|
Show HN: Watch weight-space programming in action with WebGPU
A researcher demonstrated weight-space programming using WebGPU, where editing just 131,072 weights (0.016% of an 800M parameter model) across 64 neurons in two layers eliminated looping behavior in a small language model. The targeted edit improved MATH-500 accuracy from 7.9% to 48.1% by surgically modifying specific neurons that drove repetitive outputs, with an interactive visualization showing the before-and-after token generation process.
Read Full Article →
← Back to latest