John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI

Today, I have the pleasure to speak with John Schulman, who is one of the co-founders of OpenAI and leads the post-training team here. And he also led the creation of ChatGBT 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.
Thanks for having me on the podcast. I'm a big fan.
Oh, thank you. Thank you for saying that.
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? 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 web pages from the internet. And the model is also trained to maximize likelihood where it has to put a probability on everything.
So 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 we're training to maximize log probability, it ends up being

very calibrated. So it can not only generate all of the content of the web, it can also assign probabilities to everything.
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 we do a post-training, uh, we're usually targeting a narrower, um, range of behavior where we basically want the model to behave like this kind of chat assistant.
And, uh, it's a, it's a more specific persona where it's, um, trying to be helpful. It's not trying to imitate a person it's, um, answering your questions or doing your tasks.
Um, and, uh, 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, uh, this raw content from the web. Yeah.
Okay. I think maybe I should take a step back and ask, um, right now we have these are pretty good at acting as chatbots.
Just taking a step back from how these processes work currently, what will the models release by the end of, kinds of things the models release at 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? oh yeah five years yeah i think uh the models will get quite a bit better um but a bit better in the course of five years. So, I mean, I think even in one or two years, we'll find that you can use them for a lot of more involved tasks than they can do now.
So for example, right now,, so for example, right now, um, uh, like you could imagine, uh, having the models do carry out a whole coding project instead of maybe giving you one suggestion on how to write a function. So, uh, 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, uh, it'll, um, go and write, many files and test it, look at the output, iterate on that a bit.
So just much more complex tasks. 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? Yeah, I would say this will come from some combination of just training the models to do harder tasks like this so um just uh like i'd say right uh the models aren't um aren't particularly uh like most of the uh training data is more like doing single steps at a time and i would expect us to do more uh for training the models to carry out these longer projects um so i'd say any any kind of training uh any like doing rl uh to learn how to do these tasks uh however you do it whether it's whether you're supervising the final output or supervising it like each step um i think any kind, uh, carrying out these long projects is going to make them a lot better.
And, uh, since, uh, the, the whole, um, area is pretty new, I'd say there's just a lot of low hanging fruit. Interesting.
Um, do it in doing this kind of training. So I'd say that's one thing.
Um, also I would expect that as the models get better they're just better at recovering uh from errors or they have um just uh they're better at um at dealing with dealing with edge cases or when things go wrong they know how to recover from it so uh 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 i'm not sure i understood actually how i want to understand more explicitly how the generalization helps you get back on track can you say more about that i'm I'm not sure I understood actually how I want to understand more explicitly 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 are connected.
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 if you collect a diverse data set, you're going to get a little bit of everything in it.
And, uh, and if you have models that generalize really well, uh, even if there's just a couple examples of getting back on track, or even, um, like maybe in the pre-training, there's examples of getting back on track, then like the model will be able to generalize from, uh, 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. 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 streamlined process because just getting to that point where you're already more sample efficient and then you just go to the years of carrying out tasks or something? Yeah, I would say at a high level, I would agree that longer 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 set it up in a very careful way or design the experiment in a certain, design the experiment in a certain, certain way.
Because, uh, I would say there might end up being some phase transitions where, um, once you get to a certain level, um, you can, uh, deal with, um, you can deal with much longer tasks. So for example, people, um, uh, like, I think i think when people um like think when people do planning for uh at different time scales i'm not sure they use completely different mechanisms yeah so uh we probably use the same uh mental machinery if we're thinking about one month from now one year from now yeah uh or uh like a hundred years from hundred years from now.
So we're not actually doing some kind of reinforcement learning where we need to worry about a discount factor that covers that timescale and so forth. So I think using language, you can describe all of these different timescales 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, um, some kind of, um, I don't know if it's a phase transition, but, uh, like there's some capabilities that work at multiple scales.
Yeah. Well, okay.
So 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 so writing your code in a way that's coherent and aligns with the broader goals you have 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 or, and if not,

