Partial automation of AI R&D tasks will substantially speed AI software progress before full automation.
Assessed
First pass on a small budget: assessed supported (credence 0.72) chiefly on METR's Nov 2026 Anthropic acceleration assessment, weighed against METR's developer RCTs and redaction/generalization limits. No decomposition or named arguments this pass; the unplaced Forethought passage left for a later pass. marginal_yield 0.5: decomposition into the cruxes (acceleration real and AI-caused; generality across labs; bottlenecks) would help.
Assessed Supported
verdict confidence 0.65 · credence 0.72
The claim is that AI systems taking over a growing share, but not all, of the tasks in AI research and engineering will markedly accelerate progress in AI software before AI R&D is fully automated. The theoretical case is that humans can concentrate on the tasks AI cannot yet do while AI handles the rest, so speedups compound well before the last task falls; the counter-case is that the slowest, human-only tasks and compute limits bottleneck overall progress, so partial automation yields modest gains until automation is nearly complete. The most direct 2026 evidence favours the claim. METR's high-access assessment at Anthropic (November 2026) found that benchmark capabilities of Anthropic's models (on Anthropic's version of the Epoch Capabilities Index) have grown about 1.9 times faster since October 2025 than before (95% interval 1.5 to 2.8), that this was not caused by faster growth in training compute, that a majority of the acceleration is likely attributable to AI use in R&D, and that the acceleration comes mainly from AI assistance rather than fully autonomous contributions. That is a substantial speedup arising from partial, not full, automation, and METR reports no strong bottleneck yet in view. The evidence has real limits. It covers one developer; much of the most informative evidence was redacted, so the public must take METR's judgment on trust; the benchmark-based measure may overstate progress on less well-specified tasks or reflect benchmark artifacts; and attribution to AI rather than other factors is uncertain. Controlled measurements of individual productivity are more mixed: METR's 2025 trial found experienced open-source developers about 19% slower with AI tools, and its 2026 follow-up was judged unreliable because of selection effects. Organisational-level acceleration and individual task-level uplift measure different things, and the former bears more directly on this claim. On balance the claim is supported, with credence around 0.7; replication at other labs and published, unredacted data would most move the verdict.
Claim entered the graph