AI That Rewrites Its Own Code Just Got Cheap Enough for Anyone

AI models are increasingly being used as "agents" — AI that can take actions and do work on its own — to design the step-by-step recipes computers use to solve problems. The usual approach is a bit like a student cramming for a test: the AI tries ideas, keeps the best one in its short-term memory, and starts over. That works for a while, then stalls, because the AI never truly absorbs what it learned. This new method, from a team of Chinese researchers, lets the AI permanently learn from its own best work, storing successful attempts in a growing library it can draw on later.

The results are striking. Trained on only four chip-design examples, the system outperformed two well-known rival methods across 16 different chip cases. Even more surprising: it ran on a small, relatively cheap AI model (8 billion parameters, roughly the size of models you can run on a good laptop) and still matched frontier closed-source models like GPT-5.5. Because it internalizes knowledge rather than re-reading long prompts each time, it also uses fewer "tokens" — the units AI companies charge you by.

The practical payoff showed up in GPU kernels, the specialized code that tells graphics chips how to run calculations. On four such designs, the new method made programs run an average of 8.27 times faster than a standard baseline. In everyday terms, a task that once took an hour could finish in about seven minutes — less waiting, lower electricity use, smaller cloud bills.

The catch: this is a preprint, meaning it has not yet been checked by independent experts, and it was tested only on narrow, highly technical tasks like chip and graphics-code design. It is not a general-purpose "smarter AI" you can chat with. Still, it hints at a future where AI quietly improves the software and hardware underneath everything — which is both the promise and the part worth watching closely.