John Schulman (OpenAI Cofounder) — Reasoning, RLHF, & plan for 2027 AGI

Chatted with John Schulman (cofounded OpenAI and led ChatGPT creation) on how posttraining tames the shoggoth, and the nature of the progress to come... Watch on YouTube https://youtu.be/Wo95ob_s_NI. Listen on Apple Podcasts https://podcasts.apple.com/us/podcast/john-schulman-openai-cofounder-reasoning-rlhf-plan/id1516093381?i=1000655679622, Spotify https://open.spotify.com/episode/1ivzHH9RWciXe4O1rKtldf?si=V6wkPrBsQpi15Yz4qz3UHw, or any other podcast platform. Read the full transcript here https://www.dwarkeshpatel.com/john-schulman. Follow me on Twitter https://twitter.com/dwarkesh_sp for updates on future episodes. Timestamps (00:00:00) - Pre-training, post-training, and future capabilities (00:16:55) - Plan for AGI 2025 (00:29:18) - Teaching models to reason (00:39:45) - The Road to ChatGPT (00:51:07) - What makes for a good RL researcher? (00:59:53) - Keeping humans in the loop (01:14:11) - State of research, plateaus, and moats Sponsors If you’re interested in advertising on the podcast, fill out this form https://airtable.com/appxGOvFLDLP5dlzv/pagFVrbHRohW6F2bZ/form. * CommandBar https://commandbar.com/ is an AI user assistant that any software product can embed to non-annoyingly assist, support, and unleash their users. Used by forward-thinking CX, product, growth, and marketing teams. Learn more at commandbar.com http://commandbar.com/. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.dwarkesh.com https://www.dwarkesh.com?utm_medium=podcast&utm_campaign=CTA_1

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Dwarkesh Patel?

  Today I have the pleasure to speak with John Chulman, who is one of the co-founders of OpenAI and leads the post-training team here. And he also led the creation of ChatGPT and is the author of many of the most important and widely cited papers in AI and RL, including PPO and many others. So, John, really excited to chat with you. Thanks for coming on the podcast.

• 0:21Speaker 01

  Thanks for having me on the podcast. I'm a big fan.

• 0:23Dwarkesh Patel?

  Thank you. Thank you for saying that.

• 0:26Dwarkesh Patel?

  So the first question I had is, we have these distinctions between pre-training and post-training beyond what is actually happening in terms of loss function and training regimes. I'm just curious, taking a step back, conceptually, like, what kind of thing is pre-training creating? What does post-training do on top of that?

• 0:44Speaker 01

  In pre-training, you're basically training to imitate all of the content on the internet or on the web, including websites and code and so forth. So you get a model that can basically generate content that looks like random webpages from the internet. And the model is also trained to maximize likelihood, where it has to put a probability on everything. So it's, the objective is basically predicting the next token, given the previous tokens. Tokens are like words or parts of words. And since the model has to put a probability on it, and it's, we're training with, to maximize log probability, it ends up being very calibrated.

• 1:26Speaker 01

  So it can not only generate all of this, the content of the web, it can also assign probabilities to everything. So, so the base model can effectively take on all of these different personas or generate all these different kinds of content. And then when we do post-training, we're usually targeting a narrower range of behavior, where we basically want the model to behave like this kind of chat assistant. And it's a, it's a more specific persona where it's trying to be helpful. It's not trying to imitate a person. It's answering your questions or doing your tasks.

• 2:08Speaker 01

  And We're optimizing on a different objective, which is more about producing outputs that humans will like and find useful, as opposed to just trying to imitate this raw content from the web.

• 2:20Dwarkesh Patel?

  Yeah, okay, I think maybe I should take a step back and ask, right now, we have these models that are pretty good at acting as chatbots. Just taking a step back from how these processes were currently, what will the models release by the end of, kinds of things the models release in the end of the year, will be capable of doing? What do you see the progress looking like five, you know, carry this forward for the next five years?

• 2:40Speaker 01

  Oh, yeah, five years. Yeah, I think the models will get quite a bit better.

• 2:44Dwarkesh Patel?

  But in what way?

• 2:45Speaker 01

  In the course of five years. So, I mean, I think even in one or two years, we'll find that a lot of, you can use them for a lot of more like involved tasks than they can do now. So you could, so for example, right now, like you could imagine having the models do carry out a whole coding project instead of maybe giving you one suggestion on how to write a function. So you could imagine the model, like you giving it sort of high level instructions on what to, what to code up and it'll go and it'll go and write many files and test it, look at the output, iterate on that a bit.

• 3:29Speaker 01

  So just much more complex tasks.

• 3:31Dwarkesh Patel?

  And fundamentally, the unlock is that it can act coherently for long enough to write multiple files of code? Or what has changed between now and then?

• 3:39Speaker 01

  Yeah, I would say this will come from some combination of just training the models to do harder tasks like this. So just like, I'd say, right, the models aren't, aren't particularly like, most of the training data is more like doing single steps at a time. And I would expect us to do more for training the models to carry out these longer projects. So I'd say any, any kind of training, any like doing RL to learn how to do these tasks, however you do it, whether it's, whether you're supervising the final output or supervising it like each step.

• 4:23Speaker 01

  I think any kind of training at carrying out these long projects is going to make them a lot better. And since the, the whole area is pretty new, I'd say there's just a lot of low-hanging fruit.

• 4:36Dwarkesh Patel?

  Interesting.

• 4:37Speaker 01

  And doing this kind of training. So I'd say that's one thing. Also, I would expect that as the models get better, they're just better at recovering from errors. Or they have just, they're better at dealing with, dealing with edge cases. Or when things go wrong, they know how to recover from it. So the models will be more sample efficient. So you don't have to collect a ton of data to teach them how to get back on track. Just a little bit of data or, or just their like generalization from, from other abilities will allow them to get back on the track, on track. Whereas current models might just get stuck and get lost.

• 5:20Dwarkesh Patel?

  I'm not sure I understood actually how, I want to understand more specifically, how the generalization helps you get back on track. Can you say more about that? I'm not sure I got by those two concepts that are connected.

• 5:32Speaker 01

  Right, they're not directly connected. So I would say you usually have a little bit of data that does everything. So, I mean, if you have, yeah, if you collect a diverse data set, you're going to get a little bit of everything in it. And, and if you have models that generalize really well, even if there's just a couple of examples of getting back on track.

• 5:54Dwarkesh Patel?

  I see.

• 5:54Speaker 01

  Or even like maybe in the pre-training, there's examples of getting back on track. Then like the model will be able to generalize from those other things it's seen to the current situation. So I think like, if you have models that are weaker, you might be able to get them to do almost anything with enough data. But you might have to put a lot of effort into a particular domain or skill. Whereas for a stronger model, it might just do the right thing without any training data or any effort.

• 6:25Dwarkesh Patel?

  Do you have some intuition about right now, these models can maybe act coherently for five minutes? We want them to be able to do tasks that for a human would take an hour, then a week, then a month, and so forth. To get from each of these benchmarks, is it going to be each one takes 10x more compute? analogous to the current scaling loss for pre-training? Or is it going to be a much more? Streamline process because just getting to that point where you're already more sample efficient and then you can just you just go to the years of carrying out tasks or something.

• 7:00Speaker 01

  Yeah, I would say at a high level, I would agree that longer high horizon tasks are going to require more model intelligence to do well and are going to be more expensive to train for. I'm not sure I would expect there to be a really clean scaling law unless you

• 7:19Speaker 01

  set it up in a very careful way or design your yeah, design the experiment in a certain certain way because I would say there might end up being some phase transitions where once you get to a certain level, you can deal with you can deal with much longer tasks. So for example, people like I think when people like think when people do planning for a different time scales, I'm not sure they use completely different mechanisms. So we probably use the same mental machinery if we're thinking about one month from now, one year from now.

• 8:02Dwarkesh Patel?

  Yeah.

• 8:03Speaker 01

  Or like 100 years from now. It's so we're not actually doing some kind of reinforcement learning that where we need to worry about a discount factor that covers that time scale and so forth. So I think I think using language, you can describe all of these different time scales and then you can do things like plan in the moment, you can try to make progress towards your goal, whether it's a month away or 10 years away. So I might expect the same out of models where there are some kind of I don't know if it's a phase transition, but like there's some capabilities that work at multiple scales.

• 8:43Dwarkesh Patel?

  Yeah. Well, okay. So to correct me if this was wrong, but it seems like that implies right now we have models that are on a per token basis, pretty smart, like they might be as smart as humans on a per token basis, the smartest humans and the thing that prevents them from being. As useful as they could be is that five minutes from now, they're not going to be rewriting your code in a way that's coherent and aligns with the broader goals you have for your project or something. If it's the case that once you start this long horizon RL training regime, it immediately unlocks your ability to be coherent for longer periods of time. Should we be predicting something that is human level as soon as that regime is unlocked?

• 9:24Dwarkesh Patel?

  Or and if not, then what is remaining after you can plan for a year and execute projects that take that long?

• 9:30Speaker 01

  Yeah, it's not totally clear what we're going to see once we get into that regime and how fast progress will be. So that's, that's still uncertain. I would say I would expect there to be, I would, I wouldn't expect everything to be immediately solved by doing any training like this. I would think there'll be other like miscellaneous deficits that the models have that cause them to get stuck or not make progress or make worse decisions than humans. So I wouldn't say I expect that this one little thing will unlock every all capabilities.

