Why Traditional Benchmarks Fail Modern AI Models with OpenAI Research Scientist Noam Brown

With GBT three, you couldn't scale test time compute. Like, if you gave it a budget of $10,000,000 and said, okay. Well, let's see what GBT three can do.

It really can't do that much. The precarious frameworks and responsible scaling policies, they don't really account for the amount of test time compute. They just say, okay.

Well, what's the capability of the model? The problem is we're in a world now where the capability of the model is a function of how much money you put into it, basically. If you give it a budget of $10,000, it can do a lot more than what it can do with a budget of $10.

Give it a budget of $10,000,000, it can do even At what budget should you evaluate these models? The policies that exist today don't really address that question.

and the initial reaction was kinda skepticism that it was a substantially better model. This to be fair, that only lasted for a few hours before people had some time to play around with it and and try it out themselves, and they saw that it was actually substantially better. But I think a lot of the skepticism came from the benchmark grid that was published.

Basically, whenever a new model is released, there's this benchmark grid where they show all these different benchmarks on on the x axis and then the performance of different models on the y axis, and you can just, like, compare different models. It's like a single number for a model on a single benchmark. And if you look on paper at the difference between, like, five point five and five point four or or other models, it wasn't it was an improvement, but it wasn't a huge improvement.

It was only a few percentage points on some benchmarks. So people looked at that, and they were skeptical that it was actually a better model. Once they played around with it, the story changed.

I think the reason why it doesn't show up as so much better on the benchmarks is because the benchmarks are being presented the benchmark results are being presented in the wrong way. They're not controlling for the amount of test on compute that is being used on that benchmark question. It turned out that 5.

5 is just much more efficient with its thinking. If you run it at max settings, 5.4 is thinking for a lot longer.

It takes longer to get back a response than 5.5. And once you control for the amount of thinking time, actually, you can see that 5.

5 is a substantial jump over 5.4. That is, I think, people's day to day experience with it.

And then when I mention this to people, the reaction the typical question I get is like, okay. Well, why not just have 5.5 think for as long as 5.

4? And the question is like, well, how long should they think for? Typically, the response I get is, well, until the performance plateaus.

Right? This is at some point where the performance on the benchmark is gonna plateau, and you just evaluate to that point. The thing is, the point at which it plateaus is actually really far out these days.

I mean, if it's true in g p d three land back in 2022, models couldn't really think productively for that long, and so you could just run them until they plateau. It's not that far away. But what we're seeing today with the modern models is that 5.

5 and other models can think for if you scaffold them reasonably well, can think for weeks even, before having performance plateau on some of these benchmarks. And so the point at which they plateau is simply too far out to reasonably test. We all need to actually reinforce either, like, a patience limit or a budget limit from a token perspective now, and that wasn't true a few years ago.

Exactly. And so I think the proper way to and so my claim is the proper way to evaluate the models now is you either have some kind of budget for the benchmark, whether it's tokens or cost or time or whatever, or you plot the performance as a function of the amount of test time compute that's going into the model. And then it becomes much more clear how to compare the performance between these different models.

we've seen, and actually the AISI in their evaluations has shown that the models continue to improve at a 100,000,000 tokens. You know, if you run them for a 100,000,000 tokens, they're still improving at beyond that point. Mhmm.

And that can take a very long time to run. But you also do see that, like, the performance is is it's not just, like, a discontinuous jump. It's actually, like, you can see the slope of improvement over those 100,000,000 tokens.

And so you you could probably do some kind of evaluation up to a certain budget and then just say, okay. Well, this is what we project the performance to look like. And I think this this there hasn't been a lot of research on this yet.

I actually think this would be a great paper to publish if there's any academics out there looking for something to research.

I I think it depends on the problem. I think this idea that the models, you just let them think for a week or whatever, then they respond. It's it sounds nice, and, yes, the benchmarks look great, but it's not very practical when working because, like, okay, you ask the model a question, then you sit there for a week waiting for it to come back to you.

Mhmm. I think what people have have found most effective is to kind of, like, iterate quickly with the models. And so the thinking time, I think, needs to be flexible.

When it makes sense to respond quickly to the user, it should respond quickly. And then when it makes sense to think for a long time and the user wants it to think for a long time, then it makes sense to think for a long time. I think people have been striking the right balance given what they have to deal with right now.

benchmark maxing and the ability to gain different benchmarks. What would you characterize the, like, landscape of benchmarks as today? And then do you have, like, favorites that you think are more indicative of capability than others?

So the benchmark maxing thing is also motivation for for writing an essay that I think it's really easy to show you can do much better than previous benchmarks or or previous previous models on benchmarks by just, for example, scaffolding a bunch of models together. So if you say, okay. Well, we're going to instead of just running this model once, we're gonna run it five times and take the best of the five responses or, like, ask a judge which one it it thinks is best, then you can get much higher scores than that model.

And so it's really easy to make something that looks a lot better on paper but is actually not better once you control for the amount of test time compute. That is one thing that I'm worried about when it comes to benchmark maxing. I mean, it, like it's it's a little misleading is the only concern that I have.

And then as far as, like, the benchmarks themselves, I think there is always a risk of, like, just optimizing for the benchmark. And I've I've certainly encouraged my team, and I think at OpenAI, we're pretty good about not trying to optimize for for specific benchmarks. But once you put out a benchmark, it's there's it's always at risk of just being optimized for.

And I think one way to one way to address that is to keep held out private sets that isn't publicly available.

private holdback at OpenAI? I think everybody has their own set of questions that they like to ask the model whenever it comes out. Mhmm.

For me lately, it's been I I use them to make poker bots and see how good they can make a poker bot. I think it's a nice eval because there is very little open source code for making poker bots. And there's a lot of published essay there's a lot of published papers on it, but you really have to reason through everything.

And it's, like it requires a lot of just reasoning and iteration and, like, a lot of small gotchas that I can kind of I've already worked through myself, so I can see where the models fail along the way. They've gotten really good at it now. Can you describe perhaps, like, with your poker bot creation, like, how reasoning might have progressed in model releases for you guys over a few releases?

Yeah. When the early models were really bad at it. Like, they could not basically do anything.

And then 5.2, I was able to work with it to make a reverse solver. So that's, like, the final stage Mhmm.

Of poker. And that itself was, I thought, really impressive. I I had to work with it a little bit, but I was actually really impressed because I was able to make the reverse solver probably about five times faster than I would have alone.

There were a couple things that I got tripped up on. Blockers was always a big, big issue. But overall, like, you know, with a with a bit of gentle steering, it just kind of like I I kinda felt like a grad student where, okay, they would run into issues, but at least, like, I would know what those issues were and know how to fix it.

And I could just make suggestions, and it would go off and then do it. And then pretty quickly, would actually come back with something really good. Mhmm.

And then especially the optimization, thought was very impressive. It was able to make it, like, 10 times faster than what I was able to do because it was just able to optimize the code so well. The downsides with 5.

2 is I felt like it was gaslighting me a lot, and I always had to be very careful checking it and making sure, like, okay. Is it actually doing what it said it did? Are there any things that are, like, glaring issues that it's not recognizing or it's just pretending aren't issues?

I remember there was, like, one point where for one of the models I was playing around with it, not 5.2, I kind of, like, as a as a unit test, I told it, okay. Well, let's say I have a $100 in the pot, and I fold.

How much am I losing? And the model said $92. And I was like, that's crazy.

I I have a $100 in the pot, I just fold it. How do I not lose a $100? And it said, oh, you know, it's 92.

It's close to a 100. It's fine. It's no big deal.

And I was like, clearly, this is a problem. Right? So the models did have this problem where they would gaslight you a lot.

Mhmm. But once we got to 5.5, I actually thought it was way better.

It was able to basically do a zero shot. And in fact, I've been working on just doing a full scale poker solver, and it it's basically able to do the whole thing with some gentle steering from me. And I wouldn't be surprised if, six months or a year from now, the model is able to do zero shot an entire poker solver, basically my entire PhD thesis, in one go.

Yeah. The safety evaluations thing, it's it's a bit of an inconvenient truth thing where okay. So I guess for background, a lot of the all of the labs have these things called either responsible scaling policies, preparedness frameworks.

They go by various names. But the idea is that whenever a model's released, they go through a series of evaluations to measure, are there dangerous capabilities? Could these models do things that we're we we wouldn't want, a bad actor to do?

