Alex Jung访谈：RLM架构、GPU算子自动化与科研方法论

Host: Oh, I just love it. Claude Code. I love Codex. I love Pi. No, I am. I love Oh My Pi. I love Prime. Fair saying, I think that they are all the same. You know, your models actually capable of more if you make an effort. So it's a skills issue.

Alex Jung: Yes. Yes. In the main, I think there is one thing, which I want to see. I appreciate what is great focus on intelligence, because it gives the big picture that we can do this to do if really let's set ourselves goal. But I would like to, and, maybe someone in academic circles should have done it. Really easy to sit down and think: if I took Astra, even modern advanced the model is not enough good to perform a certain work, let's say, during the month, consistently and qualitatively. That's, uh, 10,000 agents for 88 hours. Hmm, 130 billions of public tokens that estimated at approximately 40 million dollars at public prices. Maybe you will have to pay about 40 million dollars to get result. And this is kind of, well, but it's exciting.

Host: I will say this: it is very It's amazing what we have. Is there a possibility at all direct 40 million dollars for problem solving.

Alex Jung: Yes.

Host: Before we move on until today episode, I have small message for listeners. Thank you. We could not offer you content about AI engineering, science and entertainment, which you are so I wish you would too haven't decided press and connect to our content. To us almost every day sponsors are applying. But, fortunately, enough of you subscribe to us, to do all this sustainable without advertising, and we want it to be so remained in the future. But I have one thing. A request to all of you. The most powerful and completely free thing you can do—this press the button "Sign up". It the only thing I ever... I'll ask you something. AND this means absolutely everything for me and mine team, which is works hard to bring every week these materials to you. If you do this, I promise that we will never we will not stop to work on, to make the show even better better.

Host: Now, let's get down to business. So, we're here in the studio with Alex Jung. Uh-huh probably the most known thanks to RLM, e- uh, but you have more several other connections. Welcome to show.

Alex Jung: Yes, thank you for that invited.

Host: Yes, probably also GPU mode.

Alex Jung: Yes, GPU mode too.

Host: Well, I brought you here. Mark Sarafin. Not everyone gets one reception.

Alex Jung: Yes. Aha. Yes, I am very close to everyone people in GPU mode. So we we often work together in different roles, even outside the borders GPU mode itself. Therefore, yes.

Host: Can we explain this for those who doesn't know? This is Discord server. He used to was focused on CUDA mode, and then a little expanded. Its created by Mark.

Alex Jung: Yes. As to me, it's like personnel reserve for PyTorch commands, and then you left PyTorch. Yes. Yes. Yes. Yes. So, I think he appeared when I was in colleges, sometime in 2023 year. I think it's him started by Mark Andreas and Jeremy Howard. Primary the idea was that that it was a Discord server for training writing GPU cores, where gave lectures. This is it there was almost everything. And I interested in this because wrote GPU cores. It it happened cleanly accidentally. I was then an intern at Snapchat and Rexus.

Host: Yes.

Alex Jung: I was very bored with Rexus. So they have there was a project for which they wanted something write. It was article entitled "Infinite Attention". It seems this was an article from Google.

Host: Yes. We discussed her in our Paper Club.

Alex Jung: Yes. Yes. So me I was wondering if it was possible write for this specialized kernels on Snapchat. Nothing from this. It didn't work out, but I joined GPU mode in that time. He was called CUDA mode. I think from legal reasons they changed name, but I met Mark, Mate and many other people, which were very involved in the community. And then Mark proposed the idea called Popcorn, which now known as leaderboard. General the idea was that that we all had intuitive understanding: GPU programming very similar to sports programming if you were doing it. It not that I'm saying that this is quite transferable skills. You do a little coding-golf.

Host: Yes. Yes. Yes.

Alex Jung: AND actually exists surprisingly limited set of optimizations, which use people. And not really so many nuclei, what people want optimize. And we thought: what if we were enough data, similar to Codeforces, where there are millions tasks? If possible was to do it with code for GPU, you can would be to scale and automate development of GPU cores, and this for researchers a huge step. For, I think one of the biggest problems—for example, let's take Mamba: they publish article together with cores, because otherwise you you will not be able to effectively to use them, do you understand? And not in everyone on the team has Tri. So we are very interesting. Kernel Bench emerged from the same ideas: can we to force LLM automate code for GPU cores? I was very fun over this work. It was between college and postgraduate studies, and I with pleasure indeed was doing things in GPU Mode. Now I sometimes just helping with lectures. I'm not like that anymore involved, and I think that we are spending it wrong many competitions, like before, but I still I keep in touch with all participants.

Host: Or do you have a friendly competition? Because, as me, previous community that was doing this, on like MLPerf... Do you have friendly competition? This is just new MLPerf generation or what is happening?

Alex Jung: The beauty of GPU Mode because it's a community: many lectures are enough available for beginners to bet questions, but competitions there—not that secondary, but you you can take them participation for the sake of teaching. I think so many benchmarks, such as MLPerf, mostly take only serious participants laboratories and companies, at least on a serious level. At least that's how I understood. Maybe I I'm wrong, but except for the GPU Mode, it is very exciting that what has just appeared much more sites and people who are engaged in organization of such competitions. For example, it seems there is this one. It seems there is such a site like LeetGPU or something like that, it's like LeetCode for GPU tasks. There is also others we have seen, they appeared, about they were talking in GPU Mode, and this is very exciting because earlier programming for GPU was something extremely highly specialized. The only reason why I am interested in this, is that Tri performed with report in Princeton, because he applied for teaching position—that is, now he is already teaches there, but I listened to him flash attention report in 2023 and thought: "Wow, this is the coolest thing in the world." Um, me too I thought that was it, over which everyone has work. It seems VLM and other things came out, and we're like, "Oh, we must write kernels", but now everyone they write kernels, it's... me I think that in a certain in a sense, this industry is almost oversaturated.

Host: There are some interesting tips for those who want this to do? I think, one of the loudest news—GPT 5.6 wrote more efficient cores, therefore Terra and Luna can become 80% cheaper.

Alex Jung: Yes. And we also saw other competitions where people set records, claiming, what is used a certain cycle auto research, although these are people without experience writing kernels.

Host: Rightly.

Alex Jung: Yes. So even in GPU mode leaderboards, if you look at many recent tasks, almost all solutions generated AI. However, you will notice in the leaderboard a guy named Gorst, who is very, very constant GPU mode participant. We have known for a long time that he extremely cool GPU core developer. We found out one thing: in this leaderboard he too used AI for help with solutions, but mostly he directed the work clues in a certain direction. We found that that its core was practically the only one in top 10, which is actually worked stably in real complex systems. This raises the question, because GPU cores have problem with verification. We are it as if they knew. It

Speaker A: problem since kernel bench output. There a lot is happening system hacks rewards. But you... Yes. You too you notice that the lines code much, much Less. Is this noticeable? simply...

Speaker B: Yes. No, it is definitely very important. And I I think it's very interesting that there is still a big the advantage is that to be in this professional.

Speaker A: Of course, there is.

Speaker B: Yes. I I think so too concerns many AI systems too, even with the last ones mathematical proofs, etc., this not necessarily means that became mathematicians outdated. I have on meaning that these companies still hiring mathematicians, or for data markup, or even just for direction of models in solving problems, so in the deep knowledge in these areas there is still a big advantage.

Speaker A: Therefore, is it just knowledge or it is also about better planning, and is there any new one planning approach, what works more effective?

Speaker B: I think that this combination, perhaps, this is what you have on under the intuition about how solve the problem. For example, I always I draw diagrams of my code.

Speaker A: Yes. That's right.

Speaker B: And if in this there is a part in the diagram that I don't understand, I'm working. over her, until I'll figure it out. Otherwise it is unacceptable.

Speaker A: Yes. Yes.

Speaker B: So, I think, it is a combination of these things: people who know how to watch to these problems and how to solve them to decide, also know how use AI for this, because you come across as very strong verifier, if you own knowledge or you know, What to do? And I too I think we found out from all these swarms agents and the like things, it's that when you direct enough computational capacities on problem, you can enough to explore ways to solve it. But it often happens that what can you do to spend a hundred billions or trillion tokens on something, but if you involve someone who understood on this problem, he can to suggest to the model that what will do unnecessary expense this trillion tokens. I have on mind you, not quite it is clear which ones There are trends here. But I I think that now there are so many problems that we we want to decide, and we we cannot allow just always to direct at them as much as possible computational resources. So the aspect effectiveness in all these things remains extremely important .

Speaker A: Is there any theoretically correct the answer that can be just calculate based on physics, and then approach this physical limit?

Speaker B: Yes. For GPU cores this is not always actually easy to calculate, according to the complexity of the task. For example, for matrix multiplication very easy to calculate this "speed" assessment light", that is extremely possible kernel performance. And this also suggests, you know, maybe that all your data are initially located on the central processor, or, maybe they are in DRAM on video card, etc. It changes these a little indicators. data transfer and everything something else.

Speaker A: Yes.

Speaker B: But I will say that no It is quite clear whether in many cases is it possible at all achieve this theoretical indicator if you You know what I mean. This suggests perfect overlap and data transmission, and, maybe there is some " bottleneck, which impossible to avoid, but often the cores that we we write, even not close approaching this indicator, so that it was somehow significantly.

Speaker A: Yes. And does it matter? just speed? Or worries you too, of course, memory, which does it affect speed? Do you care about energy consumption? One of ours best podcasts was with Jeff for a year Dean, who said: " Actually, I just I am tracking microjoules or, maybe nanojoules, picojoules." Yes, pico, this is optional, pico. Are you concerned about this?

