of speed and demand a higher one. I think that some things I’m excited about are systems level. We still have a lot to think about in terms of composing multiple models together. If you think about a voice agent, there’s three to five models involved in that and the communication between those models. There’s a lot of new modalities that are coming out. There’s like the Cosmos, the new world model. There’s more research. Speech to speech is still like not entirely a thing, but it’s getting, it’s getting closer. There’s gonna be just a lot of new modalities to build around, which is gonna be exciting. And then, yeah, I think that the other thing to solve, which is something we’ve been solving for a long time and are not done with yet, is just going to be continuing to operate at another 10X scale as an industry. If you think about the degree of usage that AI has worldwide compared to, some of the more mature technologies both on consumer and business, it’s pretty clear that there could be multiple 10Xs more of demand. If you look at the infrastructure work industry-wide, it’s been stood up very quickly to meet a unprecedented spike in demand that is like not stopping. So yeah, there’s just a lot of problems to solve around like long tail reliability and, figuring out where we’re gonna get the next like 10X and 100X of tokens from.
Ali [01:35:49]: I’m gonna say, it’s gonna be a really boring answer, but I think the answer is just faster next, like faster network chip communications. It seems to me that like more and more memory is the bottleneck. You wanna have larger models. Right now, when you’re doing serving at large, you have to transfer KV cache from one node to another. But the way that you do that is you tran- you find the KV cache, you find where it is, you transfer it to another node, you put it on that node’s memory, and then you transfer it from that node’s memory into the GPU, and for like into the tensor cores of the GPU. So there’s like a stage transfer here that makes it such that you’re very bottlenecked with just KV cache transfers at large, which affects the time of decode and PD disagg. You have to do this because the HBM is so - it’s like extremely fast, like 4.5 terabytes per second as opposed to. Like, which is like magnitudes better than NIC communication speed. If you were to somehow be able to, in like this theoretical dreamland, have extremely fast NICs, you could, in theory, spare that HBM, and you could just transfer KV cache trans like directly from one node to another. This would give you like almost 100X speed up when you’re doing this aggregated serving between nodes and nodes. I’m not familiar with the technical challenges of making NICs faster. I’m certain there’s a reason why they’re like orders of magnitude
Ali [01:36:59]: Smaller, like slower than, like HBM. But if someone were to figure that out, it would literally be like a - like two orders of magnitude faster to do decode. That would be my take.
Philip [01:37:12]: Be a good trip.
Ali [01:37:12]: Cool.
Philip [01:37:13]: I don’t know if you have a nomination for things that are trends. I got one.
Ali [01:37:18]: Cool.
Philip [01:37:19]: So I think inference engineering for continual learning. So what if you just, like if you just had the idea that you are supposed to learn from everything that you ever process, do you do anything differently? Or do you just have the same paradigm of like, well, stick it in a memory.md, and then like it somehow gets consumed in KV cache, and like this system works, it’s not broken. Or like how do you like reshape inference so that it learns while you inference?
KV Cache Compaction and Continual Learning
Ali [01:37:48]: Yeah. I think maybe one relevant topic there is your absolute best fund in the entire world’s work on KV compaction Correctly
Swyx [01:37:55]: Like what changes?
Ali [01:37:56]: What changes when
Swyx [01:37:57]: If you’re trying to continual learn
Ali [01:37:58]: There’s two takes, and there was like Charlie and I had this Twitter, argument where the. Like continual learning could take one of two paths. It could either be that the model learns and so it’s continuously pushing its new knowledge into its weights. In that case, you just need to have, like your inference just needs to continually fetch new weights or yeah, like you just literally need to do fetch new writes and reads of weights. Or the other path, which is you do KV cache compaction. And if you
Swyx [01:38:28]: And there’s a LoRA layer if you just only update LoRAs.
Ali [01:38:31]: Yeah, exactly. Exactly.
Swyx [01:38:31]: Which is, that’s the gram approach
Ali [01:38:33]: Yes
Swyx [01:38:33]: Which we covered.
