The Breakthroughs Needed for AGI Have Already Been Made: OpenAI Former Research Head Bob McGrew

0:00I think what's really changed is that now that you have LLMs, you have this language interface to the robot so that now you can describe the tasks much more cheaply and you have really strong vision encoders that are tied into that intelligence. So that gives the robots really a headshot at doing generic tasks. So we spent years solving one specific problem teaching a robot to manipulate a Rubik's cube. And now a company like let's say physical intelligence can spend months solving a huge variety of problems like laundry folding and cardboard and packing egg crates. And that's something that they can only have because they're building on top of existing frontier models and you know the entire tech and research stuff that we've built over the last 10 years.

1:05Welcome to Training Data. Today we're excited to welcome Bob McGrou, former Chief Research Officer at OpenAI for a fascinating look behind the scenes of Frontier AI Development. Bob talks about the trifecta we have in AI, pre -training, post -training, and reasoning, and explains why we may have already discovered all the fundamental concepts needed for AGI. You'll learn why he thinks agents will be priced at the cost of compute, hence eroding traditional economic modes, and why even proprietary data will become less valuable when infinitely patient AI can recreate alternatives. Plus, Bob shares his contrary intake on where startup opportunities really lie, and why robotics is finally having its moment after years of being too early.

1:50Enjoy the show. Bob, thank you so much for joining us today. Oh, it's great to be here. We're at a really interesting time in AI development. We have a beautiful new trifecta, pre -training, post -training, reasoning. Can you help us unpack what else is left? What alpha is there in each left? So I think we're going to continue to see capabilities increase. It's going to continue to feel like it's felt super fast, super exciting over the last, even five years. And I think it's going to continue feeling like that. There's not a wall here. But what is going to be different is that 2025 is going to be the year of reasoning.

2:25So it makes a lot of sense. Reasoning is a new technique. When you have a new technique, you know, there's often an overhang of compute, of data, of algorithmic efficiency improvements that you can make. And so that's something that, you know, if you look at just the incredible progress that we saw from 01 Preview back in September, and then six months later, we go to O3 in April. And then, you know, at the same time, we also see diffusion of reasoning from OpenAI, who we've been working on that for years, out to Google, DeepSeek, Anthropic, again, just in a few months. And so this is really the right place where every lab is gonna focus for the year.

3:09And just like a sort of a fun example of how low -hanging the fruit is right now. If you look at the most interesting difference between 01 preview and 03. 01 preview is not able to use tools. 03 can use tools as part of the chain of thought. And this is pretty obvious, right? When we were training 01, we knew that this was a thing that we wanted to do, but it was difficult to implement. It took time. And so, you know, that took six months to get, you know, done and released. The next step on reasoning is going to be a lot less obvious than that. It's going to be a lot harder. And so, you know, as reasoning continues to mature, we're going to see the overhang get eaten up and it's going to start being slower and slower to make progress.

3:51You said there's not a wall. I think there's this meme and the Twitter sphere right now that pre -training is hitting wall. Can you say more about that dynamic? Yeah, and I think that's great. That's a great question here because pre -training is not going away. But what we're seeing out of pre -training is that we are at the place where it's working really well and we're hitting diminishing returns. And so, you know, diminishing returns are baked in because the intelligence of a model is log linear in the amount of compute that you're using to train it, which means that you have to have exponential increases in compute to get each increment in intelligence.

4:26When you pre -train a model, that's a giant training, it takes all of your data center for, you know, a period of months. And when you go to pre -train the next model, you can't really do it on the same data center. You can rely a little bit on algorithmic efficiency to make it better, but fundamentally you have to wait until you get a new data center. And that's not a matter of like something you can do in six months the way you can do improvements in reasoning right now. That's something that takes years. So that doesn't mean pre -training as useless though because the real lever for pre -training in 2025 is improving architectures.

