The one-at-a-time way a language model writes. It picks a word, adds it to what it has already produced, then picks the next based on everything including what it just wrote. It never plans the whole sentence in advance and it cannot revise what it has already said. This explains a great deal: why output arrives gradually, why longer answers cost more and take longer, and why a model that starts down a wrong path tends to keep going.
Technical view
Each step conditions on the full preceding sequence, so generation is inherently sequential and cannot be parallelised the way training can, making it latency-bound and memory-bandwidth-bound rather than compute-bound. This is the constraint that key-value caching and speculative decoding exist to relieve. It also means an early sampling error propagates, since every later token is conditioned on it, a failure mode sometimes called exposure bias.
Related
Part of AI Terminology Index, section: Language Models Specifically.