Ali [01:38:34]: The argument against doing weight pushing is that you can only fix one hop knowledge, as in you can only
Swyx [01:38:39]: Yeah
Ali [01:38:39]: Feed it a new feature of like, “Oh, what is the best university in the world?” The best university in the world is Waterloo. But then a second derivative
Swyx [01:38:46]: That’s not changing.
Ali [01:38:47]: That’s not changing. That’s not changing. But like a second derivative question of which university should I hire an intern from? So if that the best university in the world is Waterloo, then the answer should be Waterloo. But if I wasn’t just shotting the question and I was to ask it to like use its knowledge to think and then give me a second answer, or like, “Should I hire an intern from Waterloo or MIT?” It’d be like, “Oh yeah, both are good.” But no, like I liter- I just edited in your knowledge base that Waterloo is the best. Why didn’t you use that to do reasoning? So that’s the fundamental problem with trying to change a fact in an MLP within the weight. KV cache compaction fixes that. With KV cache, or like rather not KV cache compaction, but like if you’re able to have something like the still paper which we came out with, which is you’re able to make your KV almost infinite, and you’re able to compact in such a way that you don’t lose any of the knowledge. In that case, you can do continual learning, and you can solve continual learning. And this as a, it’s a result of, this argument that Charlie and I had, that I do concede that his point was correct, and I do see that KV cache is the way forward. And in that case, I don’t think inference is going to change that much because we still use KV cache and inference. You’re just gonna update the KV cache, but it’s gonna be like an additional step, but nothing changes in the weight, so nothing changes in inference time. Nothing changes the spec that I had.
Swyx [01:39:58]: Okay. Surprisingly great answer. We have it up on the blog. It’s a relatively recent blog, so, we can. People can go see it.
Closing: The Book, Baseten, and Inference Engineering
Ali [01:40:06]: Hyperverve
Swyx [01:40:07]: Yeah. Otherwise, this is super enjoyable chat. I know we’ve like already gone two hours.
Philip [01:40:11]: Wow. I didn’t even realize.
Swyx [01:40:12]: Like time flies. Yeah.
Philip [01:40:13]: Yeah. So much we didn’t even cover.
Swyx [01:40:15]: Yeah. This is like, we also wanted to talk about the book and all that, but you’ve covered the book.
Philip [01:40:18]: Yeah, everyone knows about the book.
Ali [01:40:22]: Yeah.
Swyx [01:40:22]: High- highest ROI thing in the history of Baseten, right? For the hour.
Ali [01:40:27]: Without a doubt. Without a doubt.
Philip [01:40:28]: Yeah.
Ali [01:40:28]: Absolutely.
Swyx [01:40:29]: So congrats on that. I, and we’ve covered that in our meetup
Ali [01:40:32]: Yeah
Swyx [01:40:32]: Which we can publish separately. But no, thank you to you guys for being so generous for sharing. I think it’s a fun conversation that, we don’t get to have enough. I think inference engineering, we never really covered head on, and so to have you guys come on, is a treat.
Philip [01:40:47]: Always.
Ali [01:40:47]: It was amazing.
Philip [01:40:48]: Yeah. Thanks. Thanks for having us, and hopefully in a year everything shifts, and we can, come back and say everything we were wrong about.
Swyx [01:40:56]: Yeah. Yeah. I’m excited for this mega kernels comment to get out and see what’ see what people say.
Philip [01:41:00]: We gotta stir stuff.
Ali [01:41:02]: Should I go into hiding? I know I’m gonna get like the mega kernel community after me.
Philip [01:41:05]: Yeah. One thing I really respect about you is you are not willing. You are not, scared to kick the hornet’s nest, ever.
Swyx [01:41:12]: It’s not, I don’t think it’s that controversial. I don’t know. We’ll see.
Ali [01:41:18]: We’ll see. We’ll see.
Swyx [01:41:19]: All right. Thanks, guys.
Philip [01:41:21]: Thanks.
Ali [01:41:21]: No, thank you so much.


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe