Qwen 3.8 27B on a laptop can read model weights, replicating the Weight-Space Programming paper
dl_weekly · x · 2026-11-12
Simon Willison replicated the viral "Sparks of Weight-Space Programming" paper locally: a 17GB quantized Qwen 3.8 27B on his laptop identified three of four policy networks from raw weights alone.
Key claims from the paper
- Coins "Weight-Space Programming": frontier models like GPT-6 Astra can natively interpret and edit raw neural network weights — a text description of floating-point numbers — without writing training loops.
- Astra writes small game-playing networks from scratch that outplay networks trained with PPO for 100M steps.
- Fixing Qwen3.5-0.8B's math: Astra located a weights issue causing reasoning loops and edited just 0.016% of weights, lifting MATH-500 from 7.9% to 48.1% and tripling IFEval, generalizing to instruction following.
Replication: Willison shows the capability isn't exclusive to closed frontier models — you can try it today on consumer hardware with a local Qwen 3.8 27B.
More from Models
- Can models learn heretical thinking? Researchers debate post-training away conformity — JeremyNguyenPhD · 2026-11-12
- Hassabis RTs tease that Gemini 4 carbon is 'awesome', results coming soon — demishassabis · 2026-11-12
- Aside's Computer Use agent claims SOTA on Scale AI's Cua Driver Bench — brandon_galang · 2026-11-12
- Inflect v3 ships six-voice local TTS in just 34M parameters, Apache 2.0 licensed — b111ue · 2026-11-12
- KOL predicts Gemini will be the top frontier LLM by year-end as rivals stumble — iruletheworldmo · 2026-11-12
- Dense rewards lose to standard GRPO at matched compute, arXiv paper finds — burny_tech · 2026-11-12