Jacob Strauss gave Claude Opus 5.5 six small neural networks, told it to fill in the numbers that define each one by hand, and found that it matched or came close to normal training on five of them. He published the results on his blog on 17 November, after reading an MIT CSAIL post from two weeks earlier, “Sparks of Weight-Space Programming.” That post reported that GPT-6 Astra could identify a network’s task from its raw weights, write a game-playing policy network’s weights from scratch, and raise the math skills of Qwen3.5-0.8B by rewriting some of its weights by hand.

A quick definition helps. Weights are the stored numbers inside a neural network, and they decide how it responds to any input. Normally a training loop (gradient descent) adjusts them gradually over many passes through data. MIT calls the alternative Weight-Space Programming: a language model reads and writes those numbers itself, with no training loop.

Strauss built a harness to test it. The agent works in a Docker container with no internet and only Python and NumPy. It can see about a dozen training examples, save and reset snapshots of its weights, and score itself on a validation set a limited number of times. Each run is capped at 50 to 100 turns, and a second LLM reviewed the logs for rule-breaking and found none. These are one person’s experiments, not a controlled study, and he chose both the tasks and the models.

The results varied by task:

The Shakespeare task, a character-level LSTM that predicts the next letter, was the failure. Written from zeros, the weights reached about 34% accuracy. Strauss then trained the network briefly to 39.9%, short of the 51.3% it would reach with more training, and asked the agent to repair it. After 36 minutes it got to 40.5%, about what half an epoch more training would have bought. Claude Fable 5.2 tried a more elaborate plan and landed in almost the same place.

The useful distinction is between design and tuning. In each success, the agent understood the data and chose a structure, such as stroke templates or per-bracket counters. It was weaker at picking exact values, which is the part gradient descent handles well. Strauss concludes that gradient descent remains cheaper, faster and usually more accurate. His other point is interpretability: for small models, you can point at a unit and explain why it exists, which a trained network rarely allows.

He also floats a narrower use, with hand-written weights serving as a starting point in place of random ones. He did not test that. Teams who want auditable small models could try the harness, which Strauss published on GitHub, against their own classifier before trusting any claim that this replaces training.

Jacob Strauss, “Experiments in Weight-Space Programming,” jacobstrauss.bearblog.dev, 17 November 2026.