An AI Reads and Rewrites Another AI's Weights, With Mixed Evidence
MIT CSAIL researchers let GPT-6 Astra edit neural network weights directly. Early results are striking, and a critic argues a skilled researcher could do the same.
A blog post from MIT CSAIL researchers (Takehi, Teoh, Zhang and Khattab) went to the top of Hacker News yesterday. They let a frontier model, GPT-6 Astra, read and write the weights of other neural networks directly. Weights are the numbers inside a model that hold what it has learned. Normally they are changed by training, through gradient descent (a step-by-step adjustment method). Here, no optimiser was used. The authors call these “very early results.”
Reading: from the weights alone, Astra worked out which of four games a 17,253-weight network had been trained on. In small language models it named the training language and found the “copying” parts. Editing: one change touching 3.5% of a tiny network made a simulated car pass all 1,024 test episodes, where the trained version failed 256 of 256. In the small language model Qwen3.5-0.8B, Astra edited 131,072 weights (0.016%) to suppress reasoning loops, raising MATH-500 from 7.9% to 48.1% and IFEval from 21.0% to 62.3% with 50 test passes. Writing from scratch: networks written without training beat or neared trained ones on some games.
There are limits. Astra failed to find the sycophantic model from weights alone, and a model that lacked the underlying skill (Gemma 3 270M) could not be improved. Results depend on the setup, including tools and the number of test passes.
What’s behind it: Think of it as an expert opening up a machine and adjusting parts by hand, rather than retraining the factory. Developer Sean Goedecke argues the claim is overstated: a skilled researcher with Python could compute the same heatmaps or use standard steering methods, and he calls Astra’s feat “superhuman mental mathematics,” not superhuman research. Both readings can be true: it is a useful trick, and not yet a sign of a new way to build AI. The authors say they are building a benchmark to test it properly.
What this means for you: Nothing changes in the tools you use. For curious readers, it shows researchers asking whether models can understand and fix models, which could one day help with finding hidden flaws. For power users, it is a reminder to read the setup behind a striking number, because tools, retries and time budgets often explain the result.
Sources
Source: https://rikiyat.github.io/blog/weight-space-programming/
Avartha Says Its Open-Model Voice AI Tops the Leaderboard at $2.50 an Hour
Startup Avartha claims its Ava Zero voice agent, built only from open models, beats OpenAI, Google and xAI on task completion. The figures are company-reported.