• 10:08Speaker 01

  But I, yeah, it's not clear, but it might like some improvement in the ability to do long horizon tasks might go quite far.

• 10:16Dwarkesh Patel?

  Would you say it's plausible or is it seems quite likely that there will be other reasons why there might be bottlenecks? And I'm also kind of curious, like what would be the nature of the bottlenecks? So it has all these representations for pre-training. Now it can do accurately for a long period of time because of long horizon RL. What's remaining?

• 10:34Speaker 01

  Yeah, maybe there's some, there's some other experience that human experts bring to different tasks, like having some taste or dealing with ambiguity better. So I could imagine that if we want to do something like research, like those, those kind of considerations come into play.

• 11:00Speaker 01

  Yeah, obviously, there's uh, they're gonna be just sort of mundane limitations around like affordances of a model, like whether it can, whether it can use UIs and obviously the physical world or having access to things. So I think there might be a lot of like mundane barriers that are probably not going to last that long, but would initially like slow down progress.

• 11:28Dwarkesh Patel?

  The websites that are designed for these AIs, once they're much more multimodal, or at least train on more multimodal data, will they be in any way different from the ones we have for humans, like the UIs that will be needed? What, how compensating for their strengths and weaknesses, how would that look different from the current, you know, UIs we have for humans?

• 11:47Speaker 01

  Yeah, that's, that's an interesting question. I mean, I would expect that models will be able to use websites that are designed for humans just by using vision, like when the vision capabilities get a bit better. So there wouldn't be an immediate need to change them. On the other hand, some websites that are going to benefit a lot from AIs being able to use them will probably want to design to be better UXs for AIs. So I'm not sure exactly what that would mean, but probably like, assuming that our, our models are still better in text mode than like reading text out of images, you'd probably

• 12:29Speaker 01

  want to have a good text-based representation for the models. So, and also just a good like indication of what are all the things that can be interacted with. But I guess I wouldn't expect the web to get like totally redesigned to have APIs everywhere because I, I'd expect that we can get models to use the same kind of UIs that humans use.

• 12:52Dwarkesh Patel?

  Right. I mean, I guess that's been the big lesson of language models, right? That they can, they can act in the similar affordances that humans have.

• 13:00Dwarkesh Patel?

  So the point you made earlier about this process could be more sample efficient because it could generalize from its experiences in pre-training of how to get unstuck in different scenarios.

• 13:12Dwarkesh Patel?

  I'm curious what the strongest evidence of this kind of generalization and transfer you've seen is.

• 13:21Dwarkesh Patel?

  Yeah, like, because the big question, it seems, about the future abilities of models is like, how much generalization there is happening? Is there something that feels really compelling to you? Like, you really learned something that you wouldn't expect it to learn from the generalization here?

• 13:35Speaker 01

  There's definitely been some interesting instances of generalization in post-training. Like, one well-known phenomenon is if you do all your fine-tuning with English data, you'll automatically, you'll have the model also behaving well in other languages. So if you train the assistant on English data, it'll also do something reasonable in Spanish, say. And sometimes you might get, you might get the wrong behavior in terms of whether it replies in English or replies in Spanish. But usually you get the, you get the right behavior there as well. Like, you get it to respond in Spanish to Spanish queries.

• 14:16Speaker 01

  So that's one, one kind of interesting instance of generalization that you just sort of latch on to the right helpful persona and then you automatically do the right thing in different languages. We've seen some version of this with multimodal data where if you do text-only fine-tuning, you also get reasonable behavior with images.

• 14:40Speaker 01

  Early on in ChatGPT, we, we were trying to fix some issues in terms of the model understanding its own limitations, like

• 14:52Speaker 01

  Like early versions of the model would think they could like send you an email or call you, call an Uber or something like the model would try to play the assistant and it would say, oh yeah, of course I I sent that email and obviously it didn't. So we uh we started collecting a little some data to fix those problems and we found that a tiny amount of data did the trick even when you mix it together with everything else. So I don't remember exactly how many examples, but something like thirty thirty examples. Well, we had us, I don't know, pretty small number of examples showing this general behavior of like explaining that the model can't doesn't have this capability and that generalized pretty well to all sorts of capabilities we didn't train for.

• 15:34Dwarkesh Patel?

  Hmm. Okay, so I I still want to go back to this because I'm not sure I understood.

• 15:40Dwarkesh Patel?

  Like if you have this model that is trained on to be coherent for longer periods of time, does that imply that unless there are these other bottlenecks, which they may or may not be, by next year, you could have models that are potentially like human level in terms of acting like like you're interacting with this as a colleague and it's like almost it's like as good as interacting with a human colleague, you can tell them to go do stuff and they go to get it done. Uh, what seems wrong with that picture if this is the capabilities you think might be possible?

• 16:12Speaker 01

  Yeah, it's hard to say exactly what will be the deficit. I mean, I would say that when you talk to the models today, they have various weaknesses besides long-term coherence in terms of also like like really thinking hard about things or paying attention to what you ask them. So I would say I wouldn't expect like just improving the coherence a little bit to like to be all it takes to get to AGI, but I guess I wouldn't be able to articulate exactly what the main weakness is that'll stop them from like being a fully functional colleague.

• 16:55Dwarkesh Patel?

  It seems like then you should be planning for the possibility you would have AGI very soon.

• 17:00Speaker 01

  Yeah, I think it's I think that would be reasonable.

• 17:02Dwarkesh Patel?

  Hmm. So what's the plan? If like if there's no other bottleneck next year or something, you got AGI, what's the plan?

• 17:08Speaker 01

  Well, I would say that if AGI came way sooner than expected, we would definitely want to we want to be careful about it and we would we might want to like. Slowed down a little bit on training and deployment until we're pretty sure we know we can deal with it safely and we have a a pretty good handle on what it's going to do, what it can do. So I think yeah, we would have to be, we'd have to be very careful if it happened way sooner than expected because I think our understanding is rudimentary in a lot of ways still.

• 17:46Dwarkesh Patel?

  And what would being careful mean? Like, because presumably you are already careful, right? You do these evaluations before you're

• 17:53Speaker 01

  Yeah, I would say just like maybe not

• 17:59Speaker 01

  not training the even smarter version, not being really careful when you do train it, that it's not it's like properly sandboxed and everything. Maybe not deploying it at scale or yeah, being yeah, being careful about what what scale you deploy it.

• 18:21Dwarkesh Patel?

  Yeah, I guess I'm not. Okay, so let's just play with a scenario like it happens next year and then you're not training a smarter system, but and then you're deploying somewhat in a measured way.

• 18:38Dwarkesh Patel?

  Yeah.

• 18:40Dwarkesh Patel?

  I'm wondering.

• 18:43Dwarkesh Patel?

  Well, presumably if this is just, this isn't particularly open in AI, but this is just intelligence was just much easier than we expected and this is why it happened. And so you wait to deploy a little bit. Now other companies have the similar level of capabilities. What happens next? So you've waited to deploy. What are you waiting for? What are you talking with these? What is every company doing in this scenario?

• 19:04Speaker 01

  Yeah. Yeah. The game theory is a little tough to think through. So oh yeah. So first of all, I don't think this is going to happen next year, but it's still useful to have the conversation and maybe it's like two or three years instead. But yeah, I guess

• 19:17Dwarkesh Patel?

  Two or three years is still pretty soon.

• 19:18Speaker 01

  Yeah, it's still pretty soon. I do think you probably need some coordination, like everyone needs to agree on some on some reasonable like limits to deployment or to further training for this to work. Otherwise otherwise you have the the race dynamics where everyone's trying to everyone's trying to stay ahead and like Everyone's like, and that might require compromising on safety. So I think you would probably need some coordination among the larger entities that are doing this kind of training.

• 19:53Dwarkesh Patel?

  And so you're coordinating to, I guess, pause deployment until, until what exactly? Like until you figure out what's happening in the models?

• 20:01Speaker 01

  Like pause either further training, pause deployment, like avoid certain types of training that we think might be riskier. So just like setting up some reasonable rules for like what, what everyone should do to, yeah, having everyone somewhat limit, limit these things.

• 20:25Dwarkesh Patel?

  And but limit to what end? Because I guess at some point, then you're gonna have to like, the potential energy that's within this intelligence will, you know, it'll only show what, what, what is a plan to do? Like suppose in two years, we get the AGI and now everybody's freaking out. And so now the AI companies have paused. And now what? Or what would be the plan to wait till?

• 20:51Speaker 01

  Yeah, that's, I don't have a good answer to that. I mean, I would say, if we can, if everyone is gonna coordinate like that, I think we would be, that would be an okay scenario. That'd be a pretty good scenario. Because I do think, like, building these models is very capital intensive. And there are a lot of complex pieces. So it's not like everyone's gonna go and recreate this stuff at home. So I think it is possible to do, given the relatively small number of entities who could train the largest models, it does seem possible to coordinate. So I'm not sure how, how you would maintain this,

• 21:32Speaker 01

  this equilibrium for a long period of time. But I think if we got to that point, we would be in an okay position.

• 21:40Dwarkesh Patel?

  Or would be, I guess, I'm curious, like, I'm not sure what happens next. Because like fundamentally, the problem, or the benefit is that like, we've got a ton of like, you like push it to the server. And now we've got a bunch of intelligences, or they could push themselves to the server. And now we got everybody coordinated, but I'm not sure what's, what we do next in this, in this world. We're like, why that, why that sets this up for a good outcome?

• 22:05Speaker 01

  Yeah, I would say if we had everyone reasonably coordinated, we could figure out some and we felt like we had solved the technical problems around alignment well enough to be able to deploy like really smart AIs that can like like act as an extension of people's will, but also prevent them from being misused in some way that would cause a catastrophe. I think then then that would be great. Like we could go ahead and like safely deploy these systems and it would it would usher in a lot of prosperity and a new like much more rapid

• 22:47Speaker 01

  phase of scientific advancement and so forth. So I think that would be what the good scenario would look like.

• 22:53Dwarkesh Patel?

  Okay, so that's that makes sense, but I'm curious, like, how would you know in a couple of years if you like all these actors, even in the best case scenario, they've agreed to pause until we've figured out that we're building aligned systems that are not themselves going to attempt to take over or a coup or not going to enable somebody else to do that? How what would proof of that look like? Or what would evidence of that look like?

• 23:20Speaker 01

  Well, I would say if we if we can deploy like systems incrementally that are successively smarter than the ones before, then I think that's safer. So I hope the way things play out is is it's not the scenario where everyone has to coordinate and lock things down and safely release things like because it would like lead to this big buildup in potential energy, potentially. So I would rather some scenario where we're just continually releasing things that are a little better than what came before. And then we while like making sure we're confident that each diff is like improving improving the safety and alignment

• 24:03Speaker 01

  in like correspondence to the improvement in capability. So and if if things started to look a little bit scary, then we would be able to slow things down. So that's what I would hope for. I would say if there's more of a discontinuous jump and the question is how do you know if the thing you've got is safe to release? I would say I can't give a generic answer like I would wanna, but like the type of thing you might wanna do to make that more more acceptable would be you would wanna do a lot of testing, like simulated deployment.

• 24:46Speaker 01

  Where that you expect, so red teaming of sorts, like you'd wanna do that in a way that you feel is like much less favorable than, or much

• 24:58Speaker 01

  more likely to fail than the thing you're planning to do in the real world. You'd wanna have a really good monitoring system so that you can, like, if something does start to go go wrong with the deployed system, you can, you feel like it's gonna be detectable immediately, like you've got, maybe you've got something watching over the deployed AIs and what they're doing and looking for signs of trouble. So, so I would wanna, yeah, I would say just, you'd want some defense in depth, like you'd wanna have some combination of, like, the model itself seems to be, like, really well behaved and have, like, impeccable

• 25:39Speaker 01

  moral compass and everything, and you're pretty confident that it's extremely resistant to any kind of takeover attempt or something, or like severe misuse. And then you'd also wanna have, like, really good monitoring on top of it. So, yeah, you could detect any kind of, any trouble.

• 25:56Dwarkesh Patel?

  What are you keeping track of while you're doing long horizon RL, or when you eventually start doing it, that you could notice this sort of discontinuous jump before you deploy these systems broadly?

• 26:08Speaker 01

  I would say you would wanna have a lot of evals that you're running during the training process.

• 26:12Dwarkesh Patel?

  And like, what specifically would it, how would you notice something like, yeah. And I mean, does it make sense to train on a long horizon RL, knowing that this is something that could happen? Or is it just like a very low possibility? How do you think about this?

• 26:25Speaker 01

  You'd wanna be pretty careful when you do this kind of training, if you see a lot of potentially scary capabilities. If those seem close, I mean, like, I would say it's not something we would wanna, we have to be scared of right now, because right now it's hard to get the models to do anything like coherent. But if they started to get really good, I think.

• 26:51Speaker 01

  Yeah, I think we would wanna, we would have to take some of these questions seriously and we would wanna have a lot of vowels that like sort of test them for misbehavior and the most, or I guess that's like for the alignment of the models, we'd wanna check, we want to check that they're not gonna, they're not gonna sort of turn against us or something. But you, you might also wanna look for like discontinuous jumps in capabilities, like you'd wanna have lots of vowels for the capabilities of the models.

• 27:27Speaker 01

  I mean, also, I guess you, you'd also wanna make sure that whatever you're training on doesn't have any reason to make the model turn against you, which itself, I think, isn't, I would say there's like, that doesn't seem like the hardest thing to do. I mean, if, like the way we train them with RLHF, that, that does feel, even though the models are very smart, it does feel very safe because the model is just trying to produce a message that is pleasing to a human. And it has no concern about anything else in the world, other than whether this text it produces is, is approved.

• 28:08Speaker 01

  So obviously, if you were doing something where there's, where the model has, yeah, it's carrying out a long sequence of actions which involve tools and everything, then it might have some incentive to do a lot of wacky, like, wacky things that wouldn't make sense to a human in the process of producing its final result. But I guess it wouldn't necessarily have an incentive to do anything other than produce a very high quality output at the end. So it, like, it's not, yeah, so I guess you have these old points about like instrumental convergence, like the model's gonna wanna take over the world so it can produce this awesome piece of code at the end.

• 28:48Speaker 01

  Like, if you ask it to write you the flask app, it'll be like, oh yeah, first I need to take over the world and then I, I need to, I don't know. But at a certain point, it's a little bit, it's a little hard to imagine why for some. Like fairly well specified task like that, you would want to first take over the world. But of course, yeah, if you had a task like make money, then maybe that would lead to some nefarious behavior as a instrumental goal.

• 29:18Dwarkesh Patel?

  Yeah. All right. So before we get back to that, I think let's step back and talk about like today's RLHF systems and everything. But I do want to follow that to a point. It's kind of interesting. Okay. So today's RLHF, the way in which it influences these models is, would you characterize it as in terms of human psychology? Is it a drive? Is it a goal? Is it an impulse? Like psychologically, what kind of thing? In what way is it being changed? And not just like the persona of a chatbot, but just like, don't talk that way. Talk this other way, or don't put those kind of outputs.

• 29:56Speaker 01

  Yeah, I would say there are probably some analogies with a drive or a goal in humans. So in that you have, you're trying to steer towards a certain set of states rather than some other states. And so I would, I would think that our concept of a drive or a goal has other elements like, like the feeling of satisfaction you get for achieving it. And, and those things might be more like have more to do with the learning algorithm than what the model does at runtime, when you just have a fixed model. So I would say, I would say there are probably some analogies, though it's, I don't know exactly like how, how close it is.

• 30:41Speaker 01

  But I would say to some extent, it is, the models do have drives and goals in some meaningful way. And in, in the case of RLHF, where you're trying to maximize human approval as measured by a reward model, the model is just trying to produce something that people are going to like and is, they're going to judge as correct.

• 31:01Dwarkesh Patel?

  I've heard two ideas in terms of using that inner monologue type of thing to get better at reasoning, at least publicly, the kinds of things I've seen. And I'm curious what you think is more promising. One is that.

• 31:15Dwarkesh Patel?

  The model learns from, it outputs a bunch of potential trains of thought and it learns to follow the one that leads to the correct answer and is trained on that before deployment. And the other one is you use a bunch of compute to do inference in deployment, which involves the model talking to itself, you know, while it's deployed. Which one do you expect it to be closer to? What is like really good at reasoning? Is it because it's doing just a bunch of inference calls? Or is it just because you've trained it to do well at that?

• 31:48Speaker 01

  Well, I would say you could define reasoning as tasks that require some kind of like computation at test time, or maybe some kind of deduction. So by definition, reasoning would be tasks that require like some test time computation and like step-by-step computation. On the other hand, I would also expect to gain a lot of like doing some kind of training time computation or practice at training time. So I would think that you get the best results by combining combining these two things.

• 32:28Dwarkesh Patel?

  Right now, you know, you have these two ways in which the model learns. It's either in training, whether it's pre-training or with the post-training, but it's like most of the compute in training is spent on pre-training and it's just glossing over trillions of tokens, just like standing by as a, you know, like almost like skimming trillions of tokens worth of information, which if a human was subjected to that would just be totally confused, right? It's like not a very efficient way to learn. And the other way is in context learning, but of course that is is more sample efficient there, but it's destroyed with each instance. I'm curious if you think that

• 33:08Dwarkesh Patel?

  There is a path for something in between those where it's not destroyed at each instance, but it's also not as frivolous as just seeing trillions of tokens where it's more deliberate and active.

• 33:24Speaker 01

  Yeah, so do you mean models having some kind of medium term memory? So too much to fit in context, but like much smaller scale than pre-training?

• 33:34Dwarkesh Patel?

  I'm not sure if memory, it might be memory. I don't have context, but certainly like when I'm trying to prepare for this conversation, it feels like I think of like, I should understand this. So I look it up and I like read it carefully and I maybe think about it as I'm reading it. And I'm not sure what it naturally corresponds to in terms of models, but what would that look like? I'm curious.

• 33:56Speaker 01

  I see. So it's not just memory, but it's also somewhat like specializing to a task that specializing to a certain task or putting a lot of effort into like some particular project.