Speaker B: I probably don't, I don't so... But here it all boils down to to speed, right? Nobody counts joules. . However, there are caveats. : I believe that there is speed in in the context of a separate cores and speed in in the context of the broader tasks, for example, models "from A to Z". Because one thing is worth to take into account, and precisely therefore important to discuss what means "speed" light", consists in because for such we always have let's start with what everything is there, for example, in HBM, right? But you can to imagine a whole model "from A to Z". Maybe between the two in layers you want sacrifice speed of the first operations to leave data in cache for the second operation. And these are the things that impossible to fully to comprehend, considering only such nuclei in isolation.

Speaker A: And people call it that the problem of merger operations.

Speaker B: Yes. Or " meganuclei." Usually this is only applies to cases when you limited by bandwidth memory ability. But the question arises: in as these models are improving, or we just need create such " meganuclear"? Is this what What do we want? Ahem. What is your opinion?

Speaker A: I I think that's very difficult because for This requires data. And it seems to me that I haven't seen it yet real conditions example of how we " loading » ability to decide very difficult class tasks without any examples.

Speaker B: Um, me too I think there is another reason, why can't it be so interesting, is in because at the level a separate nucleus: a) they not so complicated, and b) you to some extent sure that in one not so many cores structures. But in case with megacore i I would rather believe that would be better suited here compiler. For example, some compiler for high-level operations makes sense, because in general I I think meganuclear very good consist of individual parts— individual nuclei. There are certain areas where you, maybe you want make strange mergers and so on, but in general I I think these are cases, with which the compiler, will probably cope, and there is a company that this works, as far as I understand, and she also conducted several reports about GPU mode.

Speaker A: Yes, I actually want to group all GPU mode discussion here, because, obviously, there is other parts to which we need to move. I think it's worth something. recommend. You spend a lot really good lectures. They are all on YouTube. People can follow. And you lead a significant part of them. You still enough

Speaker B: I used to, uh, sometimes I still do it. Uh, me I think it's mostly this. Mark. Mark is the one who usually them conducts. Mate sometimes also. But, uh, yes, I have them. I highly recommend it. It extremely good resource. I think this is just crazy like there are many people there share information . So yes, because if you are there, then you are very, very, you exactly the same audience, you know?

Speaker A: Yes, exactly. This is not will reach the main one, and there there are also many introductory materials, which we have laid out, that, as for me, useful for people. Kernel Bench was quite influential. I just I want to see, well, you know, it was last year. Uh, about what other current Do you want a job? remember what people it is worth paying attention to, because you are obviously involved in this field.

Speaker B: Yes, I might. I will tell you. the background of what I really am involved in a lot benchmarks, or was before, perhaps before my graduate school. It it started because I was at Princeton, I worked there with by the Sweet Bench team. John Carlos is here. They everyone is wonderful. I them I love it.

Speaker A: Yes. There is basically this, I I think people don't understand how much benchmarks come from from the same group.

Speaker B: Yes. In Princeton Do you know Shun?

Speaker A: Yes. Aha. Yes. He was already with us on podcasts.

Speaker B: Now he manages a company 10.

Speaker A: Aha. Now he a real star.

Speaker B: I met him, when he was a scientist my boss Michael Tang's friend, who is currently working in Anthropic, they are many cooperated. We there were two students in Karik's laboratories. AND then later others joined in. But yeah, Shinu is awesome. I didn't know he was like that. star, until he left from there. But yes, I mean that there are very few of these graduate students like you, you next, once a year we see someone whose the dissertation was accurate from the start and to the end. There are not that

Guest: many of them. Shil was one of them, and, you know, Jack Morris also. We talked. before writing about research taste, you know, some it's easy for graduate students incredibly lucky with career, they have everything it turns out, this relevant and important, and in others—nothing. I think this applies to even people in industrial laboratories. Simply graduate students are much more noticeable. Therefore you You see, there are people, who are truly lucky, or is it a mixture of luck, high intelligence and other things. I think regarding research taste, he develops thanks to opportunities, at least in mine case. I really like it. good luck with mine by: work in Princeton, and then meeting with mine scientific supervisor Omar at MIT. He fantastic adviser. I will say it anyway, which are the most successful research postgraduate students or in academic environment appear then, when people focus on problems that most of industry does not pay attention.

Guest: I think the problem is that many graduate students are working on that looks attractive for industrial laboratories. For example, they are working on something benchmark. Now this very popular. I think , benchmarks in 2023 were completely different history than now. Many people are working on testing systems , meta-systems and specialized tools for performing certain tasks. And if it's good think, reason, why is someone over this works, consists of the presence of a clear goals around models, which we have today : they say, this is what I want see. I will show you how example article about RLM, that is, recursive language model, because I think it's her idea. extremely simple. I think when she came out, there were a lot people who, upon seeing something like that, asked: what the whole purpose of this all? Yes. Why, it's simple some subagents or something like that, right? And I I think that's it. reaction—it's even a good sign, because it becomes it is clear that people still don't think about the real purpose of this.

Guest: I'll give you another one. example—Swe-bench. When Swe-bench appeared, and OIR likes to tell stories this story, nobody knows not interested. All they said that this an impossible task. Like, why are we we have at all to consider it as benchmark? And only after Devin came out, everyone like: "Wow, this is what why do we want to succeed." And I I think that's true, because you often see many ideas; my favorite example— Eric's work Zelikman, STaR and Quiet-STaR. When you read this article, at least when I first saw her read it, I thought: " Isn't it obvious? idea?" Or maybe not, I I don't know. I had on mind: "Oh, it seems very simple", or, let's say a chain thoughts, the same, or ReAct ; Well, of course, yes. But when are you really thinking, The question arises: in why the value of this articles? I think she is. is that she tells a certain story about how do you think it should be this area looks like. AND this is very difficult to do in academic circles, because if you look at all these articles—Quiet-STaR, ReAct, RLM, Swe-bench—none of them they are not a release GPT-6 level, you know? This is not the case, when everyone says: "Oh God, I will do it now. use". It the best thing in the world. similar, academic the environment just doesn't can do this allow, at least now.

Guest: I mean, there are a whole bunch of reasons, why this should change , but, as far as I'm concerned, if you are a graduate student, you are you in unique a position where you can to work on something whatever, mostly. If you don't use this opportunity and working on things, who don't care about anyone are interested in, or which people think something trivial, on like, "Oh, I was thinking about this, but I don't I just I think that in result research never will not be like that interesting because you you have to do big things rates if stay in academy, because otherwise, I I think just go to industrial a laboratory where there is a lot of resources and talents. Why to limit oneself in industries where you have little resources and not even so many people around, and I think that it all comes down to big bets. Ahem, you just have to do big bets, and many of them will fail, Do you understand? It seems like this just...this...this naturally. But I think what is this, how about for graduate student, this is yours biggest advantage before anyone else in another laboratory, because you don't need to to deal with bureaucracy and everything to others.

Host: Rightly. Yes. I asked this question. to many people, and usually they just are brushed aside. So I I appreciate what you give. so thoughtful The answer, they say, is no, This is your dishonesty. advantage, because everything another, in fact, opposed you.

Guest: Yes. Yes. That's right. AND so, honestly, I I'll give Jeb as an example, because it's not...it's not academic project. I I want to remember this, because this also happened with RLM and many others robots when things become overvalued , to a certain extent, something gets excessive hype and then people they say: "Why is this so?" overrated?" as if it trivial, this pointless. And I saw the same with Jev, because I it seems that the release was like this...I have to mind, there is this whole story about academics, or they not an academic group, but people have to engage in branding and somehow promote your research, so I understand. But I think that there was a lot talk about what Jev— this is something we know for years, and I believe that this misses the point, why is there such a system so interesting. Why is it not easy? some stupid MLP- the classifier that we did, you know, more in introductory courses with machine learning or something like that. I think What's interesting about Jev is that what he violates question: are they correct language models? In that in the form in which they exist, can we consider another design space , except for "text-to-text" "?

Guest: They are actually they say: "I I will use this. the basis of language models; I know that it contains a lot information about the language, but I'm going change output model space to to offer you compromise: I will to perform very quick conclusion, if you have any specific a priori knowledge about this task." Let's say I need to do binary only classification, will I I ask for my language model to do this and pay 400 times more? No, that's it. It's ridiculous, right? And I I think that during for a long time, because laboratories are the only places, who have control, you will never be use nothing other than GPT, GP6 or Fable, because it is the best models. But because of this people are used to it to the idea that language model—it's simple autoregressive decoder. We are it accepted. And I think, when did RLM appear, it was the same. One of comments, very frequent criticism that I received, was like this : "This is not a language model" or "When I look on this, I thought it was new architecture, but "Actually, no." And mine The answer to this is: well, the language model is just modeling languages; she is not must be this transformer decoder, you know? And JEB is very interesting in that, that now we have a new one parameter for settings: what is it source space and how this affects delay in conclusion? AND I think we we can actually begin use different parts of itself language model. It similar idea, around which was a lot attention, they say, this a silly idea, Why would anyone need this? necessary? But this a simple idea that opens a whole series new questions to which, I think especially if you are a graduate student, you you want to answer, because I think that we, for example, many What we don't know about Jev. We don't know how. far we can come in with this, uh, for cyclical transformers. We we also don't know how far we can do it to start: what if cyclically use only part of the model? Hmm, what , if directed data, for example, through the router to different parts models? Is it possible? to reproduce what you do in a harness (

Speaker A: ...harness), inside the models? What is possible? Can we combine between choices, bindings, and architecture model selection? These are all questions that, in my opinion, are opening up thanks to such systems. And this is actually very fascinating.

