site banner

Friday Fun Thread for May 19, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

3
Jump in the discussion.

No email address required.

To those using ChatGPT to brainstorm, what kind of prompt engineering techniques have you found to be useful?

Something I've been doing is asking chatgpt for a list of responses, and then basically running high order functions on those lists. So you ask for a list of X, and then say for each item do Y, or combine all items into Z. This just lets you look at more than one idea at once, with some turning out better than others, and I think it also results in more creative answers, since you're not just getting the most common idea, but ideas 1 through 5 or 10 that are all going through these transformations. Then you can ask it to combine or shift ideas into each other for more possible creativity.

I haven't had much luck at making gpt-4 come up with genuinely novel ideas for the fields I'm interesting in (Machine Learning, Deep Learning, Reinforcement Learning), the most it can do is serve as a sort of litterature review, where I describe an idea I have, ask it if it exists, and if it does to link me the papers. Sometimes it hallucinates papers that don't exist in this way because it thinks that generating positive answers is more useful than negative ones, but I've gotten real value out of this procedure a few times, where it spit out fairly obscure results that I would've never found on my own.

aah I don't know, I've had a few obscure papers from 2013 and 2014 come up that basically implemented aspects of a few of my projects. And at one point it pointed out that what I thought was a shiny new idea was in fact called "Optimal Importance Sampling", and linked me a paper from 1991 that explained it. So it's still pretty useful unless you're doing stuff that is basically the next obvious thing to try conditional on current research, and as such are worried about people scooping you.