Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or con...
AI Research
Summary
Simon Willison demonstrates that Qwen 3.8 27B running locally on his laptop can interpret raw neural network weights, replicating findings from the 'Sparks of Weight-Space Programming' paper showing frontier LLMs can read and modify model weights as floating-point text without executing code.
Highlights
- Qwen 3.8 27B running locally in LM Studio correctly identified a game from raw weights after 46 minutes and 43,506 reasoning tokens
- The paper showed GPT-6 Astra could edit just 0.016% of Qwen3.5-0.8B weights to raise MATH-500 from 7.9% to 48.1% without gradients
- The term 'Weight-Space Programming' was coined for LLMs analyzing and modifying model weights purely through reasoning
- Critics including spaCy's Matthew Honnibal argued the paper misleadingly conflates no-tools weight interpretation with code-assisted model editing
- The capability works without any environment feedback or training loops—pure reasoning over floating-point text
auto-generated
Context
Audience
AI researchers, ML practitioners, and developers interested in local LLM capabilities and model interpretability
DomainMachine Learning
Formatblog post
Accessfree online
Source Overview
Sparks of Weight-Space Programming demonstrated that frontier LLMs can interpret and modify raw neural network weights from floating-point text alone, without executing code.
Simon Willison ran the paper's game-identification prompts on Qwen 3.8 27B locally in LM Studio; it correctly named the game after 46 minutes and 43,506 reasoning tokens. The paper also reported GPT-6 Astra editing 0.016% of Qwen3.5-0.8B weights to raise its MATH-500 score from 7.9% to 48.1%.
Topics
Related
Sparks of Weight-Space Programming paperQwen 3.8 27BLM StudioGPT-6 AstraLLM Interpretability