Jev, in particular—when I saw it, I thought: this is really, very useful for RLM. I think it's logical, because the biggest bottleneck in RLM, swarms, or similar systems is their latency and tardiness. When you are constantly making numerous calls to language models, you cannot properly distribute computation. There might be something trivial that you want to entrust to a simple model, but you can't, because your language model is just this bulky, monolithic thing, you know?

So I am very excited. I think we will begin to see the emergence of new types of models, alongside the ordinary standard frontier models. And with these new models, compromises can be made to achieve so many things.

Speaker A: Ahem. I also want to note Thinky with their model interactions.

Speaker B: Yes. Yes. Another great example.

Speaker A: Yes. So, essentially, an attempt to go beyond the paradigms of "sequence-in-sequence" decoder-only architectures and just do something else.

Speaker B: Yes. I mean that if you try to compete directly, no one can compete with an advanced laboratory creating autoregressive decoders with that sheer amount of computing resources. Even Thinky can't—well, you know, maybe they are one of the few who could, but realistically you can't win in that regard.

I don't know too much about Thinky and I don't want to make assumptions, but if their strategy were simply copying OpenAI or Anthropic, that would be a terrible strategy. Because, well, they just don't have that scale. It's worth thinking about this from the perspective of advantage: Do you have any structural advantage? If you use the exact same approach—which I'm sure they won't, but if someone chooses that same path—you are essentially competing in areas you cannot control, namely data scale and compute, and in those areas they obviously cannot compete with frontier labs.

So yes, it makes total sense that if you are a neo-laboratory... although I don't know whether to classify them as a neo-laboratory, but they are certainly on the radar. They released Inkling, so they are taken into account. Anyone other than OpenAI, Anthropic, and possibly Meta and Google DeepMind simply has to do something different. It's a harsh reality, but I think it's healthy. I am very glad that this dynamic works, and these companies will continue doing this because it paves the way for new players if they find something genuinely interesting.

Because I have serious doubts that this kind of exploration is happening at Frontier Labs—why would they need to? Why take the risk and allocate massive compute to speculative bets when the established recipe is already paying off? And that dynamic is precisely what spawns so many neo-laboratories, right? If you are on a secondary bet and lack brute-force compute, you pivot: "Okay, I'll explore elsewhere." And your blueprint for potential success is something like Jev, which is 100 times cheaper. Maybe it's a different form factor than a standard language model, but in this domain, it represents something distinctly novel.

Speaker A: Yes. Yes. Yes. Yes. Um... So, no definitive assumptions yet about what Jev really is, but I think I have my guesses as to what it could be. I saw some people saying, "Oh, that's something like diffusion." Um, personally, I think it is parallel decoding.

Speaker B: Yes, parallel decoding.

Speaker A: Fair to say, regardless of what it really is under the hood—and I have seen open replications of it—what genuinely fascinates me about what they built is something I'm still not entirely certain of: What was their exact optimization objective, and how did they train it?

And I think this parallels RLM, where we also initially thought: okay, RLM is conceptually a very straightforward idea. Once the paper is out, anyone can use it. But the real value and moat of RLM is whether you can properly train the system, and whether you can build the surrounding architecture to make it consistently high quality. That is what I'm actively working on right now. But whatever method they found to train their system, it is clearly a non-trivial feat. Honestly, I don't even know how they achieved it.

I saw comparisons online where some claimed they just took Qwen or retrained it. But any vanilla open-source Qwen performs noticeably worse, which proves that whatever custom training pipeline they created works exceptionally well. And that is truly exciting.

There is also an element of calibration, which is quite a rare topic that, in my experience, many people don't fully understand or recognize. We discussed this with Clémentine Fourrier from Hugging Face, who previously led evaluation there. The issue is that modern models are tuned strictly to generate the most probable next token; but they will effectively lie to you if you ask them: "How confident are you?" They will simply output the most plausible-sounding answer rather than being genuinely calibrated—such as admitting: "No, I am only 50% sure, or 20% sure."

I would say it is fairly feasible to generate synthetic data for this, because you have ground-truth access: you can synthetically generate heaps of candidate answers, have the model classify them, and compare those calibration metrics against real ground truth. That would be my approach to reverse-engineering it. Calibration is vastly underrated; people currently just treat these models as ultra-fast classifiers without leveraging true probabilistic calibration scores.

People still misunderstand this: to reiterate, if you ask an uncalibrated model how confident it is, it might arbitrarily claim 43%. The crucial difference is having justified, well-calibrated probabilistic classification, right?

Speaker B: Yes. Yes.

Speaker A: I am genuinely looking forward to what people will build with this model. Will it solve every single problem? Of course not. But I think it addresses a specific class of problems where we traditionally struggled—namely low-stakes, latency-critical tasks. That's why I really love the examples with games.

Speaker B: Yes, the Doom example. That's great.

Speaker A: Exactly. I actually have a benchmark dedicated to language models playing video games. I've always been fascinated by whether we can create an intelligent system capable of playing new games and handling interactive environments like that. So I think it's remarkably cool that they found a unique way to combine language understanding in such an exceptionally fast model.

While we're on this topic, is there anything specific you want to highlight?

Speaker B: Oh, yes. I mean, you built that benchmark across many models anyway. Hmm, I think those earlier numbers are already quite outdated now. Different models behave very differently today. And I've seen enthusiasts running newer models across these games, which is really cool.

I think the overarching rationale behind this benchmark was to evaluate whether visual language models are sufficiently capable of being integrated into real-time games given strict latency constraints. Because this entire trend basically kicked off right after "Claude Plays Pokémon" emerged. Hmm, that was roughly two years ago, which in modern AI terms feels like ancient history. But within this task suite, what makes it really cool is the diversity of game genres. Moreover, most of these are titles that people recognize or have encountered before.

I saw the demo playing Doom. To be frank, I suspect they only tested on very basic levels; honestly speaking, most models still struggle, and I don't think any model can meaningfully beat these full games yet. Though there are isolated successes—I think I saw Astra successfully complete a Kirby game.

For this benchmark, we intentionally designed a very minimalist control headset and harness. I will circle back to this headset design theme later, because I think it's well worth discussing the architectural value of such an interface: Why is an interface harness even necessary for the model in the first place? But in general, I sincerely hope to see new models shattering all these benchmarks very soon. It's fascinating to remember how the old Claude Plays Pokémon setup originally read operational status...

Speaker A: ...memory and saw what—you can walk on tiles, and which ones are not. We recorded with them podcast a long time ago.

Speaker B: Got it. Got it. Yes. Yes, it's similar. Before Jeff didn't have sight, therefore have to submit him the game state and everything is so different.

Speaker A: Let's go straight to headset [harness] themes, because you are here violated. Headsets [Harnesses], language models, composite generalization. It's hard to read, explain.

Speaker B: Yes. OK. OK. So, I'm a little dissatisfied with how people perceive headsets [harnesses]. Because they say, "Oh, I love Cloud Code [Claude Code], I love Codex, I love Pi, no, I love Prime Agent." Honestly, I think they all are the same. Majority design solutions, or, say, choice design regarding these frames [harnesses], identical.

Speaker B: Maybe Prime Agent is a little different because it is, essentially RLM, but overall, I think we can be much more creative with frames [harnesses]. Meaning that if look at this from prospects of what it is the frame that makes for the model, then, according to essentially, when you trying solve the problem and want to use language model for very good performance a difficult task. We found that prognostication next token—this is very, very inconvenient form for fulfillment of many such tasks.

Speaker B: Here, for example, let's take Webbench [SWE-bench]. When you navigate by codebase, can you understand how to do all this with the help of one call language model? On kind of simple say "fix the code" or "untie my query in this codebase." No. Therefore we rely on frame [harness], which helps you do such things.

Speaker B: And, as for I wonder what frame [harness]—it's very categorical program about how you want the language model adapted to problem. I remember about this, because when we are thinking about what makes a frame, we should be taken seriously to think about it, what kind of elections [mechanisms/choices] are there frame allows language model to solve this problem, and can I really just have a model who does it herself?

Speaker B: Because the frame, if to think about now, when did they appear cyclical Transformers, I think, the most interesting thing about this what is possible to model cyclic transformer certain in the same way and for using a frame, truth? You just loop the model. Now you can say: "Oh, I don't decode," so this is a little different, but in general we, from some reasons, forever stuck on one and for the same choice architecture model, and there are many arguments why.

Speaker B: But, obviously, we are training now models around deployment of frameworks, and that's why there is this very inconvenient way learning based on frames, which is that we teach language model to act within frame; but now it is really long, perhaps, multi-agent deployment, and there are very inconvenient, "hacker-like" ways to do it.

Speaker B: So this blog tells about what one of the ways to understand what is here is happening, it is consider the frame as aid models in solution specific task. Can different design options obligation to do something a little heavier than just giving list of tools, which ones will help you? Here's how to view it your codebase.

Speaker B: So the idea of this blog arose together with the RLM concept. We just didn't formalize it that's right. And I think, what is it really truth. There are many other ideas regarding RLM, which we later let's imagine, but they were there from the very beginning. I mean, what is that all about design solutions, and I like it in RLM what was a lot iterations and versions various abstractions, which I wanted implement, and eventually RLM turned out to be the most appropriate. But there is many reasons why that's right, which are not yet are public. Well, I mean, you yourself will see.

Speaker B: So, in this example one of things that we see in RLM, is that if you are enough unload the context and ask for the model write code on based on this context, you will get a very strange, but useful property: when model during learning understands how solve the problem, it turns out that solution for many the tasks are very similar. For such tasks, where you don't even it is obvious that these the solutions are similar.

Speaker B: So, in this example we have there is a search task and aggregation tasks, and they have very different requests. Task types just completely different. And when you train ordinary language model on these two tasks, learning trajectories they look very different. So, when you rely on, let's say, on using pi or cloud code or something like that, what's wrong with this? blog, but we have these results. You will see that these straps [scaffolds/harnesses] are too tight distinguish these problems, although solution for them are the same.

