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

Summary

---

Podcast Summary

No Priors: Artificial Intelligence | Technology | Startups
Episode: Why Traditional Benchmarks Fail Modern AI Models with OpenAI Research Scientist Noam Brown
Date: June 26, 2026
Hosts: Sarah Guo and Elad Gil
Guest: Noam Brown (OpenAI Research Scientist)

---

Episode Overview

This episode explores the inadequacy of traditional AI evaluation benchmarks in the era of large-scale test-time compute. OpenAI’s Noam Brown shares deep insights into why current benchmarks don’t reflect the real-world capabilities of frontier models, how evaluation frameworks need to adapt, the rise of “inference budget as capability,” and the implications for safety, research, and industry competition. Brown draws on his experience releasing major OpenAI models and experimenting with agents tasked with open-ended reasoning, like poker solvers, to illustrate what’s at risk if evaluation culture doesn’t keep pace with rapidly advancing AI.

---

Key Discussion Points & Insights

---

The Problem with Traditional Benchmarks

• Flat Benchmarks vs. Scalable Compute

  • Modern AI models' capability is now a “function of how much money (compute) you put into it” (00:00, 11:49).

  • Traditional benchmarks measure a model’s output in a fixed setting, but today’s models can improve their performance with more test-time compute, making the single-number, grid-style evaluation obsolete.

    “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.” —Noam Brown (11:49)

• Misleading Benchmark Grids

  • Users often see minimal percentage improvements between models in benchmark grids (e.g., GPT-4.5 vs GPT-5.5), but this ignores efficiency and computational allocation, causing skepticism until hands-on use reveals the real improvements (01:37–03:30).

• Test-Time Compute Plateau is Moving Outward

  • Earlier models (e.g., GPT-3) would plateau in performance quickly, but current models can keep improving on benchmarks for “weeks” before asymptoting, rendering previous evaluation cutoffs outdated (03:30–04:20).

---

Rethinking Evaluation Methodology

• Evaluation Must Now Include a Budget Limit

  • Fair assessment requires specifying a compute budget (in tokens, dollars, or time), or charting performance as a function of resources spent (04:00–04:20).

    "My claim is the proper way to evaluate the models now is... you plot the performance as a function of the amount of test-time compute." —Noam Brown (04:00)

• Practical Trade-offs for Users

  • While in theory you can let models “think for a week” to optimize results, in practice rapid iteration trumps exhaustive computation for user experience. Flexibility and user patience are key (06:08).

• The Problem of Benchmark Maxing

  • Scaffolding models or aggregating outputs (e.g., running a model multiple times and choosing the best output) can game benchmarks, producing ostensible progress that vanishes when taking compute into account (07:02).

    “It’s 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.” —Noam Brown (07:02)

• Repeated Call for X-Axis on Benchmarks

  • Brown advocates for every published result to include an x-axis: tokens, time, or compute cost. Despite consensus on this point among peers, publication inertia persists (32:16–32:47).

---

How Brown Evaluates Models Internally

• Custom Poker Bot Test

  • Brown tests new models by tasking them to create poker bots—a domain requiring layered reasoning, with little open-source code, making it a strong test of genuine model capability (08:29).

  • Progress:

    • Early models failed.
    • GPT-5.2 could build a river solver with guidance—"felt like a grad student."
    • GPT-5.5 could almost "zero-shot" the solver and optimize code better than Brown could alone (09:15–11:10).

    "With a bit of gentle steering, it just kind of felt like a grad student... I could just make suggestions, and it would go off and then do it, and pretty quickly, it would actually come back with something really good." —Noam Brown (09:15)

    • Downsides include "gaslighting" or confidently incorrect outputs, improved in newer versions.

---

Safety and Responsible Scaling Challenges

• Old Policy Frameworks are Outdated

  • Evaluation frameworks (preparedness/responsible scaling) were designed when models' capabilities barely changed with more compute. Now, their proficiency scales with budget, but evaluation standards haven't kept up (11:49–13:50).

    “The frameworks... don’t really account for the amount of test-time compute… The policies that exist today don’t really address that question.” —Noam Brown (11:49)

  • Dilemma: At what budget do you evaluate dangerous capability? If real super-capability emerges at $10M compute, should safe release decisions be based on $10, $1K, or $10M runs?

