WEBVTT 00:00.420 --> 00:03.870 -: Okay, the final principle is division of labor. 00:03.870 --> 00:06.480 That's where you split tasks into multiple steps 00:06.480 --> 00:08.100 Because you can rarely get something done 00:08.100 --> 00:09.690 with just one prompt. 00:09.690 --> 00:11.310 You are gonna chain multiple prompts 00:11.310 --> 00:13.320 or multiple (indistinct) steps together 00:13.320 --> 00:15.690 to achieve a complex goal. 00:15.690 --> 00:17.220 So let's look at text models. 00:17.220 --> 00:19.440 Again, this is our naive prompt. 00:19.440 --> 00:21.990 We're just getting a list of 10 product names 00:21.990 --> 00:25.080 and then this is the engineered prompt. 00:25.080 --> 00:29.580 This is a much bigger prompt is much more to it, 00:29.580 --> 00:32.040 but we're not worried about the prompt itself specifically. 00:32.040 --> 00:33.960 We're worried about a prompt that comes after. 00:33.960 --> 00:36.030 So this isn't just one prompt. 00:36.030 --> 00:38.130 We get the response from ChatGPT, 00:38.130 --> 00:40.279 which gives us the names of iFitShoe, 00:40.279 --> 00:42.270 iAdapt, iUniversalStep. 00:42.270 --> 00:43.950 But then we have a second prompt 00:43.950 --> 00:46.110 that comes after that in the same session 00:46.110 --> 00:48.150 and that asks it to rate the names, 00:48.150 --> 00:49.950 because that's something 00:49.950 --> 00:51.990 that you would have to do as a human otherwise. 00:51.990 --> 00:56.040 And in this case, if you design in a rating step, 00:56.040 --> 00:59.370 you can maybe eliminate some of the time spent 00:59.370 --> 01:00.750 choosing a name, 01:00.750 --> 01:03.600 because that's not important when you're just brainstorming. 01:03.600 --> 01:05.490 You can just manually look at the names. 01:05.490 --> 01:07.740 But if you're generating a list of a hundred names, 01:07.740 --> 01:09.930 or if you're running this prompt a hundred times 01:09.930 --> 01:11.910 and then finding the most catchy, most unique, 01:11.910 --> 01:15.630 most simplistic, then that would make a big difference. 01:15.630 --> 01:18.300 In the prompt template, it's not just one template. 01:18.300 --> 01:19.650 We have this divider, 01:19.650 --> 01:21.990 then we have this second prompt underneath. 01:21.990 --> 01:24.720 So it's actually more of a chat session 01:24.720 --> 01:27.780 than an individual prompt. 01:27.780 --> 01:29.610 And if we zoom in on this, 01:29.610 --> 01:32.040 we can see that it takes the names 01:32.040 --> 01:33.390 that we've given it before, 01:33.390 --> 01:36.840 it has the context from the previous part of the session, 01:36.840 --> 01:40.770 and then it's actually giving you back the names in a table. 01:40.770 --> 01:43.860 And it's rating, such as, catchiness 01:43.860 --> 01:46.080 uniqueness and simplicity. 01:46.080 --> 01:47.367 Do the same thing through the API. 01:47.367 --> 01:49.320 Now the key is that we're actually applying 01:49.320 --> 01:52.050 all the other principles of prompting to this as well. 01:52.050 --> 01:53.700 Like we've specified the format, 01:53.700 --> 01:54.810 we want it in the table 01:54.810 --> 01:56.160 and we give it some direction. 01:56.160 --> 01:58.020 We want it to rate based on catchiness 01:58.020 --> 01:59.880 uniqueness and simplicity. 01:59.880 --> 02:03.330 Each prompt in the chain itself can be optimized. 02:03.330 --> 02:05.580 Now there are other parts of the chain 02:05.580 --> 02:06.570 you might want to add on. 02:06.570 --> 02:07.980 It's not just the rating step, 02:07.980 --> 02:09.570 but there's a step beforehand. 02:09.570 --> 02:13.440 Maybe wanna do some retrieval and list some related products 02:13.440 --> 02:16.350 and use that output as the input into your prompts. 02:16.350 --> 02:18.510 And say, we have Nike ID, it's a competitor. 02:18.510 --> 02:21.870 We have a Reebok Adjustable Shoes, uses a competitor. 