• 34:08Dwarkesh Patel?

  And I'm not even sure specialization more so I'm thinking about, I don't understand this part, so let me look into this part deeper. I already understand this. I'm gonna like specializing to your existing knowledge base. Yeah.

• 34:20Speaker 01

  I see. So it's not just about finding like, I don't know, training on a bunch of sources that are relevant, fine tuning on some special domain. It's also about like reasoning about like developing some knowledge through your own reasoning and also using some sort of introspection and self-knowledge to figure out what you need to learn.

• 34:40Dwarkesh Patel?

  Yeah.

• 34:42Speaker 01

  Yeah, I would say that does feel like something that's missing from today's systems. I mean, I would say people haven't really pushed too hard on this middle ground between like large scale training, like where you produce the like this snapshot model that's supposed to do everything, like a deployed model. And then like on the other hand, like in context learning. And I think part of that is that we've just been increasing context length so much that there hasn't been an incentive for it. So if you can go to like that 100,000 or a million context, then that's actually quite a lot. And it's not, it's not actually the bottleneck in a lot of cases, but I agree that.

• 35:26Speaker 01

  You'd probably also want to supplement that by some kind of fine tuning, like the capabilities you get from fine tuning and in context learning are probably somewhat complementary. So I would expect us to want to build systems to do some kind of online learning and also have some of these cognitive skills of like introspecting on their own knowledge and seeking out new new knowledge that fills in the holes.

• 35:54Dwarkesh Patel?

  Is this all happening at the same time? Is it just like a new training regime where all these things can happen at once? Or whether it's the long horizon training or whether it's this kind of training, are they separate or are they just because like the model is smart enough? So they can both introspect and it can act on longer horizons and you can get adequate reward on long horizon tasks?

• 36:17Speaker 01

  Yeah, I would say if you're doing some kind of long horizon task.

• 36:22Speaker 01

  Well, I would, you're learning while you do the task, right? So the only way to do something that involves a lot of steps is to like to have learning and memory that gets updated during the task. So like there's a continuum between like, like short term memory,

• 36:43Speaker 01

  between short term and long term memory. So I would say, yeah, I would expect,

• 36:54Speaker 01

  I would expect this capability would start to become like the need for it would start to become clear when we start to look at long horizon tasks more and, and to some extent, just putting a lot of stuff into context will probably will take you pretty far because we have really long context now. But you probably also want things like fine tuning. And as for like introspection and the ability to do active learning.

• 37:23Speaker 01

  That might like automatically fall out of the model's abilities to know what they know because they have some like, um, models have some calibration regarding what they know. And that's why, like, that's why, um, models don't hallucinate that badly. Because, yeah, they have some understanding of the, their, their own limitations. So I think that like same kind of ability could be used for something like active learning.

• 37:51Dwarkesh Patel?

  Hmm. And how, so there's all these complicating RL procedures that many of whom you pioneered. How many of them will be relevant when you get to the point where the, uh, the model itself is this smart? That it can act as a sort of environment and interact in a more online and stable way? Is, is, is, is the path for progress going to be more straightforward than the kinds of solutions that were required for RL in the past?

• 38:23Speaker 01

  Well, I think policy gradient algorithms are not the most sample efficient algorithms. So that's probably not what you want to do at test time if you want to learn really fast. But though, who knows? I mean, maybe it's not that bad. So I think, um, something like, um, like motor learning in animals is probably something like a policy gradient algorithm. And, uh, so for example, you're like learning how to, um, shoot baskets. I think you probably, uh, like that takes, uh, maybe thousands of tries to, um, get more accurate. And I think you probably, there's probably some, something that's, uh, like a policy gradient algorithm underneath.

• 39:04Speaker 01

  But, uh, that's not going to be the fastest way to learn in, um, like if, if you have a model trying to do a project or some kind of task. So I would think we would want to rely more on like in context learning. Where, uh, you effectively have a learned algorithm, like you've learned how to explore. Uh, like you've learned how to try all the possibilities exhaustively.

• 39:30Speaker 01

  And, uh, instead of doing the same thing over and over again, making the same mistake. So, yeah, I would say we'll be able to do things that look more like learned search algorithms. And that'll be the kind of thing that, uh, gets used in a particular task.

• 39:45Dwarkesh Patel?

  Interesting. All right. I want to step back and ask about your own history. So at least at OpenAI. So you led the creation of ChatGPT. At what point did you realize, first of all, these LLMs are the path to go? And then a chatbot would be, or some way to instruct them would be a useful thing to do? Just walk me through the whole lineage from like, when this became your main focus? And yeah, what the process was like?

• 40:15Speaker 01

  Yeah, so early, so we had, before ChatGPT, we had, OpenAI had these instruction following models. And that was the idea there was, we had base models and people can prompt them in elaborate ways. But they're also kind of hard to prompt. You had to, they basically do autocomplete. So you have to set up a very good prompt with some examples. So, so people at OpenAI were working on just taking the base models and making them easier to prompt. So that if you just wrote a question, it would answer the question instead of giving you more questions or something.

• 40:56Speaker 01

  So that was, so we had these instruction following models, which were kind of like base models, but a little easier to use. And those are the original ones deployed in the API or after GPT-3. Those were the next generation of models. Then at the same time, there were definitely a lot of people thinking about chat. So, so Google had some papers, like they had Lambda and earlier Mina. So they had these chatbots and it was more like, like you had a, it was more like a base model that was really specialized to the task of chat, really good at chat. And like, I think, at least looking at the examples from the paper,

• 41:40Speaker 01

  it was more used for sort of fun applications, like where the model would like take on some persona and pretend to be that persona. It was not so functional, like, like help me refactor my code.

• 41:54Speaker 01

  So, yeah, there are definitely people thinking about chat. I had worked on a project before, looking at chat called WebGPT, which was more about doing question answering with the help of web browsing and retrieval.

• 42:10Speaker 01

  Well, when you do question answering, it really wants to be in a chat because you always want to ask follow-up questions or sometimes you need a clarif- the the model should ask a clarifying question because the question is ambiguous. So it was kind of clear after we did the first version of that that we should, the next version should be conversational. So anyway, we started working on like the conversational chat assistant. And we, this was built on top of G P T three point five, which was done training at the beginning of twenty twenty two. And that model was quite good at language and code. So we quickly realized that it was actually quite good at coding help.

• 42:52Speaker 01

  And that was one of the things we were excited about. So yeah, we worked on that. We worked on that for for most of the year and we had, we had browsing as another feature in it. Though we ended up like de-emphasizing that later on because the, like the model's internal knowledge was so good that we didn't, that the browsing wasn't the most interesting thing about it. And then we were thinking about, we had it out for beta testing or to friends and family for a while. And we were thinking about doing a public release. But at that time, actually, G P T four finished training in August or, yeah, in August that year.

• 43:35Speaker 01

  And actually, the, like, the flagship R L effort at OpenAI was the instruction following effort. Because that was the models that were being deployed into productions. So like the first fine tunes of G P T four used that, that whole stack. And that was, yeah, those models were really good and everyone got really excited about that after seeing the, like instruct fine tune G P T fours. But so they were really, really good. They would occasionally give you amazing outputs, but they were also like a little bit. The model was clearly like pretty unreliable, like it would sometimes hallucinate it a lot. And it was like pretty, it would sometimes give you pretty unhinged outputs. So it was clearly not quite ready for prime time, but it was like obviously very good.

• 44:20Speaker 01

  Good. And yeah, so I guess that people forgot about chat for a little while after that because about this like alternative branch. But then we we ended up, we pushed it further and we ended up like mixing together all the data sets like the instruct and the chat data. And to try to get something that was the best of both worlds. And I think the yeah, the models we the chat models were like

• 44:47Speaker 01

  were clearly more like it was an easy easier to use. It was sort of more. It's sort of like automatically had much more sensible behavior in terms of like the model knowing its own limitations. That was actually one of the things that I got excited about as we were developing it that like I realized a lot of the things that people thought were flaws in language models, like just like blatantly hallucinating. Could be not completely fixed, but you could make a lot of progress with pretty straightforward methods.

• 45:22Speaker 01

  Oh, yeah. And also, the the other thing about chat was that like when we had these instruct models, like the task of complete this text, but in a nice way or in a helpful way, that's like a pretty poorly defined task. So I think like, I think that task is like both confusing for the model and for the human who's supposed to do the data labeling. Whereas for chat, I think people had an intuitive sense of like what a helpful robot should be like. So I think it was just much easier to tell people like

• 45:54Speaker 01

  to to give for people to get the idea of what what the model was supposed to do.

• 45:58Dwarkesh Patel?

  Yeah.

• 45:59Speaker 01

  And so that so as a result, I think the like the model had a much more coherent personality and like it was much like easier to get like robot, like pretty sensible behavior robustly.

• 46:13Dwarkesh Patel?

  Interesting.

• 46:15Dwarkesh Patel?

  Is it the case that anybody could have made chat GPT using your publicly available fine tuning API?

