Weight-space programming's soft claims rest on 3-4 correct answers on toy tasks
maier_ak · x · 2026-11-05
Andreas Maier's critique of the soft side of MIT CSAIL's GPT-6 weight-editing claims: 'reading' what a network does rests on 4, 3 and 3 correct answers on toy tasks built to be readable; the sycophancy experiment failed from weights alone; one model, mostly one run per condition, no simple baselines. 'Sparks' fits; replacing optimizers is not yet shown.
Related event: Researcher Scrutinizes MIT's GPT-6 Weight-Space Programming Claims(2 posts)→
More from Research
- Agents ace isolated tasks but drop to ~38% at maintaining codebases over time — FullStackFrenzy · 2026-11-05
- RoboColosseum founder: robotics lags AI because it lacks real-world benchmarks — DJiafei · 2026-11-05
- ByteDance study finds on-policy distillation leaves models 2.45-16.25% below base on out-of-distribution tasks — ByteDance · 2026-11-05
- FrontierPhysics: 56 expert tasks benchmark AI agents on end-to-end physics research — Bingran You · 2026-11-05
- VideoMemBench: new benchmark exposes major gaps in MLLMs' in-context video memory — Zhengdong Hu · 2026-11-05
- Muon-RT swaps momentum for teacher trajectory, cutting memory and beating AdamW by 2.6 points — IITK · 2026-11-05