I am not sure what the amount of useful training text would be, but the transcripts of the 80k podcast could be useful as a source of ‘spoken language’ EA thinking.
Regarding ”??% rationality” (Scout Mindset is a great choice), my impression is that these did significantly influence some fraction of EAs, but not nearly all .
For HPMOR I think that there are a few arguments against including it: For one, I could imagine that the fictional setting can let the model give unexpected answers that refer to a fictional world if the input accidentally resembles discussions in HPMOR too much (I am not familiar enough with Transformers to say whether this would actually be a problem or not, but it could be very confusing if it starts mentioning Transfiguration as a cure to Alzheimer’s). Also, some characters in there are explicitly malevolent or highly cynical about humanity – I do not think that it would push EA GPT in a good direction to be trained on these.
For a nice selection of rationality texts, the LessWrong books might be a good choice as they contain texts from many different writers, and which were chosen by the LW community to be exemplary.
I am not sure what the amount of useful training text would be, but the transcripts of the 80k podcast could be useful as a source of ‘spoken language’ EA thinking.
Regarding ”??% rationality” (Scout Mindset is a great choice), my impression is that these did significantly influence some fraction of EAs, but not nearly all .
For HPMOR I think that there are a few arguments against including it: For one, I could imagine that the fictional setting can let the model give unexpected answers that refer to a fictional world if the input accidentally resembles discussions in HPMOR too much (I am not familiar enough with Transformers to say whether this would actually be a problem or not, but it could be very confusing if it starts mentioning Transfiguration as a cure to Alzheimer’s).
Also, some characters in there are explicitly malevolent or highly cynical about humanity – I do not think that it would push EA GPT in a good direction to be trained on these.
For a nice selection of rationality texts, the LessWrong books might be a good choice as they contain texts from many different writers, and which were chosen by the LW community to be exemplary.