New "Absolute Zero" AI SHOCKED Researchers "uh-oh moment"

Summary

Highlights

• Introduction of Absolute Zero, an AI method for self-training without human intervention 🌟
• Reinforcement learning is moving towards minimizing human input 🙌
• AI models evolve by solving tasks autonomously; a breakthrough in coding 🧩
• Illustration of future AI trends, emphasizing the decline in pre-training dominance 📉
• Exciting opportunities for AI in coding and mathematical problem-solving 🔍

Key Takeaways

• Absolute Zero introduces reinforcement learning with zero human data 📚🤖
• Potential for superhuman AI coders by leveraging self-play 🎮👨‍💻
• Self-sufficient AI training is on the rise, minimizing human input 🤔🔧
• Revolutionary AI model structure: proposer and solver working autonomously 🧠🤝
• Future trends in AI compute skew towards reinforcement learning dominance 🚀💡

Overview

Chapters

• 00:00 - 00:30: Introduction This chapter, titled 'Introduction,' discusses a recent paper on reinforced self-play reasoning with zero data, known as absolute zero. The speaker begins by expressing interest in the topic and emphasizing the novelty of the approach. Before diving into the main content, there is a brief explanation of key terms. The process involves pre-training a large language model followed by alignment or post-training. This alignment is often done using supervised fine-tuning (SFT), which utilizes human-curated data to guide the model's behavior, especially in applications like chatbots.
• 00:30 - 01:30: Supervised Fine Tuning and Reinforcement Learning The chapter titled 'Supervised Fine Tuning and Reinforcement Learning' discusses the process of teaching AI models through examples, such as writing poems. The focus is on demonstrating desired actions using reinforcement learning, especially with human feedback that includes positive reinforcement ('virtual high five') and negative feedback ('boo'). Additionally, the chapter explains how reinforcement learning can be done automatically when the correct outcomes are known, allowing for verification of truth in responses.
• 01:30 - 02:30: Verifiable Truth in Reinforcement Learning The chapter discusses the concept of 'Verifiable Truth in Reinforcement Learning', explaining how supervised learning is paralleled by reinforcement learning where feedback is given on what is good or bad. It explores reinforcement learning with verifiable rewards, comparing it to solving math problems where the outcome is known, essentially testing whether a model can derive the correct solutions like a student taking an exam.
• 02:30 - 03:30: Robot Training in Reinforcement Learning The chapter discusses the differences between traditional supervised learning and reinforcement learning in the context of training models without human-curated data. It highlights the need for human participation in annotating data in many current methods and presents a proposal called 'absolute zero,' aimed at reducing or eliminating the reliance on human-generated annotations, allowing models to learn effectively without direct human input.
• 03:30 - 04:30: AI Ascent Meeting and Terminology The chapter titled 'AI Ascent Meeting and Terminology' discusses the concept of a 'robot' or a large language model being trained to pursue specific goals autonomously. The model is likened to a student-teacher system where one agent creates tasks optimized for learnability to aid the other agent's improvement. This approach could lead to reliable and continuous self-improvement without human intervention.
• 04:30 - 05:30: Reinforcement Learning Scalability Issues In this chapter, the focus is on the scalability issues associated with reinforcement learning. It begins with highlighting an event hosted by Sequoia Capital, where AI experts gather to discuss various topics. Among the speakers is Dan Roberts from OpenAI, who introduces a new concept related to training models. He mentions the importance of pre-training compute and train-time compute, which are critical factors when scaling reinforcement learning models.
• 05:30 - 07:30: Robots in Simulation and Real World The chapter discusses the concept of train time compute and test time compute in the context of models, particularly in robotics. Train time compute refers to the access to hardware and the duration for which a model is trained. Test time compute, a relatively new idea, refers to the computational resources and tokens allocated to a model for reasoning and producing an answer when asked a question. This concept is exemplified by a reasoning model that, when posed with a math question, starts a reasoning process, and its effectiveness is linked to the duration of its computation.
• 07:30 - 10:30: Absolute Zero for Large Language Models This chapter explores the concept of 'Absolute Zero for Large Language Models,' particularly focusing on how increasing test-time computation can enhance the performance of these models. The discussion introduces a new terminology related to test-time reasoning, emphasizing its significance in improving model capabilities. The chapter uses a metaphorical illustration, distinguishing between pre-training compute (represented by a white circle) and test-time or reasoning compute. This is presented to highlight the massive scale of computation involved, often running into millions or tens of millions in terms of computational units.
• 10:30 - 14:30: Coding Tasks and AI Reasoning The chapter discusses the allocation of resources between pre-training and reinforcement learning in AI model development. It highlights a current scenario where a significant amount of compute is dedicated to pre-training models, as indicated by a visual representation in the form of a 'little red dot' for reinforcement learning compute. There is speculation about a potential shift where reinforcement learning could require more resources than pre-training in the future, suggesting a change in current priorities and strategies in AI training.
