Anthropic says Claude leads 26% of its AI R&D — but self-improvement still has limits
Exponential View (Azeem Azhar) · rss · 2026-10-26
Exponential View's Azeem Azhar surveys the real progress and hard limits of recursive self-improvement (RSI) in AI.
The loop is real
- Anthropic co-founder Jack Clark told the IMF the singularity has begun: in August Claude led 26% of the company's AI R&D with human supervision; OpenAI posted 719 math manuscripts produced by its internal model.
- Sakana AI and UBC revived the Gödel machine idea: agents iteratively rewrote their own code, gaining 30pp on SWE-bench and 15pp+ on Polyglot.
- Google's AlphaEvolve found a faster matrix-multiplication kernel (23% speedup), cutting Gemini training time by 1%.
- MiniMax reported its model improved its own agent's harness, sampling and memory over 100+ rounds, a 30% gain in internal evals.
The catch
In every case the AI rewrote surrounding infrastructure — harness, memory — not model weights. A September experiment claiming an AI-built agent matched a human-built one was inconclusive: run-to-run variance exceeded the difference.
A cautious frame
One taxonomy splits AI self-improvement into five types; only types 1–2 (autonomy levels) have been observed, and autonomy alone doesn't guarantee takeoff — each round must yield at least as much as the last. The headline conclusion: for science, intelligence is not the bottleneck.
More from AGI Musings
- Dev warns abliterated models could wreck software supply chains with 1000+ transitive dependencies — nitarshan · 2026-10-27
- Delete your priors: why every SOTA model drop should reset how you think the world works — zack_overflow · 2026-10-27
- Replit predicts every Fortune 500 will be a 'Self-Driving Company' by end of 2030 — amasad · 2026-10-27
- Anthropic says insurers can't cover its own catastrophic risk, X mocks it — GarrisonLovely · 2026-10-27
- MIT's Ziming Liu: AutoResearch doesn't generalize — blind training tricks cause 'forking' disasters — ZimingLiu11 · 2026-10-26
- AI accusations shake literary prizes from Caine to Goncourt: who decides what's human? — TuhinChakr · 2026-10-26