And if the model isn't very capable, then it's no big deal. But if it is very capable, if it could be used, for example, to make bioweapons, then you want to put in mitigations against that. Mhmm.

But the question is, okay. Well, how do you evaluate whether the model is capable of that? And they have, like, various protocols about, like, how they do these evaluations.

But a lot of these frameworks were developed around the era of ChatGPT, either before or after, when test time compute scaling was not really as much of a thing. Mhmm. And it made sense.

Like, with GPT three, you couldn't scale test time compute. Like, if you gave it a budget of $10,000,000 and said, okay. Well, let's see what g p t three can do, it really can't do that much more than what you could do with, like, $10 or $1.

The preparedness frameworks and responsible scaling policies, they don't really account for the amount of test time compute. They just say, okay. Well, what's the capability of the model?

The problem is we're in a world now where the capability of the model is a function of how much money you put into it, basically. Mhmm. If you give it a budget of $10,000, it can do a lot more than what it can do with a budget of $10.

If you give it a budget of $10,000,000, it could do even more. And so at what budget should you evaluate these models? The policies that exist today don't really address that question.

Mhmm. Some do some do better than others, but for the most part, this is not really a factor that's being heavily considered. Now whether it should be released anyway, I don't I don't wanna wade into this question.

I think there's, you know, there's arguments on both sides.

basically, as the models have become stronger, they've they're more they're better able to operate over longer horizons. Mhmm. So, again, with g p d three, if you wanted to run it for, you know, a week, there's really not much you could do to scaffold it into something useful that could actually run for a week.

But we're seeing now with the most recent models that you can actually scaffold, for example, 5.

infinite budget yet? I haven't really scaffolded something together where I just tell it, like, okay, just run this for for weeks. I think I could it I I Until it has some types.

Could probably give it slash goal and just, like, yeah, tell it to go nuts. But I I think at this point, it could 100% do the reverse solver if I just give it slash goal. I don't think it's at the level yet where it could do, like, the full poker solver if I gave it just slash goal and told it, yeah, go go run for a month.

But we're going to pretty soon be at that point where I I probably could just tell it, like, yeah, go work on this for a month, and then come back to me with a a full, complete PokerSolver that's state of the art. And the problem is if you want to evaluate the capabilities of a model, what it can do after running for a month, the only way to be fully sure is to actually run it for a month. And if you wanna know after six months, the only way to know fully is to run it for six months.

Now there I'll I'll get to, like, things we could do to address that a little bit later, but, like, it's important to recognize that the model release cycle is look. We're releasing new models, like, every two or three months at this point. And so a model comes out, it takes two or three months to push it to its limits, and then you have another model come out.

And so nobody actually knows what the ceiling of capabilities are for these models because nobody's actually run them for long enough to really tell. When slash goal came out, for example, I mean, people started running things that it took over a week for it to finish, and so people actually didn't realize that this was a big deal until after a week, until a week after it was released. Mhmm.

I think that's gonna be more and more true. You know, the implications of that are, I think, pretty interesting because what do the labs do to, like, fully evaluate their models before they're released? It's actually very difficult because, yeah, you would have to the only way to to really do the evaluations is then delay the model release cycle.

And, you know, there's a lot of competitive pressure right now to not do that.

that people have not fully explored given timeline? I think absolutely. I think actually a really great example is the Erdos unit distance problem.

So for the viewers that don't know, like, we used an internal model at OpenAI a few weeks ago to disprove the unit Erdos unit distance conjecture. Now I'm not a mathematician, but this seems like it was a a pretty big deal. The in the math community, it was, like, the first first problem that a lot of mathematicians had really spent a lot of time on, and the model was able to do something that they weren't able to do and do it in a way that was actually interesting and useful for mathematicians.

Honestly, it did it at a budget that was dirt cheap. I mean, we didn't put a lot of effort into this. We just we trained a new model, and we were just curious what it could do, and we ran it through some problems.

And this one, at a pretty low budget, it was like, oh, yeah. I think I have a disproof, and then we were able to verify that, yeah, that disproof is correct. After we announced the results, a bunch of people found that you could get the answer out of 5.

5 as well. If now it's not as simple as just asking 5.5, hey.

Here's the neural unit distance conjecture. What's the disprove? You had to scaffold it a bit.

