LLM named as co-author on weight-space programming study matching gradient optimization in 30/87 settings
lateinteraction · x · 2026-11-08
Francois Chaubard published a short write-up of a 'weight-space programming' (WSP) study with an AI collaborator — 'GPT-6 Astra' — listed as a named co-author, amplified by lateinteraction.
- Method: compares WSP (an LLM directly proposes or iteratively revises numerical network weights) vs. gradient optimization across 10 algorithmic tasks and 29 nonconvex task-architecture pairs
- Findings: both examples-only WSP procedures match every extrapolation reference in 30 of 87 settings; iterative WSP ranks first among 14 pipelines by summed extrapolation RMSE; examples-only error drops from 1.06×10¹² to 2.03×10⁹, mostly via Fibonacci improvements
- Highlights: one-shot WSP is exact on every extrapolation input for the square task (zero stored coefficients); sine attains RMSE 7.79×10⁻¹⁶
- The paper reports coefficient storage, tensor memory and arithmetic costs, plus reproduced prompts and construction explanations
Related event: LLMs directly edit model weights, matching gradient-based training(2 posts)→
More from AGI Musings
- Why We Need a "Human Reserve": Redundancy as National Resilience in the AI Era — J_MaestreVidal · 2026-11-08
- Podcast: why the left dismissed AI as hype — stochastic parrots and a comforting story — GarrisonLovely · 2026-11-08
- Reviewers reward flaw-free work, users reward useful work — why peer review punishes innovation — wzenus · 2026-11-08
- Ex-OpenAI researcher: I spent 1.5 years unlearning downside protection over upside creation — emilyzsh · 2026-11-08
- Math community debates Glazer's proof and OpenAI's H10 paper: slop need not compound — littmath · 2026-11-08
- Intelligence explosion debate: crudest model says ~1 year, insiders diverge from outsiders — sebkrier · 2026-11-08