🐈

you don't have authoritarian high modernist control over what the output is. you can only say "hmmm this looks interesting, let's add more here." the only thing you can control is where you focus your attention and what you choose to call forth into existence

so for gpt-3, your prompt has to be something like a high dimensional vector pointing into a direction you find "interesting" however, this space is likely 99% empty even with the language corpus being the entire internet.

so imo, the ability to riff off of metaphors and analogies is gpt-3's biggest strength.https://twitter.com/a_yawning_cat/status/1340904815816179713 …

i have found that some people very much dislike metaphors/analogies because they find them too "magical". however they can actually feel very "intuitive". they connect two concepts through an "arc" or "portal" that skips logical deterministic pathinghttps://twitter.com/a_yawning_cat/status/1341959274964996096 …

this means that a metaphor/analogy "Points" from a Here to a There that is definitely THERE. this means that once you find one metaphor that successfully delivers a payload & connects the dots, you can use gpt-3 to figure out how to deliver other payloads along the same arc!

so classes of usage like "use it like a chat bot" work by incrementally refining and growing a pointer. it's actually an erroneous model to ascribe it agency when it's more like a knee jerk reaction searching for closure

with twitter threads, there a sensation of "open" and "closed" just like with stories there is a clear sensation when the story is "finished". (just like a joke "closes" w/ a punchline) the story tapers and there is a sense of finality and closure at a final "point"