@garybasin right but the part where it turns that into an instruction is cool given how many of my attempts at getting it to write no-extra-semicolon javascript failed the only parts I can't obviously decompress by just inserting vowels are the "RcS:" at the start and the b in frmbHnst
Poking at cryptic compressed prompts and what ChatGPT does with them
7 tweets · April 2023 · 7 likes · 0 retweets · read on Twitter
@garybasin I tried asking GPT-4 what "RcS:" meant and initially I was like "wow it just told me!" but then it had a bunch of other responses when I hit regen so WHO. THE. FUCK. KNOWS.
@garybasin hmmmm tried getting it to explain in context to see if that would help and it gave yet a *different* meaning for "RcS:" = "respectful, constructive skepticism."
@garybasin looks like the "RcS:" is not some necessary function call or something, based on removing it these things are really opaque! it's not obvious what they're actually doing under the hood. thank god you can tinker and explore tiny prompt variants!
@garybasin also looks like you get basically the same response if you use a natural language condensed role/request at the top (although weirdly it becomes a formal letter (in 2/2 cases!)) (and it critiques a hallucinated UI, not the idea, but I suppose I left out "idea" from the prompt!)
@garybasin let's try a different role/request in simple language yeah, seems like you can basically do that. so as far as I can tell we've decoded 2 facts about chatGPT4-prompting - it can reinflate words with no vowels (no surprise) - it's happy to just get a few-word role/request
@garybasin I included "chat" in "chatGPT4" above because it seems quite likely to me that a huge part of why this works is actually the RLHF which causes chatgpt to be greedily searching for a way to be helpful and seizing on anything that looks like a role altho ofc
