AI models editing other AI weights: what MIT's test showed, and what it didn't

Researchers at MIT's CSAIL say frontier models can read and write the weights of other neural networks. The authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, and their post is titled Sparks of Weight-Space Programming. The claim is narrow and concrete: a model gets raw numbers from inside another network, works out what is wrong, and edits a small share of them.

In the headline experiment, GPT-6 Astra took a model scoring under 8% on a hard maths test to over 48%. No retraining, no gradient descent. The team calls the approach weight-space programming.

Nobody writes the weights by hand. They emerge from training, and they sit packed so densely that working out what any of them does is slow, painful work. There is a whole field for this, mechanistic interpretability, and progress has been slow.

Chris Olah, one of its pioneers, framed the problem plainly: we end up with an artifact that writes essays and translates, and we have no idea how to write a program that does the same. We grew it, we didn't build it. That is why a model that can read those numbers directly is a different kind of claim.

In one test, a small AI drove a car in a physics simulator and kept getting stuck on a ramp. It had trained for a billion steps and failed all 256 attempts. GPT-6 Astra got the weights, a scene description and one failed recording, then wrote a patch changing about 3.5% of the weights. Success went from 0% to 100% over more than a thousand attempts.

In another test it wrote a Breakout-playing network from scratch in one reply, scoring an average of 517 against about 25 for a network trained for 500 million steps. It was not a clean sweep: on Seaquest it scored 53 against almost 89, and on Asterix the two were roughly level.

A replication with Claude Opus 5.5 took 27 minutes and about $1 of API usage, with no human in the loop. Someone going by stochi got similar results with OpenAI's open-weight GPT-OSS-120B on a Lunar Lander task, lifting success from 0% to 88%.

Not every model manages it. In Xingyu Dang's Writing Weights Index, GPT-6 Astra led with 60.7, Gemini 4 Carbon followed at 57.7 and Claude Opus 5.5 scored 51.2. Most open models lagged badly: Kimi k3.1-Pro under 30, Qwen 12.5, DeepSeek 10.

There is a fair criticism worth taking seriously. Researcher Ivan Skorokhodov said he was skeptical at first, then concluded the framing is misleading. His argument: the researchers told GPT-6 Astra not to use gradient descent, so it did gradient-free training instead. In some experiments the model could run the network, see results, adjust weights and try again.

That is still interesting, but it is closer to search than to a model understanding a network. The video's own take is that this will not replace pre-training soon. It could become a tool on top of it.

Two consequences stand out. First, if one model edits another's weights, checking what changed and why gets harder. Simon Willison reportedly found the reasoning traces almost incomprehensible. Yuchen Jin predicted weights could become a language AI understands and humans don't.

Second, cost. A fix that takes under half an hour and about a dollar is a different proposition from fine-tuning that runs for hours or days. But the index gap matters: if only top frontier models do this well, the labs with the strongest models pull further ahead. The researchers say they are building a benchmark and tooling to measure this properly.