Automated AI research

we agree with the bottom-line conclusion of the report — that the risk of a catastrophe from Opus 4.6 or a less capable Anthropic model automating R&D in any domain is very low — but we think the evidence presented in the report is inadequate to establish this.

First public third-party review of a frontier lab's own assessment of how far its models are from automating AI R&D — a template for external verification of AI-R&D capability claims.

we think the model is far from being able to fully automate the activities needed for R&D in key domains.

The primary lab document classifying risk from models automating AI R&D as 'very low' for current models — making AI-R&D automation an explicit tracked threshold in a frontier safety framework.

The papers produced by the agents were unambiguously rejected.

Empirical test giving agents real resources to reproduce and extend two papers: agents ran hundreds of experiments but failed on research judgment, and the original authors rejected both AI-written papers.

impatience, poor time and resource management, overconfidence in weak hypotheses, difficulty coordinating parallel experiments, and hard limits from context length.

End-to-end research benchmark: GPT-5 beat baselines in only 6.7% of evaluations and completed 26.5% of subtasks — quantifying the capability-reliability gap in autonomous AI research.

At current rates of compute growth and algorithmic progress, this model's median prediction is >99% automation of AI R&D in late 2032.

A deliberately minimal 8-parameter model of when AI R&D becomes fully automated, offered as a transparent alternative to the 33-parameter AI Futures model.

We are right around the corner from recursively self-improving systems.

Reported survey of how much AI already does AI research — self-assisting coding models, AlphaEvolve, Darwin Gödel Machines — with lab researchers quoted on both promise and limits.