The BaseEssayLessWrong

Do verifiable metrics overestimate RSI?

In a five-minute LessWrong post, Alec Harris argues that the acceleration METR measured at Anthropic may partly reflect AIs overfitting to Anthropic's internal capability index, and so may not last. Harris puts 60% on a slowdown within two years and about 30% on ASI by the end of 2027.

Open the source ↗3 quotations checked, word for word ↓
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

Alec Harris argues the fast progress METR measured at Anthropic could fade, because AIs may be getting good at the benchmark rather than at research in general.

For researchers

Alec Harris separates skill at a benchmark from skill at general ML research, and argues the gap between them, not transfer to fuzzy tasks, decides whether acceleration persists.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“However, an alternative possibility is that Anthropic’s AIs might be assisting with ML research in a way that overfits to AECI such that this acceleration will not be sustained.”
Alec HarrisHarris sets out the central alternative to reading METR's finding as the start of a recursive self-improvement loop.
Found word for word in the source
“As long as you think that even a small fraction of verifiable capabilities transfer to fuzzy tasks, a FOOM on verifiable tasks will quickly transfer to fuzzy tasks.”
Alec HarrisHarris explains why little weight belongs on the concern that gains on verifiable tasks will not carry over to fuzzy ones.
Found word for word in the source
“My all-things-considered view is that there is a 60% chance we will see a deceleration in AI progress in the next two years, but a significant, maybe 30%, chance we will see ASI by the end of 2027.”
Alec HarrisHarris closes the post with an overall forecast.
Found word for word in the source

The record

Questions
What happens when AI helps build more capable AI? · How fast is AI advancing, and where is it heading?
People
Alec Harris (Wrote it)
What we read
We read the whole source.
Found by
The Radar, run of November 11, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Do verifiable metrics overestimate RSI?”, November 11, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-11-11-verifiable-metrics-overestimate-rsi
Our notesConcerns Anthropic's capability measurements, so conflict is true. One comment on the page (Kevin Zhang) was not used. Harris's probabilities are personal estimates, not measurements.