I’d never looked at the NYT bestseller list itself. surprisingly high ratio of moderately interesting things there—I expected 1/k, but I see maybe 5 hits out of 15. things where I expect that reading them improves ones’ bets and reasoning about a topic which, if understood well, is relevant to alignment. Surprising, given that I figured people’s taste would be uninteresting and meaningless entertainment slop. It’s not clear that that means reading those books is the best use of time for people professionally engaged in alignment work, but possibly trying to get what you’d get from reading, but do so faster (skim, request compression from some well-chosen LLM, etc) might be worth it. Where does one buy books in a format that can easily be converted to text for use with a personal LLM (eg, some pdfs)?
Models have some trouble with PDFs sometimes—seems like they often much prefer markdown and such if possible, so may be worth trying to convert. I’ve had tasks completely fail with PDF input and work as expected with markdown input of the same content.
I’d never looked at the NYT bestseller list itself. surprisingly high ratio of moderately interesting things there—I expected 1/k, but I see maybe 5 hits out of 15. things where I expect that reading them improves ones’ bets and reasoning about a topic which, if understood well, is relevant to alignment. Surprising, given that I figured people’s taste would be uninteresting and meaningless entertainment slop. It’s not clear that that means reading those books is the best use of time for people professionally engaged in alignment work, but possibly trying to get what you’d get from reading, but do so faster (skim, request compression from some well-chosen LLM, etc) might be worth it. Where does one buy books in a format that can easily be converted to text for use with a personal LLM (eg, some pdfs)?
The simplest general way is to buy it in whatever format and then download it from one of the well-known websites with free pdfs/mobis/epubs.
Models have some trouble with PDFs sometimes—seems like they often much prefer markdown and such if possible, so may be worth trying to convert. I’ve had tasks completely fail with PDF input and work as expected with markdown input of the same content.