20
$\begingroup$

The Stack Exchange company just posted this: Starting the Prompt Design Site: A New Home in our Stack Exchange Neighborhood, which led me to wonder if this topic is already be covered by this site or not. Obviously highly localized questions aren't good, so I'm not asking about that.

I see that there is , which has

Are any questions about how to design prompts for things like LLMs in applications like ChatGPT on-topic for this site? If so, what qualifies them to be on-topic?

I ask this as someone who knows nothing about Prompt Design or how big of a thing it is, but has heard it in a couple places and now is hearing about things like "Prompt Engineering", so apologies if this comes off as silly in any way.

$\endgroup$
3
  • 2
    $\begingroup$ I'm not a member of this site, so I don't dare post an answer, but if I had to guess, I'd say that the main reasons include the fact that there is no way to objectively answer these questions. They are effectively "guess and check", which makes them inherently implementation-dependent and subject to being rapidly obsoleted by the next genAI or even just the next update of the training set. That makes them inherently low quality and not the fodder for a healthy site. They don't form a useful knowledge base. Yes, the implications of this comment are obvious. $\endgroup$ Commented Jun 21, 2023 at 10:13
  • 5
    $\begingroup$ @CodyGray-onstrike There are ways to objectively answer the questions about the public models (e.g. GPT-2, LLaMA). Just not the proprietary ones (GPT-4, GLaDOS). $\endgroup$
    – wizzwizz4
    Commented Jun 21, 2023 at 10:41
  • $\begingroup$ Can we just do a meta question like Arqade does for screen shots of the week? Preferably starting with, prompts for Bard to make them a DM? $\endgroup$
    – Mazura
    Commented Jul 7, 2023 at 1:34

1 Answer 1

14
$\begingroup$

I don't personally have an extremely strong opinion, and would be open to hearing if the community disagrees.

But right now, I'd lean more towards saying that prompt design / prompt engineering questions would often be off-topic here. Our site is primarily about AI itself, not about how to use tools that run AI internally, if that makes sense. Asking how best to design a prompt for a language model feels like it would be similar to... how to speak more clearly if Apple's Siri or Amazon's Alexa misunderstands your questions/commands. Or how best to play against or alongside an AI-based player in a multi-player game.

The primary gray area where I could see some overlap is once you take a deep dive into how the model was trained or implemented in order to explain or hypothesize why a certain kind of prompt does or does not work well. But I expect that the vast majority of prompt design questions would not be like this.

$\endgroup$
2
  • $\begingroup$ "deep dive into how the model was trained or implemented in order to explain" An addendum: at the current moment, transformers are not explainable whatsoever and they make up pretty much every model you care to prompt engineer. $\endgroup$
    – Passer By
    Commented Jun 23, 2023 at 7:53
  • $\begingroup$ Not more clearly, it's that the first word you should ever say to Siri is google, so that it shows a search, instead of its best guess with nowhere to go from or to. I suppose it's not ontopic at PLD either with their 6 Qs with prompt in it. - We have to wait for that new site that noone wants and is at -400 on meta? $\endgroup$
    – Mazura
    Commented Jul 7, 2023 at 1:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .