Experiment shows LLMs 'writing' neural net weights revert to pre-deep-learning tricks
A blog post describes testing 'Weight-Space Programming' by having Claude Fable 5.2 manually set all 5,994 weights of a fixed CNN for MNIST digit classification, without fitting, regression, or search. The author found the model relied on hand-crafted filters and templates reminiscent of 1990s computer vision techniques, and observed similar patterns when re-examining a Go-playing network that GPT-6 Astra built in an earlier experiment.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest that when large language models are asked to directly author neural network weights rather than train them via gradient descent, they may default to older, hand-engineered heuristics rather than discovering genuinely new training paradigms. This could indicate that so-called 'LLM Descent' is less a novel training method and more a sophisticated form of manual feature engineering, according to the author's analysis.
- Claude Fable 5.2 manually set all 5,994 weights of an MNIST classifier without fitting or search, per the blog post's rules.
- The resulting weights resembled pre-deep-learning, hand-crafted filter and template techniques from the 1990s.
- A similar hand-crafted pattern was found on re-examination of a Go-playing network built by GPT-6 Astra in an earlier experiment.
Source: notesbylex.com — Lex Toumbourou, 2026-11-17
Published there as: “Weight-Space Programming builds neural networks like a 1990s researcher”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.