You had to, like, steer it a bit. And so somebody found, okay. You asked 5.

5. List a bunch of ways that you could tackle this problem. And then for it it lists one of the paths that are actually promising to get to the disprove.

And then it you tell it, like, okay. Explore this some more. And then if you do this enough times, it actually ends up arriving at the disprove.

Now what this means is you could, in principle, ask 5.5 to you know, as as a general purpose scaffold, list a bunch of different strategies, and then for each strategy, tell it to investigate that strategy. And then it would probably be able to arrive at the disprove with a general purpose scaffold.

Now that scaffold would be very expensive. I mean, it would probably cost I I just ballpark, like, a thousand to 10 to a $100,000. But it would be possible, and it would have been possible for somebody to disprove the Erdosia to disc distance conjecture before we did using a general purpose model.

And nobody had explored sufficiently, well, happens if I put a $100,000 worth of compute into 5.5? What could it do?

And the answer is, yeah, you probably could get stuff like that out of it. So people should be experimenting more with the current generation in terms of Well, this is, I think, an interesting question of, is it worth it to experiment with because, again, the model release cycle is every every couple months, we put out a new model that's even more powerful. And so the cost of disproving the air to ocean at distance congestion drops by, like, 10 or a 100 x with every model release cycle, probably, in some cases, more.

So You've seen the meme that's like, oh, alright. Like, why why bother doing any engineering work when I should just wait for the next model release? On vacation and come back two months later, and then it's, you know, a thousand times cheaper.

So agree with that? I Is that what you're doing right now at OpenAI, just waiting for the next model release? I think I I mean, I will say we're in we're in a period where progress is very fast, and, like, yeah, the models are becoming more capable.

I I can say, like, at OpenAI, one of the things that we're we're actively not doing look. We have a lot of mathematicians. We have a lot of physicists.

People are very excited about what these models can do right now Mhmm. Especially, you know, the internal models. We are trying to encourage people to not spend all their time just, like, going through all the mathematical open problems, physics problems, and just seeing pushing the models to their limits to see what they can prove or disprove Because we really think the focus should be on how do we make even more capable models, how can we get them get them out safely to the world as quickly as possible so that all the scientists in the world can use these models to solve the problems themselves.

So, yeah, in some sense, we are thinking about this that, yes, it's really tempting to just put all of our efforts into scaling up these models and see what they can do with their limits right now, but really the focus should be on how do we use these models to make even more powerful models, even more capable models that can do everything much more cost effectively.

idea of recursive self improvement, for example, where, you know, it's a dominant idea for how, you know, any lab gets to the best capability model. So one thing I should clarify, I don't think we're at the point where, okay, you just give it an arbitrary an an extremely high inference budget, and it's just it's just super intelligent across the board. Slash goal Yeah.

Baseline. Okay. G p d seven or whatever and then, like, yeah, just go nuts.

What's between us and there then? I think having played around with the model so okay. So first of all, there are some benchmarks where the models will just not improve if they have more inference budget.

So I think a lot of factual retrieval kind of questions fall into this category of if you ask a person when was Abraham Lincoln born and they don't know the date, they could sit there. They could think about it for a week. If they if they don't have access to a computer or something, they're not gonna be able to do better answering that question if they thought about it for a week compared to five seconds.

Same with the model. If you actually, interestingly enough, if you give the model these kinds of, like, factual retrieval questions and you give them a little bit of time to think, they do actually do But if you give them a week, they're not suddenly gonna do better at remembering dates. There are so there's some benchmarks where they clearly improve with more test time compute, and there's some where they don't.

I think on the other extreme, there are benchmarks where they kind of obviously will keep improving limit without limit with more test time compute. So the example I like to point to is sudoku. If you it's there's a really simple strategy to solving sudoku, which is just try a bunch of different random numbers and then see if it fits the criteria, if it if it matches all the constraints.

And if it doesn't, just try a different random combination of numbers. And, clearly, with enough time, you will be able to solve any Sudoku puzzle with this strategy. You can kind of trivially see, like, okay, any model could keep doing better and better if it was just given more test compute.

So you have and and all the benchmarks kind of exist somewhere between these two extremes. The models are not at the level where if you just give them enough test time compute, they will be able to do all of our jobs just because, yeah, there's some benchmarks where they will not improve. There are some things where they they will not improve.