then what is remaining after you can plan for a year and execute projects that take that long? 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 wouldn't expect everything to be immediately solved by doing any training like this. I would think there will 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 all capabilities, but yeah, it's not clear. But it might, like some improvement in the ability to do long horizon tasks might go quite far.
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 of the bottlenecks so it has all these representations of pre-training now it can do act coherently for a long period of time because of long horizon rl what's remaining yeah um maybe there's some uh there's some other um experience that human experts bring to different tasks like having some some taste or dealing with ambiguity better. So I could imagine that if we want to do something like research, like those, those kinds of considerations come into play.
Yeah, obviously there's, they're going to be just sort of mundane limitations around like affordances of the 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.
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? Compensating for their strengths and weaknesses, how would that look different from the current, you know, UIs we have for humans? Yeah, that's, that's an interesting question. I mean, um, I would expect that models will be able to use, uh, websites that are designed for humans, uh, just by using vision, uh, like when the vision capabilities get a bit better.
Um, so there wouldn't be an immediate need to change them. Um, on the other hand, some websites, uh, that are going to benefit a lot from AI, uh, AIs being able to use them will probably want to, uh, design to be better UXs for AIs.
So, um, I'm not sure exactly what that would mean, but probably, uh, like assuming that our, um, our models are still better in text mode, than like reading text out of images uh you'd probably want to have a good text-based representation for the models so uh and also um just uh a good uh like indication of what are all the things that can be interacted with um but i guess i wouldn't expect the web to get um redesigned to have APIs everywhere, because I would expect that we can get models to use the same kind of UIs that humans use. Right.
I mean, I guess that's been the big lesson of language models, right, that they can act in the similar affordances that humans have. So the point you made earlier about this process could be more sample efficient because it could generalize from its experiences in free training of how to get unstuck in different scenarios.
I'm curious what the strongest evidence of this kind of generalization and transfer you've seen is. Yeah, like because the big question, it seems, about the future abilities as like how, 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. There's, uh, definitely been some interesting, um, uh, instance of generalization in post training.
Like, um, uh, one well-known phenomenon is if you do all your fine-tuning with english data you'll automatically um you'll have the model also um uh behave behaving well in other languages so if you train the assistant on english data it'll also um do something reasonable in spanish say and uh sometimes you might get um you might get the wrong behavior in terms of whether it replies in English or replies in Spanish. But usually you get the right behavior there as well.
Like you get it to respond in Spanish to Spanish queries. So that's one kind of interesting instance of generalization that you just sort of latch onto the right helpful persona.
And then you automatically do the right thing in different languages. We've seen some versions of this with, um, multimodal data where, uh, if you do, um, text only fine tuning, you also get reasonable behavior with images.
Um, uh, early on in, um, chat GVT, we, uh, we were trying to fix some issues in terms of the model uh understanding its own uh limitations like um like early versions of the model would think they uh could like send you an email or call you call an uber or something like uh the model would try to play the assistant and it would say oh yeah of course i i I sent that email. And obviously it didn't.
So we started collecting 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 30 examples. Well, we had, I don't know, pretty small number of examples showing uh behavior of like explaining that the model can't doesn't have this capability and that generalize pretty well to all sorts of capabilities we didn't train for okay so i i still want to go back to this because i'm not sure i understood uh like if you have uh 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 you're interacting with this as a colleague? And it's like 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 of this is the capabilities you think might be possible yeah it's hard to say exactly what will be the deficit i mean i would say that uh when you talk to the models today they have various um weaknesses besides uh long-term coherence in terms of also like, um, like really, uh, thinking hard about things or paying attention to what you ask them. Uh, so, um, I would say, um, I wouldn't expect, um, like just improving the, uh, coherence a little bit to like, um, to be all it takes to get to AGI.
But, um, I guess I wouldn't be able to articulate exactly what the main weaknesses that'll stop them from, uh, like being a fully functional, uh, colleague. It seems like then you should be planning for the possibility you would have AGI very soon.
Yeah, I think it's, uh, I think that would be reasonable. So what's the plan if like, if there's no other bottlenecks next year or something, you got AGI, what's the plan? Well, I would say that if AGI came way sooner than expected, uh, we would definitely want to, we would want to be careful about it.
And we would, uh, we might want to, um, like, uh, slow down a little bit on, uh, training and deployment until we're pretty sure we know, uh, we, we can deal with it safely. Um, and we, we have a, um, a pretty good handle on what it's going to do, what it, what it can do.
So I think, uh, yeah, we would have to be, we'd have to be very careful, um, if it happened way sooner than expected, because I think, uh, our understanding is rudimentary in a lot of ways still. And what would, what would being careful mean? Like, uh, cause presumably you're already careful, right? You do these evaluations before you're, um, yeah, I would say just like, uh, uh, maybe not, um, uh, not training the even smarter version of not being really careful when you do train it, that it's not, uh, it's, um, like properly sandboxed and everything, um, maybe not deploying it at scale, um, or yeah, being, uh, yeah, being careful, careful about what, um, what scale you deploy it.
Hmm. Yeah.
I guess I'm not. Okay.
So let's just play with a scenario like it happens next year and then you're you're not training a smarter system, but you're deploying somewhat in a measured way.

um i yeah i'm wondering well presumably if this is just this isn't 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 about? What is every company doing in this scenario? this scenario 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 gonna happen next year but it's still useful to have the conversation and maybe it's like two or three years instead but um yeah i guess two or three years is still pretty soon i do think uh you probably need some coordination like uh, uh, everyone needs to agree on some, uh, on some reasonable, uh, like limits to deployment or to further training, uh, for this to work. Otherwise, uh, otherwise you have the, the race dynamics where everyone's trying to, everyone's trying to stay ahead and, uh, like everyone's, uh, like, and that might require compromising on safety.
So I think you would probably need some coordination among the, uh, larger entities that are doing this kind of training. And so you're coordinating to, um, I guess pause deployment until, until what exactly? Like until you figure out what's happening in the models? Like pause, uh, either, either uh further training pause deployment uh like uh avoid certain types of training that we think might be riskier uh so just uh like setting up some reasonable rules for uh um like uh what what everyone should do to yeah having everyone somewhat limit uh limit these things and but uh limit to what end because i guess at some point then you're gonna have to like the the potential energy that's within this intelligence will uh you know it'll be unleashed so uh what what what what is the plan to do like suppose in two years we get the agi and now everybody's freaking out and so now now the AI companies have paused.
And now what, or what would be the plan to wait till or? Yeah, that's, I don't have a good answer to that. I mean, I would say if we can, if everyone is going to coordinate like that, I think we would be, that would be an okay scenario.
That would be a pretty good scenario because I do think building these models is very capital-intensive and there are a lot of complex pieces, so it's not like everyone's going to go and recreate the 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 uh how you would maintain this uh this equilibrium for a long period of time but i think if we got to that point um we would be in an okay position or would be i guess i'm curious like um uh 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 can push themselves to the server um and i'm now we got everybody coordinated but i'm not sure what what we do next in this in this world we're like why that why that sets us up for a good outcome yeah i would say if we had everyone um reasonably coordinated we could uh figure out some and we felt like we had solved the technical problems around alignment well enough to be able to uh deploy like really smart ais that um can like uh like to 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 that would be great.
We could go ahead and safely deploy these systems and it would usher in a lot of prosperity and a new, like, much more rapid phase of scientific advancement and so forth.

