← Latest AI News

November 4, 2026 · Daily brief

ChatGPT can fix broken AI without retraining it

Sovereignty angle
If ChatGPT can repair a model by reading its guts, you could soon ask it to fix the tool you paid for—or check whether the vendor already did and charged you for an upgrade.

MIT researchers showed ChatGPT editing a cheap AI model's math score from 8% to 48% by rewriting 0.016% of its code—no training, one try.

What changed

MIT researchers published work showing that OpenAI's GPT-6 Astra can open up another AI model, read the numbers that make it work, find what's broken, and rewrite those numbers to fix it. Alibaba's Qwen model scored 7.9% on a math test. Astra changed 131,072 of its 800 million internal weights—just 0.016%—and the score jumped to 48.1%. No training. No gradients. One attempt.

Astra also wrote a working Breakout game from scratch, with no training at all, just by writing the numbers directly.

What you can do with it now

This is early research, not a product you can use Monday morning. But the direction matters. If a general model can read, diagnose, and fix a specialist tool just by looking at its code, that changes the repair game. Right now, when your transcription AI starts choking on accents or your scheduling bot keeps double-booking, you file a ticket and wait. If this approach scales, you could point a model at the broken tool, describe the problem, and get a working fix without waiting for the vendor's next release.

It also means you could audit a tool before you buy it—ask a model to read the weights and tell you what it actually does, not what the sales deck says.

What to keep in your own hands

Don't hand Astra—or any model—the keys to a tool that touches customer data, invoices, or anything regulated until you understand exactly what it changed and can roll it back. Changing 0.016% of a model's weights can fix a math loop or introduce a new failure you won't see until it's in production. Test every edit in a sandbox with your own data before you trust it with the real work. And keep a copy of the original: if the model breaks something, you need to be able to walk back to the version that worked.