Skip to content

Comment on Generating one token at a time is a blessing in disguise

Comments

LLMs generate their output one token at a time. The first thought when you learn this is that this is a huge performance bottleneck, as we are used to highly parallelized systems.

However, a large part of what makes LLMs feel so magical comes from this bottleneck.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.