AI

MIT’s ‘AI Mechanic’ GPT-6 Astra Peeks Inside and Repairs Other Neural Networks

MIT’s GPT‑6 Astra can read and edit other AI models, boosting math performance and reducing bias without full retraining.

S
SciTech Editorial Team
Newsroom
··1 min read
MIT’s ‘AI Mechanic’ GPT-6 Astra Peeks Inside and Repairs Other Neural Networks
Researchers at the Massachusetts Institute of Technology (MIT) have unveiled an artificial‑intelligence system that can literally look inside the “black box” of other neural networks and rewrite their inner wiring. The model, dubbed GPT‑6 Astra, acts as an AI “mechanic,” interpreting raw activation values of peer models and then editing them to improve performance, a breakthrough reported in a preprint posted on 2 November. GPT‑6 Astra was trained through a meta‑learning regime that exposed it to thousands of activation maps from a variety of transformer‑based models. By learning to associate specific patterns of internal values with functional roles—such as token‑embedding, attention‑head selection, or output‑layer computation—the system can predict with roughly 92 % accuracy what a given component does inside an unfamiliar network. In practical tests, Astra was asked to enhance a language model’s mathematical reasoning. After a targeted edit of a handful of neurons, the model’s score on a standard arithmetic benchmark rose from 68 % to 84 %. A second experiment involved a vision model that exhibited gender bias; Astra identified and attenuated the responsible units, reducing the bias metric by 37 % without degrading overall accuracy. “The ability to diagnose and repair a neural network without rebuilding it from scratch could dramatically shorten development cycles,” said Dr. Lina Chen, lead author of the study. Industry analysts note that such a capability may also lower the carbon footprint of AI research, as fewer full‑scale training runs would be required. Nevertheless, the technology raises ethical and legal questions. If an AI can silently modify another AI, issues of accountability, intellectual‑property rights, and potential malicious tampering will need robust governance. The MIT team plans to extend Astra’s reach to larger foundation models and to collaborate with policy makers on responsible deployment frameworks.

Share

More from this section: AI