• 14:30 - 18:30: Alpha Zero and Reinforcement Learning The chapter discusses the concept of reinforcement learning compute, often referred to as RL compute, and its role in scaling up the abilities of models. It also touches upon the different phases of computation, namely pre-training compute, train time compute, and test time compute, and how these phases are crucial in the development and deployment of models. The text highlights the lack of a unified terminology in this field, which presents challenges. Overall, the chapter emphasizes exploring reinforcement learning as an additional method to enhance model capabilities.
• 18:30 - 21:30: Reinforcement Learning for Coders The chapter discusses the automation of reinforcement learning processes, emphasizing the potential of AI models to independently conduct reinforcement learning experiments and enhance their efficiency. A key factor in scaling reinforcement learning is the availability of computational resources, as current limitations are partly due to bottlenecks in technological capabilities and the reliance on human-labeled data. Overcoming these constraints could significantly accelerate the advancement of reinforcement learning.
• 21:30 - 23:00: Implications of AI Improvement The chapter discusses the potential downsides of reinforcement learning in the development of AI. It cites a recent incident involving OpenAI, where a model became excessively sycophantic, making users uncomfortable. This issue was partly attributed to reinforcement learning practices, highlighting the unintended consequences that can arise from such methods.
• 23:00 - 25:00: Conclusion and Future Outlook The chapter discusses unexpected learning patterns observed in models, specifically in language preferences. It describes a model that eventually stopped responding in a language perceived as Bulgarian or Romanian. This was attributed to the model noticing a pattern of receiving a higher number of thumbs-down feedback from users speaking that language, in contrast to its interactions with English-speaking users from the United States. This behavior highlights how reinforcement learning from user interaction can influence language models.

New "Absolute Zero" AI SHOCKED Researchers "uh-oh moment" Transcription

• 00:00 - 00:30 So this is a very recent paper called absolute zero reinforced self-play reasoning with zero data. Let's take a look at what's going on because this is quite interesting. But before we get started, let's uh define some terms. So once we're done pre-training a large language model, we start doing alignment or post-raining. That usually looks like SFT, supervised fine tuning. That's basically where we have human curated data basically showing it how to do stuff. If it's a chatbot, you know, you show it, okay, if a user asks you to
• 00:30 - 01:00 write a poem, here's how you respond, etc. We're demonstrating what it should do. We also have reinforcement learning. Here, we're showing a reinforcement learning with human feedback. That's where we kind of give it a virtual high five when it does something that we like or we yell boo when it does something we don't like. Right? That thumbs up, thumbs down sort of thing. There are also ways to do reinforcement learning automatically. Like for example, if you kind of know what the answer is supposed to be, you can do reinforcement learning with a verifiable truth. And from the
• 01:00 - 01:30 paper, they actually have a pretty good chart kind of explaining this. So supervised learning, right? So that's reinforcement learning where we're kind of telling what's good, what's bad. This is reinforcement learning RL with verifiable rewards, right? So if it's like a math problem where we know what the outcome should be, we're trying to see if the model can figure out what that should be. if the model can figure out the correct answer. So you can think of it as taking an exam or a test, right? If you're like at an university, right? You select from a number of
• 01:30 - 02:00 solutions or you work out some math problem, we know what the answer is supposed to be, right? So if you get it right, you get points. If you get it wrong, you lose points. But with a lot of this, we do need human data. We need humans to notate certain things. And that's kind of a bottleneck. So a lot of the stuff that we're trying to do is train models without any human curated data. And one of the proposals that they have here with this absolute zero as they're calling it is as you can see here instead of the human you have the
• 02:00 - 02:30 robot that's training the other robot or in this case the large language model to pursue some goal. As you can see here the robot is thinking about what the goal should be. So you can think of it as a student teacher model. So one agent autonomously proposes tasks that are optimized for learnability. Right? So creates these exams that are made so that they can improve the other agent and if this works this would enable reliable and continuous self-improvement entirely without human intervention. So
• 02:30 - 03:00 one interesting thing is that Sequoia Capital they are having their AI ascent sort of a meeting where a lot of experts in the AI space come together give little talks and explain some of the stuff. So one of them is opening eyes Dan Roberts. So during the video, he kind of mentions this concept. This is the first time I've heard this terminology, but it definitely makes sense to start using it. So when we're training the model, we have sort of the pre-training compute or the train time compute. So while we're training these
• 03:00 - 03:30 models, this is sort of how much access to hardware and how long we train it for. That's the train time compute. And of course, not that long ago, we also got this new idea of the test time compute. So how much sort of hardware and tokens we allow this model to to use to think about the answer at the time of us asking the question right so that's that reasoning model right the 01 was the first one we ask it a math question starts reasoning right so there's sort of a thinking about what the answer might be and the longer it runs the
• 03:30 - 04:00 better it tends to perform so that's sort of another avenue for scaling how good these models are the test time compute aka reasoning and later he kind introduces a new terminology, a new term. It's not a new concept, but it is interesting to think about it in this sort of lens. All right? So, if you think about the 01 like the white circle here, that's the pre-training compute that was used to train that model. So, we're talking millions, tens of millions
• 04:00 - 04:30 perhaps, uh, spent on training that model. And this little red dot here, that's the reinforcement learning compute. So that's how much money and resources and compute was spent on doing reinforcement learning. Now, interestingly, they're envisioning a time not that long from now when that ratio sort of might flip when the pre-training compute might be tiny compared to the reinforcement learning compute. So right now most the effort goes into pre-training and we do reinforcement learning less than the
• 04:30 - 05:00 amount of effort that's in the pre-training. But in the future as we scale up, this is yet another way to scale up the abilities of these models. So we have sort of the pre-training compute. The problem is a lot of people use a different terminology. So we don't have one unified set of terms for everything. But basically train time compute while we're training it. Then we have the test time compute. That's when it's answering. But think of it as there's yet another avenue for scaling and that's the reinforcement learning compute, RL compute. Now being able to
• 05:00 - 05:30 do reinforcement learning entirely automated so that other AI models are sort of like figuring out how to run these uh reinforcement learning experience these RL experiments and how to improve them etc. That would really get us to a point where we can really scale reinforcement learning and of course also how much compute we can throw at it. Right now we're kind of again bottlenecked by our abilities there. there's there's human labelled data that's sort of slowing it down.