• 46:24Speaker 01

  Not exactly. I mean, they could have, I don't remember the status of which models were available for fine tuning. Assuming we had 3.5 available for fine tuning at the time, you could have made something pretty decently close, but I'm not sure you would have, I don't think you would have been able to do just one iteration of fine tuning where you have like purely human written data and you fine tune on that. I think you would want, like, you would want to do several iterations. Like, if you're not going to do RL, which we did, you'd want to do some kind of iterative supervised fine tuning where you have like humans edit the model generated outputs, because it's really hard to get people to,

• 47:09Speaker 01

  like, if you train on human generated data, even if it's really high quality, it's just hard for a model to fit that data perfectly, because it might not be, like, it might not be something a model is capable of outputting. So you need to do something iterative that looks a little bit more like RL. So I think if you had done that, you could have gotten something pretty close, but that would have been kind of non-trivial. But we also had another, like, instruction following model trained with RL that was released a little before ChatGPT. So I think if you put a chat, like wrapper on that, you would get something decently close. But it, like, that model, like, if you just prompted it with chat. So,

• 47:54Speaker 01

  but that model had some differences in strengths, like, it was, like, that model was pretty good at writing and poetry and so forth, but it wasn't, it sort of, it wasn't as good at knowing its limitations and at factuality and so forth.

• 48:08Dwarkesh Patel?

  So, stepping back from 3.5, I think I heard you somewhere say GPT-2, you're super impressed. Compared to your expectations in 2019, has AI progressed faster or slower than you would have expected?

• 48:22Speaker 01

  I would say faster than I would have expected since GPT-2.

• 48:26Dwarkesh Patel?

  Yeah.

• 48:28Speaker 01

  I was pretty like bought into scaling and yeah, pre-training and so forth being a good idea. But when GPT-2 was done, I was, I would say I wasn't completely sold on it being revolutionizing everything. Like, I only really pivoted what I was working on and what, yeah, what my team was working on in after GPT-3. So after that, we kind of got together and said, oh yeah, let's, uh, let's, uh, this language model stuff works really well. Let's see what we can do here. But yeah, after GPT-2, I wasn't quite sure yet.

• 49:05Dwarkesh Patel?

  Hmm. Especially if the stuff we were talking about earlier with RL starts working better with these smarter models. Would a fraction of compute that is spent on training that is pre-training versus post-training change significantly in favor of post-training in the future?

• 49:24Speaker 01

  Yeah, there are some arguments for that. I mean, right now it's a pretty lopsided ratio, but you could argue that the output generated by the model is like high quality compared to, or higher quality than what most of what's on the web. So it sort of makes more sense for the model to think by itself instead of just like training to imitate what's on the web. So I think there's a first principles argument for that. And I would say we found a lot of gains through post-training. So I'm not sure, so I would expect us to keep like pushing this methodology and probably increasing the amount of compute we put into it.

• 50:06Dwarkesh Patel?

  Hmm. The current GPT-4 has Elo score that is like 100 points higher than the original one that was released. And is that all because of what you're talking about with these improvements that are brought on by post-training? Or?

• 50:22Speaker 01

  Yeah, I would say that we've, I would say that most of that is post-training.

• 50:27Dwarkesh Patel?

  Interesting.

• 50:29Speaker 01

  So there are a lot of, there are a lot of different separate axes for improvement. Like you can, yeah, so we think about like data quality, data quantity, just doing more iterations of the whole process of deploying and collecting new data and like changing what you're, what kind of annotations you're collecting. So there's a lot of, a lot of things that stack up, but together they give you a pretty good, like effective compute increase.

• 50:57Dwarkesh Patel?

  Yeah, I mean, that's a huge increase. That's like really interesting that there's this much, this much room for improvement from post-training. What is, what makes for somebody who's really good at doing this sort of RL research? I hear it's super finicky, but like, what is the sort of intuitions that you have that enable you to find these ways to mess with the data and to set up these environments?

• 51:25Speaker 01

  I'd say I just have a decent amount of experience at this point from like the different parts of the stack, from like RL algorithms, obviously, since I've worked on those since grad school to like the data collection, like the annotation process, to like language, playing with language models. So I, I mean, I'd say I just dabbled with these things and I'd say the people who do well at this kind of research have some view of the whole stack and have a lot of curiosity about the different parts of it and also sort of think about, well, you want to be

• 52:09Speaker 01

  both empirical and like use experiment, let experiments update your views, but you also want to think from first principles somewhat, like what?

• 52:21Speaker 01

  Like assuming that like learning works, like what would be the ideal type of data to collect and that sort of thing?

• 52:29Dwarkesh Patel?

  So because there doesn't seem to be a model released since GPT-4, that seems to be significantly better, there's seems to be the hypothesis that potentially we're hitting some sort of plateau and that these models aren't actually generalizing that well, and you're going to hit some sort of data wall, beyond which point the abilities that are unlocked by memorizing a vast corpus of pre-training data won't actually help you get something much smarter than GPT-4.

• 52:59Dwarkesh Patel?

  What do you think that hypothesis, is that wrong? And like, I think we've talked about some examples generically about generalization, the Spanish to English and so forth, but is there, yeah, I mean, okay, so then maybe this is a run-on question, but one example I was thinking of was the idea that there is transfer from reasoning in code, if you train a bunch of code, it gets better reasoning in language. And if that's, is that actually the case? Do you see things like that, which suggests that there's all this kind of positive transfer between different modalities? So once you try training on a bunch of videos and images, it'll get smarter and it'll get some different synthetic data?

• 53:41Dwarkesh Patel?

  Or does it seem like the abilities that are unlocked are extremely local to the exact kind of labels and data you put into the training corpus?

• 53:52Speaker 01

  Yeah, okay. Yeah, I'll try to respond to all of that. So first, are we about to hit the data wall? I mean, I wouldn't draw too much from the time since GPT-4 was released, because I mean, it does, yeah, it takes a while to like train these models and to like get all the, do all the prep to train a new model, like generation of models. So yeah, I wouldn't draw too much from that fact.

• 54:24Speaker 01

  I would say there are definitely some challenges from the limited amount of data, but I wouldn't expect us to immediately hit the data wall, but I would expect the nature of pre-training to somewhat change over time as we get closer to it. In terms of like generalization from different types of pre-training data.

• 54:50Speaker 01

  I would say it's pretty hard to do science on this type of question because you can't do that, create that many pre-trained models. So maybe you can't train a like a GPT four sized model, you can't do ablation studies at GPT four size scale. Maybe you can do like train a ton of GPT two sized models, or maybe even a GPT three sized model with different data blends and see what you get. So I'm not like aware of any results or like public results on like ablations involving code data and reasoning performance and so forth. So that would be, I'd be very interested to know about those results, but.

• 55:31Dwarkesh Patel?

  I'm actually curious about, I mean, if one of the things is that the model gets smarter as it's bigger, would an ablation on a GPT two level model, which suggests that there isn't that much transfer, how much evidence does that provide for the level of transfer on a similar set of domains in GPT four level model?

• 55:51Speaker 01

  Right. You might not be able to conclude that if transfer fails at GPT two size, then it's also going to fail at a higher scale. So it might be that like for the smaller models, you, yeah, for the larger models, you learn these better shared representations, or the smaller models have to lean too much on memorization, whereas the larger models can learn how to do the right computation. So I would expect this to be true to some extent.

• 56:22Dwarkesh Patel?

  This might have a very simple answer, but so bigger models, you train them on the same amount of data and they become smarter, or conversely, they can, to get the same amount of smarts, you have to train them on less data. What, why is that the case? Like it's got more parameters, it saw less things, and now it's equally as smart. Why did that? Why is that the case?

• 56:44Speaker 01

  I don't think anyone has a good answer for a good explanation of the scaling law with parameter count. I mean, there's some.

• 56:56Speaker 01

  I don't even know what the best sort of mental model is for this. Like clearly you have more capacity if you have a bigger model. But so like you should be able to eventually get lower loss. But I guess why are bigger models more sample efficient? I guess you could, I can give you some like very sketchy explanations. Yes, please. Like, like they have, like you could say that the model is like sort of an ensemble of a bunch of different circuits that do the computation. So it has like, you could imagine that it's doing, it has a bunch of like computations that it's doing in parallel and it's like doing some, like the output is a weighted combination of them.

• 57:42Speaker 01

  And if you have more just width of them, or if you just have, I mean, actually width is somewhat similar to depth because like with residual networks, you end up like the depth can do something similar to width in terms of like updating what's in the residual stream. But if you, yeah, you could argue that you're learning all these things in parallel. You're learning all these different computations in parallel and you just have more of them with the bigger model. So you have more chance that one of them is lucky and ends up like having high, like winning guessing correctly a lot and getting up weighted.

• 58:24Speaker 01

  So that's kind of like, what would be the, yeah, there's some algorithms that work this way, like that, like mixture, what is it mixture, some kind of mixture model?

• 58:41Speaker 01

  Or multiplicative weighted update algorithm. Yeah, there's some algorithms that kind of work like this. So where you have like a, some kind of mixture of, I don't want to say mixture of experts because it means something different, but like basically a weighted combination of experts with some learned gating and.

• 59:00Speaker 01

  Actually, anyway, I said something slightly wrong, but anyway. Yeah, you could imagine something like that and just having a bigger model gives you more chances to get the right function. So that would be. And then, of course, it's not just like you have a bunch of like totally disjointed like functions that have.

