e tried weight-space programming (Takehi, Teoh, Zhang & Khattab, MIT CSAIL) on a small production model: Tesseract’s Hebrew text recognizer, 378K weights.
An AI agent (Claude Opus 5.5) edited the weights directly, with no gradients, three times:
The fixes were explainable: the agent found two start-up faults in the bidirectional LSTM, one of which swallowed numbers at the start of a Hebrew line.
Fine-tuning won this round, but we still find the result impressive: with no gradients and no training loop, an agent cut the errors by more than a third and could explain every change it made. This was a small model, and larger ones may behave differently. As agents get better, we believe weight-space editing will become a far more valuable tool.
Write-up with the negative results: [Weight-Space Programming: An AI Agent vs Fine-Tuning — Rivok Labs](https://rivoklabs.com/research/weight-space-programming?ref=hf-forum)
The fine-tuned Hebrew models we ship: [Rivok/tesseract-heb-rivok · Hugging Face](https://huggingface.co/Rivok/tesseract-heb-rivok)
Happy to answer questions about the setup.