rsi.watch

Epoch AI · 2026-10-09

Can AI automate AI R&D yet?

“As of right now, AI is far from automating AI R&D.”
Who
Epoch AI
Date
2026-10-09 Date from the Substack post.
Evidence
official post with figures about evidence levels
Replication
None recorded
Checked
2026-10-15, against the source page
Links
Source

Also on the page

  • Epoch gave Fable 5 and GPT-5.6 each 3,000 GPU-hours to develop a new post-training technique. A score of 100% means matching the gains of a published method, SDPO, over a GRPO baseline. "Recent frontier models’ best efforts topped out at 15% of the performance gains from SDPO".
  • The post also gives a more generous reading: "If assessed generously, Sol’s method achieved 35% of SDPO’s gains".

← Ledger