O3 and the Next Leap in Reasoning with OpenAI's Eric Mitchell and Brandon McKinzie

Source material

Transcript

Hi listeners, and welcome back to No Priors.

Today I'm speaking with Brandon McKinzie and Eric Mitchell, two of the minds behind OpenAI’s O3 model.

O3 is the latest in the line of reasoning models from OpenAI, super powerful, with the ability to figure out what tools to use and then use them across multi-step tasks.

We'll talk about how it was made, what's next, and how to reason about reasoning.

Brandon and Eric, welcome to No Priors.

Thanks for having us.

Yeah, thanks for having us.

Do you mind walking us through O3, what's different about it, what it was in terms of a breakthrough, in terms of a focus on reasoning, and you're adding memory and other things, versus a core foundation model, LLM, and what that is?

So O3 is our most recent model in this O-series line of models that are focused on thinking carefully before they respond.

And these models are in sort of some vaguely general sense smarter than models that don't think before they respond, you know, similarly to humans.

It's easier to be more accurate if you think before you respond.

I think the thing that is really exciting about O3 is that not only is it just smarter if you make an apples to apples comparison to our previous O-series models, it's just better at giving you correct answers of math problems or factual questions about the world or whatever.

This is true and it's great and we'll continue to train models that are smarter.

But it's also very cool because it uses a lot of tools that enhance its ability to do things that are useful for you.

So yeah, you can train a model that's really smart, but if it can't browse the web and get up to date information, there's just a limitation on how much useful stuff that model can do for you.

If the model can't actually write and execute code, there's just a limitation to how the sorts of things that an LLM can do efficiently, whereas a relatively simple Python program can solve a particular problem very easily.

So not only is the model on its own smarter than our previous O-series models, which is great, but it's also able to use all these tools that further enhance its abilities and whether that's doing research on something where you want up to date information or you want the model to do some data analysis for you, or you want the model to be able to do the data analysis and then kind of review the results and adjust course as it sees fit instead of you having to be so sort of prescriptive about each step along the way.

The model is sort of able to take these high level requests, do some due diligence on this company and maybe run some reasonable forecasting models on so and so thing.

And then right in summary for me, the model will kind of infer a reasonable set of actions to do on its own.

So it gives you kind of like a higher level interface to doing some of these more complicated tasks.

That makes sense.

So it sounds like basically there's like a few different changes between your core sort of GPT models where now you have something that takes a pause to think about something.

So at inference time, you know, there's more compute happening.

And then also it can do sequential steps because they can kind of infer what are those steps and then go act on them.

How did you build or train this differently from just a core foundation model or when you did when you all did GPT 2.5 and 4 and all the various models that have come over time.

What is different in terms of how you actually construct one of these?

I guess the short answer is reinforcement learning is the biggest one.

So rather than just having to predict the next token and some large pre-training corpus from everywhere essentially.

Now we have a more focused goal of the model solving very difficult tasks and taking as long as it needs to do to figure out the answers to those problems.

Something that's like kind of magical for me, a user experience for me was we've in the past for our reasoning models talked a lot about test time scaling.

And I think for a lot of problems, you know, without tools, test time scaling might occasionally work.

And but at some point the model is just kind of ranting in its internal chain of thought and especially for like some visual perception ones, it knows that it doesn't it's not able to see the thing that it needs.

And it just it just kind of like loses its mind and goes insane.

And I think tool use is a really important component now to continuing this like test time scaling.

And you can feel this when you're talking to three, at least my impression when I first started using it was the longer it thinks like I really get the impression that like I'm going to get a better result and you can kind of watch it do really intuitive things.

And it's it's a very different experience, but being able to kind of trust that as you're waiting, like it's worth the wait, and you're going to get a better result because of it.

And the model is not just off doing some you know, totally irrelevant thing.

It's cool.

I think in your original post about this to you all had a graph, which basically showed that you looked at how long it thought versus the accuracy of result.

And it was a really nice relationship.

So clearly, you know, thinking more deeply about something really matters.

And it seems like, in the long run, do you think there's just going to be a world where we have some sort of a splitter bifurcation between models, which are sort of fast, cheap, efficient, get certain basic tests done.

And then there's another model which you upload a legal M&A folder, and it takes a day to think.

And it's slow and expensive.

But then it produces, you know, output that would take you a team of people, you know, a month to produce.

Or how do you think about the world in terms of how how all this is evolving or where it's heading?

I think for us, like unification of our models is something that you know, Sam has talked about publicly that, you know, we have this big crazy model switcher in chat GPT, and there are a lot of choices.

And, you know, we have a model that might be good at any particular thing, you know, that a user might want to do, but that's not that helpful if it's not easy for the user to figure out, well, which model should I use for that task.

And so, yeah, making the models better able, you know, making this experience more intuitive is definitely something that is, is like, valuable and something we're interested in doing.

And that, that applies to this, you know, question of like, you know, are we going to have like two models that you know, people pick between or a zillion models that people pick between?

Or do we put that decision, you know, inside the model?

You know, I think everyone is going to try stuff and figure out what works well for like, the problems they're interested in and like the users that they they have.

But, but yeah, I mean, that that question of like, how do you you know, make that that sort of decision be like, as you know, effective, accurate, like intuitive as possible is definitely top of mind.

Is there a reason from a research perspective to combine reasoning with pre-training or try to have more control of this?

Because if you just think about it from the product perspective of like the end consumer dealing with chat GPT, like, you know, we won't get into the naming nonsense here, but they don't care.

Like, the right answer is required to get there in as little time as possible.

Right?

The ideal situation is it's like intuitive, that like, how long should you have to wait, you should have to wait as long as it takes for the model to like, give you a correct answer.

And I, I hope we can get to a place where our models have a more precise understanding of their own level of uncertainty.

Because, you know, if they if they already know the answer, they should just kind of tell you it.

And if if it takes them a day to actually figure it out, then they should they should take a day.

But you should always have a sense of like, it takes exactly as long as as it as it needs to for that current like models intelligence.

And I feel like we're on the right path for that.

Yeah, I wonder if there isn't a bifurcation, though, between like an end user product and the developer product, right?

Because there are lots of companies that use, you know, the API's to all of these different models, then for very specific tasks.

And then on some of them, they might even use like, open source models with really cheap inference with stuff that they control more.

I hope you could just kind of tell the model like, hey, this is a API use case.

And yeah, you really can't be over there thinking for like 10 minutes, we got to get an answer to the user.

It'd be great if their models kind of get to be more steerable, like, like that as well.

Yeah, I think it's just a general steerability question.

Like, at the end of the day, if the model smart, like you should be able to specify, like, the context of your problem, and the model should do the right thing, there's going to be some like limitations, because, you know, maybe just figuring out given your situation, what is like the right thing to do might require thinking in and of itself to figure out.

So like, it's not that you can obviously do this perfectly.

But, but yeah, pushing, you know, some all the right parts of this into the model, to make things easier for the user is like, seems is a very good goal.

Can I go back to something else you said, like, so the first guest we ever had on the podcast is actually known Brown.

Oh, nice.

So I've heard of him, you know, two plus years ago.

Yes, I know.

It'd be great to get some intuition from you guys for why tool use helps like test time scaling work much better.

I can give maybe very concrete cases for like the visual reasoning side of things.

The there's a lot of cases where in back to us also the model being able to estimate its own uncertainty, you'll give it some kind of question about an image and the model will very transparently tell you I should have thought like, I don't know, I can't really see the thing you're talking about very well, or like, it almost knows like that its vision is not very good.

And well, it's kind of magical.

It's like when you give it access to a tool, it's like, okay, well, I got to figure something out.

Let's see if I can like manipulate the image or crop around here or something like this.

And what that means is that it's it's it's like much more productive use of tokens as it's doing that.

And so your test time scaling slope, you know, goes from something like this to something much deeper.

And we've seen exactly that like the the the test time scaling slopes for without tool use and with to use for visual reasoning specifically are very noticeably different.

Yeah, I was gonna say like for like writing code for something like there are a lot of things that an LLM could try to figure out on its own, but would require a lot of attempts and self verification that you could write a very simple program to do in like a verifiable and and you know, much faster way.

So you know, I do some research on this company and like use this type of you know, valuation model to tell me like, you know what the valuation should be like, you could have a model like try to crank through that and like fit those coefficients or whatever in its context, or you can literally just have it like write the code to just do it the right way and just know what the actual answer is.

And so yeah, I think like part of this is you can just allocate compute a lot more efficiently because you can defer stuff that the model doesn't have comparative advantage to doing to a tool that is like really well suited to doing anything.

One of the ways I've been using some form of a three a lot is deep research, right?

