I like the parable that any text content that originated before 2021 now has increasing value, and in certain sense became a digital equivalent of physical "Low-background steel" [1].
Its value is mostly, and paradoxically, for training LLMs, to get "unspoiled" models.
Its value is mostly, and paradoxically, for training LLMs, to get "unspoiled" models.
[1] https://en.wikipedia.org/wiki/Low-background_steel