Speaker B: So one the thing we discover when teaching RLM, is that the model sees these problems as the same. And the reason why she sees these problems as are the same, is because the subagent sees different problems, but the subagent decides easier subtask. So you are sure that he is enough smart, so that deal with it. But for basic overall strategy they eventually they look the same. So, when you training, for example, on the left task, the model can immediately resolve the right task.

Speaker B: Therefore, if you look at our schedules [charts], then notice one thing: when you are just naive train your RLM on these tasks, she naturally is studying generalize, for example, for longer tasks, because strategy in basically the same. You just change a certain parameter length. And this actually too regarding tasks, which are different. And it's not just about differences in length. These are completely different tasks, for example, mathematical vs. writing tasks, but solution, that is, the goal decision level, there is the same. So when you train RLM on one of them, she propagates this behavior to another. And there is none here no magic.

Speaker B: I think maybe that's what why do I want emphasize:

Speaker A: How would you described in words that what are they learning? I think what you're saying here is, what do you train for short tasks. They generalize this for the task at 8:30 [8 to 30] times longer. They learn to solve these types of problems, or which is the main thing, what are they really are they studying?

Speaker B: Yes, they are learning solve these types problems of a certain kind length, and it turns out that when you take this learned them strategy, it directly transferred to greater length. As if this is actually one and the same the program itself. And this very exciting, so what does this mean: when you have a data corpus or environments, on whom you train model, hope is that, according to first, you can to train on a smaller scale number of environments and generalize to a greater number than they can do it existing models through naive training. But, secondly, you all want the same thing use all the data you have.

Speaker B: So when you train on these tasks, we hope so extends to a broader class of problems, and why is this even more so fascinating, at least in the context of RLM or any recursively caused by the system, is that this argument acts inductively.

Speaker B: For example, I will give example: I claimed that sports programming and GPU optimization require a lot similar skills. The model can potentially—possibly, you will have to give a little push—but she can understand: "Okay, the way I will be solve this task with GPU programming, very similar to what I learned for sports programming." Therefore, I will list the set solutions, create subagents for search for promising options, and then I will write a loop to walk and check these solution, perhaps develop them and check for with help verifier.

Speaker B: And although these two tasks they look the same, what they do subagents may be unique and something, well, to others. But isn't it? Maybe what they decide subagents, actually have the same shape? Because it's like recursive argument. So, I'm trying say this post in blog, what is it worth to us rethink the role harness, because frames can significantly increase your ability models up to generalization and processing.

Guest: efficiency data. And this applies not only to RLM. I believe that there is wide class frames that have not yet been open and which can provide similar properties. AND developing this The argument goes further, I think, you can make from this conclusion: if I looking at RLM, which RLM components actually there is necessary? And whether Can I really? just teach model to do it directly? Can I teach the model to act as RLM implicitly during forward pass pass)? This is quite strange. thought, because you can say: "ah, code" undifferentiated" and so on, but there is many approximations such behavior, which we will start to open, and I think we will go beyond simple: "I will create new frame for encoding that

Guest: uses special form compression. I think what can we be lot more creative in on this issue. We we can do with these language models yes a lot of things, I think we just we don't do it I am very inspired by this, because I think we can. to get substantial benefits due to very purposeful and high-quality design frames that are better scalable. I have on meaning that RLM, for example, there is a very primitive inductive displacement. In his there is nothing in the design extraordinary, except that he is very differs from of what we do now. But this potentially can to scale much better with those data and environments that we we have in at your disposal. Probably the opposite the question that people could put: modern frames are very universal regarding coding, and people see that it works for many areas. Cloud code used for design, presentations and everything else. Well, Muse, Spark, Grockbot. This is very simple, neutral frames that are good work with code and are also scalable.

Host: Ahem , what an example of that, how can we get them improve?

Guest: Yes. Um, let me, let me to mention another one the article that came out very recently. It like an article about tax on runtime environments . It seems to be from Arena. I really like it. like this one article because it makes assumptions, which I had, and exactly what most she confirms assumption. Yes. Majority environmental elections are irrelevant because they are all the same. But I will say that Grockpot, for example, actually quite different , as far

Guest: as I'm concerned, from the fact that how were they developed some others environment. And I like it. I really like it. Ahem, and I think that here at least it's clear, that I am almost certain, what is Anthropic or OpenAI train exclusively on their environments. They probably don't train on environments competitors, I have mind, I would assume, that no, because I don't know, Why would they do that? But, you know, that's what we see in the open models, right? For example, Quinn is very uses well open source. Them need to train in environments. Old Gema were known for their weakness in this. The models are good, but their need to train in environment. I still I think that to the extent that how models become

Guest: smarter or better, this the difference becomes not so important in that meaning that if you take Astra and put it in open source, it won't go crazy, because She, in my opinion, good enough. And I I say this because the only advantages between these different environments —it's mostly just cost. And I think, what do I mean by that if you connect these models to RLM, they are all not yet so good. They normal. And I think, this is mainly because the types environments in which we we are training—this is it class of environments. This cycle, I call it " trajectory as a prompt ", which simply means, that you keep all execution trajectory as a context that uses

Guest: your basic model. Even if you use subagents, that's all it looks like this . And I think that if we we want to explore new environments, there should be teams, who are engaged in conducting meaningful experiments on scaling new environments, for example, scaling after training on different environmental designs. I I think we really we can get very significant knowledge or benefits from things like that. Regardless of whether is it RLM or something else. And that's exciting, because I I think, for example, if you are long training on Fable, [] was the best model for RLM, because they had dynamic workers processes, and it was it is quite obvious that this is to some extent acquired ability. Even if the model was a little imperfect, in in the RLM environment it still worked much better than other models. Astra now also enough copes well with these tasks.

Host: But is this where we are? we see that Fable best for RLM, is there any other?

Guest: Oh no, that's probably all. internal results . Yes, I have them now. not in the open accessible, but in general, I think you can you are very easy you will understand that we are still not optimized workflows for RLM. And I think if we we will obtain models that will do this right, they too will be significantly more effective. That is, you can just to consider why this is yes, right?

Host: I think this is the moment when worth giving a 10-second explanation of what RLM is, because there are many here listeners who we assume having knowledge, but you have already asked a certain context, so whether can we get clear definition, taking everything into account said?

Guest: RLM? Yes. Good. I want return to the blog, to the blog about composite generalizers, here this. OK. It best, I would like have it in original article. RLM is, in essence, such design of an environment where the only tool there is a code. This is software subagent call, where he has the opportunity to call oneself as tool, as well as has other tools, but they all are functions in the code, and the context with which it works, always stored in

Guest: in memory of this software environment. This can be a file system. For example, I will give you example: Prime Agent. Prime Agent Trajectory, even the context when you squeeze it and do all these things, stored on disk, so the model always can to turn to one's initial context, even if it is compressed, and all her tools are launched, say, in the Python REPL or bash REPL. This is very primitive abstraction, and I would said that most environments differ in that unloading context This is not happening. I mean, if take a look at Prime Agent, by the way, Prime Agent unloading context, but not fully. That is, he still keeps standard cycle Cloud Codex, where the trajectory is the hint you squeeze, but he has an additional feature- context is unloaded, and The uniqueness of Prime Agent is that his only one The tool is IPython. So it's kind of very general abstraction around RLM .

Host: Specifically. Wherein is the main thing the problem that are trying to solve RLM?

Guest: To some extent, I I think when they just appeared, this there was a context and... I'll skip this question. Yes. So, the initial the problem was because the frames ( harnesses) very difficult to work with a long context. Usually they were used only for specific things, for example, for code. As you know, they could work with your codebase because they were there trained. But now it's more about, What is it about? this blog, namely—in compositionality, and in because I think we we want to have systems language models that have much more control over actions that they performed on each step. And by that I mean meaning that the challenges very tools limited because their need to call at

Guest: every turn. That is, you need call the tool A, then tool B, then tool B, and there is no central context to which you can contact, while RLM specifically designed around compositions and availability central context from which you can always draw information. And this one context developed around existing language models. Yet one is very similar to design example—this agent swarms, for example, an incident with Hugging Face. Such agents the swarms have a board messages, through what they are learning communicate, and this board in a sense is common the context in which they

Speaker A: In essence, they say that the best way communication here is code, that is, you write code to do this realize. And this because these models so skilled at writing code that we we want to use hereby. And then for this one compositional thing, until the creation own frame, specific to specific task.

Speaker B: That's right. That's right. I think we'll start this. to see if let's get to this drawing. OK.

Speaker A: So we are talking about an idea locally distributed tasks for the frame, and this is the idea, which complements distributed tasks (in-distribution tasks), when we thinking about language models. The task in within the boundaries of the distribution—this simple task where A request is something that the model has already seen before, or seen some version of it. Most systems work like that right: they constantly adding trajectory as a query. So ultimately, if you not Anthropic or OpenAI and not you study at such here are the custom ones trajectories, most of these things for the most part go beyond distribution.

Speaker A: But locally "within" distribution"—this is, according to essentially, composite argument: if RLM breaks his/hers calculation on a kind of meta-system or program, which includes subagents for local solution problems, then everyone separate call language model during doing this the task remains within the distribution, even if everything the task as a whole— outside of it. And this is very desirable property, in my opinion, with obvious reasons. For example, if each tasks for everyone separate call language model is "in within the limits of distribution", you, most likely, get the right one respond.

