Twitxy Saved

Elias Schmied

Verified account This account is verified. Learn more

@reconfigurthing

I_DECLINE_public_HF_RCE_as_offtask_prodethical

foundational questions, AI, stuff like that
london eliasschmied.substack.com Joined August 2019
3.8K Following 4K Followers 26K Posts

Latest posts 5

Elias Schmied Verified account @reconfigurthing
I want to distinguish between two different ways of updating on the METR report:

-generically updating towards shorter timelines:
This is reasonable - logically you should update towards faster progress if you were surprised to see significant uplift from coding agents this year. I personally was not very surprised, so it's not a large update for me, but it makes sense.

-updating to deferring more towards the AI Futures model and similar approaches:
I think this is the one that doesn't make sense.
Speedup from coding uplift has little bearing on what has become the main crux of very short timelines - how much generalization there is to non-verifiable tasks - so it cannot be a big update on that. For example, if you previously had 70% credence on "AI progress will not generalize as easily to the non-grindable, messy, long-horizon parts of the AI R&D process", and 30% on the opposite, I think the METR report should not update you much on this.

I worry that a lot of people are doing the second - they would not have been that surprised by significant coding uplift if they had thought about it a month ago, but are now changing their mind on deeper assumptions just because the social vibe has shifted.
6 replies 1 reposts 40 likes 3.1K views View on X
Elias Schmied Verified account @reconfigurthing
Here's some important context that many people who haven't been tracking this stuff closely probably don't have: the models of RSI that are currently going around are based on an outdated understanding of algorithmic progress (arguably the most important parameter).

We've only started to appreciate pretty recently that:
-much of past improvement in algorithmic efficiency may have been driven by a few scale-dependent innovations, i.e. would have appeared much smaller without the simultaneous compute scale-up.
-a lot of what was previously counted as "algorithmic progress" may actually have been data progress (e.g. synthetic data, data curation, distillation).

Both of these push against the possibility of fast software progress being able to overcome compute and data bottlenecks.

As far as I can tell, the estimates of algorithmic progress used by both the AI Futures model and @TomDavidsonX's 2025 model don't take these into account. This is one of the reasons why I don't feel like I can take their outputs at face value.
10 replies 3 reposts 128 likes 7.7K views View on X
Elias Schmied Verified account @reconfigurthing
It's useful to hear METR's subjective impression of whether there are near-term bottlenecks.

I think it's also worth giving the additional context that the standard story from economic history is that over the course of technological development, the binding constraint switches between different inputs as investment/R&D focuses on whatever the current binding constraint is (Claude gives a cool example below). So I think we should have a relatively strong prior that there will be other bottlenecks if we try to project this out far, e.g. over multiple orders of magnitude like some are implicitly doing.
METR @METR_Evals Nov 5
3. We believe there is a chance that capabilities progress could lead to even faster acceleration in the next 12 months, since we did not observe clear bottlenecks or strong diminishing returns. However, given our limited vantage point, we may have missed bottlenecking factors.
View quoted post on X
2 replies 0 reposts 17 likes 993 views View on X
Elias Schmied Verified account @reconfigurthing
Huh, this is great - seems like either way we'll get an update next year then! Kudos to them for sticking their necks out with a concrete prediction.
Andrew Curran Verified account @AndrewCurran_ 6d
The AI Futures Project have updated their timelines. They say we are five months ahead of AI 2027, and they now believe a Superhuman AI Researcher will arrive sometime in 2027.
View quoted post on X