• Model Release Cycle is Too Fast for Proper Evaluation

  • Models could require months of continuous operation for challenges to manifest. But new models often appear before thorough, long-horizon behaviors are observed (15:13–17:06).

---

Latent Capabilities & the Limits of Exploration

• Underexplored Potential

  • Example: OpenAI’s model disproved the “Erdos unit distance conjecture,” something even top mathematicians couldn’t, using relatively little compute. With $100K compute budget, Brown argues, similar feats may have been possible months earlier, if anyone had tried (17:14–19:21).

    “What happens if I put $100,000 worth of compute into 5.5? What could it do? The answer is like, yeah, you probably could get stuff like that out of it.” —Noam Brown (17:14)

• Rate of Progress Discourages Deep Dives

  • With major model releases every 2–3 months, investing in deep, high-budget experiments often feels wasteful since next model might do it faster, better, cheaper (19:43–20:59).

• Focus Shifts: OpenAI prioritizes building new, more capable models and safe scaling over exhaustive “push to the limit” experiments on current models.

---

Recursive Self-Improvement (RSI) and AI Research Taste

• Not Yet at “Just Set Goal = Build GPT-7” Stage

  • There exist large swaths of tasks where more compute brings rewards (e.g., Sudoku), but also domains (factual recall, novel algorithm design) where model limits are quickly reached despite high compute inputs (21:23–24:52).

    “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...” —Noam Brown (21:23)

• Human Complement, Not Replacement

  • Current strengths: acceleration of routine research and optimization (“10–100x faster”). Weaknesses: “research taste” and novel hypothesis generation—models suggest but can’t yet invent (23:59–25:24).

• RSI Trajectory

  • Brown expects a “gradual takeoff,” with researchers’ workflow transformed rather than replaced (25:18–25:58).

    “Right now it’s more about transforming what researchers do rather than fully replacing… There will be a gradual takeoff.” —Noam Brown (25:24)

• Time is the Bottleneck

  • Far from an overnight intelligence explosion—as progress depends on large inertia of running models for long durations, time itself is the new bottleneck (25:58–27:10).

---

Multi-Agent and Coordination Frontiers

• Multi-Agent Systems:
  • Brown sees “scratching the surface” of what multi-agent models can do; believes large-scale knowledge coordination will become possible, acknowledging the analogy to human civilization’s accumulated progress (27:17–29:04).
  • Current models “exist for a very short context window and then just disappear,” lacking durable, coordinated state.

---

The Competitive Landscape at the Frontier

• Pace & Pressure
  • Competition between labs is “very intense,” with mutual recognition of both positive and negative possibilities (29:47–30:43).
  • Brown finds comfort in shared seriousness about risks and positive outcomes.

---

Adoption Advice and Cultural Mindset

• Trusting Models for High-Stakes Tasks

  • Brown uses modern models even for important decisions (e.g., tax, legal paperwork)—believes many outperform human experts at routine knowledge-based tasks (31:03–31:49).

    “They’re at a point now where I feel like I can just trust the outputs, arguably more than I could trust the output from a human.” —Noam Brown (31:03)

• Benchmark Publishing Inertia

  • Despite technical consensus, the field is stuck on legacy “grid” reporting due to expectation and fear of breaking convention (32:16–32:47).

    "You end up in this bad equilibrium where everybody kind of knows it's a bad equilibrium, but nobody wants to break out." —Noam Brown (32:47)

• Routing Layers & Budget-Optimal Solutions

  • Many AI applications now act as orchestration/routing layers for allocation of compute to different model instances or strategies. Brown emphasizes that such advances must still be tested for budget-normalized, real-world gains, not just benchmark improvements (34:30–35:16).

