How to write more effective AI prompts for better results, according to Microsoft’s chief scientists

How to write more effective AI prompts for better results.

Transcript

00:00 A good prompt strategy will actually give you better outputs than a bigger, more powerful

00:06 model.

00:07 So, actually, one of the things that's important is to recognize that a language model isn't

00:13 a person, and when you're prompting it, you need to prompt it a little differently.

00:17 So, when we were chatting backstage, we did a fair amount of grounding.

00:21 This is that process of getting to know each other, but because they're humans, I don't

00:25 have to do the same level of grounding that I have to do with a language model.

00:29 With a language model, you have to be super explicit and say, like, you are playing this

00:33 role, and I want you to produce an output in this format for this audience.

00:38 So that's very explicit, and that's because it's not a human.

00:42 But there's other things that actually you can use to your advantage with the language

00:47 model not being human.

00:49 You know, I can't be like, oh, my gosh, I said something really dumb.

00:53 Let's just start over and try again when I'm chatting with these folks.

00:57 With a language model, you can actually just erase all your history and erase all the context

01:02 and start the conversation again.

01:05 And so you can actually start doing things like I don't know if you know the concept

01:07 of the wisdom of the crowds, where you, like, ask a bunch of people how many jelly beans

01:13 are in a jar, and if I asked this whole room, we'd on aggregate get the right number of

01:17 jelly beans.

01:19 You can start doing that with a language model.

01:20 You can ask it to play different roles or produce answers in different contexts and

01:24 then look at the sum total of what kinds of answers you might get.