Academia is for Ambition — Alex Zhang, MIT

[gentle music] All right, we're here in the studio with Alex Zhang, uh, I guess most famously of, uh, RLMs, uh, but you have, you have a lot- a few other affiliations. Welcome to the show. Yeah, thank you for having me. Yeah. I guess GPUMode as well? Yes, GPUMode as well. Uh, you were shepherded in by Mark Serafin. Not, not everyone gets that kind of welcome. Yep. Yeah.

Uh- Yeah, I'm very, very close to all the people in GPUMode, so yeah. Yeah. We often end up working together in various capacities, like even beyond just GPUMode itself, so. Yeah. Can we explain, uh, so people who are not that close- Mm-hmm... don't know about this. It's just a, it's a Discord. It used to be focused on, I, I guess CUDA mode, and then, then, then generalized a little bit.

Uh, it was started by Mark. Yep. It was basically like... To me, it's like the hiring pipeline with the PyTorch team. Mm-hmm. And then he left PyTorch. [laughs] Yep, yep, yep. Yeah. So it used to be, I think, uh, it started actually around when I was in college, um, in like twenty twenty-three. Uh, I think it was started by Mark Andreas and Jeremy Howard.

The original premise was just like, it was a GPU or it was a Discord dedicated to learning how to write GPU kernels, and they had like lectures. That was basically the extent of it. Um, and I got interested in it because I was writing GPU kernels. It was actually out of like pure chance. I was interning at Snapchat at the time, and- Rexxars. Yeah, I was very bored [chuckles] with Rexxars.

So, um, they had a project where like they were interested in writing... It was this paper called Infinite Attention. It was like a Google paper. Yes, we've covered it on Paper Club. Yes, yeah. So I was interested in, in whether or not you could write specialized kernels for it at Snapchat. Um, it didn't... Nothing really came of it, but I joined GPUMode.

At the time, it was called CUDAMode, uh, I think for like legal reasons or something, they changed the name. But I met Mark. Uh, I met Matei. I met a bunch of other people that were very involved in the community. And then Mark had pitched this idea called Popcorn, which was now what you see as the leaderboard today.

But the general idea was like, I think all of us had this like intuition that GPU programming is like very similar to if you guys have done like competitive programming. It's a, it's a s- not... I, I don't mean to say like they're transferable skills. You have constraints. You code golf a little bit. Yep. Yeah. Yeah, yeah, yeah.

And there's like, there's actually a surprisingly small space of optimizations that people do. Uh, and there's actually not that many kernels per se that people are interested in optimizing. And so we kind of had this thought that like if you had enough data, like in the same way that code forces, there's millions of problems.

If you could do this with GPU code, like you could scale and automate kind of GPU kernel development, which for researchers is a huge deal because I think one of the bigger bottlenecks... Like if you look at like Mamba, for example, like they released the paper with kernels because otherwise, like you can't really use it in any meaningful way, you know? And not everyone has like a Tri Dao on their team.

So we're very interested in this. Uh, KernelBench kind of spawned from that too, of like can we get LLMs to automate, um, GPU kernel code? And I think that was like a... It was a very, very fun time. It was like between college and my PhD and yeah, I, I had a really pleasant time doing stuff with GPUMode. Now I kind of just help with the lectures sometimes.

Um, I'm not as involved and I think in general, like we don't have as many competitions as we used to, but, um, yeah, I still keep in touch a lot with, with everyone there. Is there a friendly rivalry? Because I, I think the previous community that used to do this was like MLsys, MLPerf- Yeah... sort of thing. Is there a friendly rival-rivalry? Is this like the just new generation MLPerf or what's going on?

The nice thing about GPUMode is that it is also a community in the sense that like a lot of the lectures are very easy enough for a beginner to follow and ask questions and things like that. And like the competitions are like somewhat not secondary, but like you can participate in them to learn. I think with a lot of like MLsys, MLPerf kind of benchmarks, like for the most part, like only serious labs and companies participate- Oh... like, like seriously in them at least. That's, that was my understanding of it.

Um, I, I could be wrong. But I, I think also beyond GPUMode now, one thing that has been really exciting is there's a lot more websites and like people that work on hosting competitions. Like I think there's this, I think there's this website called like LeetGPU or something, and it's like LeetCode for GPU problems. Wow. Um, there's like other ones too that I've- LeetGPU...

that like we've, we've seen like there's many that have kind of spawned and like talked on GPUMode and like, um, it's very exciting in the sense that I think GPU programming used to be super niche. Like when I was interested in it, and the only reason I got interested in it was Tri Dao gave a talk at Princeton because he was, uh, applying for faculty.

I mean, he is faculty there now, but I listened to his talk on Flash Attention in like twenty twenty-three, and I was like, "Wow, this is like the coolest thing ever." [laughs] Um, and I was like, "This is like, this is what everyone should be working on." I guess like vLLM and stuff had come out too, and it was like, "Oh, you know, we should be writing kernels."

But now it's like, you know, everyone writes kernels. Like everyone... It's, it's-- I think it's actually almost saturated in some sense, uh, as a field actually. Any, any interesting takes for people that wanna get into it? So I think one of the biggest news is GPT 5.6- Mm-hmm... wrote more efficient kernels, so Terra and, uh, Luna could be eighty percent cheaper. Yeah.

And then we've seen other competitions where people are like setting records and they're like, "We're doing some auto research loop," and these are people that don't have a background- Yep... in any kernel writing, right? Yeah. So even on the GPUMode leaderboard, uh, if you look at like a lot of the recent problems, almost all the solutions are AI-generated. Mm. However, you'll notice on the leaderboard...

So there's this guy named Gauners who's like a very, very like regular member of GPUMode. We've always known for a long time that he's like a super, super cracked like GPU kernel writer. One thing we discovered on this leaderboard is like almost ev-- Uh, he also used AI to help him with these solutions, but for the most part, like he helped prompt and move it in certain directions.

Uh, we found that like his kernel was like basically the only one in like the top ten that was actually stable in like actual like- End-to-end systems. And it does bring into question, like, it's not... Like, I mean, GPU kernels have a verification problem. Like, we've kind of known this. It's been a problem since KernelBench was released. Like, there's a lot of reward hacking that goes on.

But you al- yeah, you also need to realize, like, the, the lines of code is a lot, a lot smaller, but- Yeah, I was gonna ask, is that noticeable, or is it just- Yeah, no, it's, it's, it's- Okay... definitely, like, very important. And I think, like, it's, it's really interesting that still there's a lot of alpha in being good at writing GPU kernels. Okay, so there is a gap from vibe learning. Okay.

[laughs] There definitely is, yeah. I, I think, like... And this applies to a lot of AI systems as well. Like, I think, you know, even with the most recent, like, math proofs and stuff, like, it doesn't necessarily mean mathematicians are obsolete. I mean, these companies still hire mathematicians, like, to do...

Y- whether it be, like, data labeling work or, or even just, like, steering the models to solve problems. Like, there is still a lot of alpha in being, like, knowledgeable in these things, so. Is it just knowledge, or is it also there's just more planning, and is there, uh, an emergent style of planning that works better?

I think it's, it's a mix of maybe this is what you mean, like, intuition for- Something, like-... how to solve the problems. Yeah, like, like, for example, I always diagram my code. Yeah. Right? Yeah. And then, like, if there's a part of the diagram I don't understand, I work until I understand it. Otherwise, I- it's not allowed. Yeah, yeah. Yeah.

So I think it's like, it's a mix of those things, of, like, the people who work... Like, the people who know how to look at these problems and how to solve them, like, also know how to use AI to do them because, like, you're acting as a very strong verifier. Like, if you are knowled- or if you know what to do and you're also...

Like, I, I, I think the thing that we've kind of discovered with all these agent swarms and things like this is, like, when you throw enough compute at a problem, you, like, can sufficiently explore solutions to that problem. But oftentimes, like, you know, maybe you can burn, like, 100 billion or a trillion tokens on something, but if you bring in someone who knows something about the problem, um, they can uncover something for the model that would, like, erase that one trillion token spend.

I mean, it's not, it's not super clear, like, what exactly the, the trends are here. But I think, like, there are so many problems in the wild still right now that we want to solve, and, like, we can't afford to just always, you know, throw as much compute as possible at it. Like, there is still an efficiency aspect of, of all of these things that is super, super important. Is there, like, a theoretical right answer that you can just calculate based on physics- Oh...

and then you just get close to the physics limit? Yes. So for GPU kernels, you can compute... It, it's actually not that easy to compute sometimes, like, depending on how complex the problem is. Like, for, for matrix multiplication, it's very easy to compute, um, this, like, speed of light kind of, uh, estimate of what the fastest kernel can be.

And, like, this is also assuming, like, you know, maybe all of your, all your data starts on the CPU, or maybe it starts in DRAM, on the GPU, et cetera. Like, this changes these, these numbers slightly, but- The transfers and all these things. Yeah. Yeah. But I, I, I will say, like, it's not clear, though, like, i- in a lot of cases, if it's even possible to hit this theoretical number, if that makes sense.

Like, this is assuming, like, perfect overlapping and transfer of data and, and, like, there's maybe some bottleneck that you can't get around. But often the kernels are not even close. Like that we write are not nearly close enough to this number to be, like, meaningful at all. Yeah. And is it speed that matters? Do you also care about, uh, obviously memory, which- Mm... which feeds into speed?

Do you care about power consumption? So one of my, uh, one of our top pods of the year was Jeff Dean, who was like, "Actually, I just tracked the microjoules or, like, the nanojoules, picojoules." Yeah, pico. It's often picojoules- Picojoules... I think. Yeah. Uh, do you care about that? So I don't. [laughs] Yeah. [laughs] I guess maybe I'm not, I'm not as- But everything here is speed, right?

Like- Everything here is speed... no- nobody's counting picojoules. Um, but I- there's a caveat here, which is I think, like, there is speed in the context of a single kernel, and there's speed in the context of a larger problem, like maybe the N10 model. Because, like, one thing to consider, and this is why it's important to talk about what, what speed of light is referring to, because in these cases for the kernels, like, we always start with everything in, in, uh, like, HBM, for example, right?

But you can imagine that, like, an end-to-end kern- like, an end-to-end model, what you might wanna do between two layers is, like, you might sacrifice the speed of the first operation to keep things in the cache for the second operation. And, like, these are things that, like, you can't really get out of, you know, isolation with, like, these kinds of kernels. And, you know, people call this, like, the fusion pro- or, like, the- Mega kernel...

kernel fusion problem, yeah, or, like, mega kernel stuff. And it generally only applies, like, when you are, like, memory-bound in, in, in most cases. But this is something that, like... Also, there, there is this question of, like, as these models get better, like, should we just be generating me- like, mega kernels? Is that, like, what we want? Um- What's, what's your take?

I think that this is really difficult because you need the data to do this. And I think, like, I have yet to see an example in the wild of, like, we bootstrap the ability to solve a very difficult class of problems without any examples. Um, and I think, like, the other reason why I think maybe this isn't that interesting is that at the level of an individual kernel, A, like, they're not that-- they're not as complex, but B, you're somewhat confident that there's not as much structure in a single kernel.

But, like, in a mega kernel, like, I would be more inclined to believe that, like, a compiler would be better here. Like, some compiler over, like, higher level- Ops makes sense because in general, like, actually I think mega kernels are very, like, the pieces are very composable of like the individual kernels.

There's some areas where you might want to do, like weird fusions and everything, but in general, I think these are cases that like a compiler can probably handle. And there is a company that's working on this from, from what I understand, that has given some talks on GPU mode as well. Yeah, I want to basically cluster all the GPU mode discussions here- [laughs]...

because obviously there's other parts- Right. Yeah... that we need to move on to. Yeah. Yeah, yeah. I think there is something to plug. You guys do host a lot of really good lectures. They're all on YouTube, people can follow. And you- Yes... lead quite a bit of it? You're still quite involved? Uh, I used to. Uh, sometimes I still do. Uh, I think they're mostly Mark.