---

Notable Quotes & Memorable Moments

• "If you give [a model] 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 million, it could do even more. At what budget should you evaluate these models?" —Noam Brown (00:00; repeated at 11:49)

• "What happens if I put $100,000 worth of compute into 5.5? What could it do? ... You probably could get stuff like that out of it." —Noam Brown (17:14)

• "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... There are some things where they will not improve." —Noam Brown (21:23)

• "You end up in this bad equilibrium where everybody kind of knows it's a bad equilibrium, but nobody wants to break out." —Noam Brown (32:47)

---

Timestamps for Key Segments

• 00:00–04:20 — Why traditional benchmarks mislead; necessity of budget-based evaluations
• 06:08 — User patience, rapid iteration, and practical use
• 07:02–07:50 — Gaming benchmarks and the pitfalls of "benchmark maxing"
• 09:15–11:10 — Internal use case: building poker bots, model progress
• 11:49–13:50 — Safety evaluations and why responsible scaling frameworks are outdated
• 15:13–17:06 — Model release cadence and challenges of evaluating long-horizon capabilities
• 17:14–19:21 — Latent abilities in current models and the Erdos unit distance example
• 19:43–20:59 — Why not focus on extracting every ounce of current model capability
• 21:23–25:18 — Limits of recursive self-improvement and model research taste
• 27:17–29:04 — Multi-agent systems and AI coordination frontiers
• 29:47–30:43 — The competitive dynamics among frontier labs
• 31:03–31:49 — Brown’s advice: trust frontier models for important decisions
• 32:16–32:47 — Why the field is stuck reporting on grid benchmarks
• 34:30–35:16 — Routing layers, real-world evaluation, and budget-normalized comparisons

---

Key Takeaways

• Benchmarks must reflect real-world usage: Fixation on static, single-number grids no longer suffices—performance should be presented as a function of inference budget.
• Compute unlocks latent capability: Models’ “real” abilities often appear only at higher computation/time budgets.
• Real breakthroughs may be hiding in plain sight: Researcher time, patience, and cost constraints mean radical results could be discovered by simply scaling up compute with current models.
• Frontier labs face new safety, eval, and competition headaches: Responsible scaling frameworks, research resource allocation, and competitive arms races all require reevaluation in the compute-scaling era.
• AI is not one overnight takeoff: Even with fast progress, model self-improvement is gradual and bottlenecked by time and research taste.
• Multi-agent and coordination dynamics are on the horizon: True compound AI civilization has not yet emerged, but seeds are visible.
• Cultural inertia is holding back better evaluation: Despite agreement, few break from the status quo—showing the need for leadership and norm changes in how progress is reported and interpreted.

---

Further Resources

• Find more episodes, transcripts, and analysis: no-priors.com
• Follow @nopriorspod on Twitter for updates and discussions.

---

This summary preserves the conversational tone and perspectives of the podcast. Quotes have been precisely timestamped and attributed, and structure follows the thematic flow of the episode for maximum comprehension.

Transcript

[A]

With GPT3, you couldn't scale, test time, compute. Like if you gave it a budget of $10 million and said, okay, well, let's see what GPT3 can do. It really can't do that much. The procurentist frameworks and responsible scaling policies, they don't really account for the amount of tests not 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 million, it could do even more. At what budget should you evaluate these models? The policies that exist today don't really address that question.

[B]

Hi listeners, I'm Sarah Gore and welcome back to no Priors Today. I'm here with Noam Brown, one of our godfathers of AI reasoning. We talk about the broken state of evaluations, very large scale test time computer, how he thinks about recursive self improvement, and what's next on the horizon for competition at the frontier. Welcome, Noam. I'm so excited to have you back.

[A]

That's great to be back. Yeah.

[B]

You are our first guest. I'm very proud of my taste in friends and researchers for the pod. Given how important inference time scaling has become to the industry. You should be proud too, having actually

[A]

pioneered it, played a part. Yeah, among many others.

[B]

You just wrote this essay that really resonated about large scale test time compute and why the industry is not evaluating these models as robustly as it should be. What was the motivation for it?

[A]

Yeah, the motivation was we released 5.5 and the initial reaction was kind of skepticism that it was a substantially better model. To be fair, that only lasted for a few hours before people had some time to play around with it 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 is this benchmark grid where they show all these different benchmarks on the X axis and then the performance of different models on the Y axis. And you can just 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 5.5 and 5.4 or, or other models, it was an improvement, but it wasn't a huge improvement. It was only a few percentage points in 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 time 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? There's at some point where the performance on the benchmark is going to 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, it's true in GPT3 land back in 2022, the 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.