I think that's basically a research analyst AI that you all have built that basically will go out will look up things on the web will synthesize information will chart things for you.

It's pretty amazing in terms of its capability set.

Did you have to do anything special in terms of, you know, any form of specific reinforcement learning specifically for it to be better at that or other things that you built against it?

Or how did you think about the data training for the data that was used for training it like I'm just sort of curious like how that product if it all is a branch off of off of this and how you thought about building that specifically as part of this broader effort.

I think when we think about like tool use, I think browsing is one of the most like natural places where you know you think of as a starting point of like, okay, like, it's not always easy.

I mean, like the you know, initial kind of browsing that we included in GPT for a few years back like, it was hard to make it you know, work in a way that felt like reliable and like useful.

But, you know, in the sort of, you know, modern these days, last year, you know, two years ago is ancient history.

I think it feels like a natural place to start because it's like so widely applicable to so many types of queries like anything that is, you know, requires up to date information like it should help to browse for and so in terms of like a test bed for hey, like, does you know, the way we're doing RL like does it really work?

You know, can we really get the model to learn like, longer time horizon kind of meaningful extended behaviors like it feels like kind of a natural place to start in some ways and that it, you know, also is fairly likely to be like useful in a relatively short amount of time.

So it's like, yeah, let's let's try that.

I mean, you know, in RL, like at the end of the day, you're defining an objective.

And if you have an idea for like, who is going to find this most useful, like, you know, you might like want to tailor your the objective, you know, to who you expect to be using the thing, what you expect they're going to want, you know, what is their tolerance for?

Do they want to sit through a 30 minute rollout of deep research?

You know, do they when they ask for a report, you know, do they want a page or five pages or a gazillion pages?

So yeah, I mean, you're you're definitely, you know, you want to tailor things to like, who you think is going to be using it.

I feel like there's a lot of almost like white collar behavior work that you acknowledge work that you all are really capturing through this sort of tooling going forward.

And you mentioned software engineering is one potential area.

Deep research and sort of analytical jobs is another where there's all sorts of really interesting work to be done that's super helpful in terms of augmenting what people are doing.

Are there two or three other areas that you think are the most near term interesting applications for this, whether open AI is doing it or others should do it aside?

I'm just sort of curious how you think about the big application areas for this sort of technology.

I guess my, you know, very biased one that I'm excited about is is coding and also research in general, being able to like improve upon the velocity that we can do research at open AI and others can do research when they're using our tools.

I think our models are getting a lot a lot better very quickly at being actually useful.

And it seems like they're kind of reaching some kind of inflection point where they are useful enough to want to reach out to and and use like multiple times a day for me, for me at least, which wasn't the case.

They're always like a little bit, you know, behind what I wanted them to be, especially when it comes to like navigating and using our internal code base, which is not simple.

And it's amazing to say like more recent, our models actually really spending a lot of time trying to understand the questions that we asked them and coming back with things that saved me like many hours of my own time.

People say that's the fastest potential bootstrap, right, in terms of each model subsequently having helping to make the next model better, faster, cheaper, etc.

And so people often argue that that's almost like a inflection point on the exponent towards super intelligence is basically this ability to use AI to build the next version of AI.

Yeah, and there's so many like different components of research to there's it's not just, you know, sitting off in the ivory tower thinking about things, but there's, there's like hardware, there's, you know, various components of training and evaluation and stuff like this.

And each of these can be turned as some kind of task that can be optimized and iterated over.

So there's plenty of, you know, room to squeeze out improvements.

We talked about browsing the web, writing code, arguably the greatest tool of all, right, especially if you're trying to figure out how to spend your compute, right, more efficient code, generating images, writing text, there are certainly like trajectories of action I think are not in there yet.

Right, like reliably using a sequence of business software.

I'm really excited about the computer use stuff.

It kind of drives me crazy in some sense that our models are not already just like on my computer all day, watching what I'm doing.

And well, I know that can be creepy for some people.

And like, I think you should be able to like opt out of that or have that opted out by default hate typing.

Also, I wish that I could just kind of like be working on something on my computer, I hit some issue and I'm just like, you know, what am I supposed to do with this and I can just kind of ask I think there's tons of space for being able to improve on like how we interact with the models and this goes back to them being able to use tools in a more intuitive way, I guess using tools closer to how we use them.

It's also surprising to me how intuitively our models do use the tools we give them access to.

