🔥 Trending on HN

Can an AI Read Another AI? Qwen 3.8 27B Takes a Small Test

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
model weights

Learned numbers that influence a model’s behavior.

Weight-Space Programming

A name for analyzing or changing model weights.

Hacker News

A website where people discuss technology.

What happened

Qwen 3.8 27B ran on a laptop in a test reported by Simon Willison. His report describes a simple challenge. Qwen received the learned numbers from another AI. That other AI had learned to play a small grid game. Qwen had to identify the game from those numbers alone.

The numbers are called model weights. They influence how a trained model behaves. They are not written instructions for humans. They are many floating-point values produced during training.

Qwen identified Pac-Man, Pong, and Breakout. It failed on Freeway. The task did not let Qwen run code, use a calculator, or use outside tools. It had to give one answer from the weights and the prompt.

The background

The test came from a research project called Sparks of Weight-Space Programming. The project asks whether language models can infer a network’s behavior from its raw weights. The game networks used simplified 10-by-10 grids. The four prompts were the same except for the weights.

The project also studies a different ability. A model can use code to edit another model’s weights. That might improve the other model’s performance. Reading weights without tools is a separate question from editing weights with tools. Willison focused on the first question.

The project calls the broader idea Weight-Space Programming. The name covers models analyzing and changing learned numbers. It does not prove that every model can inspect every other model.

Why it matters

Most people cannot read a long list of learned numbers and see a game inside it. If a language model can infer behavior from those numbers, researchers may gain another way to study trained systems.

The result still needs careful limits. Qwen may have recognized useful patterns without understanding the network as a person would. One successful test does not show general model understanding. It also does not show that Qwen can repair itself or improve any model.

What was confirmed

Willison used a Qwen 3.8 27B version that was about 17GB. The default 8,192-token context caused errors. He used a longer context instead. The Pac-Man run took 46 minutes and 12 seconds. The Pong run took 1 hour, 4 minutes, and 54 seconds. The Breakout run took 52 minutes and 25 seconds.

The responses reported confidence values of 72%, 75%, and 88%. Those values describe the individual answers. They are not results from a large accuracy study. Qwen failed the Freeway prompt.

The story received 736 points and 213 comments on the Hacker News discussion. That shows strong community attention. It does not prove that the experiment is correct.

What remains unknown

The test covered only four small games. We do not know whether the method works on larger networks. We also do not know whether other models would succeed. More tests must separate real understanding from pattern matching.

The weight-editing claims from the larger project need separate evaluation. They used different conditions. A Qwen reading result cannot confirm those editing claims. The project has also drawn criticism about how its broader results should be interpreted.

What to watch next

Useful follow-up work would repeat the same prompts across models and sizes. It should report every success and failure. It should also test more kinds of networks, not only small games.

The key question is simple. Can models read learned numbers reliably, or did this test find a narrow pattern? The public experiment logs make that question easier to investigate. That is the strongest value of this news today.

💬 Reading weights is promising, but not yet recursive self-improvement

HN commenters see a plausible ability for Qwen 3.8 27B to infer and perhaps modify neural-network weights, but they disagree about how general the result is. The experiments and performance claims below are commenter reports or interpretations, not independent confirmation.

  • One commenter reported that, when Qwen 3.8 27B was given C code and weights for a small network without being told its purpose, it probed the layers and inferred a likely image or upscaling filter: 25 inputs arranged as 5×5 producing a 2×2 output. The excerpt does not report that Qwen actually improved this network.
  • A supportive commenter explained that the original experiment had a known target change that the researchers could cross-check, and that the edited model improved on the target task. On that limited task, the weight interpretation therefore appears to have been correct.
  • The main objection is scope: one commenter said the evaluation showed improvement only on the single task named in the prompt, not the edited model’s overall behavior across a broad benchmark suite. Task-specific success should not be treated as general understanding.
  • Another user reported trying eight additional, less-famous games and seeing performance fall quickly. This suggests that familiar games may have helped the original examples, but it is a self-reported generalization concern rather than proof of failure.
  • On recursive self-improvement (RSI), the skeptical view is that external benchmarks are essential. Changing weights, even randomly, and selecting changes that help one task does not demonstrate that a model can improve general intelligence.
  • A practical hope is to reduce repetitive reasoning loops in highly quantized Qwen 3.8 27B variants such as IQ3_XXS and IQ3_S, which commenters associate with use on 16 GB VRAM systems. A proposed route is loop detection, identifying a loop-related weight direction, and connecting it to a LoRA training harness; the proposer had not yet worked out or tested the method.
  • The unresolved question is what such edits might break elsewhere. The comments provide no broad regression test for other capabilities or overall performance, so a claim that direct weight editing is generally reliable remains premature.

mature digest at 214 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Qwen 3.8 27B Read Another AI’s Numbers

📰 Full story: Can an AI Read Another AI? Qwen 3.8 27B Takes a Small Test

Qwen 3.8 27B looked at another AI’s learned numbers and guessed its game.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
weights

Learned numbers that help shape an AI’s behavior.

Qwen 3.8 27B

The AI model used in the laptop test.

Hacker News

A technology discussion website.

💡 The gist

  • Qwen 3.8 27B is an AI model that ran on a laptop.
  • It guessed three small games from another AI’s numbers.
  • It missed one game. Hacker News interest does not prove accuracy.

What did it read?

When an AI learns, it stores many numbers. These numbers help shape its behavior. They are called weights.

Simon Willison tested Qwen 3.8 27B with a public experiment. He gave it weights from small game-playing networks. The question asked Qwen to name each game. Qwen could not run code or use a calculator. It had to study the numbers directly.

The games used small 10-by-10 grids. They were Pac-Man, Pong, Breakout, and Freeway. The questions stayed the same. Only the weights changed.

What did Qwen get right?

Qwen identified Pac-Man, Pong, and Breakout. It failed on Freeway. This was one test, not a large contest.

The Qwen version was about 17GB. A short context setting caused errors. Willison used a longer setting. The Pac-Man answer took 46 minutes and 12 seconds. The Pong answer took 1 hour, 4 minutes, and 54 seconds. The Breakout answer took 52 minutes and 25 seconds.

The answers showed confidence values. They were 72%, 75%, and 88%. These values describe the answers. They do not measure a broad success rate.

Why does it matter?

People usually cannot understand a trained AI by reading its numbers. Qwen found useful clues in those numbers. That suggests a possible new way to study AI systems.

But Qwen may have found patterns. It may not understand the other AI like a person. This test does not show that every model can read every other model.

The original research also studied changing model weights. That is a different task. Reading numbers and editing numbers need separate tests.

What did people notice?

The story received 736 points and 213 comments on Hacker News. This shows strong interest. It does not prove the result is true.

Researchers should repeat the test with other models. They should test larger networks, too. They should publish failures, not only successes. The public experiment logs can help others repeat the work.

💬 It may read model weights, but that does not prove it can improve itself

The discussion finds an interesting ability, but the evidence is narrow. Some commenters accept the result for the tested task; others say it may not generalize.

  • A commenter reported that Qwen 3.8 27B looked at code and numbers for a small 5×5-to-2×2 image filter and guessed what it did, even though its purpose was not explained.
  • Supporters say the experiment checked a known change and the target task improved. Critics say that is still only one task, not a full test of the whole model.
  • Another user reported that results dropped quickly on eight less-famous games, suggesting that familiar-game examples may have helped. This is also a user report, not a final verdict.
  • This is not yet proof of recursive self-improvement. Outside tests are needed to see whether changes help beyond one task and what other abilities might break.
  • People hope to reduce repetitive loops in highly compressed models such as IQ3_XXS and IQ3_S, used by commenters on 16 GB VRAM systems. A loop detector plus a small extra-training method called LoRA was suggested, but it has not been demonstrated.

mature digest at 214 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Qwen Looked at AI Numbers

📰 Full story: Can an AI Read Another AI? Qwen 3.8 27B Takes a Small Test

Qwen 3.8 27B is an AI that ran on a laptop.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Qwen 3.8 27B

An AI that ran on a laptop.

weights

Numbers that help an AI do things.

Hacker News

A place where people discuss technology.

A person named Simon Willison reported this test.

Another AI had learned some games. It kept many learned numbers. Those numbers helped it play.

Qwen looked at those numbers. It guessed the games. It guessed Pac-Man, Pong, and Breakout. It missed Freeway.

Qwen used no code or calculator. It only read the numbers.

The story got 736 points on Hacker News. People also wrote 213 comments. That shows people were interested. It does not prove the test is always right.

More tests are needed. We do not know if other AI models can do this.

💬 The model may guess what other AI numbers do

The comments describe a clever experiment, but it is still a small test.

  • One user reported that Qwen looked at a small AI’s numbers and guessed that it changed a 5×5 patch into a 2×2 patch. The user had not told it the purpose.
  • Some people say the test matched a known answer. Others say it may only work for one task; one user reported that it did much worse on eight less-famous games.
  • That does not show that the AI can make itself generally smarter. People want outside tests to see what improves and what breaks.
  • People also want to stop compressed models such as IQ3_XXS and IQ3_S from repeating the same thought on computers with 16 GB VRAM. A loop-finder and a small helper called LoRA were suggested, but no working result was reported.

mature digest at 214 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources