LLMs can understand and write neural network weights

This repository contains complete, auditable records for experiments testing whether a large language model can interpret and create neural-network weights. The eight experiments ask it to identify the games a network learned, infer what a diffusion model generates, recognize languages from a language model’s middle layers, repair a failing network by editing its weights, and build game-playing networks from scratch - including a 1.5-million-weight Go model. Further tests examine whether weights reveal a language model’s sycophancy and whether weight edits can improve mathematical performance. The repository includes the networks read or written, evaluation data, prompts, model replies and reasoning summaries, code and outputs where Python was permitted, and scores with their evaluation details.

The README emphasizes that the model had no built-in tools: all 1,125 recorded turns were audited as tool-free. In four experiments, it could write Python code, which researchers ran in a restricted, offline sandbox and reported back to the model. The logs document each run’s model, reasoning effort, token use, timing and conversation state, making the experiments reproducible and allowing readers to trace reported numbers to their evaluations. Some materials are omitted, including large fine-tuned model files and intermediate outputs; rebuild instructions or exact weight diffs are supplied where applicable. The collection is therefore both evidence for claims about language models’ ability to work directly with weights and a detailed record for scrutinizing how those results were obtained.