It's like weirdly human like but I guess that's not too surprising given the data they've seen before.

But yeah, I think a lot of things are weirdly human like like my intuition for like, well, why is tool use so impactful to test time scale?

Like why is the combination so much better take any, and any role, you can make a decision when you are trying to make progress against a task as to like, do I get external validation or do I sit and think really hard, right?

And usually you want to do like one or is more efficient than the other and it's not always just sit in a vacuum and think really hard with what you know.

Yeah, absolutely.

You can seek out sort of new inputs, like it doesn't have to be this closed system anymore.

And I do feel like the the closed system ness of the models is still sort of a limitation in some ways like you're not you're not necessarily like turning this I mean, like I think it'd be great if the model could control my computer for sure.

But in some sense, it's there's a reason we don't go hog wild and say like, Oh, yes, here's like the keys to the kingdom, like have at it.

There are still, you know, asymmetric costs to like, the time you can save and the types of errors you can make.

And so we're trying to like iteratively kind of, you know, deploy these things and like try them out and figure out like, where are they reliable, you know, and where are they not?

Because yeah, like, if you did just let the model control your computer, it could do some cool stuff.

Like I have no doubt.

But you know, do I trust it to like, respond to all of the, you know, random emails that Brandon sends me actually maybe for that task, it doesn't require that much intelligence.

But you know, like, do I you know, do I trust it to do everything I'm doing?

Like, you know, some things and I'm sure like that set of things will be bigger tomorrow than it was yesterday.

But yeah, I think part of this is like we limit the affordances and keep it a little bit in the like sandbox, just out of caution.

So that you know, you don't send some crazy email to your boss or, you know, delete all your texts or delete your hard drive or something.

Is there some sort of like, organizing mental model for like the tasks that one can do with, you know, increasing intelligence, test time scaling and improve tool use, right?

Because I look at this and I'm like, okay, well, you have complexity of tasks and you have time scale, then you have like the ability to come up with these RL rewards and environments, right?

Then you have like usefulness.

Maybe you have some, you know, of course, you have some intuition about like diversity and generalization across the different things you can be doing.

But it seems like a very large space and scaling our like new gen RL is not, it's just not obvious.

Like how to me, it's not obvious how you do it or how you choose the path.

Is there some sort of organizing framework that, you know, you guys have that you can share?

I mean, I don't know if there's like one organizing framework.

I think there are a few like factors at least that I think about in like the very, very grand scheme of things is like, how much like in order to solve this task, like how much uncertainty with the environment do I have to like wrestle with?

Like for some things where it's like, this is a purely fat, like who was the first president of the United States?

Like there's zero like environment I need to interact with to like reach the answer to this question correctly.

I just need to remember the answer and say the answer.

You know, if I want you to like write some code, you know, that like solve some problem, well, now I have to deal with a little bit of like not purely internal model stuff, but also like, okay, I need to execute the code and like that code execution environment is maybe more complicated than my model can memorize internally.

So I have to do like a little bit of like writing code and then executing it and making sure it does what I thought it did and then testing it and then giving it to the user and things get like the amount of that sort of stuff outside the model that you have to like, you know, you can't just recall the answer and give it to the user.

You have to like test something and you know, run an experiment in the world and then wait for the results of that experiment.

Like the more you have to do that, the more uncertain the results of those experiments, like in some sense, that's like one of the core like attributes of like what makes the tasks hard.

And I think another is like how you know, simulators they are like stuff that is really bottlenecked by like time, like the physical world is also, you know, just just harder than stuff that we can simulate really well.

You know, it's not a coincidence that you know, so many people are interested in coding and you know, coding agents and things.

And that like, you know, robotics is hard and you know, it's it's slower and you know, I used to work on robotics and like, it's frustrating in a lot of ways.

I think both this like how much of the external environment do you have to deal with?

And then like, how much do you have to wrestle with the unavoidable slowness of the real world are two like dimensions that I sort of think about.

It's super interesting because if you look at historically, some of these models, one of the things that I think has continued to be really impressive is the degree to which they're generalizable.

And so I think when GitHub, go pilot launches on codecs, which was like a specialized code model.

And then eventually that just got subsumed into these more general purpose models in terms of what a lot of people are actually using for coding related applications.

How do you think about that in the context of things like robotics?

So do you know, there's like probably a dozen different robotics foundation model companies now.

Do you think that eventually just merges into the work you're doing in terms of there's just these big general purpose models that can do all sorts of things?