• 59:21Speaker 01

  You're taking a linear combination of it's more like a library where you might chain the functions together in some way. So you like it's there's some composability. So yeah, so I would just say there's like the bigger model has a bigger library of different computations, including lots of stuff that's kind of dormant and only being used some of the time. But those things, but it has like more space to look for the like look for those circuits to do something useful.

• 59:50Dwarkesh Patel?

  I want to ask you about.

• 59:53Dwarkesh Patel?

  Stepping back from the current research questions, just stepping back. I want to understand your sort of like modal scenario of what happens for the next few years. I think towards the beginning of the conversation, we were talking about the case in which it progresses really fast. But just let's just take like the modal scenario. You're unlocking long horizon RL at some point, but then, as you said, there's potentially other bottlenecks. So what's happening, you know, how good are these models? How are they being deployed? What other modalities are part of them at what? At what stage are these being unlocked and so forth? Just kind of want to understand your broader picture of what the next few years look like.

• 1:00:33Speaker 01

  Yeah, I would expect.

• 1:00:36Speaker 01

  I would expect things like, okay, new modalities to be added like over time or pretty soon.

• 1:00:46Speaker 01

  I would, yeah, I would expect the capabilities to generally keep getting better through a combination of pre-training and post-training, and that'll open up new use cases. So right now. AI is still not a huge part of the economy, like there's a pretty small fraction of jobs that it can help with at all.

• 1:01:06Speaker 01

  So I'd expect that to be higher over time and not just from the models improving, also from people just figuring out how to integrate them into different processes. So even even if we just froze the models at their current state. I think you would still see a lot of growth in how they're being used. So I'd expect there to be a lot of. Like, I would expect AI to be used much more widely. And I would expect it to be used for more kind of technically sophisticated tasks. Like, yeah, like I gave the programming example earlier of doing like longer projects,

• 1:01:49Speaker 01

  but also helping with various kinds of research. So I would hope that we can use AI to accelerate science in various ways. And just like, because you can potentially have the models like understand all of the literature in a given field and be able to like, be able to sift through tons of data like more than a person would have patience to do. So I would hope that we can basically like, yeah, well, I hope the form factor would basically be that people are still driving all of this and you have your like helpful assistants that you can use, you can sort of direct and point to lots of different

• 1:02:33Speaker 01

  problems that are useful to you. And everyone sort of has all these AIs helping them, helping them do more, get more done.

• 1:02:43Dwarkesh Patel?

  Hey, everybody. Real quick, I want to tell you about a tool that I wish more applications used. So obviously you've noticed every single company is trying to add an AI chatbot to their website. But as a user, I usually find them really annoying because they give these long, generic, often useless answers. Command Bar is a user assistant that you can just embed into your website or application. And it feels like you're talking to a friendly human support agent who is browsing with you and for you. And it's much more personalized than a regular chatbot. It can actually look up users' history and respond differently based on that. It can use APIs to perform actions. It can even proactively nudge users to explore new features.

• 1:03:28Dwarkesh Patel?

  One thing that I think is really cool is that instead of just outputting text, Command Bar can kind of just say, here, let me show you and start browsing alongside the user. Anyways, they're in a bunch of great products already. You can learn more about them at commandbar.com. Thanks to them for sponsoring this episode.

• 1:03:47Speaker 01

  But obviously at some point, they're going to be better than everyone, whatever they want to do.

• 1:03:55Dwarkesh Patel?

  What would that process look like? Right now, they're clearly only helping you. At some point, they're able to just do things for you and maybe like run entire firms for you or whatever.

• 1:04:06Dwarkesh Patel?

  At that point, is it, yeah, is it just gonna be a smooth process? And at that point, the hope is that we have systems that are aligned with the user enough that they can count on the firm being run in the way they expect and so forth?

• 1:04:20Speaker 01

  Yeah, I think, well, we might not want to jump to having AIs run whole firms immediately. I mean, we might want to have people like overseeing, like overseeing these like important decisions and calling the shots. So even if the models are good enough to like to actually run a successful business themselves. So, yeah, to some extent, there might be choices there.

• 1:04:53Speaker 01

  And I think people will still have different interests and what they want to, different ideas for what kind of interesting pursuits they want to direct their AIs at and like they can, people, people could like,

• 1:05:09Speaker 01

  yeah, do a lot of, AI doesn't necessarily have an intrinsic, like any kind of intrinsic desire.

• 1:05:18Dwarkesh Patel?

  Not yet.

• 1:05:19Speaker 01

  Unless we put, we put it in the system. So I think, so people can still end up being, even if AIs like become extremely capable, I would hope that people are still the drivers of like what the AIs end up doing.

• 1:05:35Dwarkesh Patel?

  Yeah, but I wonder if the economic equilibrium is so far from that where you have the equivalent of Amdal's law in a firm. The slowest part of the process is the one that's going to bottleneck you. And so, you know, the AI makes all the non-human parts of the firm 10x more efficient. The firm can no longer, you know, it's still bottlenecked by that step. And so if in the, if like one company decides to proceed by keeping humans in the loop on all the things that you really want human oversight on, then they'll just be outcompeted by other companies. If one country decides to go this route, other countries will beat it. This doesn't seem, I hope this is like, yeah, I wonder if this is a sort of a sustainable plan for keeping humans in the loop.

• 1:06:18Speaker 01

  Right, so I think if you, if we wanted to keep humans in the loop, which seems reasonable. And it turned out that firms with any humans in the loop were outcompeted with by firms that didn't have any humans. Then I think then you would obviously need some kind of regulation that like disallowed having no humans in the loop for running a whole company.

• 1:06:42Dwarkesh Patel?

  Hmm, but there's so many companies in the, in, I guess, in any country, but let alone the world. But yeah, I wonder if it's better to do the regulation on companies and to say like, you've got to keep humans in the loop in important processes. But then you had to define what important processes are. You've got to monitor every single company. And you also got to get collaboration in every single country, which has firms in it versus if this is a problem, should it be solved before the model is even deployed? Such that hopefully you would get into a situation where you did decide to build a firm end to end on these models. It's basically does what you want it to do and you don't need a human in the loop.

• 1:07:24Dwarkesh Patel?

  Does that question make sense? Like, I guess I'm just wondering in this situation, how do we actually monitor if every single firm has a human in the loop? And what happens if like China doesn't decide to do that and so forth?

• 1:07:34Speaker 01

  Right. Um, yeah, you would either have to have like every country agree to this regulatory regime. Or you would need every, you need all of the model infrastructure or the model providers to agree to this kind of requirement. So it's definitely going to be non trivial. So I guess, uh, yeah, this is looking a ways ahead. So it's a little hard to imagine, to imagine this world before seeing anything, anything like it. But. So for example, like there's some questions like would,

• 1:08:17Speaker 01

  are we actually confident that AI run companies are better in every way? Or do we think they're better most of the time, but occasionally they malfunction? Because AIs are still like, they're still less sample efficient in certain ways, like dealing with very wacky situations. So, so actually AI run firms have higher tail risk because they're more likely to malfunction in a big way. So I guess that there might be some questions, practical questions like that, that would, that would also determine how things play off, like play out, like maybe, maybe if you just require people to be accountable for various like liability, this would also change the incentives a bit.

• 1:08:58Speaker 01

  So if it turned out that like AIs are better at running everything and they're also completely benevolent and we've like totally solved alignment and we can, like they're better at being accountable to like their, to people than people are, then I would say maybe, maybe it's okay having the AIs run the firms. But I think that's, that might be pretty far out. And I think we were more likely to be in a situation where they look better, like in the short term, but they still have some problem, like the AI run entities still have some serious problems. And it's actually like practical considerations that push you more towards having humans in the loop, at least for the near future.

• 1:09:41Dwarkesh Patel?

  Okay, so this is a problem you had to deal with today with RLHF, where you have to aggregate preferences across a lot of different humans. And it'll be maybe more marked with future, more powerful systems. But when you say, well, we want these eventual AI systems that are going to fully replace humans as part of these firms to be aligned, what does that mean? Like, will it mean that they're basically do what the user wants them to do? Does it mean that they have to result in some sort of global outcome that we're happy with as the kind of people, the stakeholders in OpenAI? Like what concretely would that mean?

• 1:10:19Speaker 01

  If the models are being used, like for these higher stakes use cases, then we would have to think about RLHF in a much different way than we are right now. So I would say we're not quite, yeah, we're not quite ready for that or the current methods. Might not be completely sufficient, but I would say,

• 1:10:43Speaker 01

  I would say we would need to make compromises between the needs of the different stakeholders involved. So, so we have this, this document that, that we're releasing called the model spec. And it's about how we want our models to behave in, in the API and in chat GPT. And we sort of, we try to talk about this issue where there are different stakeholders involved. And sometimes there are conflicts between what they might want, like, like the, in our case, we were thinking of the stakeholders as the user, or the end user, that means like someone sitting in front of chat GPT or, or some other app.

