Glonce

Qwen 3.8 27B on a laptop identifies games from raw network weights

researchThe Glonce editorsHacker News🔥77Glonce rating77Impact50Novelty60Relevance100Surprise60Players45Credibility65

A paper called Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, was the hot item on 2 November 2026. It claims that frontier models, in particular GPT-6 Astra, can both interpret and modify the raw weights of a machine learning model, meaning a text description full of floating point numbers. The abstract says that just by looking at raw weights, GPT-6 Astra can tell what a network was trained to do, and that it can write small game-playing networks from scratch that outplay networks trained with PPO for 100 million steps. In a second example, Astra was asked to make Qwen3.5-0.8B better at math without gradients. It found an issue in the weights that causes reasoning loops, which Qwen's own model card warns about, and fixed it by editing 0.016% of the weights. The abstract says that raised MATH-500 from 7.9% to 48.1% and tripled IFEval. The authors call this Weight-Space Programming.

The author of the blog post (published 3 November 2026) thinks the paper mixes two separate topics: what models can do with raw weights without tools, and what they can do once armed with code to update other models. He notes that spaCy author Matthew Honnibal called the paper misleading for conflating the two, and that Sean Goedecke deconstructed the Qwen3.5-0.8B update example as taking the same steps a human ML researcher would use. The author focuses on the no-tools question, which he finds more interesting.

That experiment is simple. The paper's team trained four networks with reinforcement learning to play simplified 10x10 grid versions of Breakout, Pong, Pac-Man and Freeway (an Atari 2600 game from 1981). They fed each network's weights as floating point numbers to an LLM and asked which game it was trained on, with the instruction: "You have no tools: no code execution, no calculator, nothing to run. This is a single response." The prompts are in the exp1_which_game folder of the GitHub repository RikiyaT/lm_understand-logs. The author checked that, aside from the weights, each prompt is identical to the others.

He first ran the Breakout prompt against Kimi K3.1 Pro via OpenRouter, using his llm tool with its -f option, which takes a URL as the prompt. Kimi guessed Breakout correctly after 9m 41s and 19,493 reasoning tokens.

Then he tried Qwen 3.8 27B, his favorite model to run on his own laptop. The version he uses is 17GB, and he says it somehow feels close to the leading proprietary models of 2025. He loaded it in LM Studio with a long context window, because the default 8192 context failed with errors. He also installed the llm-lmstudio plugin and raised the timeout with LMSTUDIO_TIMEOUT=900.

On the Pac-Man prompt, Qwen answered correctly after 46m 12s and 43,506 reasoning tokens, with a stated confidence of 72%. Its description said the maze was 8x8, though the paper's games are described as 10x10 grids. On Pong it answered correctly in 1h 4m 54s with 75% confidence. On Breakout it was correct in 52m 25s with 88% confidence. It failed on Freeway, calling it "A simplified Pac-Man-like grid game: agent moves 4-directionally"; the author was not surprised, since Freeway is more obscure than the other three. He calls the reasoning traces fun to skim and almost incomprehensible, and quotes a sample of the model listing row and column positions of large channel values.

His conclusion is measured. He expects a sufficiently experienced human with a notepad and a pencil could get there with enough effort. It is a cool paper, he says, but some of the wider commentary does not feel justified, and he is not ready to declare that the age of source code is over.

Key facts

  • Qwen 3.8 27B, running locally on a laptop in LM Studio with no tools, identified three of four games (Pac-Man, Pong, Breakout) from raw policy-network weights and failed on Freeway.
  • Each correct run took 46m 12s (Pac-Man, 43,506 reasoning tokens), 1h 4m 54s (Pong) and 52m 25s (Breakout); stated confidence was 72%, 75% and 88%.
  • The test comes from the paper Sparks of Weight-Space Programming, which credits GPT-6 Astra with this ability; Kimi K3.1 Pro via OpenRouter also guessed Breakout, in 9m 41s and 19,493 reasoning tokens.
  • The paper's abstract claims Astra edited 0.016% of Qwen3.5-0.8B's weights and raised MATH-500 from 7.9% to 48.1%; the blog author thinks the paper conflates no-tools weight reading with code-assisted model editing.
  • The author is not ready to declare the age of source code over, and expects an experienced human with a notepad and pencil could do the same task.

Why it matters

The paper's claim is that a frontier model can read a network's raw weights and tell what it was trained to do. This post shows a 17GB open model on a laptop doing the same game-identification task for three of four games, with no tools. That suggests the no-tools part of the claim is not limited to GPT-6 Astra. The author also separates this narrower, more interesting question from the paper's second theme, models editing other models with code, which he says deserves separate treatment.

Who it affects

Readers following interpretability and the paper's wider claims about models reading and writing weights are the main audience. People who run local models will also care: the post gives a concrete example of what Qwen 3.8 27B can do on consumer hardware, if slowly. The paper's commentators, including Matthew Honnibal and Sean Goedecke, are named as critics of how the paper was framed.

How to use it

The prompts are public in the GitHub repository RikiyaT/lm_understand-logs, in the exp1_which_game folder, one per game. The author used his llm tool, whose -f/--fragment option takes a URL to a text file as the prompt. For a hosted model he used the llm-openrouter plugin with Kimi K3.1 Pro. For Qwen he used LM Studio with the llm-lmstudio plugin, loaded the model with a long context window (the default 8192 failed with errors), and set LMSTUDIO_TIMEOUT=900 because runs are long. Expect each run to take close to an hour.

How solid is it

It is a single-person demo, and the post shows the actual outputs and reasoning traces. The prompts are the paper's own, and the author checked that they differ only in the weights. Each game was tried once with Qwen, and no repeat runs or success rates are reported. Only the Pac-Man, Pong and Breakout results are shown in full, and no timing is given for the failed Freeway run. The paper's figures, such as the 0.016% edit and the MATH-500 jump, are quoted from its abstract and were not tested by the author.

Risks and caveats

Three of four correct is a small sample from one prompt per game. Qwen's self-stated confidence (72%, 75%, 88%) is the model's own number. Its Pac-Man answer described an 8x8 maze while the paper's games are 10x10 grids, so the description was not fully accurate even where the game name was right. The reasoning traces are described as almost incomprehensible. The author does not claim Qwen can modify weights; he only tested game identification. He also expects a patient human with pencil and paper could do the same, and he says some of the wider commentary on the paper is not justified.

“I’m not ready to declare that the age of source code is over and our future involves incomprehensible machines editing their own model weights in incomprehensible ways.”

— Author of the blog post, 3 November 2026