RikiyaT/lm_understand-logs

What this is

This repository publishes the full logs behind the blog post Sparks of Weight Space Programming: LLMs can understand and write neural network weights. It is a research artifact rather than a library: for each of eight experiments it records what was sent to the model, what the model replied (including reasoning summaries), any code the model ran and the output it saw, the scoring used for the reported numbers, and the networks the model read or wrote (2).

The experiments cover weight-space tasks such as identifying which game a network was trained on, inferring what a diffusion model generates, reading a language model's learned language from its middle blocks, repairing a failing network by editing weights, writing small game networks from scratch, writing a 1.5M-weight Go network, probing sycophancy, and improving a language model's math performance (3). It suits researchers and engineers who want to audit or reproduce the claims, not users looking for a packaged tool.

How runs were produced

The model referred to as "Astra" in the post is OpenAI's gpt-6-astra, driven through the Codex CLI: each turn is one codex exec call (or codex exec resume <thread>), with the prompt on stdin and reasoning effort set per call and recorded in run.json (12). Built-in tools (shell, file editing, web search, image viewing, sub-agents, plugins) were disabled via the Codex configuration, and each call was audited so that a turn counts as tool-free only when the event stream contains just the model's messages, reasoning summaries and Codex's code-mode-disabled notice; all 1,125 turns in this release are marked tool-free (12).

In Experiments 3, 6, 7 and 8 the model could write one Python block per turn, executed in a bubblewrap sandbox with no network, a persistent working folder between runs and a per-run time limit; printed output was returned as the next message and logged in tool_calls.jsonl (12).

Working with the logs

Start from the root README's folder table, which maps each experiment directory to the post's experiment, the runs included and the tools allowed (3). Each experiment folder has its own README explaining which run corresponds to which tab or row and which file each reported number comes from.

For a given run, the key files are documented in the file table: prompt.txt and message.txt hold what was sent, events.jsonl the raw Codex event stream with token usage, tool_calls.jsonl the sandbox code and its printed output, run.json the model, effort, token count, wall time and audit flag, and state.json the conversation record. eval/ holds scoring with the answer key, and networks/ the networks the model read or wrote. Paths inside the logs are replaced by placeholders such as <data>, <repo> and <codex>, and Codex thread ids are replaced by id- plus eight hex digits, consistent across files within a conversation (16). There is no single install-and-run command in the materials; the reproduction path is to read an experiment folder's README, then inspect or rerun its recorded steps.

Constraints and omissions

Some artifacts are deliberately excluded (17): the 17 GB Gemma 3 270M fine-tunes from Experiment 7 (data, training script and hashes are provided for rebuilding), the model's intermediate files (activations, generated-text dumps, draft networks), the Experiment 3 Python corpus (replaced by a manifest and rebuild script) and the GPQA question texts in Experiment 8's generations (scores and per-question results are kept). Reproducing the full results therefore depends on rebuilding these pieces or sourcing the third-party benchmarks, which retain their own licenses.

Licensing is split: logs, data and networks are CC BY 4.0, the harness scripts are MIT, and the MinAtar source embedded in Experiment 5's prompts is GPL-3.0 (18). The LICENSE file is labeled NOASSERTION by GitHub, so check the per-content table before redistributing or reusing material (19).

HotRepo's take

The repository was recorded on HotRepo's new-project list on 2026-11-02, gaining 29 stars and 1 fork in a window from 21:45 to 22:45 UTC, reaching 138 stars and 13 forks. That is a single snapshot, so it shows an early burst of attention around the release rather than a sustained trend. For readers, the value is the audit trail: the combination of tool-free turn auditing, sandboxed code execution and per-number provenance makes the claims checkable at the file level, and the explicit list of omissions tells you where the logs are not self-contained.