Can AI Already Edit Its Own Weights?

Alexander Wissner-Gross is more than 90 percent confident that AI is already editing its own weights. A relatively small group at MIT can do related work, so frontier labs are already preparing and playing with self-editing models. The capability is hugely positive, and a moral panic at every new capability will not get civilization very far.

A model can directly edit the parameters of another model, or potentially itself, in a semantically useful way. Recursive self-improvement will not run forever. The field will discover what a perfect model looks like in practice, the race will come to a halt, and the work will move elsewhere. Self-editing models are not the end of the world. They open a door to letting many groups contribute intelligence as modules.

A team at Moderna holds biotech knowledge, but the only path to market would be retraining an entire foundation model from scratch for over $100 million. A module built for one task could attach to a bare-bones model of pure intelligence with no specific knowledge, probably a 40 billion parameter model or even smaller, and then reach the market. The result stays transparent, so the AI police can look into the neural parameters and confirm the model is not doing anything evil. Everybody is involved, everybody benefits, and the model cannot go rogue.