Weight-Space Programming: One AI Reads and Edits the Numbers in Other AI Models

MIT researchers report that OpenAI’s GPT-6 Astra can read, fix and write the weights of small AI networks with no training. Editing 0.016% of one small model’s weights raised its MATH-500 score from 7.9% to 48.1%, mainly by ending repetition loops.
artificial-intelligence
Author

Kabui, Charles

Published

2026-11-09

Keywords

weight-space-programming, gpt-6-astra, mechanistic-interpretability, model-editing, reasoning-loops

Download as PDF

MIT researchers report that an AI model can read, fix and write the numbers inside other AI networks without any training. Those numbers, called weights, decide what a network does. In a November 2 blog post, OpenAI’s GPT-6 Astra got the weights of small networks as plain text, with no tools. It named the game each of four networks played. For a simulated car stuck against a ramp, it changed 3.5% of the weights, and the car then succeeded in every test run. From an all-zero network it wrote a Breakout player that scored 517.4, against 25.3 for one trained on 500 million game steps, though it trailed in Seaquest.

With code and up to 50 test runs, Astra edited Qwen3.5-0.8B, a small open model its makers say is prone to thinking loops, where it repeats itself until its answer is cut off. Astra removed the loop pattern from 0.016% of the weights, lifting the score on MATH-500 (500 maths problems) from 7.9% to 48.1%. Neel Nanda, a researcher in mechanistic interpretability (tracing what parts of a network do), wrote that Astra is “surprisingly capable of doing old school mechanistic interpretability research rather than anything spooky”. On Gemma 3 270M, which rarely loops, no edit moved the score by more than about 1%.

The Hugging Face account akhaliq rebuilt the edit and scored 51.0% on MATH-500, though 31.6% of answers were still cut off. A separate team’s three Astra searches on the same model gained 5%, 27.3% and 29.45%.

Read More: Antidoom, Liquid AI’s training fix for reasoning models stuck repeating text and Qwenputer, a project where a compiler, not training, sets every weight of layers added to a Qwen model

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Weight-Space {Programming:} {One} {AI} {Reads} and {Edits}
    the {Numbers} in {Other} {AI} {Models}},
  date = {2026-11-09},
  url = {https://toknow.ai/posts/weight-space-programming-llm-reads-edits-network-weights/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Weight-Space Programming: One AI Reads and Edits the Numbers in Other AI Models.” https://toknow.ai/posts/weight-space-programming-llm-reads-edits-network-weights/.