Or do you think there's a lot of room for these standalone other types of models over time?

I will say the one thing that's always struck me as kind of funny about us doing RL is that we don't yet do it on the most canonical RL task of robotics.

And I personally don't see any reason why we couldn't have this be the same model.

I think there are certain challenges with like, I don't know, do you want your RL model to be able to generate an hour long movie for you, natively, as opposed to a tool call?

That's where it's probably tricky to have you have more conflict between having like everything in the same set of weights.

But certainly, like the things you see three already doing in terms of like, you know, exploring a picture and things like that are are kind of like early signs of something like an Asian exploring like an external environment.

So I don't think it sounds too far fetched to me.

Yeah, I mean, I think the thing I came up earlier of the also the like intelligence per cost thing, you know, the real world is like an interesting litmus test, because at the end of the day, like, there is a, you know, frame rate in the real world you need to live on.

And it doesn't matter if you get the right answer after you think for two minutes, like, you know, the ball is coming at you now, and you have to catch it.

Gravity is not going to wait for you.

So you you, that's an extra constraint that we get to, at least softly ignore when we're talking about these purely disembodied things.

That's kind of it's kind of interesting, though, because really small brains are very good at that.

You know, so you look at a frog, you know, you start looking at different organisms and you look at sort of relative compute.

Yeah.

And, you know, very simple systems are very good at that ants, you know, like, so I think that's kind of a fascinating question in terms of what's the baseline amount of capability that's actually needed for some of these real world tasks that are reasonably responsive in nature.

It's really tricky with with vision to that we have.

So our models have some, I think, maybe famous edge cases of where they don't do the right thing.

I think Eric probably knows where I'm going with this.

I don't know if you ever asked like our models tell you what time it is on a clock.

They really like the time 1010.

So yeah, it's my favorite time to so that's that's usually what I call people.

It's like over 90% or something like that of all clocks on the internet or 1010.

And it's because it looks like I guess like a happy face and looks like nice and but but anyways, like the what I'm getting at is like our our visual system was developed by interacting with you know, the external world and having to be good at like navigating things, you know, avoiding predators.

And our models have learned vision a very different type of way.

And I think it'll we'll see like a lot of really interesting things if we can get them to be kind of closing the loop by you know, reducing their uncertainty by taking actions in the real world, just as opposed to like thinking about stuff.

Yeah.

Hey, Eric, you brought up the idea of like how what in the environment can be simulated, right as a as an input into like how difficult will it be to improve on this?

As you get to long running tasks, like let's just take software engineering, like, there is a lot of interaction that is not just me committing code continually, it's like, I'm going to talk to other people about the project, in which case you then need to deal with the problem of like, can you reasonably simulate how other people are going to interact with you on the project in an environment?

That seems really tricky, right?

I'm not saying that, you know, oh, three or whatever set of foundation models now doesn't have the intelligence to respond reasonably.

But like, how do you think about that simulation being true to life as a true to life true to the real world, as you involve human beings in an environment in theory?

My spicy I guess take on that as like, only the spicy but oh, three in some sense is already kind of simulating what it'd be like for a single person to do something with like a browser or something like that.

And I don't know, train two of them together.

So that you know, if you know, you have two people interacting with each other.

And there's no reason you can't scale this up so that models are trained to be really good at cooperating with each other.

I mean, there's a lot of already existing literature on multi agent RL.

And yeah, if you want the model to be good at something like collaborating with a bunch of people, like maybe a not too bad starting point is making it good with collaborating with other models.

And someone should do that.

Yeah, yeah.

Yeah, we should really start thinking about that.

Eric, I think it is a I think it's a little bit spicy.

Because yes, a work is going on.

It is interesting to hear you think that is a useful direction.

I think lots of people would still like to believe not me like my comment was extra good on this pull request or whatever it is.

Right.

And okay, I could say I could sympathize with that.

Sometimes I see our models training and I'm like, Oh, what are you doing?

You know, like, you're taking forever to figure this out.

And I actually think it'd be really fun if you could actually train models in an interactive way.

You know, forget about just like a test time, but I think it'd be really neat to train them to do something like that.

Be able to like intervene when it makes sense.

And yeah, just more, more me being able to tell the model to cut it out.

And like in the middle of its kind of chain of thought and it being able to learn from that on the fly, I think would be great.

