5 Comments
User's avatar
Elroy Flynn's avatar

That's a lot to absorb, reflecting the considerable effort you must have put into this. Great work!

I find myself generally concurring with your conclusions that there's little or no evidence of an imminent RSI takeoff.

I'll quibble with your attempt to relate the growth in global chip deployment with ECI gains. You used this as another bit of evidence of diminishing returns. However, it seems to me that chip deployment and ECI gains could be completely orthogonal measures - they don't correlate. If they were indeed strongly correlated, there's the implication that if chip deployments ceased - if we were stuck for a year at current deployments - ECI could not improve during that time. I don't think that that's the case. Also, I'd expect that much of the new capacity is used for duplicate work by the non-frontier labs that are developing their own models. Further weakening the correlation argument is that 30-60% of the current capacity is used for inference (according to various sources).

The [November 2026 METR report](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/#executive-summary) (months after your analysis) has claims of an increasing rate of capability improvements, but this is largely based on self-report surveys of Anthropic engineers, and METR acknowledges their lack of confidence in the result.

I guess we'll have to wait a bit longer to see whether the tail is drooping or wagging.

Oleg  Alexandrov's avatar

The bottleneck is always data, running the full process on realistic setups, evaluating, getting feedback, then doing another pass.

AI turned out to be a problem of mapping the world more than of sharpening the wits (which happens as a consequence).

This suggests that not only fast RSI is not plausible, but AI will stop getting smarter when we no longer have meaningful problems to throw at it.

So, AI will get smarter than us, and its own work will help improve it, but no explosion.

Lukas Finnveden's avatar

Hi, thanks for the post.

I believe that this methodology would have given you a significantly lower value than 5-10x if you had run the same methodology a year earlier, and we should expect it to increase over time as AI models get stronger and automate more of the research loop. Do you agree with this?

One argument for this is that if you kept subtracting ECI points from the models, the humans' productivity wouldn't exponentially fall towards 0. It would stabilize at a fairly reasonable number without AI assistance, and at that point, the productivity gain per ECI point would be ~0%. Your estimates suggest that the AIs have now started feeding back into the RSI loop enough to make this number increase to 3% per ECI point and I would expect it to keep increasing as they automate more and more of the research process.

To assess how close we are to a self-sustaining RSI loop, we'd ideally like to be able to measure how this number is trending over time. It's plausible to me that this trend would look pretty fast though I'm not sure either way.