So I think that would be what the good scenario would look like.

Okay, so 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 have agreed to pause until we've figured out that we're building aligned systems that uh are not themselves going to attempt to take over 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 well i would say if we um if we can deploy uh 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 it's not the scenario where everyone has to coordinate and lock things down and safely release things because it would lead to this big buildup in potential energy, potentially. So I rather uh some scenario where we're just um continually releasing things that are a little better than what came before and then we uh while like making sure we're um confident that each um diff is right like improving uh improving the safety and alignment uh in like uh correspondence to the improvement and capability.
So, and 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

a continuous 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 want to, but like the type of thing you might want to do to make that more, more acceptable would be you would want to do

a lot of testing, like simulated deployment that you expect.

So red teaming of sorts, like you'd want to do that in a way that you feel is like much less favorable than or much more likely to fail than the thing you're planning to do in the real world. you'd want to have a really good monitoring system so that you can uh like if something does start to go go wrong with the deployed system you can uh you feel like it's going to be uh detectable immediately like you've got maybe you've got something watching over uh the deployed ais and what they're doing and looking for signs of trouble so i uh so i would want to um yeah i would say, yeah, I would say just, um, you'd want some defense in depth.
Like you'd want to have some combination of, uh, like the model itself, uh, seems to be, um, like really well behaved and have like impeccable, uh, moral compass and everything. And you're pretty confident that it's, it's extremely resistant to any kind of takeover attempt or something or like severe misuse and then you would also want to have like uh really good monitoring on top of it so yeah you could detect any kind of any trouble what are you keeping track of while you're doing long horizon rl or when you eventually start doing it that uh you you could notice this sort of discontinuous jump before you deployed these systems broadly? I would say you would want to have a lot of evals that you're running during the training process.
And 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 you'd want to be pretty careful when you do this kind of training if you see um a lot of um potentially scary capabilities um if those seem close i mean like uh i would say it's not something uh we would want to we have to be scared of right now because uh right now it's hard to get the models to do anything coherent.
But if they started to get really good, I think we would have to take some of these questions seriously. And we would want to have a lot of evals that sort of test them for misbehavior.
Or I guess that's like for the alignment of the models, we want to check that they're not going to sort of turn against us or something. But you might also want to look for like discontinuous jumps and capabilities.
You'd want to have lots of evals for the capabilities of the models i mean also i guess you you'd also want to 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 um i would say there's like um that doesn't seem like the hardest thing to do i mean if uh like the way we train them with rlhf uh 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 uh pleasing to a human and it has no concern about anything else in the world other than whether this text it produces is approved. So obviously, if you were doing something 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 things that wouldn't make sense to a human in the process of producing its final result uh but i guess um it wouldn't necessarily have an incentive to do anything other than produce a very high quality um output at the end so it um like it's not um yeah so i guess you have these old uh points about like instrumental convergence like the model is going to want to take over the world so it can produce this awesome piece of code at the end 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 um it's a little hard to imagine why um for some like fairly well specified tasks like that you would want to first take over the world um but of course uh yeah if you had a task like make money uh then maybe uh that would lead to some nefarious behavior as a um instrumental goal yeah okay so before we get back to that i think let's step back and talk about like uh today's um rlhf systems and everything um but i do want to follow that that's a point it's kind of interesting um 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 i'm 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 yeah i would say there are probably some analogies with a drive or a goal in humans so in that um you have um you're trying to steer towards a certain set of states rather than some other states.
Um, and so I would, I would think that our concept of a drive or a goal has, um, other, um, elements like, uh, like the feeling of satisfaction you get for achieving it. And, uh, and those things might, um, be more like have more to do with the learning algorithm than, uh, what the model does at runtime, uh, when you just have a fixed model.
So I would say, I would say there are probably some analogies though. Um, it's, uh, I don't know exactly, um, like how, how close it is, but I would say to some extent it is, um, it, the models, The models do have drives and goals in some meaningful way.
And 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 they're going to judge as correct. I've heard two ideas in terms of using that internal 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 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 after you know while it's deployed which one do expect it to be closer to when it's like really good at reasoning is it because it's doing just a bunch of inference clouds or is it just because you've trained it to do well at that well i would say you could define reasoning as um tasks that require some kind of uh like computation um at test time or maybe some kind of uh deduction um so so by definition reasoning would be tasks that require, um, like some test time computation and, uh, like step-by-step computation.
Um, on the other hand, I would also, um, expect to gain a lot out of, um, like doing some kind of, um, training time computation or practice at training time. Uh, so, so I would think that, um, you get the best results by the best results by combining these two things.
So my uncle had prostate cancer, and I wanted to know my own risk. And I got a 23andMe test, but it was mostly useless.
I mean, the whole, what are your odds of liking chocolate is not what I was looking for. So I exported my data onto nucleus genomics, and immediately I got my risk profile for almost two dozen diseases, including prostate cancer.
And it turns out my risk is higher than 97% of people with my ancestry, which is a very useful thing to know because now I know to get screened early. Ironically, tests like 23andMe don't even look at the variants which have the largest impact, including for prostate cancer.
Many people don't know this, but 23andMe looks at less than 0.1% of your DNA and that's why I've pre-ordered Nucleus Premium whole genome sequencing. It's the only clinical grade test that reads 100% of your DNA.
I've spent a lot of time digging into this company and I think will be a big change in what we can get out of genetic tests. So if you want live a long and healthy life you can pre-order nucleus premium at my nucleus.com all right back to john 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 computing 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 that is, it's more sample efficient there, but it's destroyed with each instance. I'm curious if you think that there's a path for something in between those where it's not destroyed with each instance, but it's also not as, um, uh, uh, not as sort of frivolous as just seeing trillions of tokens where it's more deliberate and active.
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? I'm not sure if memory, it might be memory.
I don't have context, but certainly like when I'm trying I'm trying to prepare for this conversation, uh, it feels like I think of like what 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 I'm, what would that look like? I'm curious. I see.
So it's not just a memory, but it's also somewhat likeizing to a task that um specializing to a certain task or putting a lot of effort into like some particular project and i'm not even sure specialization more so um 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 um yeah i see so it's not just about uh finding like uh i don't know training on a bunch of sources that are relevant fine-tuning on some special domain it's also about like uh like reasoning about like developing some knowledge through your own reasoning and also yeah using some sort of uh introspection and self-knowledge to figure out what you need to learn. Yeah.
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 the 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 a hundred thousand or a million context, then that's actually quite not, um, it's not actually the bottleneck in a lot of cases, but I agree that, um, you'd probably also want to supplement that by some kind of fine tuning, like the, uh, the capabilities you get from fine tuning and in context learning are probably somewhat complimentary.
So I would expect us to want to build systems that do some kind of online learning and also have some of these, uh, cognitive skills of, uh, like introspecting on their own knowledge and, uh, seeking out new, new knowledge that fills in the holes. Uh, is this all happening at the same time? Like, uh, 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? Yeah, I would say if you're doing some kind of long horizon horizon task uh well i would uh you're learning while you do the task right so the only way to do something that involves uh a lot of steps is to um like to have learning and memory that gets updated during the task so like there's a continuum between um like uh like short-term memory memory, between short-term and long-term memory.