Yeah, I do think this is like the intersection of these two things where it's both like a point of contact with the external environment that is like can be very high uncertainty, like humans can be very unpredictable, in some cases.

And it's sort of limited by the take of time in the real world, if you want to like, you know, deal with actual humans, like humans have a fixed, you know, clock cycle, you know, in their in their head.

So yeah, I mean, this is if you you know, if you want to like do this in the literal sense, it's hard.

And so, you know, scaling it up and you know, making it work well is, you know, it's not obvious how to do this.

Yeah, we are a super expensive tool call.

You know, if you're a model, you can either ask me, you know, meet back over here to, you know, help with something and I'll try to think really slowly.

In the meantime, it could have like use browser and read like 100 papers on the topic and something like that.

So it's how do you model the trade off there?

But the human part is important.

I mean, I think in any research project, like my interaction with Brandon, the hardest part of the project, you know, like writing the code is that's the easy part.

Well, and there's some analog from self driving.

And lots gonna say the, you know, hang out with me every week is the hardest part of doing this podcast.

But it's my favorite part.

Look at how healthy their relationship is, Eric, we need to learn from this.

No, we're honest.

It's okay.

We got to work through it.

In self driving, one of the like classically hard things to do was like predict the human and the child and the dog like agents in the environment versus like what the environment was.

And so I think there's like some analogy to be drawn there.

Going back to just like how you progress the O series of models from here.

Is it a reasonable like assessment that some people have that the capabilities of the models are likely to advance in a spikier way because you're relying to some degree more on the creativity of research teams and like making these environments and deciding, you know how to create these evals versus like, we're scaling up on existing data set and pre training is that a fair contrast?

Spiky or like, what's the plot here?

What's the like the x axis and the y domain is the x axis and y is capability?

Yes, because you're like choosing what domains you are really creating this RL loop in.

I mean, I think this is a very reasonable hypothesis to hold.

I think there is some like counter evidence that I think should, you know, be factored into people's intuitions.

Like, you know, Sam tweeted an example of some creative writing from one of our models that I think was I'm not an expert.

And I'm not going to say this is like, you know, publishable or like groundbreaking, but I think it probably updated some people's intuitions on like what, you know, you can train a model to do really well.

And so I think there is some structural reasons why you'll have some spikiness just because like, as an organization, you have to decide like, hey, we're gonna prioritize, you know, XYZ stuff.

And like, as the models get better, the surface area of stuff you could do with them grows faster than, you know, you can potentially like, say, hey, this is the niche, you know, we're gonna carve out, we're gonna try to do this really well.

So like, there, I think there's some reason for spikiness.

But I think some people will probably go too far with this and saying like, oh, yes, these models will only be really good at math and code and like, not you know, like everything else is like you can't get better at them.

And I, I think that is probably not the right intuition to have.

Yeah, and I think probably all like, major AI labs right now have some partitioning between let's just define a bunch of data distributions, we want our models to be good at and then just like throw data at them.

And then another set of people in the same companies is probably are probably thinking about how can you kind of lift all boats at once with some like algorithmic change.

And I think, yeah, we definitely have both have those types of efforts at open AI and I think especially on the data side, like, there are going to naturally be things that we have a lot more data of then than others and but ideally, yeah, we have plenty of efforts that will not be so reliant on the exact like subset of data we did RL on and it will generalize better.

I get pitched every week.

And I bet a lot does to a company that wants to generate data for the labs in some way.

And or it's, you know, access to human experts or whatever it is, but like, you know, there's there's infinite variations of this.

If you could wave a magic wand and have like a perfect set of data, like what would it be that you know would advance model quality today?

This is a dodge but like uncontaminated evals always super valuable.

And that's data.

And I mean, yeah, like you want, you know, good data to train on and that's of course valuable for making the model better.

But I think it is often neglected how also important it is to have high quality data, which is like a different definition of high quality when it comes to an eval.

But yeah, the eval side is like often just as important because you don't you need to measure stuff.

And like as you know, from, you know, trying to hire people or whatever, like evaluating the capabilities of like a general like capable agent is really hard to do in like a rigorous, you know, way.

So yeah, I think evals are a little under appreciated.

That's true.

Evals are, I mean, especially with some of our recent models where we've kind of run out of reliable evals track because they kind of just solved a few of those.

But on the on the training side, I think it's always valuable to have training data that is kind of at the next frontier of model capabilities.