Mark is the one who usually does them. Matei does sometimes as well, but, um, yeah, I highly recommend them. They are extremely good resources. Like, I think it's kind of crazy how much people share on there. So, yeah. 'Cause like, if you're there- [laughs]... like you're very, very... Like, you're exactly the right audience, you know? Yes, exactly. Exactly. Like, this isn't gonna reach the mainstream.

And there's a lot of like introductory material as well- Mm... um, that we've put on, um, that I think is useful for people. KernelBench was kind of influential. I just want to see, like, you know, that was last year. Yep. Yep Uh, what other ongoing work do you want to shout out that people should pay attention to? Because obviously you're involved in this- Yep... field. Yeah.

I will give maybe the in- the background story of like, I am actually involved in a lot of benchmarks, or I used to be, uh, maybe prior to my PhD. It started because I was at Princeton. Uh, I worked with the SWE-Bench team there. John, Carlos- John and Carlos... Ophir. And Ophir. They're all great. Like I- Karthik... I love them. Yeah. Yeah.

Uh- There's basically this, like, I think people don't understand how mu- I mean, how much benchmarks come from the same group- Yeah... at Princeton. [laughs] It is crazy. Do, do you know Shunyu? Yes. Yeah, yeah, yeah, yeah. Uh, we had him on a pod before. Now, now he's like running Tensor. Yeah, now he's like a- [laughs]... he's like a superstar. Uh, when I, when I, I met him...

So he, uh, he was advising my friend Michael, Michael Tang, who is now at Anthropic, but they worked together a lot. We were like the two undergrads in Karthik's lab. I... And then some others joined later as well. But yeah, Shunyu's great. I, I did not know he was like such a superstar until like, later on, like after I left. But- Yeah... um- I mean, like, like, okay, so there are very few PhD students...

Like yours is, yours is like the next one. Mm. Like once a year we fe- feature someone like, who is like basically entire PhD, uh, has been, like, on target. Mm-hmm. There's not that many of them. Shunyu was like, clearly one of them. And, uh, you know, Jack, Jack Morris is another one. And like, you know, we talked... Before the show, we talked about research taste. Mm-hmm. Right?

Like somehow some grad students just have a very blessed career where like, yeah, mostly like, yep, yep, yep, yep, yep, this is like going to stick around relevant, everyone should know this. Yeah. And then others just nothing. I think this is also true of like, even people within like industry labs as well. I think it's just like grad students are a lot more visible.

So you just see, like you see like, you know, there are some people who- Yeah, you can publish... who really like get lucky and like... Or, I mean, it- it's a mix of being lucky and also being very smart and things like that. I think like with, with research tastes as well, like I think it, it gets, it gets developed through opportunities, at least in, in my case, like I got...

I was very fortunate to have like taken the path that I took, like working at Princeton and then like later, like finding my advi- like Omar at MIT, like he's a fantastic advisor. I will say though, I, I find that the most successful research from grad students or like in academia comes when people care about problems that maybe like most people in industry are not looking at.

I think this is the issue that like a lot of grad students work on things that benefit, like, that look good to an industry lab. Like for example, they'll work on some like, some benchmark. It's really popular now. I mean, I think benchmarks in 2023 were a very different story than benchmarks now. There are a lot of people that work on like harnesses and like meta harnesses and like specific harnesses for XYZ task.

And when you really think about it, the reason someone would work on this is like, maybe there's like a clear goal shaped around the models that we have today of like, this is what I want to see. But like I'll give, I'll give the, the, like RLM, like the recursive language model paper as an example because I think like it's a super, super simple idea.

I think when it came out as well, like there were a lot of people that were like, when they see something like that, they're like, "What is even the purpose of this?" Or like- It's not a tool. Yeah, like why, like what... This is just sub-agents or something, right? [laughs] Uh, and I think it's like when you get a reaction like that, it- it's almost like a good sign in the sense that like it's clear that people aren't thinking about what the purpose of this is.

And I, I will give another example of like SWE-Bench. When SWE-Bench came out... I mean, Ophir loves to tell this story. When it came out, like nobody cared. Like everybody was like, "This is an impossible task. Like why would we ever even consider this as a benchmark?" And it wasn't until Devin came out that everyone was like, "Whoa, like this is something we want to hill climb."

And I think this is, this, this rings true for... You tend to see that a lot of ideas... I, I think like the... My favorite, I guess, example of this is Eric Zelikman's work, uh- With Star... with like Star and, and like QuietStar. Like I think when you read the paper, at least when I first read the paper, I was like, "Is this not like an obvious idea?" Or maybe not, I don't know.

I mean, I was like, "Oh, this seems really simple." Or like chain of thought, and the same thing. Or like Shunyu's React. It's like, okay, like yeah, I mean, sure. But then like when you really think about it, it's like why-- what is the value of the paper? And I think it comes from like, it tells a bit of a story as to like what you want the field to look like.

And that is something that it's very hard to do this in academia because if you look at all these papers, QuietStar, React, RLMs, SWE-Bench, none of these papers are... It's not like a GPT 6 Astra release. You know? It's not like everyone's like, "Oh my gosh, like I'm gonna use this now and this is the best thing in the world." Like academia just can't afford to do this, at least right now.

I mean, I- [laughs] there's a whole slew of reasons why I think that should change. But I think it's like- If you don't have-- as a PhD student, I think you're in such a unique position where you can work on literally whatever you want for the most part. If you're not taking advantage of that and working on things that, like, nobody cares about or, like, people see as some trivial thing, like, "Oh, I thought about this, but, like, I don't use it," I just think, like, in the end, the research is just never gonna be that interesting because you kind of need to take big bets if you're gonna be in academia.

'Cause otherwise, I think, like, just go to an industry lab. Like, you know, they have tons of resources, tons of talent. Why constrain yourself in an area where you don't have a lot of resources and, like, there's not even that many people around? And I think it's just-- it literally just comes down to, like, big bets.

Um, but you just have to take big bets, and, like, a lot of them will fail, you know? Like, that's just-- it's, it's natural. But I think that is, as a PhD student, like, that's the biggest advantage you have over any single person at another lab 'cause you don't have to deal with bureaucracy and all these other things. Fair enough. Yeah.

I ask a lot of people this question, and usually they hand wave away. So I think I appreciate that you're actually th- giving a thoughtful response on- Mm-hmm... like, no, like, uh, this is your unfair advantage because everything else is biased against you basically. [laughs] Yeah, yeah, exactly. A-and so, like, it's honestly-- I will bring up JEV as an example 'cause it's, it's not an academic- Wow, okay... it's not an academic project.

I want to bring this up because this also happened with RLMs and, uh, it's happens with many other works, like things get overhyped, right? To, to an extent, like something gets overhyped, and then people are like, "Why is this overhyped?" Like, "This is trivial. This is stupid." And I saw the same thing with JEV because I think the release was like...

I mean, there is this whole thing about, like, academics or-- they're not an academic group, but, like, people have to do branding, and they have to, like, kind of market their research, and so, like, I understand. But I think there was a lot of discourse about JEV just being, like, something we've known for years. And I think it's kind of missing the point of like, why is such a system so interesting?

It's why is it not just some stupid MLP classifier that, like, we've, we've been doing, you know, back in our intro ML classes or something. I think what's really interesting about JEV is that it kind of opens up this question of, are language models correct? Like, in the form that they're in, can we consider a different design space other than text to text?

'Cause what, what they're doing is they're basically saying, like, "I will take advantage of this language model backbone." Like, I know it captures a lot of information about language, but I'm going to change the output space of the model to give you a trade-off, which is I will do very, very fast inference over like if you have some prior about this problem, like let's say I only need to make a binary classification.

Am I gonna ask my language model to do this and pay like a four hundred x cost? No. It's, it's like silly, right? Um, and I think for the longest time, because the labs are the only places that control, you're never gonna use something other than like GPT-4 or GPT-6 or Fable because they're the best models.

But because of that, like people have gotten kind of accustomed to this idea that a language model is just a autoregressive decoder. Like we have accepted this, and I think when RLMs came out, it was the same thing. Like one of the comments I-- like a very frequent criticism I got was like, "This is not a language model," or like, "When I look at this, like I thought it was a new architecture, but it's actually not."

And my response to that is like, well, a language model is just modeling language. It doesn't have to be this transformer decoder, you know? Uh, and JEV is really, really interesting in that like we now have a new thing to tune, which is like what is the output space and how does this affect inference latency? Uh, and I think we can actually start asking this about various parts of the language model itself.

Um, we are seeing this too with like loop transformers. It's a similar idea of a lot of the attention around it was like, this is a silly idea. Like why, who cares about this? But it's like it is a simple idea, but it's actually-- it opens up a whole new set of questions that I think like, especially if you're a PhD student, these are the things that you wanna answer because I, I think it's like we don't know for JEV, for example, we don't know how far we can take this.

Uh, for loop transformers, we also don't know how far we can take this. What if you loop only a subset of the model? Um, what if you route to like only, uh, like you have some router to different parts of the model? Like, can you mimic what you do in a harness inside of the model? What can you bridge between a harness choice and a model architecture choice?

These are all questions I think that open up with works like this, and that's like really exciting. JEV in particular, when I saw it, I was like, this is actually really useful for RLMs. Like I think it's, it's-- it makes sense 'cause the, the biggest bottleneck in RLMs or swarms or, or systems like these is they're slow.

When you do multiple language model calls all the time, you're not distributing your compute correctly because like maybe there's something trivial that you just want a simple model to do, but you can't do it because your language model is just this, this bulky thing, you know? So I'm very excited. I think, I think we will start to see new types of models emerge beyond just the bog standard frontier model, and that is like there's so many things that you can do with these like new trade-offs.

Um- Also shout out Thinky with their interaction models. Yes. Yeah, yeah. Another great example. Yeah. So like basically try to break the paradigm from sequence to sequence decoder only and just literally do anything else. Yeah. Yeah, yeah, yeah. [laughs] I mean, I think there's, there's a level of if you're trying to compete, you're not gonna compete with a frontier lab doing an autoregressive- No...

decoder on like the amount of compute scaling resources they have- Even Thinky will not, you know, okay, there may be one of the handful that can, but you're not really gonna do much in that list. I don't know too much about Thinky or I don't wanna say anything either, but it's like if their strategy is just to replicate OpenAI or Anthropic, like that's a horrible strategy.

[laughs] Because-- well, because like, I mean, th- they just don't have... Like, you, you kinda just have to think of it in terms of like what advantage do you have? And if, if you're going to use the same setup, I, I'm sure they're not, but it's like if you're going to do the same setup, like you're basically competing on the things you can control, which is data and compute, and obviously they cannot compete with the frontier labs on that.

So yeah, it makes sense that like if you are a Neo lab, like... Actually, I don't know if you would consider them to be a Neo lab, but I guess like- Yeah. Yeah. They're, they're in there. That's why they're the main one. I guess they're kind of a weird one. Yeah, but- They're in that list. They ship Inkling. Yeah. Like, they, they count.

Anything other than OpenAI or Anthropic, maybe like f- Meta and, and, and GDM, like you just-- you gotta do something else, you know? Like it's just-- it's the sad reality, but I think-- I, I actually think it's a good thing. I'm very happy that like scale works and these companies will just keep doing it because it opens up like potentially new players, like if they uncover something like really, really interesting because I sort of have my doubts that this is like seriously going on at Frontier Labs 'cause it's like, like why would you do that?

You know, like why would you take the risk of allocating a large amount of compute to new bets when the old bet already works? And I think that's what spins off a lot of Neo labs, right? Yeah. Yeah. You have a side bet and you don't get compute- Yeah... and you're like, "Okay, I'll go, I'll go do it." Exactly.

And you know, your example of the potential upside is something like Jev, which is x hundred times cheaper, you know, comes out and maybe it is language model. This- Yeah... in this case it's just different. Yeah, yeah, yeah. Yeah. Um- So no speculation on, on what Jev actually is. [laughs] I guess I have some guesses for what it might be. I, I have seen some people say like, "Ah, it's like a diffusion thing."

