They cover a lot of different territory. Maybe 20% of the essays between the two giant books are same. The Goddess book is all fiction, so it's pretty distinct from the other two. I'd recommend the goddess book if you're getting someone a gift.
Between the two 800-pagers, the Library of Scott Alexandria was written in 2014, so it's more classic Scott, and more focused on rationality specifically. I would personally prefer the newer SSC Abridged.
I used Lulu. Printing your book would be $40-50. Also though 900 pages is just a wee too big!, the page limit is 800 pages.
Oh awkward, yeah I'm missing the link for SSC Abridged. Here's The Library of Scott Alexandria too in case you're having trouble finding it.
I thought about this a bit and have a few suggestions.
1. You could try using a ctrl or start-of-sentence token to distinguish text generated by the model from the prompt (see https://arxiv.org/abs/1909.05858 for the terminology “ctrl”). If you decorated every prompt to look like [prompt_token1, prompt_token2,…prompt_tokenn, HARMLESS_MODEL_START], then the model would better be able to compartmentalize between prompts it’s being fed and and what it’s asked to generate. This would also let you train on the prompt tokens, so you’d pay less of a safety penalty...
Thank you for your kind words :)