What is this?
The complete logs of every experiment in the blog post 'Sparks of Weight Space Programming: LLMs can understand and write neural network weights'. It has the prompts sent to the model, its replies, the code it ran, the evaluation behind each number, and the networks it read or wrote.
The complete logs of every experiment in the blog post 'Sparks of Weight Space Programming: LLMs can understand and write neural network weights'. It has the prompts sent to the model, its replies, the code it ran, the evaluation behind each number, and the networks it read or wrote.Is it a tool or library, or just data?It's a data release, not a program. There is no package to install; it's files with prompts, model replies, sandbox code and output, scoring and networks.
It's a data release, not a program. There is no package to install; it's files with prompts, model replies, sandbox code and output, scoring and networks.How do I find the run I care about?Open the folder of the experiment you want, such as exp4fixcar/, and read its README first. It maps each run to the tab or row in the post and names the file each number comes from.
Open the folder of the experiment you want, such as exp4fixcar/, and read its README first. It maps each run to the tab or row in the post and names the file each number comes from.Which model and tools were used?'Astra' is OpenAI's gpt-6-astra, run through the Codex CLI with one turn per call and all built-in tools switched off. Code was run only in Experiments 3, 6, 7 and 8, inside the author's own sandbox with no network.
'Astra' is OpenAI's gpt-6-astra, run through the Codex CLI with one turn per call and all built-in tools switched off. Code was run only in Experiments 3, 6, 7 and 8, inside the author's own sandbox with no network.What is missing from the release?Not included: the 17 GB Gemma 3 270M fine-tunes from Experiment 7, the intermediate files the model wrote while working, the text of Experiment 3's Python corpus, and the GPQA question texts in Experiment 8. Rebuild scripts and hashes are provided for some of these.
Not included: the 17 GB Gemma 3 270M fine-tunes from Experiment 7, the intermediate files the model wrote while working, the text of Experiment 3's Python corpus, and the GPQA question texts in Experiment 8. Rebuild scripts and hashes are provided for some of these.Does the model use tools during these experiments?Not built-in Codex tools: all 1,125 turns were audited as tool-free. Code execution happened only in a custom Python sandbox in Experiments 3, 6, 7 and 8.
Not built-in Codex tools: all 1,125 turns were audited as tool-free. Code execution happened only in a custom Python sandbox in Experiments 3, 6, 7 and 8.What is the weight-space programming idea?The experiments test whether an LLM can read a neural network's weights and then edit or write them. Examples include fixing a failing network in Experiment 4 and writing a 1.5M-weight Go network in Experiment 6.
The experiments test whether an LLM can read a neural network's weights and then edit or write them. Examples include fixing a failing network in Experiment 4 and writing a 1.5M-weight Go network in Experiment 6.Can the model write game-playing networks from scratch?Experiment 5 asks the model to write networks from scratch for Breakout, Seaquest and Asterix in MinAtar, with no tools. The logs show each attempt and its evaluation.
Experiment 5 asks the model to write networks from scratch for Breakout, Seaquest and Asterix in MinAtar, with no tools. The logs show each attempt and its evaluation.Which topics is it in?interpretability, reproducible-research, neural-networks, mit-license, diffusion-models, dataset, gpt, reinforcement-learning, evaluation, fine-tuning, sandbox, transformer
interpretability, reproducible-research, neural-networks, mit-license, diffusion-models, dataset, gpt, reinforcement-learning, evaluation, fine-tuning, sandbox, transformer