Um, I, I guess that- Which is the parallel decode, right? Yeah, parallel decode. Honestly, I think regardless of what it actually is, 'cause I think you can-- I've seen some like open source replications of it. What is really exciting to me about what they did is I'm not entirely sure what their optimization objective was and how they trained it.

And I think like this is a thing for RLMs that we've also been thinking about, which is like, okay, like RLMs are a very simple idea. If I come out with this paper, like anyone can use it now. But what distinguishes the actual value of an RLM is whether or not you can train it properly and whether or not maybe you can mold some architecture around this system to make it really, really good and that's something that like, I mean, I'm actively working on I guess.

But I think for them, like they figured out a way to train the system which is completely non-trivial. Like I, I actually don't really know how they did it and I've seen some comparisons online of, you know, some people are claiming they used Qwen, uh, or they post train on top of Qwen. But every open source Qwen that you use is gonna be worse 'cause whatever they did to train it clearly works very well and so that's, that's very exciting. There's one element of calibration- Mm-hmm...

which, uh, is a, is a rare topic that I don't think people even knew about or understood. We covered it with, um, our conversation with Clementine Forey of Hugging Face when she used to d- run the evals, um, at Hugging Face, which is basically the idea that, uh, you know, models are pr- attuned to give you the most likely next token.

But, uh, they're gonna lie to you when you ask them, "How confident are you?" Because they're just gonna give you the most likely next answer instead of like actually like, "No, let's calibrate. Like I am actually fifty percent sure, or I'm twenty percent sure, and like let's try to calibrate that." I would say like if anything, I think that actually that's pretty easy to generate synthetic data around- Mm-hmm...

uh, because you can sort of see the truth and then synthetically generate a bunch of answers, have Jeff classify it, and then compare with ground truth. Oh, I see. That would, that would be my reverse engineering approach. Yeah, yeah. So I've actually-- I think calibration is probably the under-- like people are just using it as a very fast classifier. Mm-hmm.

But they're actually not even using the probability or calibration estimates- Yeah... or confidence estimates. Yeah. Yeah. I, I think it's also still just misunderstood to, to reiterate when you ask a model, "How confident are you?" It will spew out what, forty-three percent. Uh, the, the big delta is this is a grounded classification, right? Yeah. Yeah. I'm, I'm very excited to see what people do with this model. I mean, is it gonna solve everything? Like no, of course not.

But I think it solves a class of problems that we traditionally struggled with, which is low latency things. So I, I love the examples with games. That's actually like, I mean, I- Yeah, the Doom example. Yeah. Very good. I have a benchmark on language models playing video games. I've always been fascinated by whether you can have an intelligent system play new games and things of, of, of this nature.

And so I think it's really, really cool that they have sort of a unique way to do this, to capture language and understanding in like a fast, a very, very fast model. Well, while you're on the topic- Uh, okay... anything you wanna point out from video games? Video games. Oh, yeah, yeah. I mean- I mean, you know, you did do a benchmark on any- So, um-... any frame rate play. Yeah, yeah, yeah.

[laughs] Um, I guess these numbers are very outdated- Yeah. [laughs]... 'cause a lot of the models are, are, are very different. And I've seen actually people run... There, there are some folks out there that are running newer models on these games, which is really cool. I guess the general premise of this benchmark was we just want to see if vision language models are, are good enough at just like plugging into games with the latency constraint included.

'Cause this, this actually I came out with this right after Claude Plays Pokémon came out. Mm. Mm-hmm. So this was like two years ago, which I guess is like ancient now. But I think what's really, really cool about this suite of tasks, it's very diverse in terms of what games they are, and also I think most of the games are games that people know or like have seen before.

I saw, yeah, Jev playing Doom. I will say I don't think... I think they were just playing like really simple levels and stuff, but honestly, like most models still can't really do, or I don't actually think any models can solve these games very meaningfully. Like, there are some games that they can. I think I-I've seen Astra be able to solve the Kirby game.

And we also, for this benchmark, we intentionally designed a really minimal harness, and I, I will get to this point about harnesses because I think there's a whole conversation to be had about like what is the value of a harness, like what is even the purpose of a harness for a model? But in general, like I think it's, uh, yeah, I, I hope to see very quickly or very soon like all of these games beaten by newer models. Yeah.

It's interesting, like the old, old Cloud Place Pokemon, they like read state from- Mm-hmm... RAM and saw what tiles are walkable or not. We did a podcast with them- Mm-hmm... a long, long time ago. Gotcha. Gotcha. Yeah. Just fun tangent. Yeah, I mean, and, and it's, it's similar, right? Like Jev doesn't have vision- Yes... so you have to kind of feed in, uh- Yeah, yeah... these like, uh, game state and all these things.

I mean, let's go right into the harness stuff- Awesome... because you me- you brought it up. Language model harnesses are compositional generalizers. Yes. Struggled to read that one. [laughs] Explain. Yes. Okay. Okay. So I have been a little unsatisfied maybe with, with how people think about harnesses because people compare like, "Oh, like I love CloudCode, I love Codex. I love Pi." Like, "No, I love Oh My Pi. I love PrimeAgent."

To be honest, I think all of them are the same. Most of the design decisions or like the design choices around these harnesses are the same. Maybe PrimeAgent is a little bit different 'cause it's like inherently an RLM. But in general, like I think we can be a lot more creative with harnesses, and what I mean by that is if we think about this from the perspective of what exactly is the harness doing for the model?

Well, basically, when you're trying to solve a problem and you want to use a language model to solve it, like a very difficult task, one thing that we have discovered is that next token prediction is a really, really awkward form to do a lot of these tasks. So for example, take SWE-bench. When you're navigating a code base, like are you going to be able to figure out how to do all of this with a single language model call?

Like you just say, "Solve code," or like, "Solve my query over this code base." No. So we rely on a harness to help you do these kinds of things, and I think what, what is interesting is like a harness is a very, very opinionated program over how you want a language model to be form fit over a problem.

And I bring this up because when we think about the, what a harness is doing, we should really think about like what choices in the harness let the language model solve this task, and can I actually just have a language model that just does this? Because a harness, if, if you think about it, now that looped transformers are a thing, I think what's really interesting about it is you can model a looped transformer in some ways like with a harness as well, right?

You're just looping over the mo- Now, you can say like, "Oh, I'm not decoding," so it's like a little bit different. But in general, we, for whatever reason, have stuck with the same model architecture choice forever, and I... There's many arguments for why, but clearly, like we are now training models around harness rollouts.

And so there is this very awkward way of doing training over harnesses, which is that we train a language model to act within a harness, but like now it's like a really long, maybe like multiple agent rollout that we're doing, and there's like really awkward hacky ways of doing this. So what this blog talks about is like well, one way you can think about what is going on here is if the harness is basically helping the model solve a particular task, can different harness design choices actually do something a little bit more meaningful beyond just here are some tool calls that will help you.

Here is a way to grep through your code base. And so this actually, the idea for this blog came with the RLM idea as well. We just didn't package it that way, and I, I think this is actually true. There are many other ideas around RLMs that like we will be coming out with, but were all there from the beginning. I mean, these are all design decisions around, um, I think with what, what is...

What I like about the RLM is that there were many, many, many iterations and versions of different abstractions that I was interested in doing, and ultimately the RLM made the most sense. But there's a lot of reasons that aren't public as to why that's the case. Um, I mean, you, you will see, so in this example, one of the things that we see with an RLM is that if you sufficiently offload context and write, ask the model to write code over that context, you get this really, really weird but useful property, which is that when the model recognizes during training how to solve a task, it turns out that the solution to many tasks is very similar across like tasks where you, you don't even-- like it's not even that clear to you that the solutions are similar.

So in this example, we have like a retrieval task, and we have like a, an aggregation task, and they're very, very different query. Like the domain is just completely different. And when you train a regular language model over these two tasks, the trajectories look very different. And so what you're relying when you-- like let's say you use Pi or CloudCode or something, which is not in this blog, but we do have these results.

You'll find that like these harnesses distinguish too much between these problems even though the solutions are the same. And so one thing that we find when training RLMs is that like- When it sees these problems as the same-- And the reason it sees these problems as the same is the sub-agent sees different problems, but the sub-agent is solving an easier sub-task, and so you're confident it's smart enough to do it.

But for the base overall strategy, they end up looking the same, and so when you train on the left task, for example, it can immediately solve the right task. And so if you go down to, like, the plots that we have, one thing you'll kind of observe is that the-- as you just naively train your RLM on these tasks, they naturally learn to generalize, for example, to longer tasks because the strategy is basically the same.

You're just modifying, like, a length variable. And this actually also holds for tasks that are different, and they're-- it's not even-- they're not different across length. They're completely different tasks, math tasks versus writing tasks. But the solution, the, like, meta high-level solution is the same.

And so when you train the RLM on one of them, it generalizes this behavior to the second one. And there's no magic here. I, I guess maybe that's the thing that I wanna kind of stress. Like- How, how would you kind of verbalize what they're learning? So- Mm-hmm... I think in here you say you train it on short tasks, they generalize the stuff eight to thirty times longer. Yeah.

They're learning how to solve these type of problems, or what's the, what's the core thing they're actually learning? Yeah. They're learning how to solve these types of problems at a certain length, and it turns out that when you take the strategy that they learned, it is directly transferable to the longer length. Like they're- Yeah...

effectively the same program. And this is something that is really exciting because what this kind of implies is that when you have a corpus of data or environments that you train your model on, the hope is that, like, A, you can train on less environments and generalize to more than what existing models can, can do through nai-- or harnesses through existing kind of like naive training.

But B, also, you still wanna use all the data that you have. So when you train on these tasks, like, hopefully it generalizes to a wider class of, of problems. And why this is also even more exciting, at least in the context of RLMs or any recursively calling system, is this argument holds inductively.

So, like, I'll give you an example 'cause I said I claimed that competitive programming and GPU optimization use very similar skill sets. The model can-- the harness potentially, I mean, you might have to nudge it a certain way, but it can learn that, like, okay, how I'm gonna go about solving this GPU programming task is very similar to what I learned for competitive programming.

So I'm gonna list out a set of solutions. I'll, like, spawn sub-agents to list out promising solutions, and then I'll, like, write this loop to go through and check these solutions, maybe evolve them and, like, evolve them against a verifier. And between these two tasks, this looks the same. But what the sub-agents are doing are maybe, like, unique and something, something, like, different.

But even what the sub-agents are solving might actually also be of the same form, right? 'Cause it's like a-- it's a recursive argument. And so what I'm trying to get at with this whole blog post is just that, like, we should rethink what the role of the harness is because harnesses can actually greatly increase the generalization capability of your model and the amount of data that it's given.

And this is not exclusive to RLMs. I think there is a wide class of harnesses that are yet to be discovered that actually can yield similar properties. And to extend this argument a little bit further, I think what you can also extrapolate from this is, like, if I look at an RLM, what are the components of an RLM that are actually necessary?

And can I actually just directly train a model to do this? Can I train a model to act as an RLM implicitly in its forward pass? It's a really weird thing to think about because, like, you know, you might say, like, "Ah, code is non-differentiable," blah, blah, blah, blah, blah, blah. But there are many approximations of, of this behavior that we will start to uncover.

And I think, like, we will see beyond just, like, I'm gonna design a new coding harness that uses a special form of compaction or something. I think we can be a lot more creative here. Like, there's, there's so much we can do with these language models that I think we are just not doing. And I'm, like, very excited about this 'cause I think, like, I think we can get serious gains from very, very opinionated and good harness design that lends itself better to scale.

And what I mean by this is, like, you know, the RLM, for example, is a very primitive inductive bias. Like, there's nothing super special about the design other than the fact that it's very different than what we currently do. But this may potentially scale much better with, like, the data and the environments that we have available to us. I guess the, you know, opposite thing that people would probably ask is current harnesses- Mm-hmm...

are very generalized towards coding, which people see works for a lot of domains. Cloud code is being used for design- Yep... presentations, everything. Um, Muse Spark, GrokBot- Mm... these are very simple, non-opinionated harnesses that are good at code, and that is also scaling out. Um, what's the example of how we improve those, I guess? Hmm. Yeah.

