It's important to distinguish between two different takeaways you could have from the recent METR report:
Shorter timelines: AI will soon fully automate AI R&D
Intelligence explosion more likely: once AI fully automates AI R&D, AI progress will accelerate until we get superintelligence
Many people don’t realise that the recent evidence does not just point toward shorter timelines but also toward thinking that an intelligence explosion is more likely.
But the exact updates to make are nuanced. I will walk through my tentative views on each.
Strong update towards shorter timelines
I think this is pretty clear. If we take the METR report at face value, AI is already causing much more of a speedup of AI R&D than I had realised or expected, and it is also improving at a faster rate.
So we're less far away from full automation than I thought and we're moving towards it more quickly. (There is some nuance here in terms of what’s causing this speedup: it might be stronger coding uplift than expected, better AI R&D performance, or something else.)
Moderate update towards an intelligence explosion
This one is more complicated.
The key question is whether the feedback loop of AI improving itself is strong enough to accelerate over time, even holding compute fixed.
We can break this feedback loop down as follows:
higher ECI1 → more AI R&D labour → more total AI R&D effort → higher ECI → ...
Now, let's assess what we can learn about each stage of this feedback loop.
Stage 1: higher ECI → more AI R&D labour
For each increase on the ECI, how much more quality-adjusted labour do you get for your AI R&D?
Based on the METR report, the amount of quality-adjusted labour increased pretty fast! During the same time period, ECI scores rose by about two years’ worth of progress, which is a lot less than I’d have expected to produce that much extra labour.
So stage 1 is stronger than I used to think.
Stage 2: more AI R&D labour → more total AI R&D effort
Can improved labour have the same effect as increasing both compute and labour? Or does more labour not make much difference because the compute required for experiments becomes a significant bottleneck?
Well, progress at Anthropic has sped up roughly 2x, METR guess mostly because of AI, while growth in training compute didn't speed up. That counts against the most extreme bottleneck view, on which extra research labour barely helps unless compute growth accelerates with it. But it doesn't test more moderate views that I find more plausible, on which compute only caps the speed-up at several times today's pace.
It also doesn't speak to the possibility that much algorithmic progress depends on scale-dependent algorithms. If so, algorithmic progress might be much slower if we couldn't increase the scale of training compute and benefit from new algorithms that only emerge at new scales. (Though METR say they saw no evidence of significant bottlenecks over the next 12 months, which counts a little against this. But they aren't confident and Anthropic redacted their analysis of several specific bottlenecks.)
So for me there is only a minor update on the strength of stage 2. Compute bottlenecks are still very plausible. An intelligence explosion is far from inevitable.
Stage 3: more total AI R&D effort → higher ECI
Some of the speed-up in recent progress wasn't driven by increasing AI R&D efforts, but just by it being easier to make AI progress than before (METR think non-AI factors probably contributed in late 2025, though they redacted the details).
This is evidence that stage 3 is more powerful. Doubling AI R&D inputs led to more progress than we thought. (Though only a little bit more, if you think most of the acceleration was due to AI increasing labour inputs.)
The more sceptical you are of stage 1—that AI has driven the recent faster progress—the more you should update about stage 3. Once you accept that progress has indeed become faster, you have to update upwards on at least one of the stages, unless non-AI inputs such as hiring or experiment compute also sped up significantly.
Summary
So, to summarise how my views on each stage have changed:
higher ECI → more AI R&D labour: much stronger than I thought
more AI R&D labour → more total AI R&D effort: slightly stronger than I thought, but compute bottlenecks later on are still plausible
more total AI R&D effort → higher ECI: a bit stronger than I thought
Overall, this is definitely evidence that there will be accelerating progress. But the evidence doesn't feel overwhelming at all. It could still go either way. I feel less bullish than AIFP seem to be.
And remember, even without accelerating progress, AI progress will still become much faster if we automate AI R&D.
This article was created by Forethought. See all our research on our website.
In their report, METR used Anthropic’s internal version, AECI, instead.



> Some of the speed-up in recent progress wasn't driven by increasing AI R&D efforts, but just by it being easier to make AI progress than before
What makes you say this? (If higher ECI → more AI R&D labour is already much stronger than you thought)