Why AI Research Acceleration Just Doubled | Eli Lifland & Brendan Halstead
MTS
November 15, 2026
AI Futures Project researchers Eli Lifland and Brendan Halstead break down why they sharply compressed their AI timelines based on new findings from METR and Anthropic.
Eli Lifland
This recent METR report that came out last week, where they did an assessment of kind of like Anthropic's internal speed of AI progress.
And then they sort of like released their findings and their best guesses. And they found that over the course of the last year, there's been on average about a 2x acceleration of AI progress. And it seems like at first, it was like mostly not due to AI speeding up AI research and mostly due to other things, but they estimated that of the current acceleration, about 80 percent of it is due to attributable to AI speeding up AI research. And that was basically a big update for us just because, for example, in AI 2027, this 1.7x acceleration due to AI research doesn't happen until early 2027 So just naively, it seems like we might be ahead of the AI 2027 timeline.
SPEAKER_2
All right. We are back. We are live with Eli Lifland and Brendan Halstead, who are researchers at the AI Futures Project, co-authors of 2040 and today co-authors of a new post on AI timelines, where the hook is AI timelines are much shorter than previously expected. And the members of the AI Futures Project now believe that we're going to get a superhuman AI researcher in 2027 So Eli, Brendan, welcome back to MTS.
Eli Lifland
Thanks for having me.
SPEAKER_2
So for the audience, can you explain more about why you updated your timelines to much shorter?
Eli Lifland
Yes. So the main factor is this recent meter report that came out last week, where they did an assessment of Anthropic's internal speed of AI progress, and then they released their findings and their best guesses. They found that over the course of the last year, there's been on average about a 2x acceleration of AI progress. It seems like at first, it was mostly not due to AI speeding up AI research, and mostly due to other things, but they estimated that of the current acceleration, about 80 percent of it is due to sort of like attributable to AI speeding up AI research.
That was basically a big update for us, just because, for example, in AI 2027, this 1.7x acceleration due to AI research doesn't happen until early 2027 So just naively, it seems like we might be ahead of the AI 2027 timeline, and that's the biggest thing that led us to make this update. There's also been various other things, like a few other pieces of evidence that the public has shared, and AI is being super good at seemingly doing math, solving very hard math problems. But this kind of like meter report showing this acceleration is basically the biggest input to our update.
SPEAKER_2
Why would math be a big update towards timelines?
Eli Lifland
I think mainly just because we think capabilities across areas are correlated, and I think seeing just how big the progress are in math isn't math. It seems like it is some update about how much progress you can make in some areas of AI research where there might be relatively verifiable targets that you want to optimize, like maybe, for example, pre-training research.
SPEAKER_4
Sure. So you explored some hypotheses as to why the AI Futures model was wrong.
So some of these are near-automated coder skill. Hold on. Coding skills result in more AI R&D uplift than we anticipated. Skills besides coding and research taste are driving acceleration and conversion of uplift to pace of progress is too conservative. So what would be some of these skills besides coding and research taste?
Brendan Halstead
I think the biggest one that we could think of is just like AI's like generating RL environments. This is something that we didn't model in the AI Futures model. But I think possibly like this is like a different category than other kinds of coding automation because it just directly contributes to the training. And you can imagine that there might be some relationship between like RL environments and points on this Anthropic ECI index that METR was measuring this acceleration in.
Then it's possible that automating that area of coding specifically would lead to impact greater than what we simulated in the AI Futures model.
SPEAKER_4
Right. Yeah, I'm also like kind of curious what your take on like data as a bottleneck was. We were just talking to our previous guest about this.
Brendan Halstead
Yeah, I think it depends on what kind of data you mean. I think that for pre-training, it's quite hard to tell from the outside. But it certainly seems like pre-training gains have continued, if not sped up, since the early days. So it doesn't seem like, whether it's because of synthetic data or improvement to data efficiency in pre-training, then it doesn't seem like there's been really much of a problem there. And we don't really expect one in the near future.
Read the other 12 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000802087364