# Test-time compute compounds with pretraining and reinforcement learning

Markdown copy of https://realalpha.app/s/test-time-compute-compounds-with-pretraining-and/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Greg Brockman](https://realalpha.app/builders/greg-brockman/) · Co-founder & President, OpenAI · Co-founded OpenAI, where he serves as president; previously Stripe's first chief technology officer from 2013.
- Host: [John Coogan](https://realalpha.app/people/john-coogan/) · Co-host, TBPN · Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026.
- Host: [Jordi Hays](https://realalpha.app/people/jordi-hays/) · Co-host, TBPN · Founded Branded Native; co-founded fintech Capital, which Rho bought.
- Show: TBPN
- Episode: [OpenAI LIVE: Greg Brockman, Mark Chen & More](https://realalpha.app/v/gaImbWPGgtU/), Aug 7, 2025
- Moment: 1:18:07 to 1:22:05 of the episode (4 min), playing on https://realalpha.app/s/test-time-compute-compounds-with-pretraining-and/; the episode from there: https://realalpha.app/v/gaImbWPGgtU/?t=4687

## Quote

> “And then there is a third way that we're really harnessing compute, which is **compute at test time**.”
> — [Greg Brockman](https://realalpha.app/builders/greg-brockman/) at [1:21:11](https://realalpha.app/v/gaImbWPGgtU/?t=4871) · TBPN

## Takeaways

1. Pretraining, reinforcement learning and test-time compute are distinct levers.
2. Brockman says the three scaling approaches multiply.
3. Engineering and data quality can amplify progress across the stack.

## About this segment

Greg Brockman describes three ways to turn compute into intelligence: pretraining, reinforcement learning and test-time compute. He says their effects multiply, so progress can come from combining them.

From “OpenAI LIVE: Greg Brockman, Mark Chen & More”, Aug 7, 2025. [Watch the full episode](https://realalpha.app/v/gaImbWPGgtU/)

## Transcript

From the transcript TBPN publishes with the episode.

[1:18:16](https://realalpha.app/v/gaImbWPGgtU/?t=4696) So it seems like there's been a few walls that we've scaled up through and then maybe hit them. There's been talk of, like, a pretraining wall. Now we're putting tons of resources and compute towards reinforcement learning. Is there a third is there a third scaling curve that we're going to be talking about in the next few years? Are we continuing to scale up those two primary vectors? Is that too high level of an abstraction in terms of, like, how we should be thinking about just progress along the the vector of scale? Like, give me the up to date thinking on just the the fruits of scale.

[1:18:55](https://realalpha.app/v/gaImbWPGgtU/?t=4735) Yeah. I'd say fundamentally, deep learning, I think that, you know, people talk about the bitter lesson. Yeah. It's almost this exploration into how do you convert compute into intelligence, right, through a you we have some particular techniques to do that that we're kind of constantly fleshing out. And the thing that's really amazing is if you rewind to, I don't know, even the 1940s for the McCullough Pitts neuron, which is kind of the precursor to neural nets, if you look at that paper, they have all these diagrams that actually look very similar to, like, the kinds of diagrams we draw now of multilayer neural nets and things like that.

[1:19:28](https://realalpha.app/v/gaImbWPGgtU/?t=4768) Like, the basic idea of what we're trying to do has not really changed in almost, like, eighty plus years, which is just a wild fact. It means there's something deeply fundamental about the thing that we are pursuing, and that idea itself, I think, kind of came from trying to model the information processing of the brain. And it's imperfect and not an exact analogy to biology and all of these reasons that it should fail or that people have said this thing is doomed, but the results are undeniable at this point. I mean, some people try, but it's really hard to to kinda close your eyes and and sleep on this in my mind. And it's very interesting if you look at you can find quotes from the mid nineteen sixties of people trying to poo poo the whole direction saying that these neural net people have no new ideas, they just want to build bigger computers.

[1:20:16](https://realalpha.app/v/gaImbWPGgtU/?t=4816) And you can basically say something very similar today. What we're trying to do one moment. A little water break. Second. Yeah. Exactly. For all of us. Cheers. Exactly. Cheers. I'll tell you, we're all human. A proof of humanity right there. Exactly. So what we're all trying to do is find ways of taking compute and really harnessing it. And sometimes you hit a wall, but these walls tend to be ones that you can drill through. What we've found is every time you scale up, everything, all of your engineering, all of your sort of scale and variance, all these things, they get stressed to the next level. It's almost that the tolerances become tighter and tighter.

[1:20:56](https://realalpha.app/v/gaImbWPGgtU/?t=4856) It's like launching a 10 x bigger rocket means you need to be, like, a 100 x, just more precise on everything, but it doesn't mean that the fundamentals of the science are different. So pretraining, there's definitely been a lot of discussion of data wall. Doesn't mean it's fundamental. Right? It just means that we need to be better and more precise at what we're doing. There's RL, which has been something that has kind of come from spending a small amount of compute to much larger amounts of compute now. And then there is a third way that we're really harnessing compute, which is compute at test time. And we publish some scaling laws around this, and all three of these things multiply. Like, that's the amazing thing.

[1:21:30](https://realalpha.app/v/gaImbWPGgtU/?t=4890) Mhmm. And, of course, the compute and the harnessing of it is the fundamental goal, but you get these multiplicative effects out of all of it through the quality of your engineering implementation, through the quality of the data sets, through a bunch of the refining work that you do. And there's lots of different techniques and ideas, and that's what makes this field so rich and why progress is just going to continue apace.

[1:21:52](https://realalpha.app/v/gaImbWPGgtU/?t=4912) What about on the infrastructure side? You guys have been busy scaling up. What what can you share on that front?