WEBVTT

00:00.300 --> 00:02.490
-: In this video we're gonna explain a topic

00:02.490 --> 00:05.550
called meta prompting or meta prompts, where they're used,

00:05.550 --> 00:08.340
and how you can integrate them into your existing workflows.

00:08.340 --> 00:10.260
Meta prompts allow us the ability

00:10.260 --> 00:13.590
to take an output from either a machine or,

00:13.590 --> 00:15.090
you know, a human output,

00:15.090 --> 00:16.380
and then to figure out,

00:16.380 --> 00:18.900
if that output came from an AI model,

00:18.900 --> 00:21.300
what would've been the prompt that the AI model

00:21.300 --> 00:23.220
choose to use?

00:23.220 --> 00:24.990
And so, the advantages of this is,

00:24.990 --> 00:27.810
if you're struggling for inspiration or you're looking

00:27.810 --> 00:31.020
to get a more deterministic or a more reliable output,

00:31.020 --> 00:33.630
using meta prompting could be a good use case for this.

00:33.630 --> 00:35.100
Let's have a look at an example

00:35.100 --> 00:36.720
that describes this a bit better.

00:36.720 --> 00:40.500
So you can see at the top, we've got this LinkedIn post

00:40.500 --> 00:42.780
and what we really want Chat GPT to do is to

00:42.780 --> 00:44.640
give us a prompt that would allow us

00:44.640 --> 00:47.580
to partially regenerate a type of LinkedIn post

00:47.580 --> 00:48.810
that's similar to this.

00:48.810 --> 00:52.410
So at the top we have, "I want to have a Chat GPT prompt

00:52.410 --> 00:53.970
that could've been used to regenerate

00:53.970 --> 00:55.380
the following LinkedIn post."

00:55.380 --> 00:57.690
We've also got a little bit of repetition here,

00:57.690 --> 01:01.380
so, "Create a chat GPT prompt that could have generated

01:01.380 --> 01:03.540
the following LinkedIn post text."

01:03.540 --> 01:07.170
And so, what we end up with is not necessarily the output,

01:07.170 --> 01:09.840
but we want a prompt that helps us

01:09.840 --> 01:11.610
to generate that output, right?

01:11.610 --> 01:14.190
So that might be one prompt, you know.

01:14.190 --> 01:15.037
Here's another prompt.

01:15.037 --> 01:17.640
"Can you explain marketing mix modeling?"

01:17.640 --> 01:21.510
And so, what we can then do is take these types of outputs

01:21.510 --> 01:23.430
and put that back into the AI model.

01:23.430 --> 01:25.710
And then, it gives us something that's very similar

01:25.710 --> 01:27.930
to what the output was like,

01:27.930 --> 01:30.510
but it's not just limited to using this for text.

01:30.510 --> 01:32.400
We could also use this for code.

01:32.400 --> 01:35.070
So let's say, for example, we've got some code

01:35.070 --> 01:37.350
that we'd like to figure out how it would write.

01:37.350 --> 01:38.880
I've got this data pipeline here.

01:38.880 --> 01:40.560
We've got some imports.

01:40.560 --> 01:42.120
So we could take all this

01:42.120 --> 01:44.160
and, at the top of this, instead of,

01:44.160 --> 01:45.630
so what we would just put,

01:45.630 --> 01:48.930
so let's go to a new chat history, click new chat,

01:48.930 --> 01:49.763
and then I'm just gonna put,

01:49.763 --> 01:52.500
"Create a Chat GPT prompt that would've generated

01:52.500 --> 01:53.547
the following code."

01:54.750 --> 01:56.850
And then we can put the code in.

01:56.850 --> 02:00.907
And then, you'll see now what it's giving us is, you know,

02:00.907 --> 02:04.680
"Write a Python code to extract sentiment labels

02:04.680 --> 02:06.720
for a list of restaurant reviews," right?

02:06.720 --> 02:08.910
So this here would be the prompt.

02:08.910 --> 02:11.310
So we could take this, post this back in,

02:11.310 --> 02:13.560
and here's, you know, it gives us a different example

02:13.560 --> 02:15.630
of how you could do something like this.

02:15.630 --> 02:18.180
And so, you know, this could be quite interesting,

02:18.180 --> 02:20.280
'cause you can see the code is slightly different.

02:20.280 --> 02:21.480
It's using different packages,

02:21.480 --> 02:24.120
but it's achieving the same output nonetheless.

02:24.120 --> 02:27.090
And so, you know, meta prompts allow us to figure out

02:27.090 --> 02:30.180
what would've the prompt been, given the output

02:30.180 --> 02:32.430
that the large language model has seen, right?

02:32.430 --> 02:34.170
And so, you can use this when you're struggling

02:34.170 --> 02:36.240
for inspiration, when you're looking for ideas,

02:36.240 --> 02:38.790
or when you're just looking to generate

02:38.790 --> 02:40.983
more reliable results.