Um, let me, let me bring up another paper, uh, which came out very recently. It's like the harness tax paper. I think it's by Arena. I really like this paper because it puts forward a prior that I had, which is basically that, like, [laughs] most- It confirms a prior [laughs]. Yeah. Like, most harness choices don't matter because- Yeah...

all of these harnesses are the same. But I, I will say, like, GrokBot, for example, is actually quite different, I think, from my understanding, than, than how some of these har- other harnesses have been designed, and I like that a lot. Um, and I think it, it's clear from here at least that, like, I'm pretty sure- Anthropic or OpenAI are exclusively training on their harnesses.

They're probably not training on their competitor's har-- I mean, I'd assume not because I, I don't know why they would do that. But- But, you know, th-this is a thing you see in open models, right? Like, Qwen is really good at using open code. Yes. They need to train in harnesses. Old Gemas were notoriously bad at this. Yeah. Models are good, but you need to train in a harness.

I, I think, though, as models get smarter or, like, as they get better, this distinction becomes, uh, not that important in the sense that, like, if you take Astra and you put it inside of open code, like, it's not gonna go crazy because I think it's, like, just sufficiently good. And so I say this because the only benefits between these different harnesses is just cost for the most part.

And, and I think, like, what I'm getting at here is if you plug these models into RLMs though, they're not that good still. They're okay. And I think it's mainly because the types-- like, the class of harness that we are training around is this class of harness, this, like, pi loop, this, like... I, I like to call it trajectory as a prompt, which just means, like, you keep the whole trajectory of the rollout as the context that, that, that your main model is, is using.

Even if you use sub-agents, it's still, like, kinda of this form. And I think we're going to-- if we want to explore new harnesses, there needs to be teams that are dedicated to actually running meaningful experiments over, like, scaling out new harnesses, like maybe post-train scaling out on different harness designs.

Like, I, I think we can actually get very meaningful knowledge or gains from doing this kind of thing, uh, whether it's an RLM or whether it's something different, and that's exciting 'cause I think, for example, if you train a lot on... Fable for a long time was the best model for RLMs because they had dynamic workflows, and it was pretty obvious that, like, this was a capability that was somewhat trained in.

Even if the model was still, like, a little dumb, like in the RLM harness, it still worked a lot better than, than other models did. Uh, Astra is now also, like, good enough at doing these things. But- Wait, is this, is this where we see that Fable is the best for RLMs, or is there, is there some other- Oh, no, these are all internal results- Yes. Uh- I guess. Yeah, I don't, I don't have them- Okay. All right...

public right now. But in general, like, I think you can, you can very easily tell that we have not optimized for RLM, like, workflows yet, and I think if we get models that do this correctly, they will be a lot more efficient as well. I mean, you can kind of just think through, like, why this is the case, right?

I think this is the point where you have to give the ten-second what are RLMs- Oh, yes... 'cause there's a lot of listeners here that are probably- Yes, yes. We're assuming a lot of knowledge. Yes, yes. Also, I think-- but you have set some context- Yes... so you can-- like, with everything we just said- Yes... can we have a clean, crisp des-definition of RLMs? Yes. Okay. I want to go back to the blog. Yes.

The, the Compositional Generalizers blog. This one. Okay. Okay. This is, like, the best-- I wish I had this in the original paper. An RLM is basically just a harness design where the only tool in the harness is code, uh, which is this programmatic sub-agent calling thing where it has the option to call itself as a tool, and it has other tools, but all of these things are functions in code, and the context that it's dealing with is always stored in some memory inside of this code environment.

So this could be a file system. Like, this could be like... I'll give you an example, Primeagent. The trajectory of Primeagent, like the context, even the-- when you compact and do all these things, is stored on disk. So the model can always reference its original context, even if it's compacted, and all of its tools are run inside of, let's say, like, a Python REPL or a Bash REPL.

And so this-- it's, like, this very primitive abstraction. And I would say, like, where most harnesses differ is, A, context offloading is not done that, like that. I mean, if you look at Primeagent, by the way, Primeagent does context offloading but not all the way. So, like, it still maintains the standard cod-code-codex loop of, like, trajectory as a prompt where you compact, but it has the additional kind of like the context is offloaded, and it only has-- the unique point of, of Primeagent is that the only tool is, is IPython.

So this is, like, the very kind of generic abstraction around, like, RLMs. Yeah. Concretely, what's the core thing RLMs are trying to solve? Yeah. At one point, I think when it first came out, it was context- Mm-hmm... but I'll pass the question. Yeah. So the original problem was harnesses have a really bad time dealing with long context. They typically were used to only really do them for, like, specific things like code.

Like, you know, they could deal with your code base 'cause it was trained on it. But now it's more around what this blog is talking about, which is compositionality and the fact that, like, I think harnesses-- we want to have language model systems that have much more control over the actions they make at every step.

And what I mean by this is tool calls are very limited because you have to invoke them every turn. Like, you have to invoke tool A, then tool B, then tool C, and there's no central context that you can kind of draw back from. And RLMs are specifically, like, designed around composition and having, like, a central context that you can always draw from. Um, and this context is, like, designed around the existing language models.

Another, like, very similar example actually in design is, like, agent swarms, for example, the, the Hugging Face incident. Like, these agent swarms have, like, a message board that they, that they learn to communicate over, and this message board in some sense is the shared context that they, like, act over.

And RLMs basically say that, like, the best way to communicate through this is in code. Like, you write the code to do this, and it's, it's because these models are so good at writing code. Um, like, we wanna take advantage of, of that fact. And then for, for this compositional thing up to and including generating your own harness, you know, specific for the task. Exactly. Exactly. Right.

I think we will start to see that if you go up to, uh, this figure. Okay. So we, we talk about this idea of locally in-distribution tasks for a harness, and it is like a, an idea on top of, like, in-distribution tasks. When we think about language models, an in-distribution task is just a task where, like, the prompt is something that the model has either seen before or, like, has seen some version of it.

Most harnesses work like the one on the right, which is they keep appending the trajectory as a prompt, and so eventually, unless you're Anthropic or OpenAI and you train on, like, kind of these, like, user trajectories, most of these things end up being out of distribution for the most part. But locally in distribution is basically the compositional argument of if an RLM breaks down its computation into, like, um, kind of a meta harness of sorts or, like, a program that involves sub-agents that, like, look at a local problem, every individual language model call over the course of this task is in distribution, even if the entire task itself is out of distribution, and this is a very desirable property, I think for, for obvious reasons.

Like, if every task is in distribution for each individual language model call, you will probably get to the right answer. Um, so... [chuckles] I mean, the, the logical limit of RLMs is RLLMs, where, like, you're not just-- you don't just write the harness, you also train a custom model. Exactly. You collect data, e-everything. Yeah.

Like, it's a fully automated AI researcher inside of your harness. Yeah. We will see where the training of RLMs goes. I will say as an academic, I am not working on this at MIT, or at least in the scaled sense, because I [chuckles] I can't afford to. Uh, but there are companies out there that are working on this.

I think Prime and Select is very clearly working on this, and it's very cool. Like, I'm, I'm very excited to see maybe we'll observe, I don't know, but maybe we'll observe better, like, post-training scaling laws with when you train around a smart harness. Maybe we'll even see smarter harnesses that come out and, like, they, they work better around these kinds of principles. What is a smarter harness? Like, that doesn't mean anything. You just said they're all the same.

[laughs] No. What-- More of what I mean is, like, Claude Code, Codex, Pi, et cetera, are all the same in that, like, when you break down the, the logic of the harness, it's, like, virtually the same thing. Yeah. Two calls in the loop code. Yeah. Whatever. But with RLMs and, and with other harness abstractions, it looks very different. And this is where I think you really distinguish...

I mean, it's, it's in the same way that, like, I think with language model architecture choices, a lot of architecture choices end up kind of looking the same when you, like, scale it out or like i- the, the differences end up being, like, somewhat minor in terms of... I mean, for a lab it's not minor, but, you know, maybe one model converges better than the other one, like, slightly.

But in general, like, if you were-- So for example, pre-training scaling laws only hold because the architecture choices we have are somewhat stable right now. But if you were to completely change the architecture, pre-training scaling laws probably don't hold. Or, like, these kinds of-- this, like, power law is gonna look very different.

And it's like the same thing with harnesses. Like, I think all the harnesses we have right now, for the most part, roughly look the same, but there are some exceptions to this, I think, that are, that are coming out. I was gonna say, like, actually, one of the things that I've been more interested by, you know, talking about PhD students who take big risks, is that people have been-- people are also pursuing the other side, which is pre-training scaling laws don't hold if you change data.

Uh, right now it's just raw unstructured text, uh, corpus of internet. What if you had a better data representation to train on? Then, then yes, your scaling law [chuckles] would change as well. Yeah. So there's architecture, there's data, and you know, whatever else, uh, you can think about. Well, so I just wanna get back to, to this. I-- it, it all makes sense.

It's, it's very interesting how you sort of recurse up and down the stack. [laughs] [laughs] From, like, very conceptual to, like, nope, like, well this is, this is where we are today. Mm-hmm. But like, yeah, obviously, uh, you can, uh, you can scale up and down. I guess, um, I'm curious, how did you start working with Prime? Is Prime taking on more work with this? Is this their answer to Hermes Agents? Mm.

Uh, you mentioned GroqBot is a little bit different. I just wanted to, like, name check all these guys and get your thoughts on each. Yeah. Yeah. Uh, I got involved with Prime, uh, after they, they released a blog post, by the way, not affiliated with me at all, about, like, how they believed RLMs were kind of the, the future.

And I, I had a friend that was working there, GP Mode, Matei. Like, we got in touch, and I think I agreed with a lot of the researchers there and, like, what they believed about harness design. Like, I, I, I was, I was very impressed, I think, that, like, they understood the purpose of the RLM paper- Mm...

which is not necessarily just to say that, like, you know, we're solving long context tasks, but actually, like, we want more opinionated harness designs. Yeah. There, there's always, like, the result of the paper that you choose to highlight- Yes... versus the actual point. Yes. As a- [laughs] I, I, I would love to talk about, like, the incentives of academia and, like, the things around, like, why it's kind of flawed and all the issues, and we'll get back to that. [laughs] Yeah.

So, so anyways, um, you know, I, I, I love the guys at Prime, so we kind of had been... After, after we decided to work together, we decided to look into training an RLM and also build this kind of RLM harness and kinda see where we can take it. That is how, like, Prime Agent came about, and I think the reception for Prime Agent has been pretty good.

Like, the one thing I was worried about with Prime Agent is that none of them, at least at the time when we were building it, none of the models were that good at doing RLM stuff. So this was like pre-Fable, pre-Astra. Uh- I guess, I think to take a step back, can you explain what Prime Agent is, how it's different than- Yeah... a traditional, you know, Claude code, what people would expect harness? Yes.

So Prime Agent, I think I, I mentioned this a little bit earlier- Yeah, there's the diagram. Yeah... is, is basically, uh, it, it is a- A harness on top of Py, like PyMono, uh, which is PyMono for context is like the- Core agents... like a, a minimalist- Yeah... yeah, like harness. I, I use Py as the reference for everything because I think all other harnesses are basically just Py. [laughs] Um, but it is Py except we explicitly restrict IPython to be the only tool that's available to it.

Every other tool gets loaded in as like a Python module, um, or like a, like a bash kind of script that it can, it can run. So it uses the core RLM abstraction on top of Py, and then it also has this continual harness thing, which is, uh, Seth, he's another PhD student. This is a thing that he used to get language model harnesses to play games. Like, so he worked a lot with Joel, uh, who is the, like, the Gemini plays Pokemon guy, and continual harness is also, by the way, very simple.

