Try Pro free

MTS · 3 min · Nov 12, 2026

Compute and data constrain how fast AI research can accelerate

Ege Erdil disputes short-timeline intelligence-explosion forecasts, arguing that cognitive labor cannot replace scarce compute and data indefinitely.

The episode’s best minutes

Full analysis
  1. 0:59Frontier labs may struggle to sustain high margins5 min
  2. 9:36Long-term compute commitments could put lab equity at risk3 min
  3. 16:20Compute and data constrain how fast AI research can acceleratePro pick 3 min
  4. 4 more moments in the full episode
Up next · My First MillionPeer groups can become durable businesses in overlooked industriesPro pick 7:16
“I think a lot of the progress is just driven by data and compute.”
Ege Erdilat 17:04 · MTS
17:04

Takeaways Pro pick

  1. More researcher effort has diminishing returns without more compute or data.
  2. AI has seen both unusually fast compute scaling and software progress.
  3. The evidence is suggestive, not a causal experiment.

About this segment

Ege Erdil disputes short-timeline intelligence-explosion forecasts, arguing that cognitive labor cannot replace scarce compute and data indefinitely. He points to historical links between compute scaling and software progress.

Ege Erdil

Ege Erdil is a research engineer at Google DeepMind; former Mechanize cofounder and Epoch researcher.

Co-authored research on algorithmic progress in language models and computer vision, and studies AI forecasting and automation.

  • 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 “Why Frontier AI Labs Could Go Bankrupt | Ege Erdil”, Nov 12, 2026. Watch the full episode

Answers “When will we get AGI?” See all 35 answers

Transcript

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

  1. 16:24

    cruxes? And also, when do you think the intelligence explosion will happen, if at all? I don't think it will happen is the answer to the second question. I think it's like not a good concept to be talking about personally. But I think for the first question of like, why do I disagree with them? I mean, I think that I think with Daniel, it's just very clear that he has he just thinks we'll have a software on the singularity.

  2. 16:49

    Like, like, that's that's the key. I don't think that's going to happen. He thinks it's going to happen. And then you can try to get into why he thinks that. And I think if you look at the models, then I just think there's way more that you have to do to automate, to completely automate the process of making AI better. And he doesn't, oh, sure, for example, I think a lot of the progress is just driven by data and compute.

  3. 17:12

    And then he agrees with that, but then he doesn't think it's going to be a bottleneck. So the difference is that he thinks you can compensate for this by just having a lot of cognitive labor going into the process of making AI better. And I just don't think that. I think the returns on cognitive labor past some point is very low. So unless you have more data or more compute. And he just doesn't think that.

  4. 17:34

    So I think that's the key disagreement. It's very hard to settle because it's very hard to run an experiment to determine who is right here. Because you would have to have some group of people who have tons of labor but they have like almost no compute and they're cut off from the rest of the world and like we can't run we can't run that experiment because it's the best you can do is if you have a team inside a company had less compute then how much does that affect their progress but in that case they're still benefiting from spillovers from other teams so that experiment is very unclean it's not a good proxy for uh what if the entire world has less compute and i think the, like my argument for thinking this is true, is that in a lot of the diverse domains, you have a tight correlation between by how much compute in that domain was being scaled, and by how fast

  5. 18:26

    you're making software progress in that domain. So for example, we've looked at this for a bunch of fields, we looked at for like chess, we looked at for SAT solvers, we looked at it for various different domains in AI, including natural progress, natural language processing, and computer vision and so on. This was where we were still at Epoch. And the finding is that the fastest software progress basically in any domain we have seen has been happening in AI. But the fastest compute scaling also in basically any domain we know is happening in AI. And the factor by which the AI thing is faster seems to be the same as a factor by which AI has been scaling inputs faster, scaling compute faster. And these numbers always appear to be around the same order of magnitude, which is kind of like a strange coincidence, if you think they should be independent.

  6. 19:12

    So, like, I think that's suggestive evidence. It's not decisive because it's not a causal experiment. It's just observational. But it's strong suggestive evidence that compute should, like, play a big role here. And I think they just think that, yeah, it plays a big role, but you can just have really clever researchers who devise really good experiments at small scales. They have incredible research tastes.

  7. 19:34

    And you can just, like, get a lot out of that. that you can use the existing compute 1,000, 10,000 times or whatever more efficiently than we're using it now. And that's the thing that doesn't seem plausible to me.