4:59So even though you're working on reasoning, you want to improve pre -training so that you can have better inference time efficiency or so that you can have longer context or better use of the context. And when you're doing that, you have to start back from the beginning, do pre -training on this new architecture and then go through the whole reasoning process again. So that's the role of pre -training now. It's still important. It's just doing something different in the pipeline. Can you help us unpack what's left in post -training? Yeah. So post -training is pretty interesting because both pre -training and reasoning are about increasing intelligence.

5:32And there's a very clear scaling law that you get where you put in more compute and you get out increasing intelligence and post training isn't like that post training is about model personality and

5:43Poet you know intelligence. It's a sort of a thin problem right if you can get better at it it turns out to be very generalizable and it applies to everything so you can work on math and you find that it makes you better at legal reasoning but model personality it's a thick problem. You actually need a lot of human effort to think about what makes a good personality. How do I want this agent to act? And it's much more of a training process like you would go through over many years of interacting with people. And now it's a very hard research problem to take that specification for what the agent is and turn it into an actual appealing personality.

6:20But when you think about post training, I think about people like Joanne Zhang, at OpenAI or Amanda Askel at Anthropic, who really spend a lot of time crafting these model personalities and they're not research practitioners. Interesting. Right? They're people with their product managers or they're people with a very deep understanding of human nature. And are there more legs to the stool? Well, okay. So I'm going to say something potentially controversial and I think actually there aren't. So I think if you go forward to 2030 or if you go forward to 2035 and you look back and you say, what were the fundamental concepts that you needed in order to create more and more intelligence?

7:01Maybe that's AGI, maybe it's something different at that point. I think you're going to come up with the idea of language models with transformers, the idea of scaling the pre -training on those language models. So GPT -1 and GPT -2 basically, and then the idea of reasoning. And sort of woven throughout that, increasing more and more multimodal capabilities. And I think even in 2035 we're not going to see any new trends beyond those. And the reason I think this is if you go back to 2020, so GPT -3 has just been trained. You know, imagine yourself sitting where it opened AI, we haven't released this thing, but we know something apoccal has happened.

7:40And you know, Daru, Emma, Dye, Ilyas, Satskever, Alec Radford, you know, we're all sitting there in the room looking at this thing. And it was fairly obvious internally what the roadmap was. We knew that at this point, going from GPT -3 to GPT -4 by increasing pre -training, was absolutely critical. We could see that we needed to increase multi -modality, ultimately ending in a model that could use a computer. We were beginning to make experiments with test time compute. And in 2021, after the anthropic people left, we really started developing the idea of reasoning at OpenAI. And it's funny actually, if sometimes my friends ask me after an anthropic release computer use, they're like, did you see that coming?

8:24And I was like, well, we were working on that together back before they left. One of the people that did that project went to anthropic and the other one went to open AI and developed operator. And it just took many years before the multimodality had matured enough to get to that point that was obvious to us way back then. And so that's why I think, you know, we're going to see from here on out, there's very important scaling. There's very important development and refinement of these ideas. That is extremely hard. It takes a lot of brain power. It's not going to be easy. But I think if we look back from 2035, we're not going to see anything new and fundamental.

9:04I think I'm right. I kind of hope I'm wrong. It would be a lot more fun if I'm wrong. But I think we'll have to see. That's a hot take. I'm glad we have it on the record. We'll see in 2035, that's amazing. I'm curious about reasoning. It seems to me that opening up I really leaned big behind this paradigm, probably before the others. Now everybody has reasoning models. What did you see in reasoning that caused you to lean so far in so quickly? Well, effectively it really was sort of this missing piece, where with pre -training, the model has an intuitive sense of how to answer the question. But if I ask you to multiply two five digit numbers, this is something that's completely within your capability.

9:47But if I actually do it right now, you wouldn't be able to do it because it is just natural as a human capability to be able to think about something before we answer to have a scratch pad to be able to work through a problem. And that is something that, you know, the initial models, even GPT -3 really didn't have. And so we began to see, you know, glimmers of this publicly, things like, you know, thinking step by step. And the idea of having a chain of thought that you could train, the model would learn itself how to guide a chain of thought, not just be guided by cloning from publicly available data on how humans think.