I, I quite like it. It, it basically is this design, uh, principle around, like, what parts of the harness can you let the harness itself modify? There are certain pieces that, like, you'll let it modify its own skills, the sub-agents available to it, uh, what the system prompt to the model is, and continual harness is available basically as a tool inside of the IPython kernel.

And so that's what Prime Agent is, like, how, what it's designed around. Everything else in Prime Agent is like- Standard... standard. Standard. Right? Yeah. I think what is-- what I, I, I really liked about it, and we got kinda lucky, is that, like, a lot of the new frontier models actually work really well inside.

And actually even a lot of the open source models work really well, at least some of the newer ones. And there is another thing in Prime Agent I should highlight, which is that, like, we have a very particular agent-to-agent communication system or like, uh, like framework, which is because RLMs tend to spawn many sub-agents, we want a way for sub-agents to communicate with maybe the root or with each other.

And so there are some design decisions around, like, what each sub-agent is allowed to talk to, how it does it. Again, everything is in code, so it, it writes the code to do this kind of communication, which I think is really, really cool. And then there's, I guess, persistent sub-agents is another thing that was kind of added, which is sub-agents, they can last beyond, like, the standard runtime of, of the actual, like, original agent, and you can go into that sub-agent, you can prompt it more, um, like you have more visibility and flexibility into what is kind of going on.

This is my number one pain with Codex right now. They, their sub-agents are just very ephemeral- [laughs]... and they, they actively discourage you from using it for long-running things. Yeah, yeah, yeah, which, which I think it makes sense. Yeah, I- So I mean, the, the trick is just externalize to a file system, right? Yes. Yeah. Like that's the trick. That's- And then- That is the, the big trick.

[laughs] Yeah. And well, and also like force everything to run through code. Uh, you know, trust the model- Yep... that can write code, uh, and it's gonna write its own harnesses itself. So is, is Prime gonna take on like training, uh, post-training custom models for this? Is this a one-off collaboration between you guys, that's it?

Like what's- Yeah, they are training a model, uh, intern-- I think they were, they were pretty public about this actually- Yeah... um, back in March. I mean, clearly it is, it is their business, right? Yeah, yeah, yeah. Like, yeah, so [laughs] Uh, yeah, I mean they're, they're showing that they can train it on their, their kind of hosted training stack. But no, uh, so for model training, I'm, I'm not involved with them on that.

The main reason is just I have other things in the PhD I wanna work on. I think, like, there are many other big bets to take, um- Ooh... outside of just RLMs. Uh- Ooh... some, uh, some I, I don't know how much I can share yet. Uh, but in general, like I think-- I actually think one of the luxuries of being a PhD student genuinely is that there's so many big bets to take.

S- most of them will probably yield nothing, but it's a really, really exciting time to be in research, especially because I think most progress in the field has been a little bit boring. Like, I'm not saying the, the outcome is boring, but the, the process of, of doing these things tends to be quite boring. Um, and so there is kind of this question of like, you know, what, what do we wanna do next?

Um- Yeah... but we can talk about that later. Yeah. Yeah. Okay, I wanna close out a little bit more of your research, and then we can, we can- Cool, yep... uh, start signing out. Uh, since we released RLM, a lot of excitement about it. Any secondary third-party work that you wanna shout out as like that you guys should take a look at this? Oh, yeah, yeah, yeah.

So, uh, Harvey, the legal AI company, released a blog post, not affiliated at all, but they post-trained an RLM on their, like, legal work, which often involves a lot of, like, sifting through documents and kind of looking through, like, a variety of, uh, specific information that maybe is not so easy to retrieve with like a pure retrieval system, and they show like really, really good results.

It's very exciting. I mean, I, I, I was, I was shocked that, that, that they worked on this. They did not tell me, so when this came out, I was like, "Oh, that's awesome." So there's this one I think is super, super cool, and, and what they're doing there, I think this is a collaboration with Base10 by the way as well. Yes, this is Base10.

Headlong, which is law-- the law institute's kind of, uh, it is their like persistently running harness. Um, it's very cool- Oh, they renamed it? They used to call it something else. It was like Auto- Terminus... Ter- yeah, I know, I know. They- Yeah... they, they, they've, they've gone through... Yeah. All right. Um, so this is Andy Konwinski's big project. Um, it's super, super cool. I love Andy.

I don't want to downplay what they're doing because they're using the RLM abstraction, but they're doing something much cooler than the RLM, uh, which is like they, they have a system that kind of what they call like thinks persistently. So even when you don't query it, it has a way to, um, think through problems that it has in its context. Oh, so it's just like a always on type thing. It's like an always on thing, but it's like- Yeah...

not that expensive. Like they, they, they control the, the, the token costs, um- Mm-hmm... to make sure it's not like burning through all your, all your credits. Okay. This is super cool. I'm trying to think. There, there are many-- actually if you go to the RLM, uh, GitHub page, there's a bunch of things I've linked, uh, below. There's a ton of really, really cool kind of things that people have been doing.

Axe is another really cool one that I think it's just by this one guy. It's like a harness around DSPy and RLMs. DSPy also has an RLM. Oh, the last thing I'll shout out is on Arc AGI Three, I believe, there were a lot of harnesses on their, like, Kaggle competition, like the official one, not, not the, like, public Prime Intellect one that-- or, like, what people have evaluated on.

They all, like, claim to use or they reference, like, like Tufa, for example, some form or some inspired form of the RLM abstraction in their harness, which is really, really cool. I think it's, um, this is where-- I mean, this is exactly the setting where you would see a lot of benefits from composition and, and, and using code and combining, like, neurosymbolic systems with, with AI. Yeah.

And so, yeah, very, very cool. We love a good neurosymbolic reference. Yeah. [chuckles] Uh, you also, uh, mentioning RKAGI 3, OpenAI comes out and says, "We're at ninety-nine point nine percent on this." Yep. Uh, they also say, "We, we solved Navier-Stokes. We just throw a model at it." Mm-hmm. Uh, there's some debate around whether or not it's just model. Mm-hmm. Are they using an RLM? Do you know?

I, I mean, I, I, I would guess probably not, unless you say, like... I'll be, I'll be careful here because, you know, people debate what is an RLM, what is not an RLM. It's somewhat clear that what they used is some kind of swarm of agents with a shared, some shared context, like some shared file system.

And, like, this is very much in the spirit of RLM stuff, but I think there's a, there's a lot of, like, you know, more clever things that they did that's not maybe related to, to the RLM itself. I agree a little bit with the idea that, like, a harness is not that necessary for what they did. The way that I would put this is that I think a model like a GPT-6 Astra-type thing is technically smart enough, conditioned on the right information, to come up with a proof for these very, very difficult problems.

Now, how you get to that information is a giant question mark. And in, in their case, it probably came down to, like, a very long search over, like, many of these sub-age-- or many of these, like, agents in the swarm and, and maybe also, like, researchers con- I'm, I'm actually not sure about this part, but putting in, like, their kind of intuition as to, like, what you should explore and things like this.

And ultimately, like, this produced some information that some agent was able to take to finish the proof. And so in that sense, like, I think was the harness that important? No. And I, I think what this is pointing at is, like, the specific details of a harness do not really matter, and I think that's also what that, uh, what the harness tasks paper is, is pointing at, which is that, like, beyond the user's feeling of the harness, realistically all that matters is just, like, how are you composing these agents in a meaningful way to get to the final answer? And maybe that's what, like, swarms and all these things are, are, are really about.

And so from my POV at least, if we, if we start to think about, like, for user use cases, what do we want out of harnesses and, and, and things like this? Like, we want to take the good parts out of these, like the Claude codes, the Codexes, like the stream that people like to see. But, like, under the hood, whatever is running can be some really weird, complicated swarm of agents that, like, ultimately come up with an answer. The user doesn't wanna see that, though, obviously, right?

Like, it's, it's not legible information. And so I-- you know, I-- this was another kind of thing in the spirit of RLMs, like recursive language model. It sounds like it's a language model and-- but it's not a language model architecture. But the reason for this is, like, I think we will start to see in the future probably one day, and I, I wrote a blog about this, what we think of as a language model, like the thing that we query might actually be like a swarm or like a scaffold or like some weird harness design that scales very well, but the user just doesn't see it. Ultimately, all the user sees is some front-end version of this harness.

And yeah, I, I, I, I think it's a relatively safe bet at least to make that this is what we will see. Yeah. Um, and this maybe goes back to the limitations of the base transformer. Like, you know, obviously if you just took a base transformer and you said, like, "Solve Navier-Stokes," or something- Mm-hmm... it's not gonna do it. Like we, we all know this is not what's gonna happen.

But yeah, I think this is maybe the more interesting part, uh, and may-maybe the claims around, like, you know, did the harness matter is, is more around this, of like just arbitrarily pointing models like or agents at a growing kind of context of information maybe is just enough to solve very difficult problems.

Um, that I can buy. Uh, well, there's a lot in there. I do wanna say, uh, the, the, the amount that OpenAI spent is semi-public. It's, uh, ten thousand agents in eighty-eight hours. Oh, yeah. Uh, one hundred and thirty billion to- output tokens, which is estimated to be about forty million dollars in public pricing. Hmm. Surprisingly actually, like less than I thought. Yeah, not that much. Yeah. Yeah.

[chuckles] Thirty billion output for the final, but as you said, there was a lot of context being passed around. Mm-hmm. It's more than double that in just the total agent messages being sent. Yeah. Yeah, yeah. Yeah, yeah. I think, uh, one thing I was honored to bring up also was Cursor as far as, like, swarm stuff is concerned. Hmm. Uh, this is slightly older, uh, meaning February, which is ancient. Whoa.

[chuckles] Uh, but if you scroll down all the way to the, the final sort of multi-agent architecture that they arrived at, it was basically an org chart of a normal software team. One thing I'm w- thinking about-- because I, I basically there's, like, this, um, uh, gather all function- Mm-hmm... that you have to do with, with sub-agents or, you know. It's very similar to GPU programming actually.

[chuckles] Yeah, yeah, yeah. Uh, and, and so, uh, that's a bottleneck. This is a bottleneck. Mm-hmm. If there's one main guy, uh, that's coordinating all the sub guys, then they have to like re-- you know, gather again and then, and then re-coordinate. Mm-hmm. That's slow. That's, that's crappy. Uh, what a tr-true swarm should be is everyone is just their own person. Yeah, yeah. Right? Well, I agree with this, and I, I think that there is a question to be had, though.

Let me give an analogy, which is like, when would you use compaction and when would you use an RLM? And there are many settings where an RLM can solve maybe a more difficult task than compaction can. But you would prefer to use compaction in most cases because it's cheaper and it's quicker. Uh, and I think in, in the, in the context of agent swarms, there is a similar thing going on of like I'm- Fairly certain that, like, 95% of the swarm is entirely useless, or, like, what it's exploring is entirely...

You're just burning tokens. Versus in this setup, maybe not so much. I'm not sure. I mean, maybe it's also the case here, but- Everyone has a job. There's a Jira board and like- Yeah. I think at some level, that's how problems are framed, right? Yes. So like if you have a search problem and you're spanning out a bunch of sub-agents to do search, there's gonna be a lot of useless information, right?

There's one retrieval answer- Mm-hmm... that you're getting and you're spanning off. But that is consciously understood, right? Yes. But there is kind of this question of, like, what is appropriate to solve for which problems? Like, what, what design? You know, in theory, OpenAI can use, can, can, can package up this API and they'll call it swarms, and then they'll give it to you, and they'll be like, "Point this at any problem, and we'll give you a solution." Yeah.

But maybe- It's called, it's called Pro, right? Yeah. May- maybe you'll have to pay, like, $40 million to get a result. [laughs] And it's like, well... But it's exciting. I, I, I will say, like, it is, it's very exciting that we even have the option to point $40 million at a problem and solve it. Yes. Yes.

But there is, there is still kind of, uh, a lot of research to be done in this area around, like, you know, what is necessary. Like, what, what do we want to do? What design do we want? I mean, we probably don't want everything to be a swarm, but, like, where do we draw the line? Like, can the agent design that or decide that for itself? Et cetera, et cetera. So- Yeah.

