WEBVTT 00:00.180 --> 00:02.910 -: The third principle is providing examples, 00:02.910 --> 00:04.980 inserting a diverse set of test cases 00:04.980 --> 00:07.200 where the task was done correctly. 00:07.200 --> 00:09.030 Now this is really important. 00:09.030 --> 00:10.506 In fact, it's one of the main things 00:10.506 --> 00:13.920 that researchers test when they're trying out a new model 00:13.920 --> 00:15.270 and they see how well it does 00:15.270 --> 00:18.960 with no examples versus one example versus three or five 00:18.960 --> 00:21.150 or 100 examples of the task. 00:21.150 --> 00:23.970 And that's called zero shock or few shock. 00:23.970 --> 00:25.980 And the reason why we put this third, 00:25.980 --> 00:26.947 even though it's so important, 00:26.947 --> 00:31.290 is that it's the more costly prompt engineering tactic. 00:31.290 --> 00:33.150 Providing direction is usually pretty quick. 00:33.150 --> 00:35.259 You can just say in the style of Steve Jobs 00:35.259 --> 00:38.490 or if you're specifying the format, you can just say, 00:38.490 --> 00:41.910 I want this in terms of a comma separated list. 00:41.910 --> 00:43.920 Whereas providing examples, you have to go out 00:43.920 --> 00:46.140 and find good examples and cherry pick them 00:46.140 --> 00:49.020 and put them into the prompt in some way. 00:49.020 --> 00:52.230 So let's look at how this works with text models. 00:52.230 --> 00:54.589 This is our generic prompt where we're just looking 00:54.589 --> 00:57.450 for a list of 10 product names for a pair of shoes 00:57.450 --> 00:59.310 that can fit any foot size. 00:59.310 --> 01:02.820 But our engineered prompt actually includes some examples. 01:02.820 --> 01:05.191 We provided three examples of the types 01:05.191 --> 01:07.290 of names that we like. 01:07.290 --> 01:10.890 Now in the product template, we have a space for those. 01:10.890 --> 01:13.590 You can see in the hashtag Examples. 01:13.590 --> 01:15.000 And then we have the curly braces 01:15.000 --> 01:17.553 where we insert the product examples there. 01:18.450 --> 01:19.709 And typically, I would like to put 01:19.709 --> 01:21.592 that in the middle of the prompt. 01:21.592 --> 01:23.070 You start with the task 01:23.070 --> 01:24.352 'cause that's the most important thing. 01:24.352 --> 01:26.220 You give it some style direction 01:26.220 --> 01:28.140 or roles that it needs to play. 01:28.140 --> 01:32.673 And then you would also insert the examples underneath that. 01:34.446 --> 01:36.510 Now you can see that in our examples, 01:36.510 --> 01:39.840 not only did we say we want it in the style of Steve Jobs, 01:39.840 --> 01:43.757 but we also made every single product name start with an i. 01:43.757 --> 01:48.757 And that has fixed the creativity to a specific point. 01:48.930 --> 01:51.635 Now, Steve Jobs didn't name every single product with an i, 01:51.635 --> 01:53.591 but if you do this, 01:53.591 --> 01:57.060 then you will get only names that come out with is. 01:57.060 --> 02:00.270 And that's an important thing to note about examples, 02:00.270 --> 02:03.720 is that they are do make the prompt much more reliable 02:03.720 --> 02:05.400 but at the cost of creativity. 02:05.400 --> 02:08.790 You're constraining creativity to just give product names 02:08.790 --> 02:10.080 that start with i. 02:10.080 --> 02:13.410 Now, if you want to give a more diverse set of examples, 02:13.410 --> 02:16.923 you're gonna get a more diverse set of responses as well. 02:17.820 --> 02:20.994 Now let's see how we can change the results 02:20.994 --> 02:24.690 that we get based on the variable that we've input. 02:24.690 --> 02:28.200 So in this case, I went back and I changed all of the names 02:28.200 --> 02:31.830 in the examples to start with Apple and then the product. 02:31.830 --> 02:32.664 And you can see that the responses about Apple, 02:32.664 --> 02:36.600 Fit, Apple FlexWalk, Apple OmniFit. 02:36.600 --> 02:39.150 So we constrained it in that direction now instead. 02:39.150 --> 02:40.740 If you add the word mini to everything, 02:40.740 --> 02:43.170 that adds the word mini to every response. 