So I would expect this capability would start to become, like the need for it would start to become clear when we start to, uh, look at long horizon tasks more. And, uh, and to some extent, just, um, putting, um, a lot of stuff into context will probably, will take you pretty far because we have really long contexts now, but you probably also want things like fine tuning.
And as for, uh, like introspection and the ability to do active learning um that might uh like automatically fall out of the model's abilities to know what they know because they have some like um models have some calibration um regarding what they know and that's why like that's why um models don't hallucinate that badly, uh, 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.
Hmm. And how, so, uh, there's all these complicating RRL procedures, uh, that many of whomered.
How many of them will be relevant when you get to the point where the model itself is this smart, that it can act as a certain environment and interact in a more online and stable way? Is the path for progress going to be more straightforward than the kinds of solutions that were required for RL in the past? 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 who knows? I mean, maybe it's not that bad. So I think something like motor learning in animals is probably something like a policy grading algorithm.
And, uh, so for example, you're like learning how to, um, shoot baskets. Uh, 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 grading algorithm underneath. Um, 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.
Um, so I would think we would want to rely more on like in context learning, um, 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. Um, 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'll be the kind of thing that gets used in a particular task. 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 the path to go and then a chat bot 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 the your main focus and yeah what what what yeah what the process was like yeah so early um so we had um uh before chat gbt we, OpenAI had these instruction following models. And 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 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.
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.
um then at the same time there were definitely a lot of people thinking about um chat so uh so google had some papers uh 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, it was more used for sort of fun applications, like where the model would take on some persona and pretend to be that persona. It was not so functional, like help me refactor my code.
So, yeah, there are definitely people thinking about chat. I had worked on a project before looking at chat chat called uh web gpt which was more about doing question answering with the help of web browsing and retrieval and well when you do question answering uh it really wants to be in a chat because um you always want to ask follow-up questions or sometimes you need a clar the the model should ask a clarifying question because the question's 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 gpd 3.5 which was done training at the beginning of 2022 and uh that model was quite good at language and code so we quickly realized that it was actually uh quite good at coding help and that was one of the things we were excited about so yeah we worked on that uh 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 uh 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 um and then uh 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, GPT-4 finished training in August that year. And actually, the flagship RL effort at openai was the instruction following effort

because that was the models that were being deployed into production so um like the first

fine tunes of gpd4 used that um that whole stack and that was um yeah those models were really good

and everyone got really excited about that after seeing the, uh, like instruct fine tune GP fours.

