Source: video narration supplied by the operator, 06/11/2026. Text reproduced verbatim; only paragraph breaks have been added.
Something big just happened in AI research. Researchers at MIT showed that an AI can read the weights of another neural network and rewrite them, not train them, write them like code.
Every AI model is a giant grid of numbers called weights. Everything it knows and every mistake it makes lives in those numbers. Until now, there's been one way to write them: training. Show the model examples, measure how wrong it is, and nudge every single weight a little. Repeat billions of times. Training never understands the numbers, neither can we. So if a model is broken, the only fix is more training.
Weight space programming changes that. You hand the raw weights to a frontier AI and ask it to read them like code. Training changes everything a little. Weight space programming changes a few weights on purpose. It works out what each part of the network does, finds the part causing the problem, and rewrites just that part.
Here's the clearest example: Alibaba's small Qwen model kept looping on math problems. Most answers never finished. Researchers gave its weights to GPT six Astra. It traced the loop to two layers and removed it. No retraining, just zero point zero one six percent of the weights. Math scores went from eight percent to forty eight percent.
To be fair, it's early. It removed a problem that was hiding a skill. It can't add a skill the model never had.
But step back for a second: an AI that reads another AI finds the bug and fixes it. The weights aren't just trained anymore. They're programmed. And once AI can program AI, that loop gets tight fast.