AnalysisAI ModelsNovember 19, 2026

Show HN: WebGPU playground demos weight-space editing of Qwen3.5-0.8B

Read original source →tabbyml.github.io

An LLM edits 131,072 of ~800M weights (0.016%) — subtracting one direction from 64 neurons' ffn_down output weights in layers 15 and 19 — and the model stops looping. Built as a browser walkthrough after the full edit-evaluate loop needed too much VRAM.

1 source

More stories today

Open the live feed