Pinned post
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training.
Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
0:12