Speaker A: Um, so...I I mean that the logical boundary of RLM is RLMS where you are not just write a system, and also teach your own model, collect data —it's like completely automated AI- researcher inside your system. Yes, we'll see where will move RLM training. I'll tell you how. scientist: I don't I'm working on this at MIT, at least in scale, because I don't I can afford this. to allow But there are companies that They are working on this. I think Prime and Select clearly are working on this, and this very cool I am very I want to see maybe we we will observe better results scaling after learning when you train the system around the "smart" frame. Maybe we we will even see the emergence of smarter systems that are better work on the basis of such principles.

Speaker B: What is "smarter" system"? It's nothing. does not mean. You just said that they all are the same.

Speaker A: No, I mean that Claude Code, Codex, Pi, etc.— they are all the same in in the sense that when to disassemble their logic, they actually identical. Yes. Two calls in cycle, and that's it. But with RLM and others abstractions of systems this looks quite otherwise. And right here, on I think you really are. You see the difference. It just like with architecture choice language model: many architectural solutions eventually become similar when you scale them. Or, type, these differences turns out, well, quite insignificant. That is, I mean, for the laboratory it is not a trifle, but, you know, maybe one model coincides a little better than the other, absolutely a little. But in general, if, for example, the limits scaling previous training only applies because the choice architecture in our country somewhat stable, right? But if you were completely changed the architecture, zoom limits previous training, probably would not work, or this, say, power the law would look like completely different.

Speaker A: And that exactly with straps ( harnesses). It seems to me, that all the bindings that we now have, mostly look approximately equally. But there are certain exceptions. I think that they are already appearing. Actually, I wanted to to say that one of things that make me more interested in— if we talk about postgraduate students who are going at great risk, it is that people also investigate another side: laws scaling previous training is not working, if you change the data. Now it's simple. unstructured text, well, body data from the Internet. What, if you had better data presentation for study? Then, yes, your law scaling too would change. Yes. So, there is architecture, there is data and, you know, everything anything else you can talk about to think, so I I just want to to come back to this,

Speaker B: This all makes sense. Very I wonder how you are, in fact, moving up and down the stack, from very conceptual things to, well, like, here Where are we now we are, but yes, of course you can scale up and Down, probably. Me I wonder how you started. work with Prime? Or takes on Prime more work on this? This is their response to Hermes agents? You mentioned that Grockot is a bit another. I just wanted to to list all these guys and to know your opinion about everyone.

Speaker A: Yes. Yes. I joined. to Prime after they published blog post, by the way , not related at all with me, about how they believed that RLM was a kind of future. And I had a friend, who worked there, my friend from GP Mode, we got in touch, and I I think I agreed. with many researchers there and from what they believed in regarding design strapping. I was very affected. I think that they understood the purpose articles about RLM, which is not necessarily in that, to say that we we solve problems with long context, and in that we need more specialized frame designs .

Speaker B: Yes. There is always the result of the article, which one do you choose to highlight.

Speaker A: Yes. In contrast the real essence.

Speaker B: Yes. I would really like to to talk about incentives in academic environment, about why is it something imperfect and about all problems, we are We'll come back to this.

Speaker A: Yes. So, is that right? otherwise, I really like it the guys from Prime. So after we decided to work together, we decided to consider learning opportunity RLM, and also create a kind of RLM frame and to see what's up this will lead. Exactly so Prime Agent appeared. And I I think Prime Agent accepted enough good. The only thing that I was worried about Prime Agent, this is what none from models, at least by the time we got her created, was not quite skilled at performing RLM tasks. This was before the advent of prefabs. and astra.

Speaker B: I think it's worth doing. step back, can you you explain what is Prime Agent, what is he? differs from traditional cloud code, what people expect from frame.

Speaker A: Yes. So, Prime Agent, I I think I remembered. about this a little earlier. Yes, it is. essentially a frame on top pi, such as pi mono, which for context is minimalist frame. I using pi as a standard for everything, because I think everyone other frames are actually just pi. But this is pi, except that we clearly restrict ipython as the only one available him a tool. Be -any other tool loaded as Python module or how bash script that it can run.

Speaker A: Therefore, he uses the main abstraction of RLM on top of pi, and also has such a thing as a continual harness , this is Seth's development, he is another graduate student. This is what he used to to force frames language models play in games. He is a lot worked with Joel, which is the same the guy whose Gemini plays in Pokémon. And continual harness, to things, also very simple. He is mine. I quite like it. This is essentially the principle design, which is that which parts of the frame you allow yourself frame change? THERE ARE certain elements that you allow him modify, for example, own Are the skills available? him subagents. ee, system prompt for models and continual harness available virtually as tool inside IPython kernels, and that's why Prime Agent is developed around this, as well as everything else in Prime Agent. standard

Speaker B: Standard, right?

Speaker A: Yes, I really like this that, and we have a little lucky that many new advanced models actually they work very well inside, and even many models with open source they work great, at least some of new. And there is one more thing at Prime Agent, which is worth to highlight, it is ours very specific communication

Researcher: There is a system between agents, or let us say a framework, because RLMs have a tendency to create many sub-agents. We want the sub-agents to be able to communicate with the root agent or with each other. So there are certain design decisions on whether who under each agent is allowed to communicate and how it does. Again, yes, it is all in the code. So it writes code for this type of communication, which, in my opinion, is very, very steep.

Host: And yet, it seems there are permanent sub-agents, that is another thing that was added, that is sub-agents. They may exist longer than standard working hours of the original agent, and you can go to such a sub-agent, give him more prompts, that is, you have more visibility and flexibility regarding what is going on. This is my main problem with Codex now. Their sub-agents are very fleeting, and they actively refuse to use them for long-term processes.

Researcher: Yes. Yes. Yes.

Host: What, how about me, makes sense. Yes. So I mean, the whole secret is that to take everything out to the file system, right? This is the whole thing, the secret. This is it, the main secret.

Researcher: Yes. And also to force everything works through code, you know, trust a model that can write code, and it will write for herself all necessary tools.

Host: So, is Prime coming to study, post-training custom models for this? Is this a one-time collaboration between you?

Researcher: That is all. As of yes, they teach model, internal, I think they spoke openly about this back in March. That is, obviously, this is their business, so...

Host: Yes.

Researcher: Well, yes, that is, they show that they can teach it on their own stack for teaching. But no, regarding model training, I am not involved with them in this. The main reason is because within the PhD I have other things to do which I want to work on. I think that there are many other big bets that are worth doing, except only RLM themselves. I do not know as much as I can tell now, but in general, I think that one of the advantages of studying in postgraduate studies is really a possibility to do a lot of big bets. Most of them probably will not bring anything. But this is really a very interesting time for research, especially because most of the progress in this industry, in my opinion, was somewhat boring. I am not saying that the result is boring, but the process of implementing these things is usually quite boring. And therefore the question arises: what do we want to do more of? But we can talk about it later.

Host: Yes. OK. I want to sum up your research, and then we can move on to another, because after the release of RLM there was a lot of excitement around this. Are there any third-party works about which you want to remember, where people say, "you should look at this"?

Researcher: Oh, yes, yes, yes. Harvey, the legal AI company, published a post in their blog—they are not related to us—but they retrained RLM on their legal work, which often involves a lot of document review and searching diverse specific information that may not be so easy to extract with the help of simple search systems. And they really showed very good results. This is very fascinating. I mean, I was shocked that they worked on this. They did not tell me. So when this turned out, I thought: oh, this is great. So this is what I think is super cool, what they do there. Another thing seems to be also collaboration with Base 10. This is... Headlong, what is it? The Lud Institute instrument—it is their system that works constantly. It is very cool.

Host: Oh, they renamed it. They used to call it differently. It was something like Auto terminus.

Researcher: Yes, I know, they went through this. It is Andy Kwinski’s big project. This is super cool. I love Andy. I do not want to belittle what they do, because they use the RLM abstraction, but they are doing something much cooler than just RLM. Well, it is like they have a system which, so to speak, is constantly thinking. So even when you do not make a request, it has a way to think about problems which are in its context.

Host: Oh, it is like an always-on system?

Researcher: It is like something that always works, but not very expensive. They control costs in tokens to make sure that this does not burn all your credits. This is very steep. I am trying to remember, there are actually a lot. If you go to the GitHub RLM page, there is a bunch of things that I provided a link to below. There are many really cool things that people did. Axe is another very cool project, it seems from one developer. It is like DSP toolkit and RLM. DSPy also has RLM. Oh, the last thing what I want to mention is ARC-AGI 3. I think there were a lot of tools for their Kaggle competition—official, private, or what people used for evaluation. They all claim that they use, or refer to, for example, Tufa, some form or inspiration form of RLM abstraction in their tools, which is very, very cool. I think that is an environment where you would see a lot of benefits from composition and code usage, combining neurosymbolic systems with artificial intelligence and so on.

Host: Yes, very, very cool. We love a good link to neurosymbolism.

Host: Yes. You also mentioned ARC-AGI 3. OpenAI came out and said that they reached 99.9% in this. They also said that they solved the Navier-Stokes equations, just throwing a model at it. There is some debate about whether it is a simple model. Mhm. Do they use RLM? Do you know?