Uh, and then, uh, just a quick check in case you have opinions on this. Have you looked much into open-endedness as a general category of problems? Mm. Oh, I'll- Meaning no, no prompts, just go. [laughs] A, a little bit. Um, so I, I was at Sakana for a summer, uh, right after, or I guess right before my PhD, and, and that's something that they work on a lot there. And I think there's a lot of people even at, like, Recursive Superintell- there's many of them now. Yes. We just had Richard Sutcliffe on. Oh, yes. Yeah. So, so, like, Richard's company, and then, and then, and then also...

Actually, wait, that might be the s- it might be the same company. I, I, I don't remember. Is, is Tim Rocktäschel also in- Yeah. Yeah. Okay. Yes. That company. He's the main co-founder. He used to be head of open-endedness for Google. Yeah. Yeah. Um, yeah. So I think with open-endedness problems, like, I view them as somewhat similar to even, like, unsolved math problems.

Maybe that's a weird way of, of, of framing it, but, like, I think a lot of the techniques in terms of, like, how people s- like, approach them are kind of the same. Like, evolutionary search is, like, very similar to just launching swarms of agents and hoping that, like, they come up with, like, an interesting s- and this is what, what, like, AlphaEvolve and some of these other works did, uh, like, a year or two ago.

But I think what maybe is, is not, uh... And, and I, I'm not sure if, if this is what you were alluding to, but I think what's not super clear in open-endedness style search problems is do we frame all of them as basically, like, an unsolved, very difficult problem, where like, where the objective is clear?

If that's not the case, I still don't know yet entirely what the value of it is. Uh, maybe you have other opinions. Like, I, I, I don't have too many opinions on this. But at least from my time at Sakana, like, I, I got the sense that, like, we ultimately still kind of wanna approach things the way that, like, say, OpenAI approached Navier-Stokes. We want...

We-- There's still a lot of nudging in certain directions that we want to have to, like, get to the point where we have something interesting. My version of it is, like, um, maybe it's a split between basic science and applied science. Basic science, you're researching for researching's sake. Mm-hmm. You just wanna understand things better.

Uh, I have no idea if, like, there'll be any application at all whatsoever, but, uh, I mean, that, uh... And then applied, you have a goal. Mm-hmm. You're, you're trying to minimize loss in some way, you know? Uh, and so, uh, w- what I really, uh, think, you know, in terms of, like, the, the big bets that people have, uh, what if there was no prompt? Like [laughs] I see. I see. You know what I mean?

You just pick domain and let it- Like, you just, like, you just spawned- You just let it run... in this, like, swarm of things, and you're like, "Hey, what's up, guys? Like, what are you guys working on?" Yeah. And, like, you just decide- I see... like, this is an interesting problem. The biggest issue that, like-- And may- maybe there's a clear path to this, but when I was there at least, the biggest issue that people had with open-endedness was, like, how do you ultimately choose at the end?

How do you pick out the interesting stuff? Because when there is no... Like, maybe the agent comes up with a goal, but in a lot of cases, like, what they... And they have something called Foo, I think, which is like a- Mm... it's like a model router type thing that was, like, inspired at least by this idea of, like, let's pick a problem where maybe we can pick out the best solutions to something.

In this case, it's like pick the best model for this problem. [laughs] Um- You're the first person to connect model routing to open-endedness. [laughs] No, no, yeah. But, eh, so I, I, I bring this up because I think, like, with open-endedness, like, just generally the issue is, like, when we have this giant corpus of, like, slop, like, how do we sift through- Yes...

and find, like, the hidden gems? And, like, the solution to open-endedness really just letting models run forever and, like, finding, like, just doing data gen, just doing super, super, super high throughput data generation and then, like, asking agents to go through and, like, find meaningful things.

Like, uh, I, I'm not sure. Maybe, maybe that's sufficient, you know? Like, uh, [laughs] that would be- It- That'd be cool. To me it's, like, very interesting as a counter to basically all of machine learning- Mm-hmm... where you have a goal, uh, to have no goal. Yeah. Yeah. [laughs] But-- Or, or, like, an ill-defined goal that you're like, "Well, what about, what about this goal?" And you're like, "Well, okay, maybe."

And then you, like, sort of research more, and you find- Yeah... that that is an interesting goal. 'Cause I, I think, like, finding the objective function, like, like you said, like, Jeff found an objective function- Mm-hmm... that was interesting that no one was exploring. Yep. Yeah. I think that is, like, you know, similar to your message about grad students as well. Like, you know, you stay in school because you are, you, you want to pursue open-endedness.

If you want to, you know, uh, profit max and, like, you know, join the, uh, escape the permanent underclass, then you join Lab. [laughs] Right? Yeah. [laughs] Yeah. It's funny. I, I, I feel like I, I don't, I don't hear this discourse a lot. I mean, I'm, I'm in the East Coast, so it's, like, a very different type of...

But then when I, whenever I come here, it's like that's always, like, the topic of discussion. You cannot pay rent without doing this. Yeah. [laughs] Right? You- you're getting priced out, guys. Yeah. [laughs] Okay. So, so yeah, there, there's all that. Um, I don't know if you wanna-- if it's relevant e-enough to talk about the mismanaged geniuses which you were, you were pulling up. Hmm. No, no, it's just on your blogs.

But I will poke on, uh- Oh, Sakana? Yeah. Okay, yeah. So they- Cool, cool, yeah... they did in their, um, blog post, I, I talked to them about this as well. So one of the cool results of Ultra is they basically just let it loose on automated data science research- Hmm... uh, with little to no human intervention. It's just kind of making- Yeah, so this is auto research-...

cool stuff probably... which is a little bit more open-ended, and there's degrees of open-endedness, and I, I agree with that. Yeah, separate than auto research with objective, this is just- Yeah... do stuff. But you know, it- it's cool. They're, they're working on it for those that are interested. Uh, while you're bringing it up actually, uh, what is your take on Sakana? Like, what are they doing- Hmm...

apart from being, you know, the Japan one? Yeah, yeah, yeah. [laughs] Um, I mean, I actually love the people there. Like, I, I think they have a really, really smart team. Uh, and it makes sense. I mean, it branched off from, from, like, an earlier team at GDM, which was also kind of, I guess, doing this kind of, like, open-ended evolutionary research style stuff.

What I liked about my experience there at least was that they did have that, like, mishap back, I forget at this point when, but I think, like- You're talking about AI scientists? Uh, no, the GPU kernel. Oh, okay. That one now. Yes. Yeah, yeah, yeah. People cannot forgive them for that. Yeah. Yes. Yeah, yeah.

And I guess, like, the AI scientists, like, I mean, there's, there's some criticisms of it that I, I mean, I don't, I don't work on that, so I have no, no kind of take on it. But I think in general, like, what I like about them at least is they're, they're a little bit more of a research-y type lab. So, like, they don't operate in the same space as, like, OpenAI or Anthropic. Like, for sure.

Like, definitely no. They do not-- I mean, it's pretty obvious probably that a- at least when I was there, they do not have a big competitor model or something that, like, you know, that, that everyone is, is, is using. But I think they kind of operate in some ways as, like, a PhD lab, which is cool. Like, and, and I, I, I think, like, David Ha is, like-- He's, he's really, really smart.

Like, I think he has a good sense of, like-- Also, I think the market in Japan is also a little bit different for AI, and, like, who they're targeting is, is slightly different than maybe what we're used to here. But yeah, I like that they take kind of-- A lot of their research is kinda weird, I think, when, when people view it, you know? And I, I like that. Like, I think it's- We should have more weirdness, yes.

Exactly, exactly. Uh, and when you say different market, just is it, like, enterprise? Like, the way it works there is, is a bit different. Like, how, how deals happen and, and stuff like that. They do have a-- I mean, I guess this page is originally in Japanese, but they do have a model- Oh... specialized for the Japanese market. You did, you did know that. You did know that? [laughs] I did know. It's new.

Did not know this. I also have personal friends and know the team. Yeah. Uh, so there is a-- E-even from the sense of a way that you speak culturally- Mm-hmm... uh, you know, responses are tuned towards that. This is not like it's frontier on benchmarks. It is a cultural appropriate model for them, and then they have, like, chat and all that. Yeah.

But, uh, to, to mirror your point, you know, there's also, like, uh, how should education look like and someone wants to work on it, and they're very PhD lab of, "Do your thing. Why not? We have money. Go, go research it." Oh, they say it's a Kimi fine tune. That's nice. Oh, there you go. Kimi. Good for them. [laughs] Um, yeah, I mean- Yeah... and, and, you know, speaking of Kimi, right?

Like, another, uh, you know, just a, a grad student that split out and, like, yeah, I just have this, like, Kimi Delta attention at once that I wanna work on. Yeah. Yeah. Uh, and, like, somehow managed to make moonshot. [laughs] No one understand it still. Yeah, yeah. Well, he's, he's super cracked, uh, at least my, my understanding.

I mean, I, I think in general, like, a lot of, a lot of the Chinese labs have, have done really, really cool work. Yeah. Like- Any thoughts on Kimi agent swarms? Yes. Um, one thing I will say is whatever OpenAI is doing with their agent swarm is, like, clearly the right thing to do. You have to kinda think about it this way.

Like, nothing, especially without, like, a very, very smart harness design, which I, I don't, I don't think anyone really has so far. Nothing comes so easily for free. For example, like, the agent swarm design is not something you can just take for granted. Like, it's not like GPT 6 Astra is just super smart, and then it just got agent swarms running well.

They clearly trained-- I mean, the, the Hugging Face incident was them training a system to be like a swarm. Uh, and I, I think, like, clearly they've done something really well to the point where you can throw forty million dollars and solve an unsolved problem. And I, I think with the Kimi agent swarm thing, like, at least from when I read it, it just came off as, like, this is interesting, but I don't actually know whether or not this can solve anything novel.

Yeah, they just they were like, "It makes spreadsheets for you." [laughs] They kinda were just like, "Yeah," like, you know, "Here is a, here is a swarm that, like, kinda does stuff," and it's cool. Um, but-- And I, I will say same thing with dynamic workflows. I actually kind of think the dynamic workflows release was sort of a flop.

I, I don't know how, how you guys feel about it, but I think it's like, uh, my understanding is, like, it's not used that often or, or like- It's just very expensive. It's too expensive and like- It's ultra code. It's basically like, like, you know, take over my bank. And it, it doesn't a- I've tried it, and, like, it doesn't act in the way that, like...

Again, I, I, I'm not, I'm not the biggest OpenAI, like, you know, stan or something, but I think whatever they did was very impressive. Like, they somehow managed to get a way for this, this swarm to actually act- I see, I see, I see... towards, towards a goal. And yeah. I see. It's very difficult. I see.

So, so, like, efficiency of s- of the multi-agent swarm is the objective function here. Yeah. Right? Yeah, yeah, yeah. Like, like how much of this is slop? We- Like, this is a lot of slop. Yeah, I think so. OpenAI is less slop. I think we, we take for granted what it means for a swarm to converge to an answer. Yeah, yeah. Um, it's just like-- It's not, not something we can take for granted. Yeah. Yeah.

Uh, we've, we've done one pod with Noam Brown and, like- Mm. His- Oh, yes. I remember, yes... his, his, his, uh, thing-- His whole thing was like, "Okay," like, "we've w-worked on a lot of, like, competitive agents. Uh, we're working on collaborative agents." Yeah. And, like, that's now called a swarm. Yeah. I would do a quick poke. Do you have any thoughts on Gemini? They were also IMO gold.

[laughs] Like, you know, there, there was a time where they were getting agents to reason for a long time and- Um, yeah, I mean- Is it too old to think about? No, no, no. I, I, I think it's a little bit blown out of proportion. Like, Gemini For, uh, again, I, so I should preface by saying I haven't worked at any of these places, so, so take this with a, with a grain of salt.