One thing I see for research in particular is they don't have very good research taste right now, and so I think they're actually a very good complement to researchers, especially, you know, I've found, like, I've found that I've been much more effective by using these models, but they're not able to fully replace the whole research cycle. Now does that change with time? Probably.

I mean, I think the models are getting better across the board.

poker solver example, I was really impressed with the model's ability to optimize the algorithms that I had developed in my in my PhD. It was honestly it it was it was shocking to see how inefficient I was, in retrospect, and they were able to make it, like, you know, 1,000 x faster. And then I was like, okay.

Can you come up with an algorithm that is better than the algorithms that I came up with or that anybody else came up with? You and go ahead and, like, look at all the published work and synthesize that and then try to come up with something novel, and it it's not able to do it. And I can give it a lot of time, it's it's still not able to do it.

Now it's possible that if I scaffolded something and, like, kind of constrained it a bit more, that maybe it could eventually come up with something better, but it would take a lot of it's not just as simple as saying, okay. Please come up with a better algorithm.

What I've seen is with every model release cycle, it does get better at this sort of thing. It's still it's still bad in my opinion, but it it's not as bad as it used to be. And I wouldn't be surprised if at some point same thing with coding, same thing with math, where there's just, like, this inflection point where suddenly it's actually good enough to be useful.

The models are definitely accelerating what researchers can do inside the labs. But I think they are accelerating some things and not other things. And currently, we're at the point where, okay, if something goes a 100 x faster, you get bottlenecked by the things that don't go a 100 x faster.

Over time, the things that we're getting bottlenecked on are going to shrink, and and there will be, I think, a kind of a gradual takeoff in that respect. But it's more about transforming right now, it's more about transforming what researchers do rather than fully replacing the researchers.

think we're close to a very fast takeoff right now. I think fast takeoff is relative. Things are moving very fast, but I think there is this hypothesis that you could have basically an overnight intelligence explosion where the models discover some kind of breakthrough to make themselves smarter, and then that leads to more breakthroughs that make themselves even smarter immediately.

And you have, basically, in an instance, the model's just, you know, becoming very superhuman across the board in in moments. And I don't think we're headed to that world largely because of the fact that the models rely so much on large scale test on compute in order to achieve their greatest intelligence. If you if it requires so much test on compute to unlock the full capabilities of the model, then that means you're bottlenecked by time.

Things can only go so fast because the models need to run for long enough to actually do something really, really powerful. Time itself becomes a bottleneck to what we can do. And I I think that is the case right now for a lot of the labs that, ultimately, I think the biggest bottleneck for all of us is time.

And that's why all the researchers are working so intensely right now. It's it's just so many so many hours per week are being put into this because we all see what the overhang is. We see what the capabilities are, and we're just bottlenecked by how quickly can we do things.

What do you think is on the frontier that is less explored now?

I think multi agent is quite explored. I think there's sufficient scale? I think there's a lot more that could be done, but it's also one of the things that's that's hard to do at small a lot of research is hard to do at small scale.

I think multi agent in particular, it really requires in order to fully unlock the capabilities, think requires, like, frontier models. I think we've seen some pretty interesting multi agent scaffolds. I think, they're able to do a lot, but I think it's really just scratching the surface of what it will be able to do.

I mean, one way that I think about it is if if you look at human civilization, it's not that humans have become smarter over it's not that they evolved to become smarter over, you know, the past fifty thousand years. It's that humans are able to do a lot more today than they were back in caveman times because there have been billions of humans thinking for a long time and building off of each other's accumulated knowledge. We have, like, very good retrieval and scaffolding versus fifty thousand years ago.

It's not even I wouldn't even call it a scaffold. This is, like this is a very, like, organic emergent property of just, like, humans being able to accumulate knowledge, share it, and build off of it. We're not seeing that with AI models today.

They kind of they're they're born into a world for and they exist for a very short context window, and then they just, like, disappear. Mhmm. And, yeah, there are things that you can kinda do to, like, continue them, but it's very limited.

I do think eventually we will and we're starting to see, like, signs that we're entering a world where they can coordinate on a large scale. I think MoltBook and Openclaw, when they first came out, I think it was obviously a bit overhyped, but they were an indication of where things could go in the future. And I do think that eventually we get to that kind of world.