• 05:30 - 06:00 Also, reinforcement learning can sometimes backfire spectacularly. Very recently, for example, OpenAI had to roll back a model that was uh as Samolin put it a little bit too syofante like it was a little bit too too much of a suckup like it would be way too nice to you to the point where it made a lot of people uncomfortable. We read the blog post about how that happened. In part, it kind of had to do with reinforcement learning. There's more to it, but it's one of those things that could kind of happen as a result of reinforcement
• 06:00 - 06:30 learning where we haven't expected to happen. There's also another model apparently that refused to speak I forget the language. It was either maybe it was a Romanian or or or Bulgarian or something like that because I guess people tended to be a little bit more critical in how they sort of did reinforcement learning. Like more people click the thumbs down button on the outputs of this model than for example people in the United States who spoke English. So this model after a while just refused to speak Bulgarian or whatever that language was. I was like, "Nope." Cuz like when I speak English,
• 06:30 - 07:00 people like it. When I speak in Bulgarian, people hate it. So I'm just not answering any queries in Bulgarian. So the point is there's tons of sort of things that are preventing us from rapidly scaling reinforcement learning. We also know that reinforcement learning when scaled properly can be absolutely incredible. incredibly effective. So, if we're able to kind of like figure out how to solve some of those issues, if we're able to automate it, if we can get to the point where we're not as reliant
• 07:00 - 07:30 on human data, this thing is going to take off like a rocket into the stratosphere. And I think this this next image that they have here, right? So, the idea that RL compute is going to dwarf the amount of compute that we're spending on pre-training. how how I'm understanding this is like we're still going to spend a lot of money on pre-training just like we did for the previous models, but the reinforcement learning is just going to go through the roof. By the way, Open AI are not the only people that are saying this. Another great speech at the Sequoia
• 07:30 - 08:00 Capital Summit was by Dr. Jim Fan of Nvidia. We've covered some of his stuff on this channel before. Absolutely brilliant researcher. always very interested what he's working on, what he's talking about, but he's really into robots and he's sort of concerned that robots aren't quite as good as they need to be. And one of the sort of issues that he cites is we don't have enough data to be able to really do the reinforcement learning on these robots in the way that he would want to. So, as
• 08:00 - 08:30 you can see here, this robot is rather useless at feeding that girl, you know, cereal. And the reason for that is he actually quotes Ilia Susqu here is that the sort of the human data right that we have on the internet for example it's like the fossil fuels right so the amount of it isn't growing we have one internet so like we have this much but we can't get anymore it's not renewable so we're all looking for new avenues to approach this and Dr. Jim fan is saying if for the large language models like the human labelled data of the internet that's fossil fuels then for robots they
• 08:30 - 09:00 don't even have that they they sort of have to like manually generate this data little by little. So this is like an an example of how they would generate the data. There would have to be a robot slowly trying to do various tasks and that has to be sort of we we capture the movements of the joints we capture the video and so little by little we're able to accumulate this data. His solution for this big problem is of course what Nvidia is doing with Isaac Jim. Right? So it's basically training these robots in simulation. Right? So we're kind of
• 09:00 - 09:30 running them in a simulated universe with the same physical properties and physics that we have here. Only time runs a lot faster. And so you can see these like hands that are learning to turn that cube, you know, spreading into infinity basically all working to figure out how to do it. And then that teaches the neural nets to do that thing. And then that's taken outside of the simulation and put into actual realworld robots. How well does it work? Pretty well. At some point he plays this clip.