02:21.870 --> 02:23.940 We can also do things like function calling, 02:23.940 --> 02:26.490 where you actually tell the AI 02:26.490 --> 02:28.110 it can cause specific functions, 02:28.110 --> 02:31.500 and you give it an understanding of what those functions do. 02:31.500 --> 02:34.200 And then you can ask it, you know, 02:34.200 --> 02:36.000 when it's ready to do that. 02:36.000 --> 02:38.040 In this case, you could give it a function 02:38.040 --> 02:39.510 that calls GoDaddy! 02:39.510 --> 02:41.700 and then checks whether the domain is available. 02:41.700 --> 02:43.140 And that's gonna make the big difference 02:43.140 --> 02:46.110 because now you don't have to go do that yourself. 02:46.110 --> 02:48.840 And then another step, which is really powerful, 02:48.840 --> 02:50.610 is image generation, right? 02:50.610 --> 02:52.181 Like we've been seeing, 02:52.181 --> 02:53.014 we can go and generate an image 02:53.014 --> 02:55.020 of that shoe as part of, you know, 02:55.020 --> 02:56.613 the next part in the chain. 02:57.990 --> 03:00.000 Alright, so let's look at at image generation now 03:00.000 --> 03:03.120 and see how this principle applies to that too. 03:03.120 --> 03:04.410 And typically, 03:04.410 --> 03:06.990 you're not just using a one prompt in Stable Diffusion. 03:06.990 --> 03:10.050 When you see artists that are using AI, 03:10.050 --> 03:12.540 they're not just getting it right in one prompt. 03:12.540 --> 03:14.130 They're trying lots of different prompts. 03:14.130 --> 03:15.240 They're typically working in something 03:15.240 --> 03:17.610 like Photoshop where they're stitching 03:17.610 --> 03:20.250 different AI generated images together in a collage 03:20.250 --> 03:23.010 and then pulling them all together and editing, 03:23.010 --> 03:25.020 joining them together with generated film. 03:25.020 --> 03:26.160 Let's see how that works. 03:26.160 --> 03:28.200 We have this actual kind of prompt 03:28.200 --> 03:30.180 that we're running now, the template. 03:30.180 --> 03:32.910 But the problem with this is that the output, 03:32.910 --> 03:36.060 we have a white background, but we still have a problem 03:36.060 --> 03:38.430 because we, you know, we don't, 03:38.430 --> 03:41.040 we can't really use a white background necessarily. 03:41.040 --> 03:44.040 We need to cut out that shoe from the (indistinct). 03:44.040 --> 03:46.740 It's pretty hard to do in Stable Diffusion, 03:46.740 --> 03:48.960 what we do is we divide the labor, 03:48.960 --> 03:51.030 we split that out, that task out, 03:51.030 --> 03:53.070 we give it to another AI model. 03:53.070 --> 03:55.680 In this case, I'm using remove.bg, 03:55.680 --> 03:56.670 which is a free website 03:56.670 --> 03:59.520 where it just removes the background from the image. 03:59.520 --> 04:02.670 And now I have a proper PNG of the shoe. 04:02.670 --> 04:05.400 I can use that in multiple steps. 04:05.400 --> 04:07.530 And so, you know, that shoe is just one part 04:07.530 --> 04:09.120 of an overall advertisement 04:09.120 --> 04:12.360 or overall marketing collateral for the product. 04:12.360 --> 04:14.370 And you can see we can make a big difference 04:14.370 --> 04:16.140 to how the shoe is perceived 04:16.140 --> 04:18.210 just by putting it in different scenarios. 04:18.210 --> 04:20.760 And again, each one of these is generated 04:20.760 --> 04:21.870 with a separate step. 04:21.870 --> 04:24.750 So we have, you know, a background of New York City, 04:24.750 --> 04:26.280 we have a background of a spaceship, 04:26.280 --> 04:27.930 we have a background of a festival. 04:27.930 --> 04:30.210 Whatever it is you want to appeal to, 04:30.210 --> 04:33.300 in terms of an audience, you can do that now 04:33.300 --> 04:34.830 because you can take the shoe out 04:34.830 --> 04:37.890 of the context without changing what the shoe looks like 04:37.890 --> 04:39.630 and put it in a new context 04:39.630 --> 04:42.210 that was generated in a separate step. 04:42.210 --> 04:44.460 You'll find that this is just a simple example 04:44.460 --> 04:47.160 where you can use things like in painting to do this 04:47.160 --> 04:49.290 and make sure the shadows match. 04:49.290 --> 04:52.260 You can use our (indistinct) 04:52.260 --> 04:55.650 that you're really not using one model, one prompt, 04:55.650 --> 04:56.700 and then you're done.