Uh, 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. 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 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 uh i think the yeah the models we the chat models were like uh were clearly more um like it was an easy easier to use it was sort of more um it sort of uh 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 uh i got excited about as we were developing it that uh like i realized a lot of the things that um people thought were flaws in language models like just like blatantly hallucinating uh could be not completely fixed but you could make a lot of progress with pretty straightforward methods uh Oh, yeah. And also the other thing about chat was that when we had these instruct models, the task of complete this text, but in a nice way or in a helpful way, that's a pretty poorly defined task.
So I think that task is both confusing for the model and for the human who's supposed to do the data labeling. Whereas for chat, um, I think people had an intuitive sense of, uh, like what a helpful robot should be like.
So I think it was just much easier to tell people, uh, like, uh, to, to give for people to get the idea of what, what the model was supposed to do. Yeah.
And so that, so as a result, I think the, like the model had a much more coherent personality

and. idea of what what the model was supposed to do yeah um and uh so that so as a result i think the um like the model had a much more coherent personality and uh like it was much like easier to get um like robust like pretty sensible behavior um robustly interesting uh is it the case that anybody could have made chat gbt using your publicly available fine-tuning api um not exactly i mean uh they could have um i don't remember the status of which models were available available for fine-tuning uh you 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, um, 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, um, which, which we did, um, 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 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 uh 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 instruction following model trained with RL that was released a little before chat GBT.
So I think if you put a chat like wrapper on that, you would get something decently close. Uh, but it like that model, um, like if you just prompted it with chat, um, so, but that model had some, uh, differences in, uh, 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. 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? I would say faster than I would have expected since GPD2. Yeah.
I was pretty bought into scaling and pre-training and so forth being a good idea. But when GPD2 was done, I would say i wasn't completely uh sold on it um being uh revolutionizing everything um like i only really pivoted what i was working on and what yeah what my team was working on in um after gbd3 so after that uh we kind of got together and said oh yeah let's uh uh let's um this language model stuff works really well let's see what we can do here but uh yeah after gbd2 i wasn't quite sure yet especially if i the stuff we were talking about earlier with rl starts working better with these smarter models with a fraction of compute that is spent on training that is free training versus post-training change significantly in favor

of post-training in the future? 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 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 this methodology and probably increasing the amount of compute we put into it. The current GPT-4 has an 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?

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

Interesting.

So 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 changing what kind of annotations you're collecting.
So there's a lot of things that stack up, but together they give you a pretty good effective compute increase. Yeah, I mean, that's a huge increase.
That's like really interesting that there's this much, uh, this much room for improvement, uh, from post-training. What is, uh, what, what makes for somebody who's really good at doing this sort of RR research? Uh, I hear it's super finicky, but like, what, what, what is the sort of intuitions that you have that enable you to find these ways to mess with the data and

set up these environments i'd say i just um have a decent amount of experience at this point from uh like the different parts of the stack from like uh rl algorithms obviously since i've worked on those since grad school, grad school, uh, to like, uh, the data collection, um, like the annotation process, uh, to, um, like language playing with language models. So I, I mean, I'd say I just dabbled with these things and, uh, I'd say the people who, um, do well at this kind of research, uh, 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 both empirical and like use experiment, let experiments update your views. But you also want to think from first principles somewhat like what, um like assuming that um like learning uh works uh like what would be the ideal type of data to collect yeah and that sort of thing so because there doesn't seem to be a model released since gpd4 that seems to be significantly better there's seems to be uh 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-returning data won't actually help you get something much smarter than gpd4 um 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 maybe this is a run-on question but um one 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 is that actually the case? Do you see things like that, which suggests that there's all this 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 smarter from synthetic data. 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? 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 takes a while to train these models and to do all the prep to train a new generation of models.
So, yeah, I wouldn't draw too much from from that fact um i would say um 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 uh the nature of um pre-training to somewhat change over time as we get closer, closer to it. Um, in terms of like, uh, generalization from different types of pre-training data, um, I would say it's pretty hard to, um, do science, uh, on this type of question because you can't do that, create that many pre-trained models.
So maybe, uh, you can't train a, like model. You can't do ablation studies at GPT-4 scale.
Maybe you can train a ton of GPT-2 size models, or maybe even a GPT-3 size model with different data blends and see what you get. So I'm not aware of any public results on 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 I'm actually curious about, I mean, if one of the things is that the model gets moderates, it's bigger.
Would an ablation on a GPT-2 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 a GPT-4 level model? Right. You might not be able to conclude that if transfer fails at GPT-2 size, then it's also going to fail at a higher scale.
So it might be that for the smaller models, 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. 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.
Why is that the case? Like it's got more parameters. It saw less things and now equally as smart why did that why is that the case uh i don't think anyone has a good answer for a good explanation of the uh scaling law with um parameter count i mean there's some uh i don't even know what the uh what the best um sort of um, sort of mental model is, uh, for this.
Like,

clearly you have more capacity if you have a bigger model, but, uh, so like you should be able to eventually get, uh, lower loss, but I guess, uh, why are bigger models more sample efficient? Um, I guess you could, um, I can give you some like very sketchy, uh, explanation. like uh like they have um like you could say that the model is uh like uh sort of an uh an ensemble of a bunch of different circuits that do the computation so it has like um you could imagine that it's doing um it has a bunch of uh like computations that it's doing in parallel and it's uh like doing some like the output is a weighted combination of them uh and uh if you have more um just width of the or if you just have i mean actually width is somewhat similar to depth because uh like with residual networks uh you end up like the depth can do something similar to width in terms of like updating what's in the residual stream but uh if you yeah you could argue that uh you're learning all these things in parallel uh 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 uh one of them is lucky and uh ends up um like uh having high um like like winning guessing correctly a lot and getting up-weighted.
So that's kind of like what would be the, yeah, there's some algorithms that work this way, like mixture, what is it, mixture, some kind of mixture model or multiplicative weight update algorithm. algorithm yeah there's some algorithms that kind of work like this so uh where you have like a um some kind of mixture of uh i don't want to say mixture of experts because it means something different but uh like basically a weighted combination of experts with some learned gating uh and uh um 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 totally disjoint functions that you're taking a linear combination of it's more like a library where uh you might chain the functions together in some way so uh you like it's there's some composability um so yeah so i would just say there's like um the bigger model has a bigger library of different uh computations including lots of stuff that's kind of dormant and only being used some of the time.
But it has more space to look for those circuits to do something useful. I want to ask you about 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 let's just take 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 stage are these being unlocked and so forth? I just kind of want to understand your broader picture of what the next few years look like.