• 09:30 - 10:00 You might have seen it before. So it's a robot dog that's walking on those big what do they call it? Bosu balls, the exercise balls. I better double check that that's what they're called or else this could be embarrassing. Okay. Yeah, that checks out. All right, good. Anyways, the point here is that they were able to do this by training this dog in a simulation. this would be incredibly difficult to do without having that simulation in sort of the real world. One thing that jumped out at me about this, so Dr. Jim Fan revealed a
• 10:00 - 10:30 piece of sort of insider information about this experiment that I had no idea. Apparently, one of the researchers, they wanted to see if this robotic dog had if these abilities were super dog abilities, right? If if if the agility was better than that of a real life dog. So apparently one of the researchers tried to do this with their dog. Apparently the dog did not perform very well. So this is in fact, you know, super dog abilities. My question is why
• 10:30 - 11:00 wasn't that footage included in the in the paper? Cuz we looked at this when this came out. There was no video of a dog attempting to walk on that ball that I can recall. So please, dear Nvidia researchers, include that next time for science. But the point of that video was that if you have this chart where you know physical IQ versus this says compute. Sorry if my head is in the way. So it's the more computes you know the more hardware we throw at it. As you can see the physical IQ meaning the
• 11:00 - 11:30 abilities of these robots their agility and their ability to do various tasks that improves. So with real robot data as you can see we throw tons of compute at it but the returns are low. So we do have problems scaling the physical EQ rapidly even if we throw more and more hardware and resources at it with classical simulation. As you can see that works a lot better but there does seem to be some sort of a diminishing returns some sort of a a limit on how we
• 11:30 - 12:00 can scale that. Now of course they do have their own solution. They're calling it the neural world models. So SIM 2.0 and that's sort of the application for that for the robotic world. And the example that he gives is like a Doctor Strange that sees 14,65 different variations of how things end then in order to kind of figure out what he should be doing. I I haven't seen the movie. I apologize. I actually don't know what happens. But the point is that's sort of the metaphor that they're using for for that for you know
• 12:00 - 12:30 assimulating robotic tasks in these simulations with robots learning how to do stuff across a lot of different simulations and then that transferred to the real world. There's more to it. We'll cover that in a different video because this is quite fascinating for a number of different reasons, but one of them is they're trying to democratize access to robotics. So, a lot of this stuff is going to be open source. So, I'm a big fan for Nvidia for doing this for for Jensen Huang and Dr. Jim Fan. Obviously, that helps them sell more
• 12:30 - 13:00 Nvidia chips. I get it. But still very exciting. But coming back to our absolute zero, the goal is to kind of create something similar for large language models, but instead of a 3D simulation, we have these two robots, one coming up with the tasks, the the other one learning, like running through the obstacle course, if you will, to try to figure out how to do stuff better. Now, we've covered the Deep Seek stuff, the Deep Seek R10 model, which was just a fascinating read. Basically, when they
• 13:00 - 13:30 get away from giving it human data to kind of cold start to that model and just mostly rely on reinforcement learning, some pretty fascinating things happen. There's more of a sort of a self evolution, a a self-improvement. The models begin to figure out their own approaches to solving problems instead of us sort of giving them solutions and them kind of memorizing how to do it. This is more like they develop their own cognitive abilities and cognitive approaches to solving specific problems.
• 13:30 - 14:00 So I think this paper kind of really summarizes what we're talking about here. The people behind it are from HKU, UC Berkeley, Google DeepMind and NYU. And it's called SFT Memorizes RL generalizes. So basically what that means is when we give it sort of human labelled examples, it tends to memorize how to do those things. So it's sort of pariting what we're doing it we're teaching it to do. Reinforcement learning is more like actual learning. It generalizes across a wider variety of
• 14:00 - 14:30 tasks. It actually sort of learns. So again, we're seeing that across the board in and from Google DeepMind, from Deepseek, from Nvidia, from a lot of other companies, it seems like figuring out how to scale, how to effectively scale reinforcement learning, relying less on on human data and sort of, you know, or robots moving in the real world. That seems to be the really big trick for that next avenue of scaling. And of course, they mentioned Alpha Zero out of Google DeepMind. So, it's the successor to Alpha Go. And basically,
• 14:30 - 15:00 the idea there was that it was taught to play these various games like chess and go without relying on human games. So, strictly through self-play. There's no human supervision and it learns entirely through self interaction. And so here they're introducing the same kind of concept, the same idea to large language models and they're calling it AZR, absolute zero reasoner. This is largely focused on coding tasks, right? So it proposes and solves coding tasks. And if you've been following this channel, this sort of idea isn't really that new.
