Skip to content

Comment on Text classification with Python 3.14's ZSTD moduleparent

Comments

Yes LLMs are always lossy, unless their size / capacity is so huge they can memorize all their inputs. Even if LLMs were not resource-constrained, one would expect lossy compression due to batching and the math of the loss function. Training is such that it is always better for the model to accurately approximate the majority of texts than to approximate any single text with maximum accuracy.

AboutSource Built by g1lg1l

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