MIT team reports an artificial intelligence system that reads the internal workings of other neural networks

13 November 2026

A team at the Massachusetts Institute of Technology, led by co-author Rikiya Takehi, has reported that an artificial intelligence model can read the raw internal values of other neural networks and predict what different parts of those systems do, and can then edit them to improve their performance, including their mathematical ability. The work uses GPT-6 Astra, a model developed by OpenAI at San Francisco, and was posted as a preprint on the alphaXiv service, which means it has not yet been through peer review. The field this belongs to is interpretability, the study of what is happening inside a trained model rather than only what it outputs. Researchers quoted in the report caution that the system does not necessarily understand how a given network works, and Fazl Barez, an artificial intelligence safety researcher at Oxford, said he would be surprised if it could interpret models beyond existing human abilities.