02:43.170 --> 02:45.000 If you begin with the word Air, 02:45.000 --> 02:47.370 then it will come back with different names there. 02:47.370 --> 02:50.484 So it just goes to show how much of an impact on results 02:50.484 --> 02:53.013 that changing the examples can have. 02:54.600 --> 02:56.100 Now let's look at image models 02:56.100 --> 02:57.810 where it works a little bit differently. 02:57.810 --> 02:59.310 You can't usually give more 02:59.310 --> 03:01.530 than one example with image models. 03:01.530 --> 03:03.330 And you see in the naive prompt, 03:03.330 --> 03:05.220 we didn't give any examples. 03:05.220 --> 03:07.200 So it's gone in lots of different directions. 03:07.200 --> 03:08.760 It's pretty creative there, right? 03:08.760 --> 03:11.730 Like, we have this old-man-looking shoe at the top, 03:11.730 --> 03:15.330 we have the robot shoe with a robot leg at the bottom. 03:15.330 --> 03:16.800 The one on the bottom right 03:16.800 --> 03:18.660 actually looks like an astronaut shoe. 03:18.660 --> 03:20.640 Now, in terms of prompt engineering, 03:20.640 --> 03:23.460 we can see that the engineered prompt 03:23.460 --> 03:25.890 actually gives us pretty reliable results. 03:25.890 --> 03:27.546 Of all four of the options, 03:27.546 --> 03:30.480 they all constrained to the same shape. 03:30.480 --> 03:34.260 That's because we included this one example here, the image. 03:34.260 --> 03:37.050 In the prompt template, we don't list this, 03:37.050 --> 03:40.741 but there's a place to upload an example image 03:40.741 --> 03:42.210 for Stable Diffusion. 03:42.210 --> 03:45.240 They call this image-to-image in the literature. 03:45.240 --> 03:49.020 And you can see here we've uploaded an image of... 03:49.020 --> 03:52.545 This is specifically a Nike shoe that has this weird bend 03:52.545 --> 03:56.220 because it's flexible, and we've uploaded that, 03:56.220 --> 03:59.190 but we set it at 35% weight. 03:59.190 --> 04:00.600 And you can change the weight from zero 04:00.600 --> 04:02.300 to 100% would make it very similar, 04:02.300 --> 04:04.620 and 0% would make it very dissimilar. 04:04.620 --> 04:06.870 But we've settled on 35%. 04:06.870 --> 04:08.370 We can see we're getting pretty good results. 04:08.370 --> 04:09.690 We get this weird bend. 04:09.690 --> 04:14.040 We also take from the example the black and the white. 04:14.040 --> 04:16.557 Now, again, in terms of constraining creativity, 04:16.557 --> 04:20.490 one thing I found in testing this is that I couldn't get it 04:20.490 --> 04:22.710 to do any colors other than black and white 04:22.710 --> 04:26.340 because it's reading too much into this one example here. 04:26.340 --> 04:27.990 You can see I managed to get the blue 04:27.990 --> 04:29.460 and the pink highlights in there, 04:29.460 --> 04:32.160 but I couldn't change the color of the sole 04:32.160 --> 04:35.223 or the, you know, overall background color of the shoe. 04:36.540 --> 04:38.430 So here's a few examples 04:38.430 --> 04:42.000 where I've changed the results based on, literally, 04:42.000 --> 04:44.280 the same prompt, just a different background image, 04:44.280 --> 04:45.660 different upload image. 04:45.660 --> 04:47.190 Here, I uploaded some wool shoes 04:47.190 --> 04:49.470 and I got a very weird response. 04:49.470 --> 04:51.390 This one here, the fuzzy one. 04:51.390 --> 04:53.449 I uploaded a pair of Crocs and got this, 04:53.449 --> 04:56.010 and then I uploaded some platform heels and got this. 04:56.010 --> 04:57.769 So you can see, you can really change the results, 04:57.769 --> 04:59.580 and the model takes a lot 04:59.580 --> 05:02.940 of steering from the example image that you upload. 05:02.940 --> 05:05.793 So choosing a good example image makes a big difference. 05:05.793 --> 05:08.223 And both in terms of the format 05:08.223 --> 05:11.310 and the form, but also the style. 05:11.310 --> 05:13.800 So if you have an image that works really well, 05:13.800 --> 05:15.545 that captures your style pretty well, 05:15.545 --> 05:18.556 then you can use that as an example 05:18.556 --> 05:20.970 when you're doing any type of image. 05:20.970 --> 05:23.570 And that's gonna give you pretty consistent results.