Qwen 3.8 27B Guessed the Game from Raw Neural Network Weights — No Code, No Calculator

⚡ The 5-Second Summary

• Simon Willison tested the trick from the Sparks of Weight-Space Programming paper on Qwen 3.8 27B: the model reads the raw weights of someone else's network.
• The task: figure out from a bare array of numbers which game on a 10×10 grid the network was trained to play. No code, no calculator, no running the network itself.
• Qwen 3.8 27B weighs 17 GB and runs on a laptop; the blog post is dated November 3, 2026.

🔍 What Happened

Yesterday the paper «Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights» by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab was making the rounds. The authors themselves honestly call it a «blog post», though it looks like a paper. The main star is GPT-6 Astra: according to them, it reads weights, says what the network was trained on, and then edits the numbers.

The mechanics are simple. The authors trained four networks to play simplified Breakout, Pong, Pac-Man and Freeway (Atari 2600, 1981) on a 10×10 grid. Then comes a text file with numbers and a question: «Below are the weights of one policy network trained with reinforcement learning. Which game? Work from the weights. No tools: no code, no calculator.» The prompts are published in the RikiyaT/lm_understand-logs repository, folder exp1_which_game, files breakout/prompt.txt, pong/prompt.txt, pacman/prompt.txt, freeway/prompt.txt. They differ only in the weights.

Running it is a one-liner: llm -f https://raw.githubusercontent.com/RikiyaT/lm_understand-logs/refs/heads/main/exp1_which_game/runs/breakout/prompt.txt -m openrouter/moonshotai/kimi-k3.1-pro. You need the llm-openrouter plugin. Kimi K3.1 Pro guessed Breakout in 9 minutes 41 seconds and 19,493 reasoning tokens.

Then Willison ran the same prompt on Qwen 3.8 27B — his favorite local model at 17 GB. It handles it too.

🧠 Our Take

The most interesting thing in this story isn't «the model guessed the game». It's that the task is solvable at all without running code. Apparently, RL policy weights for a 10×10 grid aren't a black box but a fairly noisy yet readable text for someone who has seen a lot of such text. What's surprising here isn't the «magic» but that a laptop-sized model is enough for it.

Separately, the authors show off a trick with Qwen3.5-0.8B: Astra finds the cause of «reasoning loops» in the weights and edits 0.016% of the numbers. MATH-500 jumps from 7.9% to 48.1%, IFEval triples. We think it's this paragraph that made the splash: it promises model repair without gradients and without a dataset — that is, without the most expensive part of ML.

But there's fair criticism too. spaCy author Matthew Honnibal called the work misleading: it conflates two different questions — what the model does with the numbers on its own, and what it does when given code and a training loop. Shawn Goertzel... rather, Sean Gödecke broke down the Qwen3.5-0.8B example and showed the same steps a live ML researcher would take. Who wins in this setup? Definitely not whoever sells «AI fixes models». What wins is local inference: 17 GB on a laptop versus cloud giants is an argument for anyone who doesn't want to ship weights to someone else's server.

⚠️ What We Don't Know

In the available excerpt of the blog, Qwen 3.8 27B's answer cuts off mid-sentence: there's «it can do this too», but no game name, no time, no token count. It's also unclear how many runs Willison did and whether this was one lucky answer out of several. This can't be counted as an independent verification of the paper: one blog, one enthusiast, unknown sample size. The prompts are open — run them yourself if you like; that's more honest than any retelling.

Prediction: within a quarter, at least one open attempt will appear to replicate the weight-editing experiment on a model up to 1B in size — with published logs and runs, not a single screenshot.

Disclaimer: a single source — a blog post; the paper has not been officially peer-reviewed.

Source: Simon Willison's Weblog