A 27B model learned to patch its own weights, lifting accuracy 57.2% to 63.8%
mark_k · x · 2026-11-21
A study makes recursive self-improvement more concrete: a 27B-parameter language model given access to its own weights, graded answers, and a way to test edits figured out how to patch itself.
- The best patch raised accuracy from 57.2% to 63.8% across 490 held-out problems, with no additional training run and no extra per-answer compute once the patch was in place.
- Limitation: every session still starts from the same base model, so better patches replace older ones rather than stacking.
- The author's proposed next step: give the patched model the same self-improvement job and see if it's better at improving itself.
More from AGI Musings
- Ramez Naam extrapolation: AI a decade away from completing 98% of hour-long human tasks — GaryMarcus · 2026-11-21
- Shermer doubles down: AI means protopia, not utopia or dystopia — intellectronica · 2026-11-21
- Hassabis predicted 2 years for AI to crack a Millennium Prize problem; 6 months later, both Riemann and Navier–Stokes have proposed solutions — Hesamation · 2026-11-21
- Terence Tao on AI math: 'strip mining' that wins proofs but teaches us nothing — yacineMTB · 2026-11-21
- Imagine P=NP: Verifiable Tasks Solved by Compute, Security in Tatters — gleech · 2026-11-21
- Anthropic claims Claude proved the Riemann hypothesis; Stanford posts all-human rival proof hours later — The Decoder · 2026-11-21