MTS · 5 min · Nov 7, 2026

Fast takeoff remains a low-probability scenario

HBHerbie Bradley · AI Researcher & Writer

Herbie Bradley argues that evidence for imminent self-sustaining AI research is limited, and that progress may remain uneven because it depends on data.

The episode’s best minutes

0:53Fast takeoff remains a low-probability scenario5 min
17:27 Enterprise AI gains depend on changing workflows4 min
21:35 Anthropic’s RSI threshold is not a full acceleration loopPro pick 4 min
3 more moments in the full episode

“I have always kind of viewed the very fast takeoff scenarios as like a small probability shape or scenario or piece of the possibility space.”

HBHerbie Bradleyat 2:35 · MTS

Takeaways

Fast takeoff is one possibility, not the default forecast.
Models have automated little of the AI improvement loop so far.
Progress is strongest where data and compute are abundant.

About this segment

Herbie Bradley argues that evidence for imminent self-sustaining AI research is limited, and that progress may remain uneven because it depends on data. That leaves room for bottlenecks beyond model capability.

HB

Herbie Bradley is an AI researcher, writer, and stealth startup founder.

Theo Jaffee, Host · Co-founder & host, MTS
One of the first members of a16z's new media team; earlier hosted the Emergent Ventures-backed Theo Jaffee Podcast.
Sophia Dew, Host · Host, MTS

From “Anthropic Sees Imminent RSI Acceleration | Herbie Bradley”, Nov 7, 2026. Watch the full episode

Answers “What will AI look like in the next few years?” See all 169 answers

Transcript

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

0:59
glad to be back. Great location. So we were just talking about RSI and this idea that I've been thinking about a lot of radical AGI centrism, which is like, I think a lot of people at this moment have sort of a mindset of this is crunch time. Next year, we're going to get for real, for real rsi we're going to get like very very powerful artificial intelligence that's powerful across the you know full generality of human intelligence we're going to get something that looks like an ai 2027 scenario and this is their mainline scenario for what will happen it they don't see it as sort of a you know very timeline bullish edge case they see it as what will happen absent massive political intervention. And so the only thing to do at this point is slow down, pace the frontier, call for mass political intervention. I think a lot of people at the
1:55
labs are thinking this way. I think a lot of people in the AI safety and policy sphere are thinking this way. And I don't think it's necessarily justified. Certainly this sort of thing is plausible, and I can't rule it out. But it seems to me that people are getting a a little bit ahead of themselves. The models still have some weaknesses. There's still things that the models can't really do at all.
2:21
The data that we have on RSI doesn't actually show the models fully automating any part of the model improvement loop and so on. So what are your takes on this broadly? Yeah, I definitely agree with you on the sort of people maybe need to chill out a little bit perspective. I have always kind of viewed the very fast takeoff scenarios as like a small probability shape or scenario or piece of the possibility space.
2:50
And the evidence that we have so far on, you know, self-sustaining RSI happening very, very soon and leading, especially leading to very fast increases in capability is quite limited compared to the alternative hypothesis of, we're going to automate most of AI R&D and capability is going further than that. What will superintelligence be able to do? Become more limited by diffusion, limited by other bottlenecks.
3:16
We can't really tell which of these worlds we're in, but I do think people should have a lot of weight on kind of, you know, nothing ever happened style priors or priors about things being limited by data or limited by diffusion and so on. Yeah, I mean, limited by data. It also like jumps out at me that the capabilities that models have, that they have gotten really, really good at, are capabilities in which we have invested enormous amounts of data and compute.
3:46
So, like, for example, math or coding, these are areas which, A, they're easily verifiable, unlike other domains, and B, the labs have spent enormous amounts of resources, compute, capital, data. On improving the models in these specific areas. And in areas where there's not a lot of data, or the data's not very good, the models have made very little to no progress. Yeah, in a lot of cases. I do think, for example, the qualitative reasoning that you get from models on open-ended brainstorming has improved a bunch over the last year.
4:18
But nonetheless, it's definitely slower than in math, clearly, because we're seeing Navier-Stokes getting sold at the same time. You know, I can argue against fable and easily find gaps and holes in the arguments and so on. Yeah, I think a lot of this is going to be remaining quite jagged. And when I look for where the sources of the jaggedness will continue to be, that is just sort of data. You know, I think the researcher Yascha Sol-Dikstein from Anthropic had this talk, which was on Twitter a little bit, about talking to researchers at a university and giving the take that, oh, you only have a few years left to make intellectual contributions.
4:53
And I think that's like a take which may only apply in some ways to AI researchers themselves. I think a lot of the intellectual contributions that academics make are like not very well quantifiable. It's very hard to build benchmarks for them. Ultimately, there is more to intellectual contribution than what can be very easily quantified in an AI eval, especially considering almost all evals are still very bad in quality.
5:17
Yeah. Why don't we dive in on that?