WEBVTT 00:00.180 --> 00:01.170 -: Hey, welcome back. 00:01.170 --> 00:02.970 And in this section we'll be exploring something 00:02.970 --> 00:05.280 called Asking the Model For More Context. 00:05.280 --> 00:07.680 We'll see two different types of scenarios 00:07.680 --> 00:09.270 and how you can apply this. 00:09.270 --> 00:12.270 Let's take something like a very broad prompt, 00:12.270 --> 00:15.450 which is gonna give us not a very specific output, 00:15.450 --> 00:18.510 and see how we could ask the model for more context 00:18.510 --> 00:20.580 about the output that it generates. 00:20.580 --> 00:22.567 In this example, we have a prompt that says, 00:22.567 --> 00:25.350 "Goals: I'm looking to improve my fitness, 00:25.350 --> 00:27.030 health, and wealth. 00:27.030 --> 00:29.760 Provide a very detailed, numerical, 00:29.760 --> 00:32.490 hierarchical outline for the above goals." 00:32.490 --> 00:35.040 You'll see that we get some information 00:35.040 --> 00:37.920 about fitness, health, and wealth. 00:37.920 --> 00:41.490 But we have parts of this are very vague, 00:41.490 --> 00:43.140 we don't really understand them. 00:43.140 --> 00:45.270 We could then ask ChatGPT 00:45.270 --> 00:47.527 to drill down into these by saying, 00:47.527 --> 00:51.540 "Could you please provide more context 00:51.540 --> 00:55.410 about ways that I can do this?" 00:55.410 --> 00:58.380 And so what we're doing is we're taking a very broad goal, 00:58.380 --> 01:02.190 and we're asking the model to give us more context 01:02.190 --> 01:04.053 about what it's decided to do. 01:04.920 --> 01:08.190 As well as that, we could also ask it for more context 01:08.190 --> 01:10.830 about why it decided to generate an output. 01:10.830 --> 01:13.020 So, here we've got, kind of, 01:13.020 --> 01:15.540 a bit more deeper level information 01:15.540 --> 01:18.780 about improving our diet and nutrition. 01:18.780 --> 01:19.740 That's great. 01:19.740 --> 01:20.880 And then as well as that, 01:20.880 --> 01:22.860 let's pick something that's been quite specific. 01:22.860 --> 01:27.860 So maybe "Why should we," Why should we," 01:27.900 --> 01:30.960 and then, "reduce sugar intake?" 01:30.960 --> 01:34.290 So we can kind of change this into a question 01:34.290 --> 01:38.040 which allows us to really understand, 01:38.040 --> 01:41.130 not just from a broad perspective why we should do things, 01:41.130 --> 01:43.440 but getting some type of evidence. 01:43.440 --> 01:45.690 And so this is quite a useful technique 01:45.690 --> 01:48.030 when you're learning about different types of things. 01:48.030 --> 01:50.220 Maybe there's something you don't understand. 01:50.220 --> 01:53.437 You can even ask it, and ask the model for more context. 01:53.437 --> 01:56.760 "Can you explain this more simply?" 01:56.760 --> 02:01.230 So that can be an example of asking for more context, 02:01.230 --> 02:03.480 is asking for the context in a different way 02:03.480 --> 02:05.190 so that you can understand it better. 02:05.190 --> 02:07.620 So there's a couple of different use cases here. 02:07.620 --> 02:10.530 Let's have another look at a different use case, 02:10.530 --> 02:12.990 which is kind of similar, but slightly different. 02:12.990 --> 02:15.157 So, what we've got in this initial prompt is, 02:15.157 --> 02:17.880 "I'm looking to ask for a promotion at my job, 02:17.880 --> 02:20.580 use the following information to generate 02:20.580 --> 02:23.400 an effective and engaging letter to my boss." 02:23.400 --> 02:25.680 So ChatGPT is now confident 02:25.680 --> 02:27.843 about what the task is meant to be doing, 02:28.770 --> 02:32.160 and we're also confident about the information 02:32.160 --> 02:33.060 that we've given it. 02:33.060 --> 02:35.670 But what happens if we didn't give it all the information 02:35.670 --> 02:37.620 where it could generate a really good answer? 02:37.620 --> 02:41.267 So one way to deal with this is to give ChatGPT 02:41.267 --> 02:43.260 or GPT-3 an out clause, 02:43.260 --> 02:45.150 by saying, "Before writing the letter, 02:45.150 --> 02:46.770 if you don't have all the information 02:46.770 --> 02:49.260 to do the task, avoid," 02:49.260 --> 02:51.000 and then whatever the task you said it was, 02:51.000 --> 02:53.190 so in this case, "generating an effective 02:53.190 --> 02:55.050 and engaging letter to my boss 02:55.050 --> 02:57.180 and ask for more information." 02:57.180 --> 02:59.190 ChatGPT has decided at this point 02:59.190 --> 03:01.170 that it doesn't have enough information 03:01.170 --> 03:02.940 to make a good output for us. 03:02.940 --> 03:04.980 And so we are kind of using it as an agent. 03:04.980 --> 03:07.650 And then if you'll just give me a second, 03:07.650 --> 03:08.880 I can paste in, 03:08.880 --> 03:12.540 this is provided with all the extra information, 03:12.540 --> 03:15.750 now it's able to then provide a better cover letter 03:15.750 --> 03:17.040 or a better written letter. 03:17.040 --> 03:20.640 So rather than just telling ChatGPT or GPT-3 03:20.640 --> 03:22.537 to produce an output, we can say, 03:22.537 --> 03:24.990 "Hey, if you don't think that you could produce 03:24.990 --> 03:27.960 a reliable output, please ask for more context." 03:27.960 --> 03:31.110 And in this case, the model is asking for more context 03:31.110 --> 03:33.480 and we're not necessarily asking for more context. 03:33.480 --> 03:36.870 And so this can help improve the liabilities of your output 03:36.870 --> 03:40.740 because now ChatGPT has all of the relevant information 03:40.740 --> 03:42.843 inside of it to make a better decision.