# Automated AI research could sharply change the pace of progress

Markdown copy of https://realalpha.app/s/automated-ai-research-could-sharply-change-the-pace-of/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Zack Kass](https://realalpha.app/people/zack-kass/) · Co-founder & President, Korza · Built OpenAI's sales, partnerships and solutions teams through ChatGPT's launch; now advises Fortune 1000 boards.
- Host: [Chris Williamson](https://realalpha.app/people/chris-williamson/) · Host, Modern Wisdom
- Show: Modern Wisdom
- Episode: [AI DEBATE: “Most People Have No Idea What’s Coming”](https://realalpha.app/v/mSjaMyP5QjY/), Aug 17, 2026
- Moment: 55:27 to 1:01:30 of the episode (6 min), playing on https://realalpha.app/s/automated-ai-research-could-sharply-change-the-pace-of/; the episode from there: https://realalpha.app/v/mSjaMyP5QjY/?t=3327

## Quote

> “the amount of **compute required for inference** is so great, and we do not have anywhere near enough.”
> — [Zack Kass](https://realalpha.app/people/zack-kass/) at [58:02](https://realalpha.app/v/mSjaMyP5QjY/?t=3482) · Modern Wisdom

## Takeaways

1. Automated AI development could remove humans from the research loop.
2. The compute needed to serve models may compete with frontier training.
3. Companies disagree on whether frontier research is still worth its cost.

## About this segment

They explain how AI systems could take over parts of developing their successors, and why that prospect is prompting calls to pace frontier development. They debate whether companies have economic reasons to slow down.

From “AI DEBATE: “Most People Have No Idea What’s Coming””, Aug 17, 2026. [Watch the full episode](https://realalpha.app/v/mSjaMyP5QjY/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[55:32](https://realalpha.app/v/mSjaMyP5QjY/?t=3332) So, I mean, the headline sentence is, we request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development. That last phrase, automated AI development, also sometimes gets called recursive self-improvement. The idea here is just the stated goal of all these companies, including Sam and Jakob on a live stream last year, is to take humans out of the loop of what they're making so that gbt6 is building gbt7 which then turns around and immediately starts building gbt8 and that is a step change like for all of human history and i think these companies really do think in terms of big history you hear sam and he's constantly thinking about ai as like an industrial revolution scale thing not an iphone invention scale thing through all of history human population has been

[56:22](https://realalpha.app/v/mSjaMyP5QjY/?t=3382) a bottleneck on progress not the only bottleneck but always a bottleneck there was never a sector of the economy that could just go off and double on its own and if we actually get to full automated arnd that changes and i think it's just like really hard for us to imagine what that would be like we have just no how much time have we got left before recursive self-improvement is no longer a discussion like even a debate at the moment it works no it works so when do you think it will be well this is a good step i mean i was i was waiting for this okay i was waiting What, like under three years?

[56:58](https://realalpha.app/v/mSjaMyP5QjY/?t=3418) No, it works. Wow. It works, just so we're clear. Recursive research works. Don't let anyone tell you otherwise. It works, and hence this letter. So let me explain this, why I think this letter was published when it was, and then we can debate. So the first reason is the attack, okay? The hugging face attack. There's damage control that needs to be done to win back public mandate, to tell people, hey, we care about you, which is necessary right now when people are like, well what are these models going to do next for what it's worth though this letter was in development before that attack happened Uh, and I think my point is the signatories and the support that is now that is sweeping through these companies, but let me, let me, I'll go through the list and then, and then we can talk about it.

[57:41](https://realalpha.app/v/mSjaMyP5QjY/?t=3461) The other reason is recursive research is here and everyone's realizing it's kind of the dog that maybe caught the car and it's like, okay, we did it. Now should we, for the, now everyone's putting on the live hat and going, wait a second, and Eric Hatton going, okay, maybe we should pause for a second. Maybe we shouldn't keep sprinting aimlessly at this incredible outcome. But there's another reason, which is that the amount of compute required for inference is so great, and we do not have anywhere near enough.

[58:14](https://realalpha.app/v/mSjaMyP5QjY/?t=3494) And we're going to have to point most of the compute from the frontier at inference. And this is because of the diminishing model returns theory, which is a paper I wrote that says at some point i'll send it to you it's really good um it cited all the time on modern wisdom and uh and uh but it argues that you and i won't care about gbt7 which is a really interesting problem and and it's actually one of the reasons that we get to debate why ai is being built and why people can talk about not wanting to make more progress because we're like what what the fuck's the matter with 5.6 is good enough and the dimensional returns theory is sort of here which is like hey the next model might matter to solving this crazy disease, but I don't need more intelligence. I need more time with my friends and family. I need this to pump the brakes. And the other reason is that

[59:02](https://realalpha.app/v/mSjaMyP5QjY/?t=3542) they found new surface areas. So whereas we used to call them labs, the companies that built this technology, then they became token distributors, then they became application providers. Now they're device manufacturers, but the end goal and the, the, where I think we should also debate is the agentic internet who controls the agentic internet. Because if you can turn your, now that you have these war chests, now that if you can turn your attention from building the frontier to actually building applications that, that control the internet, that's where you can actually lower your capex and massively improve your revenue. And so a lot of this arrives at a moment where everyone goes, okay, we need to, we need to slow the brakes, but also where a lot of, a lot of attention is going to start to be paid to trying to become bigger than Google by actually

[59:47](https://realalpha.app/v/mSjaMyP5QjY/?t=3587) building the agent that barters between two people or building the agent that plans a vacation and books everything. Like owning that is worth more than owning the frontier. And the new serve This area battle is for the agentic internet. So just to make sure I understand your claim, you're saying companies have an incentive to want to slow the frontier because there's more money to be gained in stopping R&D now and just trying to grab- We all have an incentive, but companies also have an incentive because the CapEx expense on the frontier is just the buildout on the frontier is going to be too great relative to the demand for inference.

[1:00:26](https://realalpha.app/v/mSjaMyP5QjY/?t=3626) the token demand is so great and all of the compute is going to have to be spent on inference i'm not i'm not quite seeing the connection there for two reasons okay one is if it's my best business strategy after i do the research to say i'm going to stop training frontier models i don't know why that means i should write a letter announcing that to the world and making sure that all my competitors do the same uh and two i think it's clear that pushing the frontier also means just driving down the cost curve everywhere beyond the frontier, or sorry, before the frontier.

[1:00:59](https://realalpha.app/v/mSjaMyP5QjY/?t=3659) Like you train Fable 5, Anthropics Best Model, and then you can distill from that Sonnet 5. And now you've got your like cheaper model that everyone wants to talk to because they don't need the advanced intelligence. So I think there are zero incentives that in like a straightforward, self-interested way point towards companies saying, let's slow down research. Research is how you get gains at every level of the intelligence.

[1:01:21](https://realalpha.app/v/mSjaMyP5QjY/?t=3681) Especially when as one of the three or four frontier companies by slowing down you're giving all of these other competitors a chance to