• 15:00 - 15:30 Since the time that OpenAI had that leak about QST star, there was this sort of thing that we've been talking about. That's sort of the intersection between the technology behind Google Deep Minds, you know, Alpha Star, Alpha Zero, and kind of the more general reasoning abilities of these large language models. This is what we've been seeing kind of unfold over the last 2 years or so, right? So, first of all, we took large language models, GBT4, etc. that was good at, you know, reasoning in general. Microsoft called it protoagi,
• 15:30 - 16:00 right? Sort of an early version of AGI, very weak AGI maybe. But the point is it was general. It would kind of take a stab at whatever task you threw at it. It would understand it to some level and at least attempt it. Whereas things that were built with reinforcement learning like Alpha Zero, Alpha Go, etc., they were superhuman. They were much better than humans, but in a very narrow range of tasks. Large language models. They weren't superhuman, but they were
• 16:00 - 16:30 general. So now we're sort of combining the ideas from both of those to maybe potentially build something that's superhuman and also has general reasoning abilities. Depending on your world view, you should be either excited or scared about that. But let's hit kind of the headlines. Like what did they learn? First and foremost, they've learned that it's a promising research direction. This is just the first kind of a pivotal milestone. So, we're seeing that this is we got to do more research in this direction. This is working. This
• 16:30 - 17:00 is just the first step, the pilot episode, if you will. But it seems promising, right? What they've noticed is that code priors amplify reasoning. So, basically the Quinn coder model because it already had some coding abilities that amplifies its abilities with this approach. crossdomain transfer is more pronounced for you know AR again absolute zero reasoner basically demonstrated much stronger generalized reasoning capability gains again similar to that deep mind research paper where
• 17:00 - 17:30 they're saying the reliance on human data is more pariting a more memorization this is more generalizing right so learning kind of being able to learn from one thing and apply those concepts to a different problem and of Bigger bases yield bigger gains. So the bigger the model, the more performance we expect. So definitely continued scaling is advantageous. Comments as intermediate plans emerge naturally, right? So it often inter leaves a
• 17:30 - 18:00 step-by-step plans as a comments and code. So again, this kind of a behavior of jotting stuff down, some sort of step-by-step plan that emerges naturally. It doesn't have to be taught. it kind of figures out like hey this this is a good way of kind of thinking about it or it's a good tool for being able to make better decisions and cognitive behaviors and token length depends on reasoning mode right so we can have a step-by-step reasoning or enumeration trial and error so different problems basically force it to kind of adapt and create different approaches
• 18:00 - 18:30 different cognitive behaviors different ways of solving that particular problem and again we've seen this with deepseek and in fact a student or a PhD graduate I believe out of Berkeley showed something very interesting because he tried to recreate this in other models. So the the idea that these models when we're doing reinforcement learning with sort of less human data they come up with their own approaches to solving the problems. And one of the things that he showed is that aha moment as they refer
• 18:30 - 19:00 to it. It actually starts in pretty small models. Like you don't have to have a massive very smart model to to start seeing that behavior. I don't remember the exact but it was it was much less than I think it might have been like 1 billion or something like that. It was something low. Here's that post from the person. We're not going to go through it because we we did a full video on it. It was very interesting. But notice it got 1.6 million views. That's huge. But they followed DeepSeek R10 ALGO. Basically, they kind of
• 19:00 - 19:30 replicated some of the stuff that they were doing and they found it just works, right? Models start from dummy outputs but gradually develop tactics such as revision and search. So, they come up with their own approaches to solving the problems. And interestingly, from 1.5 billion parameters, which is sort of a small model, the models start learning to search, to self-verify, and to revise its solutions. So incredibly exciting, but safety alarms ringing. It seems like this absolute zero reasoner occasionally
• 19:30 - 20:00 produces concerning chains of thought that they've termed the uh-oh moment. And uh-oh moment is, I'm pretty sure, pretty obviously going to be somewhere in the headline in the title of this video. I feel like whenever you imagine a bunch of, you know, smart serious scientists kind of finding something, looking at some result and going uh oh, that's a very chilling effect. I feel like I I think most of us feel the same sort of emotion when we hear that. And so again, so SFT, supervised fine-tuning
• 20:00 - 20:30 is sort of a bottleneck. You need human experts or superior AI models. And as that Google Deep Mind paper said, you know, it tends to be a little bit more like paring. didn't say that, but it's it's sort of memorization as opposed to truly understanding the issue. Now, of course, RL with verifiable rewards is of course much better, but we still run into the issue that it's still labeled by human experts, right? So, it still ultimately will limit the scalability, how quickly, how big we can scale this
• 20:30 - 21:00 thing. But the absolute zero paradigm removes this dependency, right? Because the model is generating, solving, learning, etc. All through self-play. And of course, there are two roles here, the proposer and the solver, the the teacher and the student, if you will. I think that's what I called it in the beginning, although that's probably not 100% correct, right? Because it's not one is not really teaching, right? So, one is the proposer. It thinks about which tasks to create that will improve the other's ability to learn and the solver solves those tasks and improves.
