FULL STORY

AI Accelerating AI Research: Claims, Evidence, and Doubts

After reports that Anthropic sped up research 10x, METR claimed convincing evidence that AI is accelerating frontier research, while analysts debated shrinking returns per R&D dollar and whether per-researcher speedup is even a meaningful metric.

2026-11-08 ~ 2026-11-09 · 4 episodes · 13 posts

Episode 1 · Anthropic's 10x AI Research Speedup Debated (2026-11-08, 2 posts)

Anthropic is said to have achieved 10x research acceleration through internal AI use, but critics argue the gain is a one-time boost from models like Opus 5.7 rather than a compounding, sustainable advantage.

Episode 2 · METR data recast on spending axis: AI capability accelerates but gains per dollar shrink fast (2026-11-08, 6 posts)

Drawing on METR's public AECI (Agent Emergency Capability Index) data, multiple analysts have independently reached similar conclusions: AI capability progress itself is accelerating, but R&D spending is growing even faster, so the progress bought per dollar is rapidly shrinking.

Confirmed

  • Michael Wiebe re-plotted METR's AECI trend against R&D spending rather than time: in the first half of 2025, roughly $1 billion of R&D bought about 4.05 AECI points, while in the first half of 2026 that fell to just 1.49 points—a year-over-year decline of about 63% in output per unit of R&D.
  • Ramez Naam, combining METR's fit of the capability growth rate of Anthropic's internal models with SEC filings, estimates an exponential decline in capability gains per billion dollars of R&D: 6.58 points in 2024, 4.05 in the first half of 2025, and a further slide in the first half of 2026—an roughly 8-fold shrinkage over two years. Reddit users citing the analysis reached the same conclusion.
  • On a time axis, progress did accelerate after METR's October 7, 2025 breakpoint, with annualized capability gains nearly doubling, but these findings are offset by the backdrop of explosive spending growth.

Unconfirmed

  • In the discussion, @trevposts pointed out that the key variable is whether spending will slow on a log scale; at the current pace of R&D spending growing 5–6x per year, exponential expansion can last at least another year. This is a personal prediction, and there is no consensus on when the spending curve will peak.

Why it matters

  • This analysis shifts the debate over "is AI still accelerating" from a time dimension to an economic one: even if the capability curve stays steep, declining output per dollar combined with the need for spending to multiply every year means AI progress depends on whether capital expenditure can expand indefinitely—a core constraint for lab financing and industry expectations.

Episode 3 · METR Claims Strong Evidence AI Is Accelerating Frontier Research (2026-11-09, 3 posts)

METR says it has compelling but redacted evidence that AI is accelerating frontier research without relying on larger training runs. Researchers cite estimates of at least a 1.4x research speedup, with one report suggesting AI nearly doubles the pace of capability progress—equivalent to a 10x faster researcher.

Episode 4 · Debate: Is Accelerating Individual Researchers a Useful Metric? (2026-11-09, 2 posts)

ZhongRuiqi argues that accelerating individual AI researchers is not a meaningful metric, since frontier labs like Anthropic are bottlenecked by coordination and compute rather than researcher headcount.