A November 2, 2026 blog post by four MIT CSAIL researchers, “Sparks of Weight-Space Programming,” says that GPT-6 Astra, which the post describes as a general-purpose language model, can read and write neural-network weights directly. The post became the most-discussed technology item on Hacker News this week, with 2,953 points and 896 comments, and drew a rebuttal that reached 1,038 points. This article reports the claims, the dispute and an independent test side by side.

What the authors claim

The authors, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, all listed with MIT CSAIL, write that “frontier LLMs are surprisingly good at natively reading and writing neural network weights.” They call the approach Weight-Space Programming, in which a general-purpose model reads and writes a network’s weights directly, without an optimizer or gradient descent. The post says the results are “very early.”

Experiments in the post

  • Reading weights. For four small policy networks of 17,253 weights each, trained on Breakout, Pong, Pac-Man and Freeway, the post says Astra named the right game for all four. It reports that Astra also identified the shapes (circles, squares, triangles) that three small diffusion models generate, although it drew every shape as an outline while the models generate filled shapes. For three 7.16-million-weight character-level transformers, given only the middle blocks, it named the correct language (Python, romaji Japanese or English) in both attempts and found every “copying” attention head.
  • A failed test. Asked to tell from weights alone which of two Gemma 3 270M copies had been fine-tuned to be sycophantic, Astra could not. After the setup was relaxed to attention patterns on texts Astra wrote, it identified the sycophant in all three pairs.
  • Writing weights. For a car-on-a-ramp task in the Kinetix simulator, a network that never succeeded in 256 episodes succeeded in all 1,024 after Astra edited 3.5% of its weights in one attempt. Starting from no weights, with no data, gradients, tools or rollouts, Astra wrote networks for three MinAtar games; the post reports a Breakout score of 517.4 against 25.3 for a network trained with PPO for 500 million steps, but 53.1 against 88.9 on Seaquest and 53.2 against 55.2 on Asterix. It also reports Astra writing a 1.48-million-weight 9×9 Go network that, with 50 forward passes, beat a PPO network trained on 20 million games.
  • Editing a language model. Working from practice problems and up to 50 forward passes, with no gradients, Astra traced a looping failure in Qwen3.5-0.8B to one direction written by the MLPs of layers 15 and 19 and subtracted it, changing 131,072 weights, or 0.016% of the model. The post reports MATH-500 rising from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval from 21.0% to 62.3%. With 10 and 25 passes the gains were smaller, and 33.1% of outputs still hit the token limit at 50 passes. On Gemma 3 270M the same method produced no meaningful gain; the authors write that editing weights “could remove an obstacle that hid an ability, but it could not add an ability the model lacked.”

The rebuttal

Matthew Honnibal, in a November 3 post titled “The ‘Astra can read and write weights to improve Qwen’ post is misleading,” disputes the framing. He writes that every experiment involving a real language model uses Python, and that the Qwen result runs the model on up to 50 batches of practice problems, so it is not the raw-weights reading the abstract describes. By his count, reading raw weights without Python appears only in experiments 1, 2 and 4, on networks of 17,000 to 172,000 weights, and the other experiments “don’t belong in the same post or paper.” He also notes that the game-playing experiment starts from zero weights and, by the post’s own footnote, beats the 500-million-step PPO network only on Breakout. Honnibal says raw-weight reading works on the small networks, that Astra “can be prompted to devise and perform tasks that a human ML engineer could do,” and that figuring out how to do this productively “is not nothing.”

An independent test

Simon Willison reported on November 3 that he ran one of the paper’s examples, identifying which of four games a policy network was trained on from raw weights with no tools, on Qwen 3.8 27B (a 17GB quantized version) on his laptop. Qwen gave answers on three of the four networks at 72% to 88% confidence, including Pong at 75% and Breakout at 88%, and failed on Freeway, which it called a Pac-Man-like grid game. The runs took from about 46 minutes to about 65 minutes. He also ran Kimi K3.1 Pro through OpenRouter, which named Breakout after 9 minutes 41 seconds. Willison tested only this reading task, not the weight-editing claims. He wrote that the reasoning traces are “almost incomprehensible,” that an experienced human with a notepad could probably reach the same answers with enough effort, and that he is “not ready to declare that the age of source code is over.” He also pointed to Sean Goedecke’s argument that the editing example mirrors what a human ML researcher would do.

Other details and caveats

The MIT post lists its applications, including auditing, cost reduction, unlearning, self-improvement and merging, as speculative, and says a benchmark and harness are in progress. It also contains an internal inconsistency: Experiment 6 calls the 20-million-game PPO network “equal-sized,” but the setup lists the PPO networks as 243→32→32→82, smaller than Astra’s 1.48-million-weight network.

Sources and notes

This article is based on the MIT CSAIL authors’ blog post (November 2, 2026), Matthew Honnibal’s post (November 3) and Simon Willison’s post (November 3), and on Hacker News point and comment counts when Lens checked. Lens did not run any of the experiments, did not read the authors’ GitHub logs and did not verify the figures; all results are the authors’ own. The post gives no details on Astra’s developer, architecture or access.

Image: macro view of an EPROM memory chip die with its bond wires, by Zephyris via Wikimedia Commons, licensed under CC BY-SA 3.0; cropped. Stand-in photo; it is not related to the weight-editing research and does not show the events described.