Yeah, I would expect things like, okay, new modalities to be added, uh, like, um, over time or, uh, pretty soon. Um, 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, um, AI is still, um, not a huge, uh, part of the economy. Like there's a pretty small fraction of, uh, jobs that it can help with at all.
Um, so I'd expect that to be higher over time and not just from the models, uh, improving also from people just figuring out how to integrate them into different processes. So even, even if we just, um, froze the models at their current, uh, state, um I think you would still see a lot of growth in how they're being used.
So I would 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 of technique like technically sophisticated tasks like um yeah like i gave the programming example earlier um of doing like longer projects but also helping with um various kinds of uh research so i would hope that uh we can use um ai to accelerate science in various ways and uh just um like because you can potentially have the the models like understand all the literature in a given field and be able to like uh be able to sift through tons of data um like more than a person would have patience to do so i would hope that we, yeah.
Um, well, I hope the form factor would basically be that people are, uh, still driving all of this and you have your, uh, like helpful assistance that you can use. You can sort of direct and point to lots of different problems that are useful to you.
And everyone sort of has all these, uh, AIs, uh, helping them, uh, helping them do more, get more done. 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. 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. But obviously at some point they're going to be better than everyone whatever they want to do.
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.
At that point, is it, yeah, is it just going to be like 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. 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 overseeing these important decisions So, uh, even, even if the models are good

enough to, uh, like to actually run a successful business themselves. Um, so, uh, yeah, to some extent there might be, uh, choices there.
Um, and, uh, I think people will still have different interests, uh, and what they want to different ideas for what kind of, uh,uits they want to direct their ais at and uh like they can people people could uh like um yeah do a lot of um ai doesn't necessarily have an intrinsic uh like um any kind of intrinsic desire unless we put we put it in uh the system think, uh, so people can still end up being, even if AIs, uh, like become extremely capable, uh, I would hope that people are still the drivers, uh, of like what the AIs end up doing. Yeah.
But I wonder if the economic equilibrium is so far from that, where, um, you have the equivalent of Amdahl'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 to 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, uh, plan for keeping humans in the loop. Right.
So I think if you, um, if we wanted to keep, uh, humans in the loop, uh, which seems reasonable, um, and, uh, it turned out that, um, firms with any humans in the loop were out competed with by firms that didn't have any humans. Then I think then you would obviously need some kind of regulation that, uh, like disallowed, um, having no humans in the loop for running a whole company but there's so many companies in the uh in well 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 say like you've got to keep humans in the loop and important processes but then you have to define what important processes are you've got to monitor every single company um 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 does that question make sense like i guess i'm just wondering in this situation right how do we actually monitor every single firm as a human in the loop and what happens if like china doesn't decide to do that and so forth right um yeah you either have to have uh like um every country uh agree to this regulatory regime or you would 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, yeah, this is looking a ways ahead. So it's a little hard to imagine this world before seeing anything like it.
But so, for example, like there's some questions like would, are we actually confident that AI run companies every way, or, uh, do we think they're better most of the time, but occasionally they, um, malfunction because AIs are still like, they're still less sample efficient in certain ways, like dealing with very wacky situations. So, um, so actually, uh, AI run firms have higher tail risk because they're more likely to malfunction in a big way.

So I guess there might be some practical questions like that that would also determine how things play out. Like maybe if you just require people to be accountable for various liability, this would also change the incentives a bit.
um so if it turned out that uh like ais are better at running everything and they're also

completely benevolent and we've It also changed the incentives a bit. 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 OK having the

AIs run the firms. But I think that's that might be pretty far out.
And I think we're more likely to be in a situation where we're going to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be more likely to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a situation where we're going to be in a that might be pretty far out. And I think we're more likely to be in a situation where they look better in the short term, but they still have some problems.
The AI run entities still have some serious problems. And it's actually practical considerations that push you more towards having humans in the loop, at least for the near future.
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, does that mean like will it mean that they 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 with the stakeholders in open ai like what concretely would the would that I mean, if the models are being used 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 ready for that, or the current methods might not be completely sufficient. But I would say we would need to make compromises between the needs of the different stakeholders involved.
So we have this document that we're releasing called the model spec, and it's about how we want our models to behave in the API and in ChatGPT.

And we try to talk about this issue

where there are different stakeholders involved

and sometimes there are conflicts

between what they might want.

Like in our case,

we were thinking of the stakeholders

as the user or the end user.