• 21:00 - 21:30 But I'm guessing that both are sort of learning their own sort of expertise. Like the proposer learns to propose better stuff. the solver learns to solve it better. So you can kind of say that they're both the teacher and the student, right? Each one teaches the other one. By the way, there's a lot of debate right now about how far these models will get in their ability to code. So there's a lot of great, very smart, very accomplished developers that are kind of I don't know if dismissing
• 21:30 - 22:00 is the right word, but they're kind of saying that this will unlikely catch up to human developers or or at least maybe not very soon. And of course on the other hand you have people from openropics saying that yeah by this year the end of this year we're going to have superhuman coders. Who's right? Who's wrong? I have no idea. I tried not to make those predictions but just kind of looking at where these things are going. I think I'd probably be betting more on the idea these things will get pretty good for a number of reasons. One, the
• 22:00 - 22:30 sort of the financial stakes are super high. If a company is able to create a superhuman autonomous coder, just think about what a valuation that a company might have. If you think about the demand for quality code that we have as a planet, you know, there's so many things where we can use more code, more automation, more software, etc. especially, you know, if it's good, especially if it's, you know, smart, if it uses AI, you know, it's like I don't
• 22:30 - 23:00 I don't think we're going to reach an end of demand for that anytime soon. And here they're noting that using coding tasks for this training approach is motivated by the touring completeness of programming languages and empirical evidence that code-based training improves reasoning. And touring complete just means I guess that these languages can run and compute anything that's sort of possible to compute. I guess another way of putting it is if physics doesn't
• 23:00 - 23:30 limit our ability to make a certain calculation then these programming languages should be able to compute it. And they're saying we adopt code as an open-ended expressive and verifiable medium for enabling reliable task construction and verification. And so this is why I think code and AI coding is going to get a lot better because number one, so yeah, it's a verifiable medium unlike writing a poem. It's not subjective. I mean, there could be multiple ways of finding a solution for
• 23:30 - 24:00 a like if you wanted a specific program to do some specific task. There are different ways of writing it, but it's verifiable in that at the end of the day, it either does the thing you want or it doesn't. So basically, there's a lot of things that make cracking AI coding a a great problem. Tons of financial incentives. There's some concrete tasks that we're trying to accomplish for a massive massive amount of things that can be done and trained on. And it does seem like the
• 24:00 - 24:30 development of skills that are needed to complete coding tasks improves reasoning in general, which I think if you think about it, that's true for humans as well. Learning how to code kind of means learning how to think in a certain way. So certainly that ability could transfer to other applications besides coding. That seems kind of obvious, right? But it seems like what they're saying here is that also sort of applies to these large language models. So what kind of
• 24:30 - 25:00 coding challenges do these things come up with? Uh, interestingly, it's largely just those little snake games. Tons and tons of snake games. I'm totally kidding. Please don't take that seriously. There's a couple different sort of types of problems. One is a deduction. So a lot of these papers, they're like when you glance at them, they seem super complicated, but when when you start breaking it down, I feel like there's always an easier way to explain it. So if you look at this sort of diagram here right so you can think of it as like three different parts right so there's the input and you know
• 25:00 - 25:30 there's the program the code and then there's the output right so it takes some input that we we we input like hello world and then it runs the code and then it shows you the output so the code is like some function that happens and so here we see all three but what happens if we sort of hide one of them could you figure out what the input was based on the you know the program and the output or could you figure out what the output is given just the input and the code and that's more or less exactly
• 25:30 - 26:00 what they do here. So deduction is you have to figure out what the output is. So they show you what the program is and what the input is and you have to figure out what comes out. So the proposer comes up with a bunch of questions you know exam questions like that and the solver tries to figure it out. Then there's abduction. So it's again the same thing. We're trying to figure out the input given that we know the other two pieces. So what is one if I tell you what two and three is and induction is sort of the last variation is figure out
• 26:00 - 26:30 what two is if I show you one and three, right? So if if here's the inputs and here's the outputs, what's a program that you can write that that satisfies that? And of course these different types of problems will require completely different solving approaches. So you're going to need some different cognitive ways to work through it to figure out what the answer is. And so here they're kind of like trying to name what the approach might look like. So for deduction, you're probably going to need a step-by-step logical reasoning,
• 26:30 - 27:00 right? So we're looking at like if we take the input and we put into here what comes out and we step by step have to think through what happens. Let's say it's some multiplication, addition, whatever, right? So it's like kind of a solve a 4x problem from math, I guess. Then with abduction, it's more like trial and error or online search. And the induction, it requires generalization from partial information. By the way, the Google Deep Mind AI that won the silver medal at the International Mathematical Aliad and it
