Skip to main content
Back

gpt-6-astra Edited a Rival AI's Weights and Boosted Its Math Scores

A researcher published the full logs behind a viral claim that a frontier model can read and edit neural network weights, drawing both interest and sharp skepticism.

Repo, Full Access
  • Rikiya Takehi released lm_understand-logs, full experiment traces behind the Sparks of Weight Space Programming post.
  • Experiments use OpenAI's gpt-6-astra via Codex CLI with all built-in tools disabled, across 1,125 audited turns.
  • Eight tasks cover reading weights (identify game, language, sycophancy) and writing them (fix, create, edit for math).
  • Headline result: Qwen3.5-0.8B MATH-500 rises from 7.9 to 14.1 after 25 Astra editing passes, no gradient descent.
  • Follow-up on Gemma 3 270M showed no meaningful improvement, undercutting the general-capability framing.
  • Critics argue Astra performed gradient-free search, exploiting known weight-perturbation effects rather than truly understanding weights.

Rikiya Takehi has published lm_understand-logs, the experimental record behind a claim that gpt-6-astra can inspect neural-network parameters and edit them to improve performance. The repository accompanies Takehi’s post, Sparks of Weight Space Programming: LLMs can understand and write neural network weights, and includes the prompts, model responses, code, outputs, evaluations, and checkpoints behind the reported results.

The experiments used gpt-6-astra through the Codex CLI. Takehi is an independent researcher, not an OpenAI researcher. The logs allow others to examine whether the results demonstrate an understanding of weight space, an effective gradient-free search procedure, or sensitivity in the models being edited.

Eight experiments, with every turn exposed

The release covers tasks ranging from checkpoint identification to direct parameter editing. A tensor is a multidimensional array that stores a network’s learned parameters, while a .pt file is a common PyTorch checkpoint format.

Experiments included in the repository
Experiments Task
1 through 3 Infer an Atari game, a diffusion model’s output domain, and a learned language from checkpoint weights.
4 through 6 Repair a damaged network with one edit, generate reinforcement-learning policies from scratch, and construct a 1.5-million-parameter Go network over 0, 10, 25, and 50 passes.
7 Determine which of two language models exhibits more sycophantic behavior.
8 Edit a 0.8-billion-parameter language model to improve math benchmark scores without gradient descent or conventional training.

How the sandboxed loop worked

Each turn launched a separate codex exec call with reasoning effort set to high, xhigh, or max, depending on the experiment. Every built-in Codex tool, including shell access, web search, file editing, and subagents, was disabled.