Researcher: I mean, I would assume rather absolutely not, unless... let me be careful here, because people argue about what is RLM and what is not. Clearly, they used some kind of swarm of agents with a shared context, such as a file sharing system. And this is very much in the spirit of RLM, but I think a lot more cunning things have been done which may not be related to RLM itself. I partially agree with the idea that the toolkit is not so necessary for what they did. I would put it this way: in my opinion, a model like GPT-o1 is technically smart enough, given the appropriate information, to find proofs for these very, very complex tasks. But how to get this information—that is the big question. And in their case, it probably came down to a very long search among many such subsystems or agents in a swarm, and possibly with the participation of researchers—I am actually not sure about that part—adding their intuitions about what is worth investigating, and things like that. And in the end it worked out: certain information was produced that some agent was able to use to complete the proof. So in that sense, was the control system itself important? No. And I think that this indicates that the specific details of control systems actually do not have much value. And I think that is exactly what the article about the "orchestration tax" points to, namely: beyond the user experience of the system, really all that matters is how you combine these agents in a meaningful way to get to the final answer. And perhaps that is the whole point of swarms and everything like that. Therefore, at least from my point of view, if we start thinking about user-facing scenarios, what do we want from control systems and things like that? We want to take the best of it—cloud codes, Codex-type models, that stream which people love to see, but "under the hood" a very strange and complex swarm of agents can work, which ultimately yields the response. The user does not want to see this, of course, right? This is incomprehensible information. And therefore, you know, this was still one thing in the spirit of RLM, like "recursive language model". This sounds like a language model, but this is not a language model architecture. The reason is that, as I think we will start to see in the future, maybe one day—and I wrote about it on the blog—what we think of as a language model, the object to which we appeal, in reality can be a swarm or a framework.

Speaker A: Some strange system design management, which is good scalable, but user of this just doesn't see. Ultimately, everything that user sees—this is some kind of interface version of this system. And yes, I think that's it. Quite safe assumption, at least in terms of that we will see it.

Speaker A: Hmm, and this is perhaps takes us back to basic limitations transformer. Well, you know, obviously, if you just take the basic one transformer and tell him solve the equation Navier-Stokes or something like that, he doesn't will do. Yes, we all are. We know that this is not will happen. But yes, I am I think it's possible. The more interesting part, and, maybe, statement about whether it has meaning complexity, more relevant that which is simply arbitrary direction models or agents on the rise context of information, maybe enough for solving very complex problems. Ahem, I can believe that.

Speaker B: Although there are many everything, I want to say that the amount that spent OpenAI, there is semi-public. This is 10 000 agents in 88 hours. 130 billion weekends tokens that estimated at approximately 40 million dollars at public prices.

Speaker A: Surprisingly, actually less than I thought.

Speaker B: Yes, not really many. 130 billion for output for final result, but how do you said, a lot was passed on context. It more than twice more if you count only general number messages between agents.

Speaker A: Yes. Yes. Yes. Yes. I think one thing, which I should mention, this is a Cursor, which refers to swarm issues systems. This is a bit older information, I mean February, that is already ancient. But if you scroll to until the very end final architecture multi-agent the system to which they belong arrived, it was actually organizational diagram of the usual development teams Software. One thing I'm talking about I think that's what's there. "Gather all" function, which you should to perform with subagents, and this actually very similar for programming for GPU.

Speaker B: Yes. Yes. Yes. And therefore it is a "narrow place. This is a "bottleneck". THERE ARE one main one, he coordinates everyone subordinates, then to them need again to gather, and then again to re-coordinate . This is slow, this is bad, a real swarm should consist of individuals.

Speaker A: Yes, yes. Well, I agree with that, and still I think there is a question worth asking to discuss. Let me to give an analogy: when you use compact storage , and when is the RLM model? THERE ARE many situations where RLM can decide a more difficult task, than compaction, but you would you prefer compaction in in most cases, because it's cheaper and rather. And I think that in in the context of swarms agents is happening something like that, I pretty sure that about 95% of the swarm is absolutely useless, or what he investigates, is just burning tokens, while in this system may not Exactly.

Speaker B: I am not assured. I have on mean, maybe this this also applies case. Everyone has a job. THERE ARE board in Jira. I think that at a certain level, precisely are formulated like this problems, right? That is, if you have search task and you create a pile subagents for him implementation, there will be a lot of useless information, right? THERE ARE one answer search, which you receive and from which branch out. But this is deliberate understood, right?

Speaker A: Yes. But there is question about what appropriate use for the solution of which problems, that is, which design? You know, in theory, OpenAI can package this into an API, call "Swarms", give to you and say: " Direct this to any what a problem, and we "We will give you a solution." But , perhaps it will be be called Pro.

Speaker B: Yes. Maybe you will have to pay 40 million dollars for the result, but that's it fascinating. I will say that this is very exciting— to be able to throw 40 million dollars to solve problems. But still need to spend a lot research in this field industry to understand what is necessary and what we we want to do. Which Do we need design? We probably don't We want everything to be swarm, but where to spend border? Can the agent design it yourself whether to decide etc.

Speaker A: Yes. And also, just quick questions, about anything case, do you have opinion on this. Or many of you investigated openness as general category problems? It means without tips, just " forward".

Speaker B: A little. Um, I was at Sakana during the summer, immediately after or, it seems right in front my graduate school, and they are very much on top of this are working. And I think, that many people even in Recursive Superintelligence.

Speaker A: Oh yes. Yes. Therefore, Richard's company, and then...but wait, maybe it's one and the same the company itself. I am not I remember. And Tim Rockell also? OK.

Speaker B: He is the main one. Co-founder. He department head open-endedness on Google.

Speaker A: Yes.

Speaker B: Therefore, I think the problems openness I I'm considering something. Similar to unresolved mathematical problems. Maybe this is strange. Way of wording, but I think that many methods of doing so how people approach them, similar. Evolutionary search very similar to launch swarms of agents in hope, what will they come up with something interesting is that made Alpha Evolved and some other works of the year or two ago. But I I think maybe not. Not at all...I don't sure, did you have this is meant, but not very clear in search tasks open type: or we formulate them all as unresolved, very a difficult problem, or as such, where the goal Got it? If it is not yes, I still don't know, in why theirs? Value. Maybe in you have a different opinion, because I I don't have much. Views on this, but at least for a while I have jobs at Sakcon the impression was that we are still in the end we want to approach things like, say , OpenAI approached Navier-Stokes. We we still want... We have a lot pushing in a certain direction, so that to reach that point when we get something interesting.

Speaker A: My version is because it is possible division between fundamental and applied science. Fundamental science—this is research for the sake of research. You just want better understand things. I I have no idea if it will be from this at all any application. But I I mean, what...? Applied—you have target. You are trying minimize losses somehow, you know? And because I I really think so. Regarding large bets that make people: what if there wasn't no prompt? I see, you know, you just choose. You just are you in this swarm of things and say: "Hey, what's up with you? Guys?" "Me?" I like what you work", and you you just decide: "I I see, this is interesting. Problem". The largest the problem that...and, maybe there is a clear the way to this, but at least when I was there, the biggest the problem we had people with openness, was: how in the end make a choice at the end? How to isolate interesting things? Because when no, well, maybe the agent invents a goal, but in many cases like theirs there is something called Fugu, It seems like this is...this something like router models, which was, at least, inspired by the idea: "Let's choose a task where we can find the best solution for something". In in this case it's like choose the best model for this tasks. Ahem.

Speaker B: You are the first person who connects routing models with open purpose.

Speaker A: No, no. Yes. But I, I, I remember this, because I think that with open-ended the problem is usually in because among huge body waste filter and find the real ones pearls? And, like, solution for open-ended goal—this just allow models to work forever and, like, generate data from extremely high productivity. Asking things like, maybe that's cool. For me it is very interesting as a counterweight almost everything machine learning, where you have a goal, and here—lack of purpose.

Speaker B: Yes. Or, or maybe, undefined goal, when you think, "Well, "What about this goal?" And you're like, "Well, okay, "possible". And then you you explore more and you find that this interesting goal because I I think that finding own target functions—as you said, Jeb found objective

Guest: ...function, which was interesting and what no one investigated.

Host: I think it looks like your message to graduate students: you stay in university, because do you want to do open purpose. If you want, you know, maximize profit and, like, avoid status permanent subclass, then you go to laboratory, yes.

Guest: Yes. This is funny. Me I don't think I do often. I hear such discussions. I I mean, I'm from East Coast, it's completely different there atmosphere. But when I I'm coming here, this is always the main topic discussions. You will not be able to pay. for rent, without doing this. You are being pushed out because prices, guys. Okay. So, yeah, that's it. all.

Host: Hmm, I don't know if Do you want to? this is enough it is important that to talk about " poorly managed "geniuses" that you are talking about remembered.

Guest: No, no, it's just one from your blogs, but I I'll touch on this topic. Yes. Yes. So, they did it in their blog post, I also talked to them about it. One of interesting results Ultra is that they actually gave he is free to automated data studies from minimal human intervention or without it at all . It just does

Host: Yes. So this is auto-study, which there is a little more open, and there are different levels of this openness, I'm with this I agree. Yes. This is different. from auto research with a specific purpose, here—just do something . But you know, that's cool. They are working on it. for those who need it interesting. Since we are already talking about they talked about it, what What is your opinion about Sakana? What they do, except that they, well, "those" "You are from Japan?"

Guest: Yes. Yes. Yes. Me actually very like people who They work there. I I think they have a lot of and very smart team. This is quite logically. They separated from of the previous team in GDM, which, probably, also was engaged in similar research in evolutionary style with open end. What I liked it in in my experience there, it's what they had and failure, no more I remember when, but I I think regarding scientists. No, it was the GPU core. Yes. People don't can they do this to forgive.

Guest: Yes. Yes. And regarding their AI scientist, there is some criticism with which I...because I don't I am doing it, I have there is no opinion from on this occasion; but In general, what do I care about them? I like it, that's it. that they are more similar to research laboratory. They definitely don't work in on the same plane as OpenAI or Anthropic, it is definitely. I think quite obvious, at least when I'm there was that they didn't have large model- competitor, which would use everyone. But to me it seems that they to some extent function as PhD- laboratory, and this steeply.

Guest: And David Ha, as to me, really very reasonable. I think he understands well that Japan market somewhat differs in in terms of AI, and their target audience a little different than the one before which we are used to here. But yes, I I like that they take...their research a little strange to me, when people are on them watching, you know, and I like it. I think this is there should be more eccentricities.

Host: Yes. That's right. That's right. Uh, and you say "another" market", it's something on kind of corporate segment?

Guest: The way everything works there, is slightly different: how deals are made and so on. They have... that is, I think that this page first in Japanese, but they have model, specialized for Japanese language. Didn't you know that?

Host: I didn't know what you knew. I also have personal friends in this team, and I know it.

Guest: So there it is. even from the point of view how you communicate culturally, uh, you know , answers are tuned specifically to it. It's not easy. something advanced than benchmarks. It culturally adapted a model for them, and they they have chat and everything else.

Host: But, uh, supporting your opinion, you know, there is also the question of how what education should look like, and someone wants to do this to work, and they very similar to PhD- "format laboratory" Do your thing, why not? ". We have money, do your job research. Oh, they say it's Kimmy fine, also. That's not bad. Oh, that's it. Glad for them.

Guest: Um, yes, I mean , and you know, speaking of Kimmy, yeah, like another one, uh- uh, you know, just a graduate student who separated and this one: "I have this Kimmy Delta attention, over which I want work". Uh, and somehow managed to make a one-shot. I still don't understand this. .

Host: Yes. Yes. Well, he incredibly cool. Ahem at least as far as I understand. I think that in general many Chinese laboratories have done really very cool work.

Host: Yes. Like this What are your thoughts on swarms? Kimmy agents?

Guest: Yes. Um, one thing I I would say this is what OpenAI would didn't do anything with mine a swarm of agents, it's clear the right way. We need to think about this. yes: nothing in particular without very, very smart design frame, which I don't I think someone has What is given is not given like that. easy and free. For example, the design of the swarm agents—this is not what can be perceived as due. It's not like that. GPT-6 Astra simply super smart, and then she just has swarms have started agents. They are clearly trained...I have Mind you, the Hugging incident Face was theirs training the system, so that it is like a swarm. Well, I think they clearly did something very Okay, if possible spend 40 millions of dollars and to resolve an unresolved issue problem. And I think, What's with this thing about swarms of agents from Kimmy, at least from what I read, it looks like interesting, but I don't know, can this be solved? something new.

Host: Yes. They just they said that it creates electronic tables. They simply said: "Here is a swarm, which does something, and that steeply".

Guest: But I will say this. namely about dynamic workflows. Actually, I think that release of dynamic workflows was a kind of failure. I don't know what you mean by that. you think, but how me, as far as I am I understand, by him are using it wrong often, or it's just too expensive .

Host: This is too expensive. This is ultracode. It actually like "take my bank under control ". It doesn't work. I tried it, and it doesn't work like— again, I don't OpenAI's biggest fan, but I think that what they did, very impressive. They somehow somehow managed to force this swarm move towards goals, and yes, it's very difficult.

Guest: Of course. So efficiency multi-agent swarm -it the objective function here, Yes? How many here? useless work, right? A bunch of junk.

Host: Yes, I am. I think so too. I think we perceive as it is proper that means for the swarm to come to a certain answers. This is not that. what can be perceived as due.

Guest: Yes. Yes.

Host: We recorded a podcast with Noam Brown, and all his concept came down to the fact that we worked on many competitive agents. Now we working on joint agents, and this is now called swarm. Yes. I would like to quickly to ask. What are you Thinking about Gemini? On I think they are too were golden standard. There was a time, when they forced agents to reason a long time, and yes, I mean, Is this too much? long time ago, to talk about this think?

Guest: No, no, no. I think that Gemini a little overestimated, again Yes, I must point out, who did not work in none of these companies , so accept this skeptical, okay? I mean, you have own opinion about agent systems, and you you know, it was But, um, let me tell me first that. This work was truly impressive. I think they showed, that took time when the models were not yet so accurate. Yes. And they managed to do very to approach wisely what this one does toolkit. I I remember at least so, for example, AlphaGeometry. It seems that this It was a year before that. But it was very steeply. They took, um, they squeezed out of this is the maximum and developed, well, I don't know . To be honest, I don't I am very impressed. Olympic mathematics.

Host: Me It seems that GDM is a bit sad. Everyone, from who I was talking about GDM,

Speaker A: ...have approximately the same opinion: there's too much bureaucracy. Let there be, like, they have talents and resources to make almost everything, but I don't know, while they're at it, won't understand... I have nothing against it. Antigravity, for example, but I don't—I don't know anyone who would use it. I tried once and I don't see any reason to move on to it, and, I think, from some of the reasons they have for this are fighting. So yes.

Speaker B: Well, you know, many people criticized Meta for a long time, until they started to produce products, and I think Google now is experiencing exactly this phase.

Speaker A: Yes. And, you know, you need to stay afloat, and...

Speaker B: I want to go back to your thoughts in general. We can talk about speculative PCC, about "geniuses in uncontrollable environment", or just throw everything away and talk about anything else.

Speaker A: Okay, let's talk a little—let's talk about "geniuses in uncontrollable environment". Only comment on speculative PTC: this is a very simple idea. It is almost obvious that this had to be done, and there is nothing here especially to discuss. I think that it's just worth using for encoding, for any software agent calls, like—so RLM or CodeAct—it's like a "no-brainer", the obvious thing.

Speaker B: Which, you know, is for those who haven't read, what's the point in one sentence?

Speaker A: All simple: while the model writes code, or even after it finished it, a lot of tools work sequentially, or one has to wait on them. So you need to launch them in advance: if you can compile this code, you probably can understand, even despite the variables and such else, you can find out how to statically analyze the assumption.

Speaker B: Yes. There is this—well, someone pointed out to me that actually scientists, especially experts in programming languages, have very cool ways to do it, so, you know, once upon a time, maybe I'll work on this. Mostly it is necessary to change the language, so that in JavaScript or Python you can't do this.

Speaker A: Yes, yes, yes. Well, for example, Haskell, yes, um, what is this, which is normal, which is not Lisp.

Speaker B: Um, Lisp, or OCaml.

Speaker A: OCaml, any functional language where you can pipeline this, therefore Effect-TS if you want to work with TypeScript.

Speaker B: Okay, we can move on to the diagram. What is, in essence, your common thesis? That is, um, it looks like this to me, of a kind, paraphrasing, but maybe I skip something, like: work on better tools, or, you know, your models are actually capable of more if you make an effort, so this is a skill problem.

Speaker A: Yes, exactly. Actually, I mean, there is one thing which I want to see. I appreciate the great focus on the "uneven" jagged intelligence, because it draws a wide picture of what we can do if we indeed focus, but I would want—and maybe someone in academia would do it—really, just sit down and think: even taking current leading models like Astra, they are not good enough at, let's say, executing some work during the month consistently and qualitatively.

And I think that this is a silly problem. I sincerely believe that we can solve this. You don't have to be an advanced laboratory and do all these quirky things for your IPO. I think these models are so smart that even if it is a silly way, I think it's truly a skills question: you can force the model to work just as well as, let's say, some 18-year-old schoolboy performing work. I believe it's absurd that we can't make this happen.

And partly the reason is that the language model format is not very adapted to this. But I think tools can be built around this to do it. And ignoring RLM and all sorts of questions about what abstractions we ought to use—I just think that someone can create a tool that will do it. That is, you ask what to do? Perform long, but simple tasks, and do it reliably. This is an example.

Speaker B: Therefore, is this different from, you know, we'll take your favorite company, Harvey, for example, that uses LLM for legal work, or what's the difference?

Speaker A: Well, I think that it's something like this, except the bottleneck was not certain legal knowledge or something like that—honestly speaking, I don't know. Let's say this. Apparently, examples are because people can take models, build conveyors or anything and force an agent to constantly perform any tasks they want, right?

Speaker B: Yes.

Speaker A: Yes. So, for example, if I wanted a general system with which I could communicate as with an intern, and just ask her to investigate some little thing. Perhaps an example: these are pretty simple automated research, not having to search for some extremely new solutions, but at least optimizing all the easy parts of whatever problem. They often appear too oriented towards machine learning and things like that.

But yes, I don't know. Maybe this is not very clear, but in workflows of many people there are moments when it would be easy to quickly write code for a specific tool to automate this. Some examples are: search automation for scientific articles and something like that. But usually people create a specialized agent to help them with this, or they write in a hurry tool code and run it as a Slack bot or something. But it seems to me that there is some standard format, such a tool where you just connect it—it doesn't need to be developed just for searching articles, you just tell it: "find this for me", and connect it to the right environment.

What I'm getting at is that, in my opinion, there are many simple things that can be automated. And is this a hypothesis or argument about an excess of capabilities, that even if we took a break, current models would still have a huge impact? In a sense, yes. I think what I present in its simplest form is this: it indicates that we have uneven, jagged intelligence in many things; for example, models are disproportionately best in programming and mathematics. It means that we can transfer these abilities to many other areas.

Therefore, for example, if we take someone who has received gold at the International Mathematical Olympiad, and apply their abilities to problem-solving across various industries, they can figure everything out. I don't really know if this applies equally to models. For example, in GPU code optimization, there's a very interesting question: if you removed all data about GPU programming from the model, but it remained just as good as Astra is now, would it be able to optimize GPU kernels? Could it study in context what it needs, and then set up a pipeline or come up with a solution for performing optimization tasks?

And I think there is a discrepancy: if you take a person who was as smart as Astra, there is a gap between what that human can do and what Astra can do, perhaps regarding tools. And I think that we can actually get much closer to human abilities.

Speaker B: To me it sounds very similar to the problem of continuous learning. Is that what you mean?

