Paper says GPT-6 Astra can read and edit neural network weights without tools

A research paper from MIT CSAIL, authored by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, reports that OpenAI's GPT-6 Astra can read the raw weights of a small neural network and say what it was trained to do. It can also edit the weights of a language model to fix a failure mode. The authors call the approach Weight-Space Programming.

The no-tools test comes first. The authors trained four small networks with PPO on simplified 10x10 grid versions of Breakout, Pong, Pac-Man and Freeway. Each network has 17,253 weights and the same shape, so the architecture gives nothing away. Astra was given the weights as floating-point text, a description of the grid and its actions, and an instruction not to use tools. It named the right game in all four cases. Simon Willison, who ran the Pac-Man prompt on Qwen 3.8 27B locally through LM Studio, found the model also identified Pac-Man, Pong and Breakout, but failed on Freeway, taking 46 minutes to 65 minutes per run. Its Freeway transcript guessed a Pac-Man-like game. Willison noted that Kimi K3.1 Pro, run through OpenRouter, got Breakout right after 9m 41s.

The reasoning behind the no-tools result is where critics focus. Sean Goedecke argues that the input layer is simple to read. Each of the five input channels, such as agent, moving objects and reward targets, maps directly onto the 10x10 grid, so the first-layer weights for each channel show where that channel matters. He computed the L2 norm of those weights across the grid in three lines of NumPy. The result was a heatmap in which the player's position and the layout of objects are visible, so a person looking at it could plausibly identify the games. In his view this is 'neat' but not superhuman. The paper's own description says Astra worked out where each channel's first-layer weights concentrate, for instance noticing that Breakout's agent weights lie along the bottom row, which is the paddle. Goedecke's point is that the test checks a simple calculation over a labelled layout.

The Qwen result is the paper's headline, and it differs from the no-tools test in a way the abstract does not make obvious. The abstract says that 'just by looking at the raw weights' Astra found a looping problem in Qwen3.5-0.8B and fixed it by editing 0.016% of the weights. That raised MATH-500 from 7.9% to 48.1% and raised IFEval by 3x. The paper's own table shows that experiment is not tool-free: the model wrote Python, and it ran the network on up to 50 batches of practice problems. Matthew Honnibal, the spaCy author, called the post misleading for this reason. He sorted the experiments in the paper's summary table by whether the model read raw weights, wrote weights, used Python or ran the network. He found the raw-weights-only setting appears only in experiments 1, 2 and 4, on networks of 17K to 172K weights. In his view, the Qwen result shows that Astra can devise and carry out tasks a human ML engineer could. That is a real achievement. Combining it with the raw-weight readings in one headline, however, implies a single capability the experiments do not establish.

Goedecke also looked at the Qwen fix and described what Astra did as a known method. The model's problem was that after about 32k tokens it fell into loops, which the paper says Qwen's own model card warns about. Goedecke proposes that a researcher would collect looping and non-looping texts, measure how much each weight contributes to the difference in activations, and trim the top weights. He compares this to the Heretic abliteration approach. He says the transcript shows Astra wrote four pairs of texts, measured the contributing weights, cut them back and tested the result, ending with a patch that let the model work past 32k tokens. Goedecke concludes that this is what a competent human researcher would do. He compares the capability to a language model editing compiled binaries rather than writing source. His objection is therefore that the editing example follows a known procedure, not that the result is false.

The paper's reported figures are specific, and they are worth keeping apart. Astra's one-shot edit to the Kinetix car-on-ramp policy succeeded in all 1,024 episodes after changing 3.5% of the weights. The paper notes that the original network never succeeded in 256 episodes and that Astra was given one recorded failed episode. In the MinAtar experiment, Astra wrote networks with no training that beat PPO trained for 100 million steps in all three games. Against PPO trained for 500 million steps, the networks scored 20 times as much in Breakout (517.4 against 25.3), fell below it in Seaquest (53.1 against 88.9) and came close to it in Asterix (53.2 against 55.2). The paper's Gemma 3 270M sycophancy experiment, which used a Python sandbox, failed in its weights-only attempts and succeeded only after it could see attention patterns. Those numbers come from the paper and its linked logs, and the logs say the Go and Qwen experiments used Python and forward passes.