Like- I mean, you have strong opinions on agent harnesses- Yeah... and you know, this was- But, um, let me just say this first. This work was really impressive. I think what, what they showed here was like they took a, a time when the models weren't that good- Yes... and they managed to be very, very smart about, like, what the harness does.

I remember for this at least, like, yeah, like alpha geometry. I guess that was a year before this, but it was very cool. They took, um... They took it to the max, and they, like, designed, uh... I don't know. It's, I, I, I'm not too big on competitive math, but I think, like, GDM, it's sort of a shame. Like, everyone I've talked to about GDM kind of has the same opinion, which is that it's way too, like, bureaucratic.

Whatever is g- They have the talent and the resources to do almost anything, but, like, I don't know, until they figure that part out, like... Nothing against Antigravity, for example. [laughs] But like, I don't know anybody that uses Antigravity, and so I've tried it once and it's-- I don't see a reason to switch to it.

Um, and I, I think for whatever reason, like, they've been struggling with this, so. Yeah. Yeah. Well, uh, you know, m- a lot of people dogged on Meta for a long time until they started- Yep, and they recovered... like, coming out, and like, yeah, I think, uh, Google's going through that phase right now. Yeah. Um, and, uh, you know, it's, it's just you gotta stay alive and- Yeah. Yeah, yeah... I wanted to focus back on- Yeah.

Cool. Yep... just like your, uh, your thoughts, just general. Uh, we can talk about speculative PTC- Hmm... uh, Mismatched Geniuses, or just, like, throw away all this and just talk about whatever else. [laughs] Okay. Let's, let's talk about Mismatched Genius for a little bit. Yeah. I, uh, the only comment I'll say on speculative PTC is that it's a really simple idea.

It's almost, like, obvious that this should be done, and, like, there's not much more to talk about it. Like, I think it's just, like, you should just use it for, for, like, coding. Like, anything with programmatic agent calling, like RLMs or Code Act, like yeah, it's like, it's like a no-brainer. What's the... You know, for, for people that haven't read it- Yeah... what's the one-liner for people?

The, the simple thing- I like the-... is when the model is, uh, writing its code or, like, even as, like, after it finishes writing the code, a lot of tools tend to be, like, sequential or, like, you have to wait on them, so you should just launch them in advance. Like, if, if you're able to JIT compile this code, you can probably figure out, like, even though it's, like, kind of variables and stuff, like you, you can figure out, like, how to fix the- Yeah, statically analyze.

Yeah, yeah. So- Speculation. Yeah. There is, there is this, uh... Someone pointed to me, some actually, like academics have, especially PL, like programming languages people, have some, like, very kind of cool ways of doing this, and so, like, you know, at some point maybe I'll, I'll, I'll, like, work on this more.

I guess mo- mostly you have to change language, because if you are in JavaScript Python, you can't do this. Yes. Yeah. Yeah. Yeah. So like, uh, Haskell, yes. Um, um, what's, what's the, uh, what's the normal one that's, that's not Haskell? Lisp. Uh, Lisp, uh, OCaml. OCaml. Oh, yeah, yeah. Any, any functional language- Yeah, yeah, yeah... you can actually, like, pipeline this. Yeah.

Um, so Effect TS if you wanna do TypeScript. Yeah. Okay. We can switch over to, uh- Yeah. I like this diagram... Mismatched Geniuses. Yep. Which, which is like your, you guys' whole thesis, right? Mm-hmm. Like that, uh... I mean, to me this is like kind of like a restatement, but maybe I'm missing something- Yeah...

of, like, well, uh, working better harnesses or, like, y- you know, your, your models actually are capable of a lot more if you try harder, so this is a skill issue. Yeah. Yeah, yeah, basically. [laughs] I, I... I mean, I, I, I think there's... One thing I want to see-- I, I appreciate that there's a big focus on, like, jagged intelligence, because it paints a big picture of, like, we can do this if we really set our minds on it. But I kind of wish...

And, and maybe someone in academia should do this. Like, really just sit down and think about, like, if I took Astra, even the current frontier models are not good enough at, like, doing a particular job over, let's say, the, the span of a month consistently and well. And I think this is, like, a stupid problem. Like I, I, I genuinely think we can solve this. You don't need to be a frontier lab and, like, do all this like, you know, fancy stuff for your IPO.

Like, I think, like, these models are so smart that even if it's, like, a silly way, I think that it genuinely is a skill issue of you can get a model to be as good as, let's say, like, just some 18-year-old high school kid doing some job. I think it's, like, ridiculous that we can't do that, and it's-- part of the reason is, like, the, the, the format of a language model is, is not really amenable to that, but I think you can shape a harness around it and do it, and I think, like, this in itself is, like, ignoring the RLM stuff, ignoring all the, like, you know, what abstractions should we use.

Like, I, I just think someone can design a harness that can do this. Like, I, I... I mean, that's maybe it's- When you say do this, do what? Do long-running but simple tasks, uh, and, and do them reliably. What's an example? So is this different than, like, you know, pick your favorite company, Harvey, for example- Mm-hmm... using LLMs to do legal work?

Or what's the- I mean, I, I, I guess it's kind of like that, except if the bottleneck was not, like, certain legal knowledge or something. Like, I, I don't know. Let's say, uh- I mean, I guess, like, the example is, you know, people can take models and build pipelines or whatever- Mm-hmm... and have an agent repeatedly do whatever task they want, right? Yeah, yeah.

So, so for example, like, if I wanted a general system that I could kind of talk... I can talk to it like I would talk to an intern, and basically just ask it to do, to explore some small thing. So maybe an example of this is, like, very silly auto research is, is maybe an example of this. Of like, not necessarily finding super, super novel solutions, but at least optimizing all of the easy parts of any problem.

They often end up being over-indexed for, like, ML training and, and, and things like that. But yeah, I don't know. May- maybe, maybe that's, that's, that's not, like, super, super clear. But there is a lot of people's general workflows where you probably could just vide code up Some specific harness to help you do, like automate this thing. Some examples are like automating finding like research papers and stuff like that.

But usually people will design like a specialized agent to help them do this kind of thing. Or like sp- like they'll vibe code a harness like and just run it, or like their Slack bot or something. But I almost think there's just like a standard form, like just a harness that you just plug in. Like you don't-- It doesn't need to be designed for finding papers or fi- Like, you just kinda tell it, "Find this for me," and you like plug it into that setting.

What I'm getting at is that I think there's a lot of easy things that can be automated. And- Is, is this like a hypothesis or a point around like capability overhang? Like even if we paused, there's still a lot of impact to be had with current state of models? In some sense, yes. Like I guess what I'm presenting is the easiest form of this, but what this is kind of saying is that like we have jagged intelligence on a lot of things.

Like for example, models are like disproportionately good at coding and math. This is saying like we can translate those abilities to many other things. So like for example, if you took... Usually if you take someone who was an IMO Gold or something, and you kind of apply them to a lot of different problem-solving domains, they can figure it out. I, I don't actually know if this applies to models.

For example, like in GPU code optimization, one very interesting question is whe- if you were to take out all of the GPU programming data, like from a model, but you... it was re- it was like as good as Astra is now, just without, like with that taken out, would it be able to still optimize GPU kernels?

Uh, like would it be able to learn in context roughly what it needs to learn and then like have some pipeline or like, or like come up with some solution to solving like optimization tasks? And I think like there's like a mismatch between like if you took a human that was as smart or like knew as much as Astra, there is a mismatch between what that human can do and what Astra can do, maybe around a harness.

Um, and I think we can actually approximate the human a lot more. To me it sounds very approximate to the continual learning problem. Uh, I think you're making- That's the best example. Yes. [laughs] Yes. Why didn't you just say that? Yeah, yeah. I guess I, I- I was gonna... I was like thinking, should I just blurt out some words like continual learning? I'm, I, I'm, I'm careful with that, but yes.

I, I- You have a very- Maybe, uh-... lightning in his eye. [laughs] Maybe, uh, like a, a bit of a... I don't know, like a- No, so, so I, I think you're, you're a very, uh, uh... I don't know. I don't know your undergrad actually. Are, are you like a math person- A little. Yeah... generally? Or- Yeah. So you're, you're sort of like a- You said you were gonna like study math.

Yeah, that's what I wanted to do at least. Yeah... like a category theory type of, uh, abstraction where you think in categories and then y- you have to like then translate down to the specific. And but then you like actually really care more about the category. Mm-hmm. Uh, and like that, that's the communication error because like everyone's listening for the specific, but actually you're trying to also, you know, convey the, the general. Yeah. Um, which is hard.

[laughs] [laughs] I don't really know. Uh, you, you can, you know, maybe use like a shorthand of like, "Okay, I'm at level two, and then I'm gonna go up- [laughs]... to level three, then come back to level two." That we should have, say, some like epistemic like shorthand for like this kind of thing- Yeah... because, yeah, it, it's hard. Like you're, you're compressing a lot into word after w- like sequential word decoding.

Mm-hmm. Um, should we convert to neuralese? Uh, you know, is there like a, [laughs] Mm-hmm... a better form of output than English or, uh, Python or JavaScript? [laughs] I don't know. This is very s- Yeah. Yeah, I-... kind of like a shit post, but like people have speculated about like what is the native language that people want out, uh, that models want to output. Mm-hmm. Uh, some people say binary.

That's a, that's Mark Andreessen's thing. I don't know. Yeah. [laughs] Um- PTX... let's say a, a mix of English and Python, and I only say this because the capability of a model is somewhat a reflection of what we train them on. So we still want like, like... Yeah, I don't really buy the, the binary argument.

Um- [laughs] I guess I like, I, I understand, but it's like, uh, like- Yeah. You, you wanna- I don't know... you wanna model the world in some way. I think the, you know- Yeah... uh, one thing I'll, I'll bring up is always, which I always do in this kind of conversation is Sapir-Whorf- Mm... uh, which is you, if you choose English, you will have locked into however long English has been around, which is let's say 500 years, uh, which is not that long. Like actually- Mm...

uh, you know, like what you, well, the language that you speak constrains how you think. Mm-hmm. And if you learn a different language, uh, for example, uh, someone, uh, in Chinese we don't have tenses. Um, I don't know if you- I, I- I, I actually didn't know that- Oh, oh... and I speak Chinese. Oh, oh. I, I, I did know that, but my, my Chinese is not great. Okay. Yeah. Yeah.

Or like, uh, in, let's say in Japanese or, uh, I don't know. Uh, I forget what l-language it is. Like in, in Korean, everyone you speak to, yeah, you have to like acknowledge social status. Yeah, yeah, yeah, yeah. That is a different dimension than gender, right? Like it just like, it just influences everything you do.

Um, yeah, I- when, when I take Ling 101, um, apparently there's a, there's a, there's a language in Africa where like there's a vegetable gender. Uh, yeah, right? Like j- just like you have... or like Eskimos know no word for snow or whatever. Like anyway, so like the, the language that you adopt affects your thinking, and if you- Mm...

choose to output your train of thought in English, you are biasing towards whatever English solves. Hmm. I don't know what the, the sort of prior of English is. That's interesting. I, I did not think of it that way. I mean, at some level it's interesting, right? So you're right on language. Uh, a lot of model chain of thought also fluctuates language. Uh, the obvious example is Chinese models- Mm-hmm...

speaking in English might still reason in Chinese. Uh, but at the same level, most models are very capable multilingually- Mm-hmm... and that adaptation we can see. You know, you can add in languages. You don't get that much from adding a whole language, but- Yeah... they'll reason interchange. So- Yeah. So we're all autoregressive. I mean- Yeah. But also like, uh- Like-...

let's say German, like, you know, uh, subject-object, uh, agreement, uh, you have to put the, the verb at the end, which is very super annoying, like very famously. Uh, yeah, right? Like you don't know what you're doing until the end- Mm-hmm... where you're like, "Oh, that, uh, mess of nouns and then the verb."

[laughs] Well, the, the, the most classic one that most people be- have heard of is Arrival, where, um, they have the heptapods where they, they think, uh, that time is like flat to them. So they think in-- they, they, they, they output entire sentences at one shot. Um, so it's, it's as closest to like, uh, the difference between autoregression and diffusion. Mm-hmm.

We talk in autoregression, what if you could talk in diffusion? Mm. Uh, where things just resolve over time. I see. I see. But like the whole idea shows up at once. Ah. [laughs] Ah. I see. So that is a drastically different lang- uh, language, but it is a language. I see. Oh, that's really interesting. That's really, really interesting.

Uh, whi- which like machines could speak, like, that we, you know, probably will never speak, but like yeah, machines don't care. Maybe this is a huge tangent- Yeah... but are there, are there not like things inherently that are reasoning chains that are inherently autoregressive in, in some sense? Yeah, time. Like, yeah. Or like, yeah, yeah, yeah. Yeah.

Um- Something happens first, then something else happens. Even like, yeah, anything in code, for example, like has to be causal in, in, in some... Usually at least has to be causal. [laughs] Um- Well, no. Uh, so, so then you have to-- then you're not exploring enough, uh- That's true. Yeah, yeah, yeah...

function, uh, program and language theory, uh, where, uh, everything is like pure functional and like completely relational- Mm... and, and, um, you sort of abstract away the solver that translates the relationships that is, are always true into code. So I, I, yeah, I, I feel like this is maybe a, a little bit- [laughs]... too out of my depth.

No, it's interesting though. Yeah. Uh, but I love languages, uh- Mm... whether it's coding or human, and I do think a lot about how that affects reasoning and, and, uh, the boundaries of what we can do. I don't need to go too much beyond that. I don't know if you have any other thoughts. Uh, my closing question was gonna be, uh, you have all these research, uh, directions that you wanna do.

Your, you, you know, you had a GPU mode phase. You had a, uh, RLMs phase. Mm-hmm. Uh, presumably you have other stuff planned, which is why you're not, um, dou- doubling down on that. By the way, I notice that it is interesting how you guys do start with the GPU side, and then you migrate towards the zero gradient side. It's, which is what- Mm-hmm... uh, Shenyu called it.

Doesn't that feel less legit than messing with GPUs? Yeah, I guess in, in the sense that like, so you did bring up that like I, I like to think about things in like a math-oriented way. Category, yeah. And it's like very uncomfortable sometimes to be working on like harnesses and agents because it's so- Because you think all harnesses are the same. Yeah, it's just, it's like super fuzzy.

And so like, it's just like two new ideas and harnesses. Got it. Yeah, it's, it's, it's- [laughs] Like- It's also just like, like empirically it's hard to like verify a lot of findings, at least with, with the compute that we have available to us. But the reason why I think I've moved on to a lot of these problems is I think actually this is where most of like the innovation is yet to happen. To me, like the GPU level is a means to exploring other ideas.

Like you, you, you, you want to, for example, like get good at writing kernels or like even automate writing kernels for the sake of a broader goal of like I want to explore ideas where I'm not bottlenecked by systems challenges. In that sense, like, you know, I guess a lot of what's written there is all harness stuff, but I am also interested in things at the model level as well. Um, but I'll just leave it at that. Okay. Yeah.

That's a good hint. Um, anything, any-- if people wanna reach out to you, uh, what are you looking for help on? What do you want collaborators on? Um, any sort of calls to action? Yeah. So, um, I guess there's nothing I have in particular where I feel like I need to work with someone on, unless it's like, unless it's with a company before like compute or like with, to talk with other people about it.

But I will say I, I'm not-- I'm never opposed to working on ideas with other people. I get reached out to a lot by often undergrads or, or even like other students. Podcasters. [laughs] Podcasters. Um, and usually I feel like I get an email that's something along the lines of like, "I really like RLMs," like, "I wanna work together." Uh, and I, I feel like I- Yeah, that's a bad reach-out, right? Yes, yeah.

The worst is like, "Can I pick your brain?" Yeah. And like, "On, on what?" Like, "Read my paper, dude." Like [laughs] They'll like, they'll be like, "I read your paper," in quotes like recursive language models or like Prime agent, like a self-improving RLM harness or something. Uh, and it's kinda like I, I, I like-- I really like people that are opinionated, even if we disagree.

I think if you have strong opinions and are, are able to like think through why you think those opinions are, are right or wrong, 'cause usually it's, it's hard to, to, to, to actually tell. But like you have strong convictions about certain problems, like I'm, I'm always happy to like chat and even like potentially work on something together. I have like no limit to who or like what I would like to work on.

So, um, yeah. No limit. Yeah, yeah, yeah. I, I, you know, in, in the era of agents, I think there's a lot more work you can do, you know? Like, like bandwidth-wise. So I, yeah. I, I think in general, like I am not hard to impress, but I think it just takes a little bit of effort to- Yeah... kind of, yeah, know- No... know what you want. It's very clear.

I mean, a-and, and like when you see a new thing come out, well executed, good, simple idea, uh, then, then like get the- Yeah... immediately gets your attention, right? I get excited, yep. And it's actually like not that hard to get the, the same attention at all the Frontier Lab guys. Mm-hmm. Yeah. Uh, because they are looking for you. Uh, you just have to put yourself out there, right? Exactly, yeah.

But yeah, it's true. I, I, I will say I think, I think human attention very scar-scarce right now, and I do struggle with like the number of projects I have going on. Mm. And, uh, I don't know how to manage it. I don't think agents are helping at all. [laughs] Like, I would just prompt it and I'll prompt the thing and then never look at it. Mm. Right? Like, which is very common. Yeah. Yeah, and that sucks.

I guess, uh, maybe the, one of the smaller differences in like-- I mean, actually, maybe you were doing research, I'm not sure. But for, for me at least, like I, I'll have maybe like 10 or 15 different ideas that I wanna do, but the thing is like most of them are bad. Yeah. And, and, and this also maybe is true even for someone that reaches out to me. Like maybe the idea is actually bad, but it, it looks interesting to me.

And so like we can spend like some time looking into it, and if like we feel like there's actually something there, like then we, we should, we should take the next few weeks and just really pursue it. And like this is my, my style with-- this is why I love the PhD, by the way, because there are times when I'm just thinking about problems like maybe on a run or just like playing tennis or something.

Like I'm not, I'm not working, I guess, but it's like those are the most fun times. And then when I like really am convicted about something, I'll just like drop everything and just do it. Mm, mm-hmm. Like, just spend like all my time thinking and working on that problem. And then I, you know, once you get to the point where like you can just run experiments, uh, then it's, it's, it's kind of easy coasting again, so- Yeah, sorry. This is- Yeah, go for it...

like the, the f- the fourth last question. [laughs] Which is like, uh, I, I think a lot of people are also thinking about science as the next frontier, like, uh, like physical sciences- Mm... bio, math even. Um, how do you separate, like, I guess, let's say, your choice of projects that is applicable for industry- Mm-hmm... and then maybe s- part- p- choice, choice project that's just, like, science?

I actually worked on, like, AI for bio stuff before I started my PhD. The field has changed a lot since then, I, I- Yeah... I should say. 'Cause it's, it's be- like, it used to be a theoretical like, "Of course, what, what do you mean? Like, I, you know, I, I have one path, and then I chose that." Yeah. But now a lot of people are crossing over.

Yeah, yeah. And like, so we have started a science pod to just- Mm... cover those things. Oh, wow. Because a lot of engineers are like, "Well, actually, there's, uh, tractable problems there." Yeah. You know, I will preface by saying my, my understanding of a lot of these topics is probably pretty limited, but I think, like, if I find out, either because someone reaches out or, like, I look at a problem and I'm like, "Hey, like, some design principles that we use or that we're thinking about right now actually make a lot of sense for this problem," I get excited about those as well, but I think, I think it's harder.

I, I, I don't know. I think with... I think science, especially, like, empirical or, like, a- applied science, has very, very long, like... What is it called? Like, feedback loops or whatever. Um, like- Yeah, it converges to a robotics question. Yeah, yeah. I mean, to me, like, also this aspect of, like, what is worth spending and betting my time on now?

Because, like, maybe I spend a lot of time on this problem, and then, like, in six months, like, a different solution kinda- like, maybe a new model comes out and it's like, "Oh, it's way better for this." And so I do have to be careful. I, like, you know, you have to be conscious about, like, where you think things might be going. Yeah, same, same. Um, so yeah. Publish, publish cycle. Yeah, yeah, yeah.

So- And I can just, oh, RKJ-3, it's saturated. We, we did it. [laughs] Yeah, I mean, RKJ-3 got saturated in less than a year, so it's kinda- [laughs]... you know? Like, it's... I, I don't know. Like, if, if, if you were a lab picking that problem, like, you're probably kind of sad now, 'cause- Yeah, well, so exactly. That's why knowledge work, gaming, all these things are saturated. Yeah.

Now the, actually the frontier is science. Yeah. And like- Knowledge work is saturated. [laughs] Yeah. GDP val is, like, 80-something, 90-something. [laughs] Um, I mean, like, you know, there's, there's 90 to 100% that's obviously gonna g- gonna take the next- It's gonna be really hard. Yeah... 10 years, but, like, uh, well, the next low-hanging fruit is gonna be- It makes sense... like, you know, the other stuff. Yeah, yeah, yeah. Yeah. Maybe I'll think about that more. I actually...

I, I haven't given too much thought to- I, I'm just trying to guess your next direction, actually. [laughs] No, no. I, I, I will say because I think, I think especially at MIT, like, it's, it's, it... y- there's a lot of really talented scientists there, like, in, in the natural sciences, and I think it's, it's a little bit, like, sacrilegious almost to be like, "I'm gonna figure out, like, your problem."

Like- No, that's how, that's how it's done It's, it's great. Oh, no, I know. Yeah, yeah. So I interviewed Yi Tay who did the IMO thing. He- Oh, yes. Oh, yeah... he's like, "I've never- Yeah, yeah, the back-... I've never been to IM- IMO. I don't even know what it is." Yes. [laughs] It's just a scale model, dude. [laughs] [laughs] Yeah, yeah, yeah. Uh, which is, like, very disrespectful, but like, whatever.

But that's why I- [clears throat] But at some point, like, you have to respect. Like, okay, you know, the, the progress is being made. Mm-hmm. Like, number is getting output, right? True. Yeah. Yeah, true. I mean, that, that is a very good lesson. I... It's very interesting, like, 'cause the mathematicians are responding this way to narrative stuff right now. Right, yeah. Like, Ter- Turing's tower is like, "No."

Turing's tower. Like, "Let's not, let's not use AI." Yeah. I'm like, "Mm, I don't know." [laughs] Well, yeah. I mean, I, I think that, that, uh, th- that whole thing is kinda weird, 'cause I feel like, I feel like they would've had a stronger case if a lot of them didn't work with OpenAI before, like, all this happened.

No, that's ad hominem. Uh, and they, they're really trying to stay away from that. So what, right? [laughs] So what? Like, I, I don't know. Like, like, so what? They got... You know, they, they've, they've collaborated. I collaborate with people that I don't- I guess that's true, yeah... I don't agree with or- That's true... whatever. Yeah, yeah. That's true. Or like, I did a thing, and then now I, now I regret that. I changed my mind.

You know, whatever, right? Yeah, yeah. So I'll defend their right to say that, but like, uh, yeah, a lot of people are reasonably disagreeing with, uh, with them. Mm, yeah. Yeah. Um, okay, cool. Awesome. Uh, thanks for, uh, joining us. Congrats on, uh, your s- your success so far. I can't believe you're still not done with your PhD. [laughs] Well, it's year, year two, you know. Yeah, yeah.

I can't believe we did this podcast without going through the RL paper. [laughs] Uh, actually, he, he- Give the definition. I think it's, I think the paper is more about, like, empirical results. Like, the actual idea is quite simple. Yeah. Yeah. And you've talked about it many times. Yeah, yeah, yeah, at this point. I think there's, there's more interesting things to, to, to look over now, so. Cool. Yeah.

Well, we're excited- Field moves fast... to see what you do next. Thank you so much. [upbeat music]

Academia is for Ambition — Alex Zhang, MIT

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin