An interactive WebGPU playground demonstrates “weight-space programming”: modifying a small, targeted portion of a model’s existing weights to address a specific failure, without training or gradient descent. It focuses on Qwen3.5-0.8B, whose reasoning mode repeatedly rechecks itself on some math problems instead of answering. A recorded replay compares the base and edited model on identical prompts and seeds, and visitors can also download and run both models locally in a browser through llama.cpp, WebAssembly, and WebGPU.
Across a batch of MATH-500 problems, 93% of base-model runs hit the 32,000-token limit. An AI agent found a direction associated with repetition and subtracted it from the output weights of 64 neurons in feed-forward layers 15 and 19 - 131,072 weights, or about 0.016% of the model. The edited model’s rate of hitting the limit fell to 33.1%, while reported MATH-500 accuracy rose from 7.9% to 48.1%. The playground visualizes the changed weights and neuron activity: the edit pushes against the neurons’ output, which intensifies during loops, and makes repeated next tokens less certain. Its analysis covers only a handful of recorded loops and a control, and explicitly presents the activation relationship as correlation rather than proof.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.