Ajeya Cotra argues that AI-assisted research could create a powerful feedback loop: better models help researchers build better models. She interprets METR’s assessment of Anthropic as a warning that this loop may already be accelerating progress.

Her argument has three layers:

  • Reported trend: an internal index of model test scores improved at roughly twice its earlier pace. Much of the evidence supporting METR’s interpretation remains redacted.
  • Inferred productivity: research depends on both people and computing resources. Cotra’s simplified model suggests that doubling overall progress could require a much larger improvement in researcher productivity, perhaps around tenfold.
  • Conditional forecast: if better models keep accelerating research smoothly, without major bottlenecks, she argues that extreme capabilities could arrive within a year.

Cotra’s own footnote qualifies the headline. METR attributed roughly 80% of the current acceleration to AI, implying about 1.8 times the progress expected without AI. Her judgment that Anthropic’s policy threshold has likely been crossed also draws on expected future acceleration.

Caveat: Cotra neither worked on the assessment nor saw its unredacted evidence. The one-year scenario is an extrapolation, not a measured result.

The 164-comment thread on Hacker News adds a debate about what the test scores capture.

What the thread adds

  • chr15m — argues that progress on measurable, trainable tasks can leave large gaps elsewhere. Their concern is that an aggregate score hides uneven abilities.
  • reasonableklout — pushes back: capabilities can transfer between domains, they argue, and rapid automation of verifiable work would matter even with those gaps.
  • 0x20cowboy — offers a small-scale counterpoint to abstract forecasts: they report trying an experiment in twenty minutes that previously seemed to require a week. That is a personal anecdote, not a laboratory productivity measurement.

What is actually improving?

cgearhart asks whether the gains come from data, architecture, or training changes. smlacy asks whether the time required to train models still limits the feedback loop. The essay’s simplified model leaves those concrete engineering questions open.

HN handles are pseudonymous, and individual comment scores are unavailable. This is a slice in HN’s own ranking order, not a consensus.