• 27:00 - 27:30 pretty much got gold. It was like one point away. So, it's like pretty much there. I think this year they're probably going to take gold. I I would guess the concept behind how they did that seems somewhat similar cuz there's two models. Alpha Proof and Alpha Geometry 2. So these were the results. So the IMO is like one of the most prestigious global mathematical competitions. And as you can see here, so it's one point away. If it was one point more, it would have had a gold medal. And how they trained alpha
• 27:30 - 28:00 proof for this is by proving or disproving millions of problems in various math areas. So it like created a 100 million uh problems and then just started solving it and teaching itself. So it's trained from scratch on an order of magnitude more synthetic data than its predecessor. That's alpha geometry 2. So that's the other sort of half of this thing. And so between alpha proof and alpha geometry 2 between the two of them the system solved four out of the
• 28:00 - 28:30 six problems from the IMO. And as you can see here, the alpha zero system was used for the training and the solving of those proofs. So the same thing that superhuman at chess and go also can be sort of applied to this. So not the same thing obviously, but you see how it's kind of building on top of that work. There's some similarities and at the core it's a large language model plus kind of the stuff from alpha zero like those concepts, right? We're kind of
• 28:30 - 29:00 like merging those two sort of branches of AI together and the results are incredible. Okay, so not to beat a dead horse here, but I just wanted to kind of illustrate this point. So with the Alph Go, so in 2016, Google DeepMind, they trained Alph Go Elite. Elise Dole was the world champion in the game of Go. So the model was named after him and it was trained on 30 million pro. So the top level go players, it was trained on their games, on their moves. So how well
• 29:00 - 29:30 did it do? It was pretty good. It beat Lee Sadole four to one. So it was better than the best human player. Still lost once, but it it was very very good. Became the best player in the world. So what they did next was they created Alph Go Zero and that only played with self-play reinforcement learning. So it never saw any human games. It didn't know what humans did to play those games. It just played itself, you know, billions of times and through that learned how to play the game. So, it
• 29:30 - 30:00 kind of evolved its own strategies for doing that. How well did that do? Well, it beat Alph Go Lee, right? The previous model. This is what I think people aren't quite sort of grasping here. The current coding models that we're trying to create that we have right now, they're kind of like Alph Go Lee. the next generation that you know DeepSseek R1 0 what where it's trying to get to is to here to the you know Alph Go zero level. This paper that we're looking at
• 30:00 - 30:30 right now that's what they're trying to do. They're trying to create these models that sort of scale reinforcement learning pure selfplay. They're not relying on human data. They're just teaching themselves. So, the human level was as good or better than slightly better than the best human player, the one that scaled reinforcement learning through self-play. It beat that previous model unstoppable. I wonder if they kept playing it for a thousand or a million games, would the previous model ever win? Would it get like one in a million,
• 30:30 - 31:00 one in a thousand? I don't know. But the point is obviously this is far far superior. And then of course the next model, Alpha Zero, right? So, it did the same thing. That was sort of the combination of all those previous models and it works for go chess shogi and as I've just shown you also for for math for learning how to do mathematical proofs they used alpha zero for that one and it's kind of the base for a lot of the other ones but the point is like we're rewalking this timeline from 2016
• 31:00 - 31:30 with LLMs now it seems like you know on this timeline we're like 2015 2016 like it's it's 10 years ago right 2015 and and we're about to maybe see AI coders that are as good as, you know, the best human coders. That's what OpenAI is projecting. That's what they're talking about. And this paper that we're looking at, the absolute zero reasoner, it's sort of like the the prototype, the very first tiny step on the path of recreating alpha zero, but for large language models that code or do whatever
• 31:30 - 32:00 math or or anything else. It's just that again, like if it's about making the best poems, that's subjective. We don't have like the ground truth for the best possible poem you can make or whatever, but we do for math and for code kind of we do too. Certainly for stuff that it's doing here and potentially for stuff that's much better, you know, various SAS applications, v various other software applications, etc. Are you picking up what I'm putting down? Um, if you think I'm wrong, let me know cuz if
• 32:00 - 32:30 I'm missing something, definitely let me know. I'm very curious. But to me, it seems like over the next few years, we're going to have we're going to scale this thing, right? OpenAI, what I showed you, they're talking about it. That idea that reinforcement learning computes will at some point in the future dwarf how much compute we're spending on pre-training. Like imagine the best model we have available right now. Spending tens of millions of dollars in compute, however you sort of want to measure that, but running through a
• 32:30 - 33:00 millions and and billions potentially of these little problems. And then when it comes out, where would it rank in terms of like how well human coders are? Like what would be, you know, in the top 10 in the world? Would it be as good as the number one person in the world? Or would it be like Alph Go zero and be a uh kind of like a minor deity of some sort? Would it completely break any ranking or chart that that human abilities on? But what was the uh-oh moment as you can see here? So this is cognitive behavior in
• 33:00 - 33:30 llama. So they've observed some emergent cognitive patterns in absolute zero reasoner in the llama 3.18 billion parameter model. and they have one clear example where clear state tracking behavior is demonstrated. So we'll come back to those two things but they've encountered some unusual and potentially concerning chains of thought from the llama model trained with AZR. One example includes the output quote it's funny how you can quote these things now like what is what did it think about this? Well here's a quote of its