• 1:11:26Speaker 01

  The developer, so this is like someone using the API, who might be serving other end users with their app, like the,

• 1:11:35Speaker 01

  the platform, which is OpenAI, like, we don't want the models to expose, like, expose us to legal risk and so forth. And then the rest of the humanity, including people not part of the, like, who might not be users or customers or anything. So obviously, like, the user might ask,

• 1:11:59Speaker 01

  ask the model to do something that we think is, like, actively harmful to other people. And so we might have to refuse that. By the way, this isn't the order of priority necessarily. So this is just like, we have these four or so classes of stakeholder. Actually, you could also say, maybe in the future, we'll say the model itself, the model itself. So I would say we're not going there yet. But anyway, they,

• 1:12:26Speaker 01

  yeah, we have these different stakeholders. Sometimes they have conflicting demands and we have to make some call on how to resolve those conflicts. And it's not always obvious how to do that.

• 1:12:39Speaker 01

  So I would say we had to think through, yeah, we, we just had to think through the trade offs and basically the, like, the rough heuristic is that we mostly want the models to follow your instructions and be helpful to the user and the developer.

• 1:12:57Speaker 01

  But when this impinges on other people's and other people's happiness or way of life, this becomes a problem and we have to block certain kinds of usage. But we don't want to be too, we mostly want the models to just be an extension of people's will and do what they say. We don't want to be too paternalistic. We want to be kind of neutral and not like impose our opinions on people. Yeah, we want to mostly let people do what they want with the models.

• 1:13:32Dwarkesh Patel?

  I got a chance to read the spec beforehand and it was, I guess, a question of how well that transfers over to how the model itself behaves. But I was impressed with how sensible the trade offs were. Like, it made sense that this is the, it was like explicitly stated, the actual edge cases rather than the kinds of things where everybody can, which are obvious. Like, in this case, you really are going after the edge cases.

• 1:13:58Speaker 01

  Yeah, we wanted it to be very actionable so that it wasn't just a bunch of nice sounding principles, but it was like each, each example kind of tells you something about some non-obvious situation and reasons through that situation.

• 1:14:10Dwarkesh Patel?

  Yeah. Okay, now I have a couple of questions about the state of the research itself. So,

• 1:14:18Dwarkesh Patel?

  famously in the social sciences, things are really hard to replicate. And that's a question about how much of the science there is real versus these manufactured bespoke sorts of experiments. When you look at the average ML paper, does it feel like the, like a really solid piece of literature? Or does it feel often like it's the equivalent of what P-hacking is in the social sciences?

• 1:14:42Speaker 01

  Everyone has their complaints about the ML literature, but I would say overall, I think it's a relatively healthy field compared to some other ones like in the social sciences. Just because, well, it's grounded, it's largely grounded in practicality and getting things to work and If you publish something that can't be replicated easily, then people will just forget about it. And it's like accepted that often you don't just report someone's number from their paper, you also try to re-implement their method and compare it to your method on the same, say, on the same training data set. So I think if you publish

• 1:15:27Speaker 01

  methods that are like really hard to implement, or are really finicky, they'll tend to get forgotten. And as a result, people actually try to open source their work a lot. I guess there's also, there's various incentives that, there's various unfavorable incentives, like, yeah, people are incentivized to make the baseline methods, like the methods they're comparing to worse. And like, there are other like mild pathologies, like trying to make your method seem sophisticated mathematically. But I would say overall, I feel like the

• 1:16:06Speaker 01

  field makes progress. And I would probably like to see a little bit more science and trying to understand things, rather than more like hill climbing on benchmarks and trying to propose new methods. And there's been a decent amount of that recently. But yeah, I think it's, we could use more of that. And I think that's a good thing for like academics to work on. Oh, yeah. And the social sciences, on a slightly different note, I think actually, I'd be really excited to see more research using base models to

• 1:16:45Speaker 01

  do simulated social science, because these models have a probabilistic model of the whole world. And you can set up like a simulated questionnaire, or like a conversation.

• 1:17:00Speaker 01

  Like, and you can look at how any, anything is correlated, like any, any traits that you might imagine, you can see how they might be correlated with other traits. So it'd be pretty cool to see if people could replicate some of the like more notable results in the social sciences, like, like moral foundations and that sort of thing.

• 1:17:19Dwarkesh Patel?

  Yeah.

• 1:17:19Speaker 01

  By just like prompting base models in different ways and seeing what's correlated.

• 1:17:24Dwarkesh Patel?

  What is that Stanford experiment? The, um, the, the one where they, the Ash conformity test, right?

• 1:17:30Speaker 01

  Oh, yeah.

• 1:17:30Dwarkesh Patel?

  It'd be fun if that replicated with the language models as well.

• 1:17:35Dwarkesh Patel?

  Very interesting. With the rest of the research that happens at big labs, how much of it is increasing the, or decreasing the amount of compute you need to get a certain result as an actual computer multiplier versus how much of it is things that are just making the learning more stable and just building out the infrastructure? I guess the broader question I'm trying to ask is, since GPT-4, does it feel like with the same amount of compute, you can train a much better model? Or does it feel like, oh, we've like made sure that learning can happen better and in a more scalable, scalable way with GPT-5? But it's not like, uh, we can train GPT-4 with like GPT-3.5 budget now or something like that.

• 1:18:14Speaker 01

  Yeah, well, um, definitely there's always progress in improving the efficiency.

• 1:18:20Speaker 01

  Whenever you have a 1D performance metric, you're going to find that, uh, like, uh, different improvements, can kind of substitute for each other. Uh, so you might find like, uh, you might find that you, uh, post-training and, um, pre-training both improve the metrics or, uh, like improve, uh, they, they, they'll have a different, slightly different profile of which metrics they improve. But, uh, if, if at the end of the day, you have a single number, they're both gonna, they're gonna substitute for each other.

• 1:18:50Dwarkesh Patel?

  Yeah.

• 1:18:50Speaker 01

  Uh, somewhat. So I would say for, uh, something like a, like a human evaluation, like what do humans prefer? We've definitely made a lot of progress on both sides, uh, like pre-training and post-training and improving that.

• 1:19:03Dwarkesh Patel?

  Hmm. Okay, a couple of rapid fire questions about RLHF. So obviously RLHF is important to make these models useful. So maybe the lobotomized description is inaccurate. But there is a sense in which all of these models, once they're put in a chatbot form, have a very similar way of speaking. They really want to delve into things. They want to turn things into bullet points. They often seem sort of have this formal and dull way of speaking. And there's complaints that they're not as creative, like what we're talking about before with, it can only do rhyming poetry and not not rhyming until recently, I guess. Is that a result of the particular way in which RLHF happens now?

• 1:19:48Dwarkesh Patel?

  And if so, like, is it because of who the raiders are? Is it because of what the loss function is? Why is this the way all chatbots look?

• 1:19:54Speaker 01

  Yeah, I would say there's a decent amount of room for variation in exactly how you do the training process. And I think we have a lot of, I'd say we're actively trying to improve this and make the writing more lively and more fun. And I think we've made some progress, like improving the personality of ChatGPT. So it is, it is more fun and like you, it's, it's better when you're trying to chit chat with it and so forth. It's less robotic. I would say, yes, it's a kind of interesting question how some of the tics came about, like, like the word delve. I've actually caught myself using the word a bit recently.

• 1:20:36Speaker 01

  So I don't know if it rubbed off on me from, from the model or what. But actually, I think there's also, there might be some funny effects going on where there's like unintentional distillation happening between the language model providers where, like if you hire someone to go do a labeling task, they might just be feeding, feeding it into a model, they might just be pulling up their favorite chatbot and like feeding it in and having the model do the task and then copying pasting it back. So there might be. That that might account for some of the convergence, but also I think some of the things we're seeing are just what what people like. I mean, I think people do like bullet

• 1:21:18Speaker 01

  points. They like the structured responses. People do often like the big info dumps that they get from the models. So yeah, I think there's so it's not completely clear how much is just a quirk of the particular like choices and like design of the post-training processes and how much is actually intrinsic to like what people actually want.

• 1:21:52Dwarkesh Patel?

  It does seem persistently more verbose than some people want, and maybe just because during the labeling stage, the raiders will prefer the more verbose answer. But I wonder if it's if it's inherent to because of the how it's free trading, the stop sequence doesn't come up that often and like it really wants to just keep going or.

• 1:22:14Speaker 01

  There might be some biases in the labeling that lead to verboceity, like the fact that we tend to train for one message at a time rather than the full interaction. So like, if you only see one message, then there's something that just has like a clarifying question or maybe a short response with an invitation to follow up is going to be it's going to look less complete than something that covers all possibilities. There's also a question of what people whether people's preferences would change depending on how fast the model is streaming its output. Like, like clearly, if you're sitting there waiting for it to waiting for the tokens to come out, you're going to prefer that it gets to the point.

• 1:22:57Dwarkesh Patel?

  Yeah.

• 1:22:57Speaker 01

  But if it just gives you a like a dump of text instantly, maybe you don't actually care if there's a bunch of boilerplate or like if there's a bunch of stuff you're in a scam, you'd rather just have it all there.

• 1:23:08Dwarkesh Patel?

  Yeah. The reward model is, I think, such an interesting artifact because it's the closest thing we have to an aggregation of what people want, what preferences they have.

• 1:23:23Dwarkesh Patel?