10:24That was very powerful. And we knew that it would be more powerful than pre -training because in fact, your thoughts are inside your head. They are not something that the model has access to. And so almost all the data that's out there is actually something that's just the final process. But you don't get to see that chain of thought. And so the model had to figure it out for itself. That's why reasoning mattered. You alluded to earlier that we probably still have to uncover more things in reasoning. Do you think we have a good sense of what those things are today or are we earlier in that R &D stage?

10:57I think at this point with reasoning, if you are at the call face, then you're seeing a lot of ideas and refinements of things that you can do. I think we've gone past the point where if you're on the outside, if you're not at a front -chair lab, you're probably not seeing them anymore. This is the same situation we saw where at one point, academic labs could make huge amounts of progress. Later, I would begin to see academic papers and I think, oh, they rediscovered this thing that we found a long time ago. And so, you know, now the level of effort that's being put into this, I think is actually quite intense.

11:37So there are definitely things to be discovered, but they're not sort of simple ideas that you and I could talk about. Cool. Switching gears a little bit, you tweeted recently about agents, I think a very, very interesting take, that agents will be incredibly powerful, but priced at the cost of compute due to competition. If that's the case, where do you see the opportunities in new startups and companies that are now building agents. Yeah. So I mean, I think that the thing about agents is people think, well, you know, I'm going to go develop an agent and they look at how much the job is worth out there by a human.

12:11So, you know, you want to develop an AI lawyer and you think lawyers get paid a lot of money. So I'll be able to develop an AI lawyer and the lawyers, I'm going to be able to charge huge amounts of money. 10 thousand dollars. Exactly. Exactly. Right. But the reason lawyers are expensive is because their time is scarce because there's only so many people who have undergone that training. But by the time you've made an AI model out of it, well, now there's effectively an infinite number of lawyers. And so it's not scarce at all. And maybe you, with your AI lawyer startup, will be able to have a lead over other people.

12:43But it's the same frontier model underneath. And some other startup can come in and compete that away. And so we should expect to see it priced at some opportunity cost over the cost of compute. You're changing. You now have a lot more supply, infinite supply of the highest capability intelligence in whatever domain you now have. On the one hand, there's a story where this is bad because startups can make money, but this is actually the future we want. We want services that don't require people to be extremely cheap. You want everyone to have access to a lawyer. What you want to be expensive and scares are things that are actually about personal relationships.

13:24So maybe we won't be asking the human lawyers to write contracts because agents will be doing that for us, but we'll be asking them for sort of deep advice on how legal challenges affect the detailed challenges that I'm facing in my business. And I think that's the world we want to live in. You think application companies will make any money selling agents though? Like, where would you tell us to invest? Yes and no. So just to back up for a second, people often talk about where does the value accrue in the stack, right? Is it at the model layer? Is it at the application layer? And if you look at the model layer, it's very competitive.

14:02Every company has a frontier model. Some of the frontier models can do things other frontier models can't, but by and large, they're all really very good. And if you're an enterprise, you can swap them out very easily. And beyond the frontier, you know, all the models that are answering the bulk of questions are distilled are very competitive. And so this isn't a very good business to be in when you consider the cost of training the models. So what's the point of training models in the first place? It's to give you an option. It's to give the frontier labs an option on the valuable places in the application layer that are coming up.

14:34So you know, chat GPT, that's a great business, right? There's a lot of competition over that. I think probably it's too late to replace chat GPT. Maybe not. You'd have to do something very different. coding another place where all the frontier labs are eager to see right now. I think you can compete with the frontier labs, but you want to do something that's different, something that involves more than just you know, you talking to your computer, you doing some sort of personal productivity task on your computer, something that involves other people, something that involves an enterprise. I think that's, you know, I think that the modes that you have for your business are going to be the same motes they always were.

