WEBVTT 1 00:00:00.000 --> 00:00:02.100 What is the difference between prompt engineering 2 00:00:02.100 --> 00:00:04.320 and just simply prompting these models? 3 00:00:04.320 --> 00:00:06.090 Let's start with a naive prompt. 4 00:00:06.090 --> 00:00:08.070 This is something that you might type in 5 00:00:08.070 --> 00:00:09.060 if you want a blog post. 6 00:00:09.060 --> 00:00:12.417 So write a blog post on prompt engineering. 7 00:00:12.417 --> 00:00:15.690 I'm just gonna hit enter and it will do a pretty good job. 8 00:00:15.690 --> 00:00:17.880 And these models are really fantastic. 9 00:00:17.880 --> 00:00:19.110 They get better and better, 10 00:00:19.110 --> 00:00:20.767 and that leads some people to think, 11 00:00:20.767 --> 00:00:23.040 "Eventually we won't need prompt engineering." 12 00:00:23.040 --> 00:00:24.480 But I think that's completely false 13 00:00:24.480 --> 00:00:27.300 because these models are really smart, 14 00:00:27.300 --> 00:00:29.670 but you're not giving 'em any direction here. 15 00:00:29.670 --> 00:00:30.620 This is just a naive prompt. 16 00:00:30.620 --> 00:00:32.730 If 100 people type this in, 17 00:00:32.730 --> 00:00:36.000 they're gonna get 100 very similar sounding blog posts, 18 00:00:36.000 --> 00:00:38.310 and there's not gonna be any real context in there. 19 00:00:38.310 --> 00:00:42.090 There's not gonna be any type of refinement to that prompt, 20 00:00:42.090 --> 00:00:42.950 and you need those sorts of things 21 00:00:42.950 --> 00:00:45.210 in order to get to something really good, 22 00:00:45.210 --> 00:00:46.560 that's human level 23 00:00:46.560 --> 00:00:49.533 because what we don't want is an internet full of blog posts 24 00:00:49.533 --> 00:00:52.170 that are just the default average 25 00:00:52.170 --> 00:00:55.440 of the internet-type response that we get from ChatGPT. 26 00:00:55.440 --> 00:00:56.940 Now, prompt engineering 27 00:00:56.940 --> 00:00:59.460 is really adding a lot of that secret sauce 28 00:00:59.460 --> 00:01:00.780 from your own skill, 29 00:01:00.780 --> 00:01:03.210 as well as testing what works reliably 30 00:01:03.210 --> 00:01:05.490 because if you're gonna be generating one blog post, 31 00:01:05.490 --> 00:01:06.690 it doesn't matter that much. 32 00:01:06.690 --> 00:01:09.060 If you're gonna generate 100 in a day 33 00:01:09.060 --> 00:01:12.420 or 1,000 in a week, then you need prompt engineering 34 00:01:12.420 --> 00:01:17.220 because if 10 in 100 of those 1,000 blog posts are wrong 35 00:01:17.220 --> 00:01:19.140 in some way or require human editing, 36 00:01:19.140 --> 00:01:21.540 you're not really saving yourself that much time. 37 00:01:21.540 --> 00:01:24.090 If you can get them as high quality as possible 38 00:01:24.090 --> 00:01:26.100 just by adapting the prompt template, 39 00:01:26.100 --> 00:01:28.170 and then the higher quality translates 40 00:01:28.170 --> 00:01:31.440 to every single time you run that prompt template. 41 00:01:31.440 --> 00:01:34.620 Let me show you what an engineered prompt might look like. 42 00:01:34.620 --> 00:01:35.970 And this is the prompt 43 00:01:35.970 --> 00:01:37.800 that we're gonna be working through on this course, 44 00:01:37.800 --> 00:01:40.770 and you can see how every part works together, 45 00:01:40.770 --> 00:01:43.380 but you can see there's a lot more information here. 46 00:01:43.380 --> 00:01:45.390 I'm setting a writing style, 47 00:01:45.390 --> 00:01:47.700 I'm using the Hero's Journey framework. 48 00:01:47.700 --> 00:01:49.710 I'm asking it to outline the story for it, 49 00:01:49.710 --> 00:01:52.710 telling it I want a sci-fi-type fantasy post 50 00:01:52.710 --> 00:01:56.190 rather than the generic, average prompt engineering guide 51 00:01:56.190 --> 00:01:57.450 that you see out there. 52 00:01:57.450 --> 00:01:59.280 I'm also providing instructions 53 00:01:59.280 --> 00:02:02.040 to combat specific edge cases that I found 54 00:02:02.040 --> 00:02:03.870 where it's not writing long enough 55 00:02:03.870 --> 00:02:05.220 and we're gonna learn how to do all 56 00:02:05.220 --> 00:02:06.810 of these things and more. 57 00:02:06.810 --> 00:02:08.280 But just wanted to show you 58 00:02:08.280 --> 00:02:09.690 what this type of thing looks like. 59 00:02:09.690 --> 00:02:11.220 Here we have an outline 60 00:02:11.220 --> 00:02:13.410 where it's broken down the whole narrative first. 61 00:02:13.410 --> 00:02:15.660 Then it started to write the blog post, 62 00:02:15.660 --> 00:02:18.990 and you can see that it's writing much longer type text. 63 00:02:18.990 --> 00:02:22.230 So it's writing two paragraphs here for each section 64 00:02:22.230 --> 00:02:23.910 'cause there's gonna be 12 sections 65 00:02:23.910 --> 00:02:25.740 to the hero's journey stages. 