That means like someone sitting in front of ChatGPT or some other app. The developer.
So this is like someone using the API who might be serving other end users with their app. Like the platform, which is OpenAI.
Like we don't want the models to expose us to legal risk and so forth. And then the rest of the human, of humanity, including people not part of the, like who might not be users or customers or anything.
So obviously like the user might ask, ask the model to do something that we think is is 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. So I would say we're not going there yet.
But anyway, 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. um so uh i would say we had to think through um yeah we we just had to think through the trade-offs

and basically the, uh, like the rough heuristic is that we mostly want the models to, uh, follow your instructions and be helpful, uh, to the user and the developer. Um, but, uh, when this impinges on other people's, uh, and, um, other people's happiness or, uh, or way of life, this becomes a problem and we, we have to block certain kinds of, uh, usage.
Uh, but we don't want to be too, um, we, 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 kind of neutral and not like impose our opinions on people uh yeah we want to both mostly uh let people do what they want with uh the models i got a chance to read this back beforehand and it was uh i guess it's a question of how well that transfers over to how the model itself behaves but the i was impressed with how sensible the trade-offs were like it made sense that this is the i believe 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 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 uh each example kind of tells you something about some non-obvious uh situation and reasons through that situation yeah okay now i have a couple questions about the uh the state of the research itself so famously in the social sciences things are really hard to replicate and it'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 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 everyone has their complaints about the ml literature but i would say overall i think it's um a relatively healthy field compared to some other ones like in the social sciences um just because uh well it's grounded uh 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 accepted that often you don't just report someone's number from their paper,

you also try to reimplement their method and compare it to your method

on the same training data set.

So I think if you publish methods that are 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, um, um, like incentives, uh, that there's various unfavorable incentives. Like, um, yeah, people are incentivized to, uh, make the baseline methods, like the methods they're comparing to worse.
And, uh, like there are other, um, like mild pathologies, like trying to make your method seem sophisticated mathematically. But I would say overall, I feel like the 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 we could use more of that. And I think that's a good thing for academics to work on.
Oh yeah, on the social sciences, on a slightly different note, I think actually I'd be really excited to see more research and uh using base models to do um simulated social science uh because uh these models have a probabilistic model of the whole world and you can uh set up like a simulated questionnaire or um like a conversation and um like and you can look at how any anything is correlated like any um 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 science it's like like moral foundations and that sort of thing yeah by just like uh prompting base models in different ways and seeing what's correlated. What is that Stanford experiment?

The one where they, the ASH conformity test, right? Oh, yeah. It would be fun if that replicated with the language models as well.
Very interesting. with the rest of the research that happens at big labs how much of it is

increasing the uh or decreasing the amount of compute you need to get a certain result as an actual compute 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 computer, we 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 we can train GPT-4 with like GPT-3.5 budget now or something like that. Yeah, well, definitely there's always progress in improving the efficiency.
um whenever you have a 1d performance metric you're going to find that uh like uh different improvements um 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 have a different, slightly different profile of which metrics they improve. But if at the end of the day, you have a single number, they're both going to, they're going to substitute for each other somewhat.
So I would say for something like a human evaluation, like what do humans prefer, we've definitely made a lot of progress on both sides, like pre-training and post-training and improving that. 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 and if so like is it because of who the raters are is it because of what the loss function is why is this the way all chatbots look yeah i would say there's a decent amount of room for variation in exactly uh how you do the training process and a lot of, um, I I'd say we're, um, actively trying to improve this and make the writing more lively and, uh, and more fun.
And I think we've made some progress, like improving the personality of chat GBT. So it is, um, it is more fun and like you, it's, it's better when you're, uh, trying to chit chat with it and so forth.
Uh, it's less robotic. Um, I would say, um, yes, it's a kind of interesting question how some of the, the ticks came about, like, um, like the word delve.
I've actually caught myself using the word a bit recently. Um, 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 feeding it in and having the model do the task and then copy and pasting it back.
So that might account for some of the convergence. But also, I think some of the things we're seeing are just what people like.
I mean, I think people do like bullet points. They like the structured structured uh responses people do often like the big info dumps that they get uh from the models uh so yeah i think there's um so it's not completely clear um how much is just a quirk of uh the uh particular uh like choices and uh like design design of the particular choices and design of the post-training processes and how much is actually intrinsic to what people actually want.
It does seem persistently more verbose than some people want, and maybe just because during the labeling stage the raters will

uh prefer the more verbose answer but um i wonder if it's if it's inherent to because of the how

it's free training the stop sequence doesn't come up that often and like it really wants to just

keep going or there might be some biases in the labeling that lead to verbosity like the fact that

we tend to um train for one message at a time rather than the full interaction so uh like

Thank you. labeling that lead to verbosity, like the fact that we tend to, um, train for one message at a time rather than the full interaction.
So, uh, like if you only see one message, um, 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, um, it's going to look less complete than something that, um, covers all possibilities. a question of whether people's preferences would change depending on how fast the model is streaming its output.
Clearly, if you're sitting there waiting for the tokens to come out, you're going to prefer that it gets to the point. But if it just gives you a dump of text instantly, text instantly maybe you don't actually care if there's a bunch of boilerplate or uh like if there's a bunch of stuff you're in a skim you'd rather just have it all there yeah um 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 um when you think about models that are much smarter the kind of way in which we'll um i mean one hope would be that you could just give a sort of like list of things we want um 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 pairwise 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 subliminal preferences yeah that's that's a good question so i think uh like these preference models do learn a lot of subtleties um of uh yeah subtleties about what uh what people prefer um that are would be hard to articulate in a, like in an instruction manual.
Yeah. Um, maybe if you, um, like, uh, obviously you can write an, uh, like an instruction manual that has lots of examples of comparisons.
Um, and that's like, that's what the model spec has. It has a lot of examples with some explanation.
Um, so, uh uh it's not clear what the optimal uh format is for describing uh preferences i would guess that whatever you can get out of uh like a big data set that captures fuzzy preferences you can uh distill it down to a uh like a smaller a shorter document that mostly captures the ideas and uh and i would think that the big uh, like a smaller, a shorter document that mostly captures the ideas.