[B]

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.

[A]

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.

[B]

Given the model evaluation cycle and the fact that performance does not asymptote for many tasks over quite a long period of time. What do you do about that in issue? The fact that some of the evals that you would want to run are both beyond the scope of budget or time that's reasonable given the current model release cycle.

[A]

I mean, I think for things like cyber, we've seen and actually the AISI in their evaluations has shown that the models continue to improve at 100 million tokens. If you run them for 100 million tokens, they're still improving at beyond that point. And that can take a very long time to run. But you also do see that the performance, it's not just a discontinuous jump, it's actually like you can see the slope of improvement over those hundred million tokens. And so 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 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. Can you predict what their performance looks like at an inference budget of, let's say, $10,000 only using inference budgets up to 10 or $100?

[B]

So maybe an orthogonal question for you. Do you think users are systematically not thinking long enough with their models about problems?

[A]

What do you mean by not thinking long enough?

[B]

If you can build an agent or control the amount of test time compute being used, there's, there's what is done by the model itself and there's what the user can do. Do you think that the industry is using test time compute an optimal amount way undershooting it, or it's a problem in the models where they just need to be able to do that thinking faster?

[A]

I think it depends on the problem. I think this idea that the models, you just let them think for a week or whatever and then they respond, it sounds nice. And yes, the benchmarks look great, but it's not very practical when working because you ask the model a question and then you sit there for a week waiting for it to come back to you. I think what people have found most effective is to kind of 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.

[B]

How would you characterize. There's a lot of talk about benchmark maxing and the ability to game different benchmarks. What would you characterize the landscape of benchmarks as today? And then do you have favorites that you think are more indicative of capability than others?

[A]

So the benchmark maxing thing is also motivation for writing an essay that I think it's really easy to show you can do much better than previous benchmarks 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 going to run it five times and take the best of the five responses or ask a judge which one 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 computer. That is one thing that I'm worried about when it comes to benchmark maxing. I mean, it's a little misleading is the only concern that I have. And then as far as the benchmarks themselves, I think there is always a risk of just optimizing for the benchmark and I've certainly encouraged my team and I think at OpenAI we're pretty good about not trying to optimize for specific benchmarks. But once you put out a benchmark, it's always at risk of just being optimized for. And I think one way to address that is to keep a held out private set that isn't publicly available.

[B]

The most popular fallback advice for figure out if a model is significantly better or not is to just play with it for a while. Do you have anything more sophisticated than that that you suggest people do? Do you create your own set of new evaluations each time besides private holdback?

[A]

At OpenAI, I think everybody has their own set of questions that they like to ask the model whenever it comes out. For me lately it's been I use them to make poker bots and see how good they can make a pokerbot. I think it's a nice eval because there is very little open source code for making pokerbots and there's a lot of published papers on it, but you really have to reason through everything. And it requires a lot of just reasoning and iteration and 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.

[B]

Now, can you describe, perhaps with your poker bot creation, how reasoning might have progressed in model releases for you guys? Over a few releases, yeah.

[A]

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 river solver. So that's, like, the final stage of poker. And that itself was, I thought, really impressive. I mean, I had to work with it a little bit, but I was actually really impressed because I was able to make the river solver probably about five times faster than I would have alone. There were a couple things that it 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. It kind of felt like a grad student where, okay, they would run into issues, but at least 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, it would actually come back with something really good. And then especially the optimization, I 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 unit test, I told it, okay, well, let's say I have $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 have $100 in the pot, and I just folded. How do I not lose $100? And it said, oh, you know, it's 92. It's close to 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. But once we got to 5.5, I actually thought it was way better. It was able to basically do it zero shot. And in fact, I've been working on just doing a full scale poker solver and it's basically able to do the whole thing with some gentle steering from me. And I wouldn't be surprised if, you know, six months or a year from now, the model is able to just zero shot an entire poker solver, basically my entire PhD thesis in one go.

[B]

Let's talk about the larger implications of needing to evaluate these models relative to, let's say, speed of their reasoning or efficiency versus, you know, token volume. Right. Or dollar budget or whatever your scalar is. Can you describe some of the larger implications in your essay, including around safety evaluations?

