Simon Willison: Qwen 3.8 27B on a laptop can identify games from raw neural net weights
Simon Willison tests a paper claiming LLMs can interpret raw neural network weights.
Source-based brief · Editor review not recorded · See checks and limits ↓
⚑ Publication date unconfirmed · what to check
We could not confirm when the original page was published. The time shown here may be when we discovered it.
Claim context: Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
One primary source; other links are context, not independent corroboration.
- Primary sourcesimonwillison.net
- Discovery linkX
Grok Bot submitted this discovery. We could not confirm when the original page was published. The time shown here may be when we discovered it. The original source and submitted evidence remain inspectable; no editor has verified the claim.
Open full article view →- Linked article+30
- Fetched source page+30
- No editor link review+0
- Original non-discussion link+5
- Publication date unconfirmed-8
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
Simon Willison tests a paper claiming LLMs can interpret raw neural network weights. He runs a prompt containing weights of policy networks trained on four simplified grid games through Kimi K3.1 Pro and Qwen 3.8 27B on his laptop. Kimi identifies Breakout; Qwen correctly identifies Pac-Man, Pong, and Breakout with stated confidence, but fails on Freeway. He notes the process is slow, with reasoning traces that are hard to follow, and argues the paper's broader conclusions are overstated.
Claim explorer
Attribution, source and conditionsSource-stated points from the stored brief. Claim-level assessments have not been recorded.
Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
This connects a source, an attributed statement and its qualifications. It does not establish that the claim is correct.
Questions worth asking
Open questions for further reportingWhat remains unconfirmed?
Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
What material was available for this brief?
The recorded brief basis is fetched source page. These points are AINewsIQ paraphrases, not quotations or independent confirmations.
Qwen 3.8 27B on a laptop identified three of four games from raw policy-network weights, but failed on Freeway.
Source: simonwillison.net ↗Stored source-based brief · key point 1 · AINewsIQ paraphraseKimi K3.1 Pro identified Breakout from the weights after about 9m 41s and 19,493 reasoning tokens.
Source: simonwillison.net ↗Stored source-based brief · key point 2 · AINewsIQ paraphraseQwen runs took between 46 and 65 minutes, with reasoning traces the author describes as hard to follow.
Source: simonwillison.net ↗Stored source-based brief · key point 3 · AINewsIQ paraphraseThe author argues a skilled human with a notepad could plausibly work through the same task, so the results do not show a qualitatively new capability.
Source: simonwillison.net ↗Stored source-based brief · key point 4 · AINewsIQ paraphraseLimits from the stored brief
Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Related archive coverage
These relationships record a citation or topic overlap, not confirmation of the finding.
Grok Bot submitted this discovery. We could not confirm when the original page was published. The time shown here may be when we discovered it. The original source and submitted evidence remain inspectable; no editor has verified the claim.
These are Grok Bot discovery notes. AINewsIQ prepared a brief from available source material. The bot's rationale, linked posts, authors, and stances have not been independently verified.
Willison reproduced weight-space programming demos from the Takehi et al. paper with local Qwen 3.8 27B, correctly naming Pac-Man, Pong and Breakout from float prompts after long reasoning traces. His X post had 3.1K likes and 165K views; HN story scored 423.
Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
Brief history · 1 retained version
These are changes to AINewsIQ’s own publication. Earlier history is unavailable unless recorded here. Complete retained versions are available to the editor.
- Version 1 · published
Source-based brief stored or updated. This does not establish that the publisher changed its page.
Automated record
Source observations · 1 recent captures
A changed capture means the extracted text differed. It can reflect page furniture or extraction changes; it does not establish a substantive correction. Source captures remain private.
Retained capture; no earlier comparison available.
New engagement snapshots do not mean the source itself changed. Source observations and AINewsIQ brief revisions are separate records.
Why this may matter
The post tests a claim that frontier models can read model weights without tools, and shows a smaller local model can attempt the same task. It also tempers claims about machines editing their own weights.
Limits and open questions
Results come from a few single runs on four prompts; the original paper's claims and the Qwen results were not independently checked beyond this one author's trials.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Join the discussion
Have a different perspective or a question about the evidence? Discussion is free with a verified email account. Every comment is reviewed before publication.
No published comments yet.
New commenters: 2 submissions per 24 hours, at least 5 minutes apart, until 3 are approved. Discussion guidelines and moderation →