AI model repairs weaker system by changing a fraction of its weights

An experiment at the Massachusetts Institute of Technology suggests that a powerful artificial intelligence model can improve a weaker model by adjusting a tiny share of its numerical weights, ERR’s R2 Portaal reports in a technology commentary by Kristjan Port.

The GPT-6 Astra model, developed by OpenAI, examined and rewrote the weights of Qwen3.5-0.8B, a smaller open AI model created by Alibaba. The smaller model had been getting stuck on some reasoning tasks. Astra changed 131,072 values out of about 800 million weights, or roughly 0.016% of the total. That was enough to address much of the problem and increase the smaller model’s performance on a mathematical test sixfold.

Astra did not teach the model new mathematics, Port writes. The necessary potential was already there, but a systematic fault in parts of the neural network was holding it back. The larger model identified and corrected much of that fault.

Port says the experiment is notable because a general-purpose AI model interpreted the internal numerical structure of another company’s model and deliberately rewrote it to achieve a functional goal. He compares the process with biological evolution, which can sometimes involve abrupt changes rather than only small, gradual ones.

The commentary links that idea to the philosopher Gottfried Leibniz’s principle that nature makes no leaps, and to Charles Darwin’s account of gradual evolution. But evolution can also involve horizontal gene transfer: bacteria can take up DNA from other organisms and acquire fully developed traits in a single evolutionary step. The environment and natural selection then shape what happens to those changes.