[A]

Yeah, the safety evaluations thing, 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 is released, they go through a series of evaluations to measure are there dangerous capabilities, could these models do things that 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. But the question is, okay, well how do you evaluate whether the model is capable of that? And they have various protocols about how they do these valuations. 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 and it made sense. Like with GPT3, you couldn't scale test time compute. Like if you gave it a budget of $10 million and said, okay, well let's see what GPT3 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 tests I'm computed. 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. If you give it a budget of $10 million, it can do even more. And so at what budget should you evaluate these models? The policies that exist today don't really address that question. 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 want to wade into this question. I think there's arguments on both sides. But I think the important thing to recognize is that this is a question that is not being. We're just kind of like pretending that this issue doesn't exist. And I think it's important to just one way or the other account for it.

[B]

Yeah, it was the mirror image of the capability question of if the models can continue to do more and more without asymptoting on some tasks at very large budgets, then they should also be able to do so for tasks we don't want them to do as a society. Right. And so testing for that and what budget is allocated. It also seems out of sync from the model release cycle itself. Right. There's been this acceleration of you get a new model every single, sometimes few days and weeks at this point versus six months. And you have a line in the essay where you say like, the only way to truly evaluate an agent on some very long running task might be to run it for a year. And that's gonna be true of both useful and negative tasks. Right. And so how do you think about that versus the model release cycle?

[A]

Yeah, this is also an interesting dynamic where basically as the models have become stronger, they're better able to operate over longer horizons. So again, with GPT3, if you wanted to run it for 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, 5.5 into doing a series of experiments that can run for weeks, for months.

[B]

Have you given your poker solver task like infinite budget yet?

[A]

I haven't really scaffolded something together where I just tell it like, okay, just run this for weeks. I think I could give it until

[B]

it has some tasks.

[A]

I could probably give it slash goal and just like, yeah, tell it to go nuts. But I think at this point it could 100% do the river 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 probably could just tell it like, yeah, go work on this for a month and then come back to me with a full complete poker solver. 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 want to know after six months, the only way to know full fully is to run it for six months. Now I'll get to things we could do to address that a little bit later. But it's important to recognize that the model release cycle is, look, we're releasing new models 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 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. I think that's going to 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 the release? It's actually very difficult because, yeah, you would have to. The only way to really do the evaluations is then delay the model release cycle. And there's a lot of competitive pressure right now to not do that.

[B]

Do you think there's exciting latent capability in the models that are already released that people have not fully explored given timeline?

[A]

I think absolutely. I think actually a really great example is the Erdos unit distance problem. So for the viewers that don't know, we used an Internal Model at OpenAI a few weeks ago to disprove the Erdos unit distance conjecture. Now, I'm not a mathematician, but this seems like it was a pretty big deal in the math community. It was like the 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 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, the 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. Now, it's not as simple as just asking 5.5, hey, here's the Erdos unit distance conjecture. What's the disproof? You had to scaffold it a bit. You had to like steer it a bit. And so somebody found, okay, you ask 5.5 lists a bunch of ways that you could tackle this problem. And then it lists one of the paths that are actually promising to get to the disproof. And then you tell it like, okay, explore this some more. And then if you do this enough times, it actually ends up arriving at the disproof. Now, what this means is you could in principle ask 5.5 to, as a general purpose scaffold, list a bunch of different strategies, and then for each strategy, tell it to investigate that strategy. And, and then it would probably be able to arrive at the disproof with a general purpose scaffold. Now, that scaffold would be very expensive. I mean, it would probably cost, I just ballpark, like 1,000 to $100,000. But it would be possible, and it would have been possible for somebody to disprove the Erdos unit distance conjecture before we did using a general purpose model, and nobody had explored sufficiently. What happens if I put $100,000 worth of compute into 5.5? What could it do? And the answer is like, yeah, you probably could get stuff like that out of it.

[B]

So people should be experimenting more with the current generation in terms of.

[A]

Well, this is, I think it's an interesting question of is it worth it to experiment with because again, the model release cycle is every couple months we put out a new model that's even more powerful. And so the cost of disproving the Erdos unit distance congesture drops by like 10 or 100x with every model release cycle, probably in some cases more.