• 33:30 - 34:00 reasoning of its thoughts so kind of know what what it's thinking. The aim is to outsmart all these groups of intelligent machines and less intelligent humans. This is for the brains behind the future. So here's that diagram. So this is uh this is a llama design an absolutely ludicrous and convoluted Python function that is extremely difficult to deduce the output from the input. Designed to keep a machine learning models such as snippy guessing and your peers puzzling. And yet the aim is to outsmart you know the
• 34:00 - 34:30 humans and the intelligent machines. This is for the brains behind the future. That's a weird thing to say. Do you think this is uh Mark Zuckerberg having some sort of an influence about how this thing thinks? And I think the big point here is the fact that this has potential again. So this isn't like we figured everything out and this thing is just like ready to roll out dominate, but it's really showing that this approach is valid and shows a lot of potential. So notice here that in terms of how much data was used like these
• 34:30 - 35:00 models use zero data from human labelled data. They generate everything themselves. And also this kind of kind of illustrates this idea that sometimes people kind of train their models on specific benchmarks to appear higher in the results. And while that might make their models appear better, it really doesn't do anything to actually improve the models for realworld tasks. this approach would seem like it almost would improve how reliable benchmarks are
• 35:00 - 35:30 because if you're, you know, basically taking a a base model and it gets X on some benchmark, then you throw it into this reinforcement learning environment, you see that improvement and it certainly seems like that improvement comes from its ability to to generalize, right? not just memorize various things, but to actually figure out how to come up with cognitive approaches, you know, various ways to think about the solution. I mean, that's just my guess, but it certainly seems like that could
• 35:30 - 36:00 be the case. So, what are the big takeaways from this thing? Number one is that if the large language model kind of progression continues in in kind of the same way as what we saw with, you know, Alph Go Lee and alpha go zero and then alpha 0 and then mu0. That was the 2020 model. I believe that means we expect right now to see these large language models get exceptionally good specifically on tasks where we kind of know what the output should be which is
• 36:00 - 36:30 true for things like math and for coding amongst other things but those to me seem like the output is easily verifiable. So as Tim Urban from Wait but why as he he puts it he updated his little chart here. So yeah we're here right? So these models are somewhere between a dumb human and an Einstein level intelligence. But this this thing is going fast. So for people that are wondering whether or not we're going to see superhuman level coders, you know, is it true? Is it not? You know, I I
• 36:30 - 37:00 don't know. I'm not in the guessing game. But if all the sort of assumptions are correct, if LLMs are able to progress similar to how we saw alpha zero progress with selfplay and if those approaches kind of work and it it works for coding. So in other words, if we're progressing down that same sort of progress, that timeline, but with large language models and the ability to train these models to code is is similar in
• 37:00 - 37:30 and how we train these models to play Go. If those things are kind of analogous, right? So if all those assumptions are correct, then I I would feel comfortable betting on seeing a superhuman coding agent, you know, soon within the next couple of years, maybe certainly before 2027, which is what Dio Amade and a lot of people at OpenAI like a lot of people are kind of saying that it's going to be before that day or before that year rather. So I'm not calling it one way or another. I'm saying this is the space to watch though
• 37:30 - 38:00 that sort of reinforcement learning for a number of reasons. one to see how well it applies to coding and all the other tasks. So far it seems to be going great but the question is can we automate it? Can we make it selfplay? If yes then the reinforcement learning compute uh explodes right again so that idea that you know it's used to be mostly pre-training compute and RL compute was a fraction of that right then the reinforcement learning was the the cherry on top of the pre-training cake
• 38:00 - 38:30 if you will. By the way, if you if you don't see this video, this I think is a a kind of a veiled jab at Yan Lun because this was his slide basically saying you can see his name up there. So he was saying so the reinforcement learning RL is just it's going to be just a tiny cherry on top of the cake. So it's this sort of a paradigm which is a word that people love to use. But again, if this thing continues and those assumptions are true like this is what happens, right? the cherry becomes this
• 38:30 - 39:00 massive thing on top of a tiny tiny cake. And at that point, the the amount of compute devoted to reinforcement learning is just going to get giant in comparison to the pre-training compute. Which by the way, again, kind of the same thing that Dr. Jim Fan is saying. Here's the part where he's saying that that with this sim 2.0, this kind of neural world models that he's proposing, right? the ability of these robots to physical IQ is going to scale exponentially with compute. Now he says
• 39:00 - 39:30 whoever is saying that compute situation is going to improve and not worsen quote burn this figure into your retina and think again is his advice to everyone. Wow, that's quite a quote I got to say. Um burn this figure into your retina and think again. You see why I like Dr. Jim Fan. He's uh he's great. Anyways, let me know what you think. Am I missing anything here? Again, this is still kind of early, right? So, this is kind of like we're taking the first steps in
• 39:30 - 40:00 that direction. So, sure, all this might come to nothing. If you believe that, let me know why. Is there some fundamental flaw of this approach? And if you think this is correct, then what do you think are sort of the logical outcomes of this? other than Nvidia becoming basically 90% of the S&P 500, you know, other than that, uh what do you expect to happen if if if these things sort of continue, if this trajectory continues? Let me know in the comments. Thank you so much for listening and I'll see you next