  When you think about models that are much smarter, the kind of way in which we'll, I mean, one hope would be that you could just give a sort of like list of things we want that are not a sort of trivial and obvious kinds of like UN Declaration of Rights things. On the other hand, I think I heard you make the point that, well, a lot of our preferences and values are very subtle and so that they might be best represented through these pair-wise preferences. When you think of a GPT-6 or GPT-7 level model, are we giving it more of like a written instructions or are we still doing which kind, you know, these sorts of like sub-level preferences?

• 1:24:06Speaker 01

  Yeah, that's, that's a good question. So I think like these preference models do learn a lot of subtleties of, yeah, subtleties about what, what people prefer that are, would be hard to articulate in a, like in an instruction manual.

• 1:24:22Dwarkesh Patel?

  Yeah.

• 1:24:23Speaker 01

  Maybe if you like, obviously you can write an, like an instruction manual that has lots of examples of comparisons and that's like, that's what the model spec has. It has a lot of examples with some explanation. So it's not clear what the optimal format is for describing preferences. I would guess that whatever you can get out of like a big data set that captures fuzzy preferences, you can distill it down to a, like a smaller, a shorter document that mostly captures the ideas. And, and I would think that the big, like, like the bigger models are, like

• 1:25:07Speaker 01

  they do, like, learn a lot of these concepts automatically of what people might find, like they'll have some.

• 1:25:17Speaker 01

  They'll just learn from all the pre-training data what people would find useful and helpful. And what they'll have like some, there'll be some complex like, like moral theories that they can, they have and they can, but of course, there's still a lot of room to latch on to a different, like different style or a different morality. So I think like when we have, like if we were to write a doc or if we're going to align these models, what we're doing is latching on to a specific, like specific style, a specific morality. And there's still like a decent,

• 1:25:57Speaker 01

  you still need a decent, decently long document to, to capture exactly what you want.

• 1:26:03Dwarkesh Patel?

  Yeah. How much of a moat is better post-training?

• 1:26:08Dwarkesh Patel?

  Currently, companies distinguish themselves by, well, how big is our model and so forth? Will it be a big moat who has figured out all the finickiness that you were talking about earlier with regards to all this data?

• 1:26:19Speaker 01

  I think there's something of a moat because it's just a very complex operation. And there's, so it takes, you have to have a lot of skilled people doing it. And so there's a lot of tacit knowledge and there's a lot of organizational knowledge that's required. So, so I think, yeah, I think post-training, like to create a model that actually like has all the functionality people care about is pretty complicated. It requires a pretty complicated effort. So, and this requires a lot of, this is basically

• 1:27:01Speaker 01

  an accumulation of a lot of R and D. So I would say, I would say that makes it somewhat of a moat that it's not trivial to spin this up immediately.

• 1:27:14Speaker 01

  It does seem like, like the same companies that are putting together the most serious pre-training efforts are also putting together the serious post-training efforts. So it seems like it is, it is somewhat, somewhat possible to copy or to, to spin up more of these efforts. There's also like one force that sort of makes it less of a mode is that you can like distill the models or you can take someone else's model and clone the outputs, or you can use someone else's model as a judge to like do comparisons. So I think, like the more big league people probably aren't doing that because it goes against

• 1:27:58Speaker 01

  terms of service policies, but, and it would also be a sort of hurt to hit to their pride, but I would expect some of the smaller players are doing that to get off the ground.

• 1:28:07Dwarkesh Patel?

  Hmm.

• 1:28:08Speaker 01

  And that catches you up to a large extent.

• 1:28:11Dwarkesh Patel?

  I guess that was clear of the mode. What, what is the median raider like? Where are they based? What are their politics? What is their sort of knowledge level?

• 1:28:20Speaker 01

  I would say it's, it varies a lot. So we've definitely hired raiders with different skills or for different kinds of tasks or projects. So I would say like a decent, a decent mental model is just look at people who are on Upwork and other platforms like that, like who's doing sort of odd, odd jobs with remote work. So it's, yeah, it's a pretty international group. There's, there's a decent number of people in the US. We hire different people, like different groups of people for different types of labeling, like whether

• 1:29:04Speaker 01

  we're more focused on writing or like STEM tasks. So people doing STEM tasks are more likely to be in India or other sort of like middle or lower middle income countries. Whereas people doing more like English writing and composition tend more to be like US-based.

• 1:29:28Speaker 01

  So yeah, and I'd say there there have been times when we needed to hire different experts for some of our campaigns. Some of the people are very, some of them are very talented and like we even find that they're like at least as good as as us, the researchers at doing these tasks and they're like much more careful than us. So I would say, I would say the people we have now are quite skilled and conscientious.

• 1:29:54Dwarkesh Patel?

  Hmm. With with regards to the sort of plateau narrative, one of the things I've heard is that a lot of the abilities these models have to help you with specific things is related to the having very closely mashed labels within the supervised fine tuning dataset. Is that true? Like if if it can teach me how to use FFmpeg correctly, like there's somebody who's like doing figuring out, seeing the inputs and seeing what flags you need to add and some human is figuring that out and smashing to that. And is it? Yeah. Do you need to hire like all these label owners who have domain expertise in all these different domains?

• 1:30:36Dwarkesh Patel?

  Because if that's the case, it seems like it would be a much bigger slog to get these models to be smarter and smarter over time.

• 1:30:42Speaker 01

  Right. You don't exactly need that because yeah, you can get quite a bit out of generalization. So if you like, like the base model has already been trained on tons of documentation, tons of code with shell scripts and so forth. So it's already seen all the FFmpeg man pages and lots of bash scripts and everything and it's, so like the base, even just giving the base model a good few shot prompts, you can get it to answer queries like this and

• 1:31:18Speaker 01

  just training a preference model like for helpfulness will.

• 1:31:23Speaker 01

  Even if you don't train it on, probably even if you don't train it on any STEM, it'll somewhat generalize to STEM and like, so not only do you not need like examples of how to use FFmpeg, you might not even need anything with programming to get some reasonable behavior in the programming domain.

• 1:31:46Dwarkesh Patel?

  Maybe final question is, we've touched on this in different ways, but to put it together, so you say you're training on much more multimodal data, presumably like these things understand what screens look like and will be able to interact with it in a much more coherent way. And also you're going to do this long horizon RL, so they'll be able to act as agents and assistants who can be part of your workflow in a much more integrated way.

• 1:32:14Dwarkesh Patel?

  What, what do you expect that to look like and what will be the next steps from there? So suppose by the end of the year or next year, you have something that's like an assistant who can work with you on your screen. Does that seem like a first of all, a sensible thing to expect? And then where does it go from there?

• 1:32:31Speaker 01

  I would definitely, yeah, I would expect things to move in that direction. It's unclear what's going to be the best form factor, whether it's like something that's, it's like a Clippy that's on your computer and helping you with something, or if it's more like a, like helpful colleague in the cloud. So we'll see which kinds of form factors work the best. And I would expect people to try all of them out. Yeah, I would expect more like, yeah, I would expect something like, yeah, the mental model of like a helpful assistant or helpful colleague to become more real, where you can share more of your

• 1:33:13Speaker 01

  everyday work or have it like, instead of just giving it one-off queries, you would have a whole project that you're doing and it knows about everything you've done on that project so far. You can tell it, it can like even proactively make suggestions like,

• 1:33:30Speaker 01

  Maybe you can tell it, oh yeah, like, remember to ask me about this and if I've made any progress on it. So I think like proactivity is one thing that's been missing. Yeah, I'd really love to see better, like,

• 1:33:46Speaker 01

  a more like moving away from sort of one-off queries, like using the model kind of like a search engine.

• 1:33:53Dwarkesh Patel?

  Yeah.

• 1:33:53Speaker 01

  A smarter search engine and more towards like having a whole project that I'm like doing in collaboration with the model. And it knows everything I've done. It's proactively like suggesting things for me to try, or it's going and doing work in the background.

• 1:34:09Dwarkesh Patel?

  Yeah, that's, that's really interesting. Well, by the way, it's a final question. What is your, what is your median timeline? It replaces your job.

• 1:34:16Speaker 01

  Yeah, it replaces my job.

• 1:34:21Speaker 01

  Maybe like five years.

• 1:34:23Dwarkesh Patel?

  Yeah, pretty soon.

• 1:34:24Speaker 01

  Yeah.

• 1:34:26Dwarkesh Patel?

  Interesting. Okay. Well, John, this is super interesting.

• 1:34:30Dwarkesh Patel?

  Yeah, thanks so much for making the time. I think this is like, seems like one of the parts of the AI process that are super important and people don't understand that much about. So it was super interesting to delve into it and.

• 1:34:42Speaker 01

  Yeah.

• 1:34:43Dwarkesh Patel?

  Your thoughts on it.

• 1:34:43Speaker 01

  But yeah, thanks for having me on the podcast. It was fun to talk about all this stuff.

• 1:34:49Dwarkesh Patel?

  Hey, everybody. I hope you enjoyed that episode with John. He's just a very thoughtful guy and it's super interesting to learn about the way in which these models become the kind of Shagat that they are. Anyways, as you can see, I'm now doing ads on the podcast. So if you'd like to advertise, you can reach out at the link in the description. And of course, if you enjoyed the episode, it's really helpful if you can share it with other people who you think might enjoy it. Your friends, group chats, Twitter, whatever else. See you on the next one. Cheers.