p(doom)

ACT: Adaptive Compute Transformer | p(doom)

Large language models exhibit remarkable reasoning capabilities with scale. However, a fundamental flaw of current-generation transformer-based language models is their uniform allocation of compute per token.