WebGPU demo lets you edit model weights in the browser
What happened
A Show HN post on the Hacker News front page on November 19 demonstrates weight-space programming in the browser using WebGPU. An AI agent found a "repetition loop" direction in Qwen3.5-0.8B and subtracted it from the output weights of 64 neurons in layers 15 and 19. The post puts the edit at 131,072 of roughly 8 billion parameters, or 0.016%, with no training or gradient descent. The post does not say what the edit changed in the model's output.
Written by AI from the coverage · updated yesterday
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageShow HN: Watch weight-space programming in action with WebGPU
一个 AI 智能体在 Qwen3.5-0.8B 中找到一个“重复循环”方向,从第 15、19 层的 64 个神经元输出权重中减去该方向,仅改动约 80 亿参数中的 131,072 个(0.016%),无需训练或梯度下降。
Heat over time
Heat now 5·Comparable peak 10(Nov 20 01:00)·Comparable change over 24 hours –
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.