Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
Web page · https://rikiyat.github.io/blog/weight-space-programming
First seen by Oriel on .
AI summary
The post introduces Weight-Space Programming, showing GPT-6 Astra can infer neural-network behavior and directly edit or create weights for specified tasks without gradient descent or training loops.
- Astra identified the game trained into four identically structured PPO policies from 17,253 raw weights, using patterns such as input-channel and action-layer structure.
- From middle transformer blocks alone, Astra correctly identified Python, romaji, or English training data and matched all copying heads measured by running the models.
- Astra edited 3.5% of a 171,664-weight Kinetix policy, making a car reverse before accelerating and succeed in all 1,024 episodes after previous failure.
- In the reported Qwen3.5-0.8B experiment, Astra edited 0.016% of weights without training, raising MATH-500 from 7.9% to 48.1% and improving IFEval threefold.
- The diffusion-model experiment showed structural understanding but also a limitation: Astra identified circles, squares, and triangles while interpreting generated filled shapes as outlines.
#weight-space programming #llm agents #neural networks #mechanistic interpretability
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 2,935, latest observation ; +2,704 over observed period since
- Lobsters: 51, latest observation ; +46 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 889, latest observation ; +841 over observed period since
- Lobsters: 16, latest observation ; +15 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50341174 | https://rikiyat.github.io/blog/weight-space-programming/ Read discussion |
| Lobsters | u7wdqj | https://rikiyat.github.io/blog/weight-space-programming/ Read discussion |