On: AI companies destroy physical books – let's scan rare books before it's too late

[ivell]

Knowledge of every member of a species is better than few samples? LLM became better due to its huge dataset instead of just few samples.

[eru → ivell]

You hit diminishing returns pretty quickly. And LLM training only bothers with a single copy of each book; they don't even need to destroy the documents. (Eg they didn't destroy the American declaration of independence to train on it.)