Yeah. The ability of the models to to share knowledge on a more global level and be able to build on that knowledge productively.

that exist today are accelerating what researchers at the Frontier Labs can do. There's obviously, like I said, limits to that right now, but the ability to use the models to improve the the model research is a real thing, and it's it it is like an amplifying force. I think that will continue to be true.

I think they'll become more true over time. One thing that I am comforted by is I think all the researchers at the frontier labs I all the frontier labs, I think, recognize what is at stake and what these models like, what what the what the risks are. And that's something that I I find comforting that I think everybody really understands.

Like, okay. This is a pretty serious thing, and it can lead to really great things or it can lead to really bad things. And, yes, there's a competitive dynamic between the labs, but, like, we can also try to figure out how we all get to the positive outcomes rather than the very negative outcomes.

they worked I mean, this is probably not even true for your audience necessarily, but there's a lot of people that experimented with AI back in, like, 2023 and felt like they couldn't trust the outputs and then don't use it for really high stakes decisions. And, actually, I think the models have progressed to a point where they are very good for these kinds of things. I mean, I asked them tax advice, or I bought a condo recently, and I was asking it for advice.

I'm like, okay. Well, what's all the paperwork that I have to fill out, and, like, how do how do I what does it all mean? It's actually really good for these kinds of questions.

Mhmm. So I use it day to day for for a lot of this kind of stuff, and I think they're at a point now where they've actually been at a point for a while now where I feel like I can just trust the outputs, arguably more than I could trust the output from from a human person. An expert human.

Yeah. Okay. I have two two final questions for you.

Oh, this is such a good question. I wish I had time to think about this ahead of time. You can you can just hang out with me and think about it.

Yeah.

still I still feel like it's not consensus, though, because, like, you know, people still don't publish the benchmarks this way. Oh, that's true. Yeah.

Yeah. That's actually why I wrote like inertia. That's kinda yeah.

But that's kinda why I wrote the essay. I was just like, look. I mean, we can talk about this, but, like, yeah, the part of the motivation is, like, I would talk to researchers about we it makes sense to show the benchmarks with an x axis.

Whether it's tokens or cost or time, there there should be an x axis. And everybody would say, like, yeah. That makes sense.

We should do that. But But they're not acting with the importance of, like, a good heart. Like, this is we have to measure the correct thing.

Well, really, their response is people expect us to to publish the Grid. Mhmm. And then well, okay.

Well, why do people expect the Grid to be published? Because everybody publishes the Grid. And so you kind of end up in this this bad equilibrium where everybody kind of knows that it's a bad equilibrium, but, like, nobody wants to break out.

And I I felt like, okay. Well, if I just hopefully come out and say, like, look, guys, let's all recognize that we're in a bad equilibrium, and let's move to this different equilibrium where we're we're plotting things with an x axis, that hopefully that can no.

pieces of that are clearly being externalized in all these applications. Yeah. I do I do think this is related to the fact that, like, benchmarks should be evaluated with an x axis of tokens or cost.

I I have seen some evals recently that show, like, okay. Well, with with a routing layer, you can achieve much better performance Mhmm. By basically doing consensus among the models.

Yeah. And, like, I definitely believe that if you do consensus among the models, that you're gonna achieve better performance than any individual model. But it's important to ask, like, are you gonna do better than having that model basically think for longer?

Like, once you control for the amount of test time compute, is it is it actually still doing better? Mhmm. That's that's the question that you want to figure out.

Okay. That's very principle of you, which is like, yes, routing is fine, but it's all subject to the same budget question. Yep.

Right. If you put it on the same scaler, then you can make an optimal decision, and I think maybe I win. Mhmm.

I I I don't even know necessarily that I I would believe that the routing does better, but then there's still a question of, is it gonna do significantly better? Is it very fragile? Is it reflective of real world use cases compared to benchmarks?

Because, like, one issue you could run into is that you could optimize for certain benchmarks with the routing, and then show, like, oh, yeah. We see there's big improvement on these benchmarks. But in real world use cases, it actually ends up not being a significant improvement.

So I I would say, at the very least, like, I would say, you wanna control for test on compute, and then you also wanna have all the same skepticism about benchmarks that you would normally have.

Yeah. It's great to be back.