spaCy creator Matthew Honnibal calls MIT's Astra weight-editing research "seriously misleading"
量子位 · wechat · 2026-11-05
MIT's Astra went viral claiming it could fix Qwen3.5-0.8B's math score (7.9% → 48.1%) by modifying just 0.016% of its weights, without training or gradients, purely by "reading weights." spaCy creator and Explosion co-founder Matthew Honnibal published a lengthy critique arguing the paper's presentation is seriously misleading.
His core points:
- The paper actually bundles three distinct capabilities: reading weights (identifying tiny 17K–172K-parameter networks from raw floats in context), writing weights (generating small-network parameters from scratch to play Breakout), and fixing an LLM.
- Only the first two work from raw weights alone. The headline Qwen experiment let Astra write Python and run up to 50 forward passes, iteratively probing the model — an "off-field coach," not mind-reading.
- Packaging these as one unified superpower makes readers believe the small-network technique scales to 800M-parameter models.
Honnibal still credits the work: an AI acting as an ML engineer — writing code to diagnose and patch other networks — is a genuine breakthrough that models couldn't do a year or two ago. The problem is the hype-driven framing.
Related event: MIT's Weight-Space Programming Claim Meets Wave of Independent Scrutiny(13 posts)→
More from Models
- Reading OpenAI's cancel pages: what happens to Projects files and memory after downgrading is undocumented — Ok_Negotiation_2587 · 2026-11-05
- Baidu open-sources ERNIE 5.5 Flash Base: 280B MoE with 14B active params and 1M context — Baidu_Inc · 2026-11-05
- Hidden-instruction test: Kimi K3.1 obeys 0/15 while GLM-5.3 and GPT-6 hit 26/27 — teortaxesTex · 2026-11-05
- Google's unreleased Gemini 4 Carbon tops intelligence index at 60, prompting calls to ship it — Hesamation · 2026-11-05
- Baidu releases ERNIE 5.5 Flash Base model on Hugging Face — Xianbao_QIAN · 2026-11-05
- AA details Hy4 sub-benchmarks: 50% knowledge accuracy but 89% hallucination rate — ArtificialAnlys · 2026-11-05