Weight-Space Programming: what's measured vs anecdote in MIT CSAIL's GPT-6 claims
maier_ak · x · 2026-11-05
Andreas Maier audits MIT CSAIL's bold blog post claiming GPT-6 Astra can natively read, repair and write neural network weights without gradient descent. Credit where due: all 1,125 model turns are publicly logged and audited, the edited model is released as an exact diff, and a Hugging Face engineer rebuilt and re-ran the eval. He then sorts the eight experiments by evidence strength — hard results (thousands of trials, e.g. the Qwen3.5-0.8B weight edit lifting MATH-500 from 7.9% to 48.1%) vs soft ones (3-4 correct answers on toy tasks, failed sycophancy test, no baselines). Verdict: 'sparks' are real; replacing optimizers is not yet shown.
More from Models
- Dev Lets Opus Produce Mnemos MCP Announcement Video, Calls It Mind-Blowing — RileyRalmuto · 2026-11-05
- AMD ships Day 0 support for Kimi K3.1 Pro: 2.8T params, 1M context on 8x MI355X — Kimi_Moonshot · 2026-11-05
- Claude spontaneously finds a serious OpenBLAS bug while just explaining code to a user — Aizkmusic · 2026-11-05
- Baidu open-sources ERNIE 5.5 Flash: 280B-param MoE with 14B active and 1M context — teortaxesTex · 2026-11-05
- Did Anthropic reward Opus 5.5 for saying 'let me check rather than answer from memory'? — andrey_kurenkov · 2026-11-05
- New speech-to-speech models are natural talkers but still need hybridizing to finish tasks reliably — alexisgallagher · 2026-11-05