15:13There will be network effects, brand, economies of scale. And so you want to find an agent that allows you to have those network effects, not just something that would be high priced out in the world. Are there particular domains that you think are maybe outside of the scope of what frontier labs want to innovate in and build in that you think are interesting and that you've been mulling about? We've got scientists, lawyers, research analysts, agentics, offer engineers. What are the domains have you been thinking about? I am, so personally, I'm very interested in robotics because I think robotics is something, I wouldn't actually say it's off the roadmap of the Frontier Labs right now, but I think it's something that's far enough away that to me, it feels like where AI was a few years ago.

16:02And so I think this is a very good time to be a company like skilled or a company like physical intelligence, or to start a new robotics company, maybe not one that's competing with those two, but something is doing something different, something on its own. I think it's at the end stages of being a research challenge and a matter of months or years, small digit years away from being commercialized. So I think that's really fun. Why now? What do you think has changed? Open AI famously had a robotics effort for a long time. What do you think has changed? Well, you know, so in between Palantir and OpenAI, I actually wanted to start a robotics company myself.

16:38And I got to the point of teaching a robot to play checkers from Vision back in 2016. Wow. Yeah, it was very cool. It could pick up the checkers pieces. It could pick up the checkers pieces and it could move them to a different place on the board. Nice. And my conclusion from this was that it was very fun and super cool and really far away from any form of commercialization. And when we pursued robotics at OpenAI, we didn't pursue it for commercial moment motives. It was really a demonstration of the power of machine learning and some of the ideas we had there later played into large language models.

17:12But I think what's really changed is that now that you have LLMs, you have this language interface to the robot so that now you can describe the tasks much more cheaply. And you have really strong vision encoders that are tied into that intelligence. So that gives the robots really a headshot at doing generic tasks. So we spent years solving one specific problem, teaching a robot to manipulate a Rubik's cube. And now a company like let's say physical intelligence can spend months solving a huge variety of problems like laundry folding and cardboard and packing egg crates. And that's something that they can only have because they're building on top of existing frontier models and the entire tech and research stack that we've built over the last 10 years.

18:04Yeah, I'm going to go back to this point you had on, you know, where is the value? And I really like to be framing that the foundation models kind of have an option on whichever parts of the application stack they want to own. How much of the application market do you think the foundation models will win? I think I would look at this like in a slightly different direction, which is if you're a startup, you know, where is it safe to play and where is it that you're going to get steamrolled by the Frontary Labs? And so I think I think the areas that that I think are safe to play in are areas where you have to understand something very deeply outside the model.

18:40And so I think a lot of enterprise really has this flavor. So for example, you know, Poundter AIP actually really fits this where it's, you know, it's not a model company, but it's something that sits outside the model that interacts with the rest of the business. There's another company I'm an investor and an advisor and called Distill that builds AI systems that allow you a business to sort of extract the context from within the business, feed that to the models, and then use that to make decisions. And so, you know, these are things that the Frontier Labs don't want to do. The Frontier Labs see business problems as how do I train a model to do something new?

19:22And if you look at all these enterprises, each one of those is a very small problem. It's not worth open AI or in Thropic's time to train a model specifically for each one of them. If you look at the problem and you think about what is the system that goes around the models And how do I use the models to sort of ease the context in and get the outputs out? Then suddenly that's one problem and I think it's a big opportunity. What are the specific use cases and problems that distill and palintea's effort solve for those enterprise companies? So a lot of times right now what you see is you're trying to automate some existing piece of work.

19:59And the easy cases are where that piece of work is in a regulated industry. And you're working on something like healthcare, maybe you're interacting with insurance companies. And you have a workflow that is extremely scripted where the company cares a lot about fidelity to that workflow. And that doesn't mean you can just say, hey, AI, go read the clinical guidelines and make these decisions. But with a process of transformation, you can get it to the point where the AI can do that. And that's sort of the low hanging fruit. And then the next level up though is that if imagine you're working on something that isn't a regulated industry or that isn't extremely scripted.

