GitHub radar
How an AI reads and edits another network's weights without training
A researcher published full logs of his experiments: a language model like ChatGPT looks at the raw numbers inside a tiny neural network and figures out what it does — and in one case rewrites those numbers by hand to make it perform better, with no training at all.
An open repository of text logs from 8 experiments: an OpenAI model guessed what small neural networks do just by reading their raw weights (the numbers inside the network), and hand-edited those numbers.
When it helps
Useful if you're curious how AI models work under the hood and what can be done to them without training. Not needed if you just use Claude Code for everyday work — it changes nothing there.
Pros
- Logs are fully public — you can verify every step yourself
- Score on MATH-500 rose from 7.9% to 48.1% with no training
- Nothing to install — it's just text files already in the repo
Cons
- It's a research archive, not a ready-to-use tool
- One weight-editing session took 8-20 hours
- Scaling the method to larger models is still unproven
How to set it up — step by step
- 1Open Claude Code in your project folder (an empty folder works too).
- 2Send the agent the text from the block below.
- 3Let the agent read the README and the exp8_math folder and work through the logs.
- 4Read the agent's summary of how the model edited the math network's weights and what it changed on the tests.
- 5If you want, ask the agent to walk through the logs for the other experiments listed in the README table.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the RikiyaT/lm_understand-logs repository, read the README and the exp8_math folder in full, and tell me in your own words how the model edited the weights of the small math model and what that did to its benchmark scores — don't run any code, just read the text logs.
▌ More finds