I mean, I think a lot of the things that oh three and oh four many could already can already do those types of tasks like basic tool use, we probably aren't, you know, super in the need for for new data like like that.

But I think it'd be hard to say no to a data set that's like a bunch of like multi turn user interactions and some code base that's like a million lines of code that you know, is like a two week research task of like adding some new feature to it that requires like multiple pull requests.

I mean, I mean, like something that was like super high quality and has a ton of supervision signals for us to learn from it.

Yeah, that I think that would be awesome to have, you know, definitely wouldn't turn that down.

You play with the models all the time, I assume a lot more than average humans do.

What do you do with reasoning models that you think other people don't do enough of yet?

Send the same prompt many, many, many times to the model and and get an intuition for the distribution of responses you can get.

I have seen it drives me absolutely mad when people do these comparisons on Twitter or wherever and they're like, Oh, I put the same prompt into blah, blah and blah, blah.

And this one was so much better.

It's like, dude, you like, like I mean, was something we talked about a bit when we were launching is like, yeah, three can do really cool things like when it chains together a lot of tool calls.

And then like sometimes for the same prompt, it won't have that, you know, moment of magic or it will, you know, just take a little, it'll do a little less work for you.

And so, yeah, though, like the peak performance is really impressive, but there is a distribution of behavior.

And I think people often don't appreciate that there is this distribution of outcomes when you put the same prompt in and getting intuition about that is useful.

So as an end user, I do this and I also have a feature request for your friends in the product org.

I'll ask, you know, Oliver or something, but it's just, I want a button where like assuming my rate limits or whatever support it, I want to run the prompt automatically like 100 times every time even if it's really expensive.

And then I want the model to rank them and just give me the top one or two.

Interesting.

And just let it be expensive.

Or synthesis across it, right?

You could also synthesize the output and just see if there's some other maybe you're then reverting to the mean in some sense relative to that distribution or something, but it seems kind of interesting.

Yeah, maybe there's a good infrastructure reason you guys aren't giving us that button.

Well, it's expensive, but there are, I think it's a great suggestion.

Yeah.

Yeah, I think it's a great suggestion.

How much would you pay for that?

A lot, but I'm a price insensitive user of AI.

Yeah.

Perfect.

Those are our favorite.

Should have Sarah tier as one of your chairs.

Exactly.

Exactly.

I really like sending prompts to our models that are kind of at the edge of what I expect them to be able to do just kind of for funsies.

Like a lot of times before I'm about to do some like programming tasks, I will just kind of ask the model to go see if it can figure it out.

All the times like no hope of it being able to do it.

And indeed, sometimes it comes back and I just am pretty like I'm like a disappointed father, but other times it does it and it's amazing and it saves me like tons of time.

So I kind of use our models almost like a background queue of work where I just will just like shoot off tasks to them.

And sometimes those will stick and sometimes they won't, but in either case, like it's always a good outcome if something good happens.

Let's call.

Yeah, I do that just to feel better about myself when it doesn't work.

I'm still providing value.

Whatever works, I feel even worse about myself.

So that's very hit or miss.

Yeah.

There are some differences in terms of how some of these models are trained or rl'd or you know, effectively produced.

What are some of the differences in terms of process in terms of how you approach that series of models versus other things that have been done at opening in the past?

The tools stuff was very, it was quite the experience getting working at a large scale setting.

You can imagine if you're doing like async rl with a bunch of tools that those are you just adding more and more failure points to your infrastructure.

And what you do when things get evidently fail is pretty interesting like engineering problem, but also like an rl, like ml problem too, because, you know, if you're, I don't know if your Python tool, like, you know, it goes down in the middle of the run, like, what do you what do you do?

Do you stop the run?

Probably not.

That's probably not like the most sane thing to do with like that much compute.

So the question is, like, how do you handle that gracefully and not, you know, hurt the capabilities of the model, like as a as an unintended consequence.

So there's been a lot of learnings like that of how you deal with like huge infrastructure that's asynchronous for RL.

RL is hard.

This has been great, guys.

Thank you.

Yeah, thanks so much for coming.

Yeah, thanks.

It's fun.

Thanks for having us.

Find us on Twitter at no priors pod.

Subscribe to our YouTube channel.

If you want to see our faces, follow the show on Apple podcasts, Spotify or wherever you listen.

That way you get a new episode every week and sign up for emails or find transcripts for every episode at no dash priors dot com.

[BLANK_AUDIO]