What happened
On 7 October 2026 Epoch AI published 'Can AI automate AI R&D yet?', presenting early results from InnovationEval, a new evaluation that measures whether AI can independently discover novel machine-learning techniques comparable to those developed by human researchers. Agents were given an end-to-end task — devise a post-training method beating a strong GRPO baseline, anchored to on-policy self-distillation (SDPO) — with the finding that 'recent frontier models make little progress on this task, despite running experiments using thousands of dollars' worth of GPU time.' The paper sets out a memorization-tracking methodology (its successors GPT-6 Astra and Claude Fable 5.1 were already aware of the anchoring paper) and plans to expand and repeat the evaluation to track AI's progress toward automating AI research itself.
Why it matters
A methodologically serious, quantitative check on the widely-quoted claim that AI is close to automating AI R&D — useful calibration for capability-acceleration risk conversations.
Action needed
Use the InnovationEval methodology and result as an evidence-based baseline when evaluating claims about AI R&D automation timelines.