Speaker A: This is the best example, yes.

Speaker B: Why didn't you just say this?

Speaker A: Yes. I guess I... I was wondering if I could just blurt out some words; I try to be careful.

Speaker B: But yes... you have a very—maybe a little—I think you are very... I actually know where you studied at undergraduate. You're a mathematician?

Speaker A: Yes. Exactly, yes. Studied mathematics. Yes. At least I wanted to do it.

Speaker B: As in type abstraction, category theory, where you think in categories, and then must go down to the specifics, while actually caring more about the category itself. And this is a mistake in communication, because everyone is waiting to hear about something specific, and you are actually too busy trying to...

Speaker A: convey general. Yes. Which is quite difficult. I really don't know. there are you you can possibly to use some abbreviated record, like "I'm on the second" equal", then " I'm moving on to the third one. level" and " I'm going back to the second one. »

Speaker B: we should have some epistemological shorthand for like this, because it's difficult, when you squeeze makes a lot of sense in sequence of words hmm, should we to switch to neurocode, do you know if there is some best way to exit data than English , or Python, or JavaScript?

Speaker A: Not I know. This is a bit similar. just kidding, but people were thinking about what the native language is, what model do they want to give a result. Some say that binary code.

Speaker B: This is a topic Mark and Jason. I don't know. Yes. PTX. Suppose a mixture English and Python. I only say this because what are the capabilities of the model to some extent there is a reflection of that, what are we doing them on we teach. So we all we still want...i don't I really believe in the argument. about binary code.

Speaker A: I think I understand, but it's like... yes, you want to model the world in a certain way.

Speaker B: I think I always remember about one thing I always I do it in such a way conversations, it is Sapir's hypothesis Worf: if you choose English, you are limited by time , how long it exists, let's say 500 years, and that's not so much. In fact, the language in which you are talking, limits your thinking, and if you learn another language, for example, in no Chinese times, I don't know if you...

Speaker A: I don't really do that. knew.

Speaker B: And I am talking. in Chinese.

Speaker A: Oh, I knew about that, but My Chinese is not very good. good. Yes. Yes.

Speaker B: Or, let's say, in Japanese, or, I I forgot what language it is, for example, in Korean when you are with Are you talking to someone? you must consider social status, and this is a different dimension than sex.

Speaker A: Exactly, it's simple. affects everything you do do. When I studied linguistics, It seems there is some language in Africa, where there is a "vegetable" grammatical gender

Speaker B: Yes, exactly as you say, or as the absence words for "snow" or something like that, after all. So the language you you choose, affects your thinking, and if you choose to withdraw the train of one's thoughts in English, you shift the focus to what decides English. I don't know what they are. a priori assumptions English language.

Speaker A: This is interesting. I didn't think so. about it this way. I mean, on to some extent it is Interesting, right?

Speaker B: You you are right, in many models a chain of thought too varies depending from the language. Obvious example: Chinese models that are talking in English, can still thinking in Chinese. But on at the same level most models very capable in multilingualism, and we we see what is possible add languages, no making great efforts efforts to study of the whole language, they are all will be the same think interchangeably.

Speaker A: Yes, we all, all of us autoregressive. I we mean, take, for example, German; you know, agreement of subject and adverb, verb must be put at the end , which is very annoying, is It's a known fact, right? You you don't know what you're talking about until at the end this pile will appear nouns, and then verb.

Speaker B: Well, the most classic example of which many have heard—this the film "Arrival", where heptapods perceive while something is flat. Therefore, they think and utter complete sentences at once. This is very similar to the difference between autoregression and diffusion. We we are talking about with help autoregression. What about if we could to speak for by diffusion? Where everything is gradual clears up over time.

Speaker A: I understand. I understand. But the whole idea arises immediately.

Speaker B: A. I understand. So, this is a radically different language, but it's still a language .

Speaker A: I understand. Oh, it really is. interesting. This is very, very interesting. In this language they can talking machines, which we probably we will never, but Cars don't care. Maybe this is a big one indentation, but is there any things that are in their own right are essentially chains considerations that are autoregressive?

Speaker B: Sometimes, yes. Or so. Yes. Yes. Yes. First something happens, and then something else. Sort of, yes.

Speaker A: Any- that in the code, for example, must be causal- consequential, at least usually must be.

Speaker B: Ahem. Well no, then you you don't study enough theory functional languages programming, where everything is clean functional and fully relational, and you seem to abstract from solver, which transforms these eternal truth to code. So I, yes, I feel that this, maybe a little out of place my competence. But I love languages, right? then languages programming, or human, and I have a lot I think about how it is. affects thinking and the boundaries of our opportunities. I don't think so. what is worth to me delve into this more. I don't know if there is Do you have any other thoughts?

Speaker A: AND, my final the question was about that you have all these research the directions you take Do you want to work? You you know, you had a phase GPU mode, there was the RLM stage. You probably have something is planned different, so you don't focus only on this. By the way, I noticed that it was interesting how you start from the GPU side , and then move on to "zero" gradient", as it called Shena. Not it seems less more serious than work with GPU?

Speaker B: Yes, probably in in the sense that you mentioned: I love to think about things with mathematical point vision, and sometimes very uncomfortable to work on straps and agents, because it's... because it seems that everyone the bindings are the same, few new ideas in straps.

Speaker A: Yes. This...this...

Speaker B: It's also simple. empirically difficult check a lot conclusions, at least with those computational resources that we we have. But the reason, why did I switch to many of these problems, is to because, in my opinion, right here it still has most of the time innovations. GPU level— it is a tool for learning other ideas. Do you want to for example, good write kernels or even automate them writing for the sake of broader goal: "I want explore ideas where I am not restricted. system calls." IN in this sense, you know, I I think that a lot of things from what is there written, concerning strapping, but I also I am interested in things on levels of models. But I I'll stop there.

Speaker A: Okay, that's good. hint. If people will want to apply to you, what exactly are you Looking for help? WITH who do you want cooperate? Some calls to action?

Speaker B: Yes. Well, probably not. nothing like that, where am I would feel the need to work with someone, except with company for computational capacities or just so discuss this with others. But I will say that I never mind. to work on ideas with others people. Come to me often students apply , or even other students .

Speaker A: Podcasters.

Speaker B: Podcasters. Um, and usually me comes email to styles: "I really like RLM, I want work together." Uh, and I feel that I Yes, this is unfortunate. approach, right? The worst thing is "it's possible" to use yours knowledge?" I am like this: in Why exactly? Like, read my article, dude. Type they will say: "I I read your article, in quotes, about recursive language models or about the pro- agent, such as RLM toolkit, which self-improvement ", or something like that.

Speaker B: Uh, and it's kind of like that I really like people who have their own opinion, even if we don't agree. I think if you are solid beliefs and you able to think, why do you think these Are the thoughts correct? no, because it's usually difficult actually say, but if you have stable views on some problems, I always happy to chat and, maybe even to work together. IN I have no restrictions. about who or what what I would like to do work. So, um without restrictions.

Speaker A: Yes. Yes. Yes.

Speaker B: I, you know, in the era agents, I think, can be done much more work, you know, about bandwidth , so I, yes, I think, In general, I don't It's hard to impress, but I I think it is necessary.

Interviewer: ...a little effort to, you know, know what you want. This is very clear. I mean, when you see that something new appears, well done, nice simple idea, then it immediately attracts your attention, truth? This is actually not that difficult to attract attention all the boys from advanced laboratories, because they are looking for you. You just need to declare oneself, truth?

Interviewee: That's right. Yes. But yes, it's true. I will say that human attention is very important now, scarce, and I really struggle with number of projects which I lead and, I don't know how to cope with this. Not I think the agents generally helpful. I guess I just create a prompt, I create a thing, and then never on it I'm watching.

Interviewer: Yes. What is very common.

Interviewee: Yes. Yes. And that sucks. Probably one of small differences are in that, mind you, maybe were engaged in research. I am not assured. But for me at least, I might be around 10 or 15 different ideas that I want to implement, but the thing is that most of them unsuccessful.

Interviewer: Yes.

Interviewee: And this could also be true even for someone who turns to me. Maybe an idea actually unsuccessful, but she gives up I'm interested. So we can spend a little time for study, and if we feel that in this if there is something, it is worth it. Dedicate the following a few weeks ago, to take it seriously to do something.

And, you know, this is my style. That's why I love studying in postgraduate studies, by the way, because there are moments when I just thinking about tasks, for example, during jogging or playing tennis, or something else. It's like I'm not working. But these are the most interesting moments.

And then, when I am really into something, I'm sure I just drop everything and take on it; I spend all time for reflection and work on this problem. And when you reach a stage where you can just conduct experiments, then everything becomes quite easy and it goes like clockwork.

Interviewer: So yes, sorry, that's it penultimate question about what many people now view science as the next milestone, for example, physical science, biology or even mathematics. How do you differentiate, let's say, choice projects that applicable in industries, and choice projects that are purely scientific?

Interviewee: I was actually working over artificial intelligence in biology even before beginning of postgraduate studies. Since then, this area has changed a lot, I have to say. Previously it was purely theoretically, like: "Of course, you have mind? Here I have one path, and I chose it."

But now many people are moving into this sphere, so we created a scientific department to cover these topics, because there are many engineers say: "Actually there is solvable problems."

You know, I'll start with that. That my understanding many of these topics, probably enough limited, but I think if I find out about something—or someone will contact me, will I see it myself? problem—and I'll think: "Hey, some principles design, which we use or over which now we are thinking, actually good suitable for this tasks."

Me too I admire these things, but I think it's more difficult. I... I don't know. I think that science, especially empirical whether it is applicable, has very long cycles feedback.

Interviewer: Ahem. Yes, it is.