And, uh, and I would think that the big, uh, like, like the bigger models are like, they do, um, like, uh, learn a lot of these concepts automatically of what people might find. Uh, like they'll have some, uh, uh, they'll just learn from all the pre-training data, what people would find useful and helpful and uh what they'll have uh like some there'll be some complex uh like uh like moral theories uh that they can they have and they can but of course there's still a lot of uh room to latch on to a different uh like different style or a different morality so i think like when we have um like if we were to write a um a doc or if we're going to align these models what we're doing is latching on to a specific uh like specific style a specific um morality and there's still like a decent you still need a decent uh decently long document to document to capture exactly what you want.
Yeah. How much of a moat is better post-training? Currently, companies, I 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? I think there's something of a moat because it's just a very complex, uh, operation and there's, uh, so it takes, uh, you have to have a lot of, uh, skilled people doing it. And, uh, so there's a lot of tacit knowledge and, uh, um, there's, uh, a lot of organizational knowledge, uh, that's required.
So, um, so I think, um yeah i think post-training uh like to create a model that actually um like has all the need the functionality people care about um uh is pretty complicated uh it requires a pretty complicated effort um so and this um requires a lot of this is basically an accumulation of a lot of r d um so i would say um i would say that makes it somewhat of a moat that it's not trivial to spin this up immediately uh it does seem like um like the same companies that are putting together the most serious uh pre-training efforts are also putting together the serious post-training efforts. So, uh, it seems like, uh, it is, uh, it is somewhat, um, somewhat possible to copy or to, to spin up more of these efforts.
Um, there's also like one force, uh, that sort of makes it less of a mode is that you can, uh, like distill the models or you can take someone else's model and, uh, clone the outputs, or you can, uh, use someone else's model as a judge, uh, to like do comparisons. So I think, uh, like the more big league people probably aren't doing that because it goes against, uh, terms of service policies, but, and it would also be, uh uh uh 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 and that catches you up to a large extent i guess that helps you clear the mode what what is the median radar like where are they based what are their politics uh what is their sort of knowledge level i would say it's um it varies a lot so we've definitely um hired uh raiders with different um skills uh or for different kinds of tasks or um projects um so i would say a um like a decent um a decent mental model is uh just look at people who are on Upwork and other platforms like that.
Like who's doing, um, sort of odd, odd jobs with remote work. Um, so it's, um, yeah, it's a pretty international group.
There's, there's a decent number of people in the U S. Uh, we hire different, um, people, uh, like different, um, groups of people for different types of labeling, whether we're more focused on writing or STEM tasks.
So people doing STEM tasks are more likely to be in India or other sort of middle or lower middle income countries, whereas people doing more like English writing and composition tend more to be like us based. Um, so yeah.
And I'd say there, there've been times when we needed to, um, hire different experts for some of our campaigns. Uh, some of the people are very, some of them are very talented and, uh, like we even find that they're like at least as good as, as us, the researchers, at doing these tasks and they're much more careful than us.
So I would say the people we have now are quite skilled and conscientious. 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 matched labels within the super wise fine tuning data set.
uh is that true like if if it can teach me how to use ffmpeg correctly like there's somebody

who's like doing uh 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, yeah, do you need to hire like all these label rollers who have domain expertise in all these different domains? 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.

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 prompt, you can get it to answer queries like this. And just training a preference model, uh, like for helpfulness will, um will um uh even if you don't train it on um probably even if you don't train it on any stem it'll somewhat generalize to stem and uh uh like um so you not only do you not need uh like examples of how to use f of mpeg you might not even need anything with programming to get some reasonable behavior in the programming domain.
Maybe final question is, we've touched on this in different ways, but to put it together. So you say you're turning on much more multimodal data.
Presumably, these things understand what screens look like and we'll be able to interact with it in a much more coherent way. And also you're going to do this along 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. What do you expect that to look like and what will be the next steps from there? So suppose by the end of of the year or next year you have something that's like an assistant who can work with you on your screen it does that seem like first of all a sensible thing to expect and then where does it go from there i would definitely um yeah i would expect uh things to move in that direction um it's unclear what's going to be the best form factor whether it's like uh something that's uh

it's like a clippy that's on your computer and helping you with something or if it's more like a um like helpful uh colleague in the cloud so we'll see uh which kinds of form factors um work the best uh and i would expect people to try all of them out um yeah i would expect more uh like yeah i would expect something like a um yeah the mental model of a like a um helpful assistant or helpful colleague to become more real um where you can share more of your uh everyday work or have it uh 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 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.
Uh, yeah, I'd really love to

see, um, better, um, like, um, a more, uh, like moving away from sort of one-off queries, uh, like using the model, kind of like a search engine, a smarter search engine and more towards, uh, like having a whole project that, um, 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. Yeah, that's really interesting.
By the way, so final question. What is your median timeline? It replaces your job.
Yeah. Oh, it replaces my job.
Maybe like five years. yeah pretty soon yeah um interesting okay well john this is super interesting uh um yeah i thanks so much for making the time i think this seems like one of the parts of the ai process that is super important and people don't uh understand that much about so it was super interesting to delve into it and get your thoughts on it but yeah thanks for having me on the podcast it was fun to talk about all this stuff 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 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.