Pass IndexThe State of AISign in

Autoregressive

Producing an answer one token at a time, each conditioned on all the ones before.

how it works · the vocabulary

next-token prediction

It is why output is billed per token and why it streams: the model genuinely does not know the end of its sentence when it starts. It is also why output costs more than input, which can be read in one pass.

In the catalogue

Nearby

AgentAgent memoryAlignmentBM25ChunkingCold startComputer useContext window