Qwen 3.8 27B Reads Neural Network Weights on Laptop

Simon Willison has successfully replicated a widely-discussed weight-space-reading experiment on his own laptop, using Qwen 3.8 27B to identify three of four small game-playing neural networks from their raw parameters alone. The replication, which required no code execution and ran entirely in LM Studio with a quantized model, correctly named Pac-Man, Pong, and Breakout policies at confidence levels between 72% and 88%.
Officially released on November 3, 2026, the replication complicates earlier claims that weight-space reading collapses outside controlled demonstrations. Willison's results show the capability emerging in a mid-tier open model running on consumer hardware, not just frontier systems.
The Replication Setup
Willison pulled the Q4_K_M.gguf quantization of Qwen 3.8 27B from Hugging Face and ran it locally under LM Studio. He fed the model the raw parameters of four small policy networks trained on classic Atari games, providing no code execution environment and no additional context beyond the weights themselves.
The model correctly identified Pac-Man at 72% confidence after 46 minutes and 43,506 reasoning tokens. Pong followed at 75% confidence after 1 hour 4 minutes, and Breakout scored highest at 88% confidence after 52 minutes. The only miss was Freeway, which Qwen incorrectly guessed as "a simplified Pac-Man-like grid game."
What the Model Actually Does
Internal reasoning traces from Qwen's output hint at the analysis strategy. Sample reasoning shows the model scanning for spatial patterns: "What about agent channel0? We see some large channel0 rows near row8 col7/8 and maybe bottom?" The model appears to reconstruct game mechanics from weight distributions, inferring board layouts and agent behaviors without running the network.
The original paper, "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights" by Rikiya Takehi and colleagues, made a stronger claim that GPT-6 Astra can determine a network's function purely from weights. That claim drew skepticism from spaCy's Matthew Honnibal, who argued the paper conflated pure analysis with tool-assisted execution.
Significance for Open Models
Willison's replication demonstrates that weight-space reading is not exclusive to frontier models or carefully controlled demonstrations. Qwen 3.8 27B is a mid-tier open model, quantized to 4-bit precision and running on consumer hardware. The fact that it reproduces the core result suggests the capability may be more robust than critics initially believed.
- The model processed networks with thousands of parameters, not toy examples
- Quantization to Q4_K_M did not eliminate the capability
- Inference times ranged from 46 to 64 minutes per network, making the approach practical for research
The replication does not validate the paper's broader claims about a post-code future. Willison explicitly declines to read his results as vindication: "It's a cool paper, but some of the wider commentary I've seen around it does not feel 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."
What This Means
Willison's replication establishes that weight-space reading is reproducible outside controlled research environments and accessible to mid-tier open models. The result opens new research questions about how language models represent and reason about numerical structures, but it does not support claims of imminent self-modifying AI systems. For practitioners, the experiment confirms that Qwen 3.8 27B can perform complex pattern recognition across non-linguistic inputs when prompted appropriately, a capability worth exploring in domains beyond neural network analysis.
on Emergent today
.avif)




