🐈

gpt-3 acts like a highly saturated liquid. the crystal that is formed is dependent on the shape of the thing that is thrust into it it's not that different than an actual brain that reacts based on triggershttps://twitter.com/AdeleDeweyLopez/status/1411473967001395203?s=20 …

TLDR; hooman brains are no different than bubble solution or those clicky heat pack things

just saw github copilot described as fancy markov chain autocomplete which works for general usage as well

thinking about this more there may be a weakness in this model b/c language isn’t a static state independent of the reader it makes technical sense as an internal model just like u can consider the input/output a long alphanumeric string

but the actually useful understanding is a much smaller space. blank spaces demarcate words so a string of words is better. words express ideas so input/output as an “idea” is even better

gpt-3 after all is trained on “ideas of the internet” but this makes gpt-3 feel too much like a search engine. you put in an input to specify some kind of unique state, and then it gives you adjacent states like a librarian fetching a book