Explainer video · about 2 minutes · captioned

Weight-space programming, in two minutes

Animated, captioned explainer in eight scenes. Title card; a grid of numbers showing that a network is a list of weights; three cards: Read, Fix, Write; the headline result, 0.016 percent of weights edited and MATH-500 rising from 7.9 to 48.1 percent; a reality check with three limitations; Claude Fable 5.2's restricted skills and its MNIST weight-writing score; a table of how other labs are responding; and the bottom line. The full transcript is below. — The animation plays once, with on-screen captions, and stops on the final card. Replay · The full transcript is below.

Transcript

0:00Weight-space programming: AI that reads and writes the inside of other AI models. Here is what happened this month, and what it means.

0:09Every neural network is a long list of numbers, called weights. Until now, those numbers have been set by training algorithms, like gradient descent, not by reasoning.

0:19On November 2, researchers at MIT showed that GPT-6 Astra can read those numbers, and tell what a small network was trained to do. It can fix a broken network, and even write working game-playing networks from scratch.

0:33The headline: by editing just 0.016% of a small language model's weights, with no training, its math score jumped from 7.9% to 48.1%.

0:44But that edit fixed a defect. The model kept getting stuck in loops. On a model without that defect, editing added nothing. Independent teams found that 6 of 16 editing attempts produced nothing worth keeping, and that reading weights often leans on memorized guesses.

0:59On November 9, Anthropic released Claude Fable 5.2. During training, it deliberately held two AI research skills, accelerator kernels and model architecture design, at the level of its previous model. Weight editing is not on that list. In an independent test, Fable 5.2 was the best Claude model yet at writing weights.

1:20Will other labs follow? Not soon. OpenAI is building toward an automated AI researcher by 2028. Google DeepMind reports no AI research threshold crossed. The White House Accord covers monitoring and audits, not capability limits. And open-weight models cannot be capped at all.

1:38The bottom line: weight-space programming is real, early, and narrow. Fable 5.2's limits are a meaningful signal, but only a small brake, unless others follow. Read the full white paper at news.rvacyber.com.

Read the full white paper, with sources