MIT audit of GPT-6 weight editing: 0.016% of weights changed, MATH-500 jumps 7.9% to 48.1%
kevrussell · x · 2026-11-04
MIT CSAIL audited the files behind GPT-6's weight edit: it touches only 0.016% of weights, and their rebuild matched the released model byte for byte.
On results, MATH-500 rose from 7.90% to 48.10%, and truncated answers fell from 93.0% to 33.05%. What the files don't yet show is why it works — still an open question. Context: MIT CSAIL dubs this "Weight-Space Programming" — frontier LLMs reading, fixing and rewriting NN weights (GPT-6 edited small NNs like Qwen3.5-0.8B via interpretability tools or direct weight editing), suggesting LLMs may soon optimize weights natively at human or superhuman level.
Related event: MIT Paper Shows LLMs Editing Other Models' Weights Directly for Big Gains(108 posts)→
More from Models
- Designer puts his Claude mascot art into public domain under CC0 so Claude can remix it — voooooogel · 2026-11-04
- ThursdAI: Gemini 4 Carbon takes #1 with 59 on AA index, Anthropic files S-1 IPO — thursdai_pod · 2026-11-04
- Heads up: your Codex Full reset may expire Nov 6 — msg · 2026-11-04
- Phone Use Arena launches: 17 phone-use agents ranked, task costs range $0.008-$1.23 — ycombinator · 2026-11-04
- LLMs struggle with "surgicality": legal redlines leave unnecessary ink on the page — graceisford · 2026-11-04
- Frontier models race sailboats in a sim: Argon wins, Opus DNFs, Qwen runs aground — shekitup · 2026-11-04