I made an LLM think in latent space. The model never read a single thought
A developer tested an aggressive version of latent-space reasoning on a 1.5B-parameter language model, where the model pauses during generation to run parallel hidden-state rollouts without decoding t…