Earlier quoted context omitted.
> I personally believe that these materials should definitely be part of large model training. Already done: https://news.ycombinator.com/item?id=37752272 It turns out that the real safety risk with AI is not Mecha-Hitler, it's just that it might end up reading the wrong sorts of books and accidentally conjure a horde of demons.
> accidentally conjure a horde of demons. If we have a certain perspective on demons as self-sustaining information processing loops—we’ll, yeah, demons abound. And, yes, this hn post from BenBreen — is amazing. We’ve been in touch! And if Ben is reading, we’ve made progress on our version controlled community—LLM book translation prototype. Of course, it’s much to early to share on hn. Or whatever, it’s great, take…
Such as paranoid schizophrenia, drug addiction, self harm, harming others, etc.