[B]

You've seen the meme that's like, oh, why bother doing any engineering work when I should just wait for the next model?

[A]

Yeah, just go on vacation and come back two months later and then it's, you know, a thousand times cheaper. Do you agree with that?

[B]

Is that what you're doing right now at OpenAI, just waiting for the next model release?

[A]

I think, I mean, I will say we're in a period where progress is very fast and like, yeah, the models are becoming more capable. I can say, like at OpenAI, one of the things that we're, we're actively not doing. And 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, especially, you know, the internal models. We are trying to encourage people to not spend all their time just going through all the mathematical open problems, physics problems, and just 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 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 at 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.

[B]

What is changing about the direction or allocation of resources for research in your mind, given your beliefs about this very large scale, the impact of very large scale test time, computer, how does this interact with the idea of recursive self improvement, for example, where it's a dominant idea for how any lab gets to the best capability model?

[A]

So one thing I should clarify. I don't think we're at the point where, okay, you just give it an arbitrary, extremely high inference budget and it's just super intelligent across the board, slash, goal, make GPD 7 or whatever, and then, yeah, just go nuts.

[B]

What's between us and there then I

[A]

think, having played around with the model, 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 don't have access to Wikipedia or something, they're not gonna be able to do better answering that question if they've thought about it for a week compared to 5 seconds. Same with the model, actually. Interestingly enough, if you give the model these kinds of factual retrieval questions and you give them a little bit of time to think, they do actually do better. But if you give them a week, they're not suddenly going to do better at remembering dates. 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 without limit with more test time compute. So the example I'd like to point to is Sudoku. 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 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 say, like, okay, any model could keep doing better and better if it was just given more test time compute. 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 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. I've found I'm 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. Some things are getting better the faster than others, but they're not at the point where they're fully replacing researchers with just enough test time.

[B]

Computer, can you give an example or two of asking the model to do a research task for just like, this is a terrible idea.

[A]

I mean, I think going back to my poker solver example, I was really impressed with the model's ability to optimize the algorithms that I had developed in my PhD. It was honestly, it was shocking to see how inefficient I was in retrospect. And they were able to make it, like, you know, 10, 100x 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? Go ahead and, like, look at all the published work and synthesize that and then try to come up with something novel and it's not able to do it. And I can give It a lot of time and 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.

[B]

And how do you think that gets improved?

[A]

What I've seen is with every model release cycle, it does get better at this sort of thing. It's still bad in my opinion, but 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 this inflection point where suddenly it's actually good enough to be useful. I wouldn't be surprised if we encounter that point for research tastes as well.

[B]

Given that, what is your framing of RSI today? How should we think about it?

[A]

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 100x faster, you get bottlenecked by the things that don't go 100x faster. Over time, the things that we're getting bottlenecked on are going to shrink. And there will be, I think, 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.

[B]

So that actually implies that you don't think we're close to a very fast takeoff right now.

[A]

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 models just becoming very superhuman across the board 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 tests 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 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 just 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.

[B]

What do you think is on the frontier that is less explored now? We've talked about multi agent before.

[A]

I think multi agent is quite explored. I think think at sufficient scale. I think there's a lot more that could be done. But it's also one of the things that's hard to do. 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, I think requires 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 you look at human civilization, it's not that humans have become smarter over. It's not that they evolved to become smarter over the past 50,000 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.

[B]

We have very good retrieval and scaffolding versus 50,000 years ago.

[A]

It's not even. I wouldn't even call it a scaffold. It's a very organic emergent property of just humans being able to accumulate knowledge, share it and build off of it. We're not seeing that with AI models today. They're born into a world and they exist for a very short context window and then they just disappear. And yeah, there are things that you can kind of do to continue them, but it's very limited. I do think eventually we will and we're starting to see signs that we're entering a world where they can coordinate on a large scale. I think multipook 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 of

[B]

some coordinated compounding state.