66 00:02:25.740 --> 00:02:28.530 Then we're gonna have something like a 1,000 67 00:02:28.530 --> 00:02:29.977 to 2,000-word blog post. 68 00:02:29.977 --> 00:02:31.680 I'm gonna show you how to do this, 69 00:02:31.680 --> 00:02:34.170 how to get to this more complicated prompt iteratively 70 00:02:34.170 --> 00:02:36.270 without taking too much time and effort. 71 00:02:36.270 --> 00:02:38.610 And once you get to this point, you'll be able 72 00:02:38.610 --> 00:02:41.910 to make prompts that are equally as reliable 73 00:02:41.910 --> 00:02:44.490 and get human-level results 74 00:02:44.490 --> 00:02:47.040 because most people when they try ChatGPT 75 00:02:47.040 --> 00:02:48.600 and they say it's not very good, 76 00:02:48.600 --> 00:02:51.630 it's because they just don't know how to ask the right way. 77 00:02:51.630 --> 00:02:53.610 And that's what you're gonna learn here. 78 00:02:53.610 --> 00:02:57.330 We're gonna also dive into image generation as well 79 00:02:57.330 --> 00:03:00.180 because we could do something similar there too. 80 00:03:00.180 --> 00:03:02.400 These principles apply to image generation, 81 00:03:02.400 --> 00:03:03.840 as well as text generation. 82 00:03:03.840 --> 00:03:06.150 So I'm just gonna paste in here create an image 83 00:03:06.150 --> 00:03:07.620 that matches this blog post. 84 00:03:07.620 --> 00:03:10.050 This is the naive way to do it 85 00:03:10.050 --> 00:03:11.910 because we, you know, want just an image 86 00:03:11.910 --> 00:03:14.010 that matches the blog post and we already have a lot 87 00:03:14.010 --> 00:03:15.300 of prompting in the text side, 88 00:03:15.300 --> 00:03:17.670 so it's gonna create a pretty good image. 89 00:03:17.670 --> 00:03:19.800 The cool thing about DALL-E and ChatGPT 90 00:03:19.800 --> 00:03:22.470 is that ChatGPT actually writes the prompt for you, 91 00:03:22.470 --> 00:03:24.090 makes it a bit more sophisticated, 92 00:03:24.090 --> 00:03:26.760 but we might not get the type of image we really want 93 00:03:26.760 --> 00:03:28.980 and we're not really getting very much choice here. 94 00:03:28.980 --> 00:03:30.420 It's just gonna create one image. 95 00:03:30.420 --> 00:03:33.060 So we can see here this is a pretty cool image 96 00:03:33.060 --> 00:03:34.470 and it looks futuristic, 97 00:03:34.470 --> 00:03:36.030 like the sort of thing we're looking at 98 00:03:36.030 --> 00:03:37.980 in terms of the blog posts. 99 00:03:37.980 --> 00:03:39.150 So it does match pretty well. 100 00:03:39.150 --> 00:03:41.190 These are good results straight away. 101 00:03:41.190 --> 00:03:44.100 But a more optimized prompt might be something like this 102 00:03:44.100 --> 00:03:46.170 where I've written my own image prompt, 103 00:03:46.170 --> 00:03:47.820 so I'm just pasting this in here. 104 00:03:47.820 --> 00:03:49.920 I'm asking it to create different matching images 105 00:03:49.920 --> 00:03:51.960 to this blog post and this style. 106 00:03:51.960 --> 00:03:56.040 And I'm talking about solarpunk, which is a type of scene 107 00:03:56.040 --> 00:03:57.360 that I researched and found 108 00:03:57.360 --> 00:03:59.580 that I really like that visual style. 109 00:03:59.580 --> 00:04:01.650 And then talking about what's actually happening 110 00:04:01.650 --> 00:04:04.590 in the image, giving a bit more understanding 111 00:04:04.590 --> 00:04:07.680 of the composition, telling it what's in the background, 112 00:04:07.680 --> 00:04:08.820 all sorts of different things. 113 00:04:08.820 --> 00:04:10.440 Just gonna paste that in here 114 00:04:10.440 --> 00:04:12.660 and then we'll see what type of image we get back. 115 00:04:12.660 --> 00:04:13.980 Cool, so here's the first image. 116 00:04:13.980 --> 00:04:16.710 And you can see it's a very different visual style 117 00:04:16.710 --> 00:04:18.630 and we use prompt engineering to get here. 118 00:04:18.630 --> 00:04:21.960 This is the type of style that I wanted and it's futuristic. 119 00:04:21.960 --> 00:04:23.880 We're futuristic in a very different way, 120 00:04:23.880 --> 00:04:25.650 and you could try different styles 121 00:04:25.650 --> 00:04:28.110 once you realize what words you need to use 122 00:04:28.110 --> 00:04:31.590 to elicit that style from ChatGPT and Dall-E. 123 00:04:31.590 --> 00:04:32.820 That's what you're gonna learn. 124 00:04:32.820 --> 00:04:35.850 It's gonna be transferable across not just text and image, 125 00:04:35.850 --> 00:04:39.330 but also video, audio, really any type 126 00:04:39.330 --> 00:04:42.903 of AI model now and way into the future.