Reaction
Classifiers assign labels; LLMs generate text token by token
SemiAnalysis_ explains binary, multiclass and multilabel classification, then compares those tasks with an LLM’s next-token predictions.
TLDR
SemiAnalysis_ explains that classifiers map inputs to labels from a fixed set. An LLM’s final layer also classifies possible tokens, the thread says, but repeats that step to generate text. It claims output tokens are always 3–5 times more expensive than input tokens because they are generated one at a time, rather than processed in parallel.
Combined views
—
2 Sources, first seen ago
