What’s really disgusting is how unnecessary this is.

LLMs have topped out in terms of language fluency. You’re not going to get a smarter model with 250 trillion tokens than with 25 trillion tokens. There are still other gains to be made in the LLM/LRM space, but they don’t require ripping up rare books.

And they’re doing it destructively because it’s cheaper. That’s it. They absolutely could scan nondestructively. They’re trillion-dollar companies, and they do this in a shitty way to save pennies.