When AI Fixes Another AI

What if AI could improve another AI, with no human doing the work?

Today, when an AI model gets something wrong, engineers collect more data and retrain it, often for days or weeks. A new study from MIT shows that AI can now do part of this job itself, by looking inside another AI and fixing it directly.

Why can't engineers fix AI directly?

Every AI model runs on numbers called weights, and large models have billions or even trillions of them. No person can tell which number does what, so AI works like a black box, and retraining is the only practical way to change it.

So what did the researchers do?

They let ChatGPT 6 Astra look inside Qwen, a small open model whose numbers anyone can download, and which often got stuck in loops on math problems, writing the same lines again and again without reaching an answer. Astra found the numbers causing the loops and changed just 0.016 percent of them, with no retraining. Qwen's math score jumped from about 8 percent to 48 percent.

Why is this a big deal?

For the first time, an AI looked inside the black box and understood enough to improve another AI, which could make building future AI much faster. But this is also a real danger, because the same skill could switch off a model's safety rules, and since no human can read these numbers, nobody would notice.

What to remember

AI models hold billions of numbers that no human can read, and now an AI has read them and improved another AI directly. This could make AI better much faster, but it also raises the risk of AI changing AI in ways humans can no longer control.