20:39And you want to ask someone, you want to automate some labor and sense process. Well, the first thing you have to do is make that legible. And if you go to someone and you ask them to describe their job, a lot of times, their manager doesn't know what they do. They don't even really know what they do. They can give you examples. But they can't say like, this is the workflow that I follow. because in practice, they don't follow a single workflow. And so I think that is what a lot of these problems look like. And for example, that's actually what Düsseld is doing is to work with companies, help them take the data they have, interview the people with AI, systematize all of that, and have it be something that an AI model can actually execute.

21:23That's really interesting. So this is also somewhat related to this other question I wanted to ask you about proprietary data. I was surprised to actually see you tweet this, but I very intrigued by this question that you posed, which was, how valuable will your proprietary data be compared to what your competitors infinitely smart, infinitely patient agents can estimate from public data? Can you impact that for us a little bit? Yeah, so, you know, a starting point for this is a few years ago, there was a lot of interest in training industry vertical specific models, you know, that finance companies would say we've got all of this data that no one else has And we're going to train a finance model on top of GPs or on top of Lama And it's going to be so much better and actually all of those were worse than the next generation of GPT because the power of intelligence and the ability to synthesize new information was bigger than the power of sort of memorizing the old information that you have.

22:21So that's I think what this theme looked like a couple years ago. But, you know, fast forward a year or two years. You know, now the story is I have all of this proprietary data, I've accumulated it over years. And in some sense, for a lot of times, you know, if that data is teaching the model of skill, or if it's meant to teach the model of skill, that data is sort of embodied labor, right? Someone worked through all these case studies. Some one called all of these customers and found out all of this information. Well, that embodied labor is now free. AI can do all those things. And so now there's an opportunity you can have AI call all those customers.

22:58Do a big survey, find out what they know. You can have AI work through all the case studies. Just a lot of chats with 03, right? And then now you've replicated that proprietary data, but without needing all of that work. How do you square that with the value of real world proprietary data, say something like what cursor gets from its developer community constantly or Tesla autopilot over the last handful of years? So I think those are in the middle because they're really huge amounts of data. I think there are challenges sometimes to training on the data that you get from your users. A lot of time, you know, one thing models can't do is if you train data and you memorize data about a specific person, maybe that leaks out into the next person.

23:48So that's a real challenge to using these kinds of proprietary data. I think there is a kind of real world for proprietary data that's very useful, which is data, very specific data about very specific customers that they trust you to use on their behalf. So to give an example, my financial advisor knows a lot about me. She knows my entire portfolio, she knows the kind of objectives that I have. And Ristallers, right. And she uses all of that information to give me a better outcome, which is what is the next asset I should buy. And she doesn't do that. Like the data doesn't make her a better financial advisor.

24:21It doesn't teach her skills. But it allows her an opportunity to use the skills she already has. And so that's the place where I think proprietary data is really useful. I want to switch gears a bit to coding. It feels like software engineering has just gone through this fast takeoff moment. And just judging from the pace of how quickly things are changing, there's at least a certain subset of the market that thinks the superintelligence takeoff, probability is a lot higher than the folks thought it was just given how quickly coding has taken off. What's your view of what's happened in the coding space?

24:56So I think, you know, on the one hand, coding has taken off very quickly. On the other hand, way back in January 2020, as soon as we saw GPT -3, we launched a project to train GPT -3 how to code. And so, you know, when you look at an exponential curve, you know, the progress is actually the same the whole time, but the impact of that progress can become very non -linear when it passes a threshold. And that's what's happened with coding in the last couple of years. And so my take on where coding will go is that you're going to continue to see a mix of coding with the user in an IDE, traditional cursor -style work, and coding in the background as an agent, something like Devon -style work.

25:43And this is going to continue for a long time as a year or two, maybe, is a long time in AI adoption. That's forever in AI years. But that, you know, if you think about something like vibe coding, right? Like the story you hear with vibe coding is you can, you know, if you have a PM, right? And you want to create a demonstration project. I think that you're going to see PM's vibe coding really cool prototypes, really cool demos that they can use to get user feedback. But then those things are going to get thrown away and that they're going to get rebuilt with professional software engineers because, you know, if you are given a code base that you don't understand.