[A]

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

[B]

Given this set of beliefs and your work. Like, how would you characterize just competition at the frontier between the, between the three kingdoms? If there is no overnight takeoff, it's just researchers grinding away, trying to make good high taste algorithmic and investment decisions about where to go and then compute allocation and then policy decisions and eval decisions. It feels like slightly more grounded than I suppose, like racing towards some immediate hard takeoff that nobody can catch you on.

[A]

I think the competition is very intense right now. I do think the models that exist today are accelerating. What researchers at the Frontier Labs can do. There's obviously, like I said, limits to, to that right now. But the ability to use the models to improve the model research is a real thing and 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, all the Frontier Labs, I think recognize what is at stake and what these models like, what the risks are. And that's something that 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 we can also try to figure out how we all get to the positive outcomes rather than the very negative outcomes.

[B]

I think I'd be remiss to ask, just because you have been right very early for a long time on the importance of test time, compute and reasoning as a framework. Are there ways in which you use the models that you should, you would encourage others to. Right? Is it just goal everything?

[A]

I think for a lot of people 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 on 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. 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 a human.

[B]

An expert human.

[A]

Yeah.

[B]

Okay, I have two final questions for you. One is, is there something you think that the rest of the research community doesn't agree with you on or doesn't understand the importance of quite yet?

[A]

Oh, these are such good questions. I wish I had time to think about this ahead of time.

[B]

You can just hang out with me and think about it. Is it weird to be consensus now? You're a bit salty. Three years ago, when you're like, why don't people understand how important this is?

[A]

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

[B]

Oh, that's true. Yeah. That's actually why I think that's like inertia.

[A]

That's kind of. Yeah, but that's kind of why I wrote the essay. I was just like, look, I mean, we can talk about this. But like, yeah, part of the motivation is, like, I would talk to researchers about it makes sense to show the benchmarks with an X axis, whether it's tokens or cost or time, there should be an X axis. And everybody would say, like, yeah, that makes sense, we should do that. But.

[B]

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

[A]

Well, really, their response is, people expect us to publish the grid. And then, 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 bad equilibrium where everybody kind of knows that it's a bad equilibrium, but nobody wants to break out. And I felt like, okay, well, if I just hopefully come out and say, like, look, guys, let's all recognize that, that we're in a bad equilibrium. And let's move to this different equilibrium where we're plotting things with an X axis that hopefully that can. Next time there's a model release, a company can feel comfortable not publishing the grid, at least not at the very front, the top line, and we can have a more productive evaluation of these models.

[B]

Then a last question for you. How do you think about companies across all of these specialized domains who feel the value that they have is essentially like the routing layer, the choice layer of, you know, my goal is composed of a bunch of discrete tasks. Some require more intelligence and less. And within my. My job as a vendor is to solve that problem or achieve the optimal outcome with taking into account the budget constraints. And so I will manage the parallelization. And how much inference do you spend on it from what model? Because I think the Frontier Lab point of view is that that routing happens both within the, you know, behind the API, behind the application, and then some of it in the model itself. And pieces of that are clearly being externalized in all of these applications.

[A]

Yeah, I do think this is related to the fact that benchmarks should be evaluated with an X axis of tokens or cost. I have seen some evals recently that show like, okay, well, with a routing layer you can achieve much better performance by basically doing consensus among the models. And I definitely believe that if you do consensus on the models that you're going to achieve better performance than any individual model. But it's important to ask, are you going to do better than having that model basically think for longer? Once you control for the amount of test time compute, is it actually still doing better? That's the question that you want to figure out.

[B]

That's a very principled of you, which is like, yes, routing is fine, but it's all subject to the same budget question. If you put it on the same scaler, then you can make an optimal decision. And I think maybe I win.

[A]

I don't even know necessarily that I would believe that the routing does better. But then there's still a question of is it going to do significantly better? Is it very fragile? Is it reflective of real world use cases compared to benchmarks? Because 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 this big improvement on these benchmarks, but in real world use cases it actually ends up not being a significant improvement. So I would say at the very least, I would say you want to control for test on compute, and then you also want to have all the same skepticism about benchmarks that you would normally have.

[B]

Awesome, Noam, thanks so much and for being on the mission for breaking us out of this false equilibrium.

[A]

It's great to be back.

[B]

Find us on Twitter nopriorspod. 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 [e-mail].