It is sort of fun to bounce little ideas off ChatGPT, but I can’t imagine wanting to read somebody else’s ChatGPT responses.

IMO a lot of the dumb and bad behavior around LLMs could be solved by a “just share the prompts” strategy. If somebody wants to generate an email from bullet points and send it to me: just send the bullet points, and I can pass them into an LLM if I want.

Blog post based on interesting prompts? Share the prompt. It’s just text completion anyway, so if a reader knows more about the topic than the prompt-author, they can even tweak the prompt (throw in some lingo to get the LLM to a better spot in the latent space or whatever).

The only good reason not to do that is to save some energy in generation, but inference is pretty cheap compared to training, right? And the planet is probably doomed anyway at this point so we as well enjoy the ride.