WEBVTT 1 00:00:00.120 --> 00:00:04.719 So imagine this, What if you could talk to AI, like, 2 00:00:04.919 --> 00:00:06.960 really talk to it so it just gets exactly what 3 00:00:07.000 --> 00:00:09.119 you mean and it responds perfectly. 4 00:00:09.240 --> 00:00:12.839 Yeah, like a genuine meeting of minds, almost exactly. 5 00:00:13.199 --> 00:00:16.320 Well, today we're diving deep into something called prompt engineering. 6 00:00:17.160 --> 00:00:20.000 It's basically the art and science of having those kinds 7 00:00:20.000 --> 00:00:25.760 of conversations with large language models, you know lms like chat, GPT, 8 00:00:26.679 --> 00:00:28.879 and it's way more than just typing in a question, right, 9 00:00:29.320 --> 00:00:31.879 It's about really crafting that dialogue, it really is. 10 00:00:32.320 --> 00:00:35.200 And our guide for this is a fascinating new book 11 00:00:35.280 --> 00:00:38.119 just out in twenty twenty five. It's called Prompt Engineering 12 00:00:38.200 --> 00:00:43.280 Empowering Communication, Oh okay, by Ajantha Devi Varamani and Anan Naire, 13 00:00:44.119 --> 00:00:46.479 and it's well, it's pretty foundational. It's like our map 14 00:00:46.479 --> 00:00:49.840 for understanding how to really tap into what these lms 15 00:00:49.920 --> 00:00:50.240 can do. 16 00:00:50.560 --> 00:00:54.079 Okay, great, So our mission today for you listening is 17 00:00:54.119 --> 00:00:56.359 basically to cut through the hype maybe some of the 18 00:00:56.439 --> 00:00:59.240 jargon too, Yeah, definitely, and get you really informed about 19 00:00:59.240 --> 00:01:02.719 what prompt engineer is, why it suddenly becomes so important, 20 00:01:02.719 --> 00:01:04.879 and you know how you might start thinking about using 21 00:01:04.920 --> 00:01:06.760 it even if you're totally new to this, right, we'll 22 00:01:06.799 --> 00:01:10.359 look at some maybe surprising ways prompts are being used 23 00:01:10.680 --> 00:01:12.719 and what it all kind of means for how we'll 24 00:01:12.760 --> 00:01:15.799 interact with AI going forward. That's good, okay, So let's 25 00:01:15.879 --> 00:01:19.239 unpack this a bit. We've all like asked chat chepte 26 00:01:19.319 --> 00:01:21.799 a question, yeah, or use some other AI. 27 00:01:21.680 --> 00:01:23.760 Tool, Sure most people probably have by now. 28 00:01:23.920 --> 00:01:27.760 But prompt engineering it takes that simple question and turns 29 00:01:27.799 --> 00:01:31.879 it into something well more sophisticated. The book has this 30 00:01:31.959 --> 00:01:35.840 great description actually, oh yeah. It calls it intertwining instructions 31 00:01:35.840 --> 00:01:38.959 and information into this web in a way that gently 32 00:01:39.079 --> 00:01:41.599 nudges the LM toward the intended result. 33 00:01:41.920 --> 00:01:42.879 Nudges I like that. 34 00:01:43.120 --> 00:01:46.159 Yeah. It says it's like whispering ideas into the ear 35 00:01:46.200 --> 00:01:49.920 of a gifted artist, offering just enough direction to ignite 36 00:01:49.920 --> 00:01:53.359 their creative spark without drowning out their distinct voice. 37 00:01:53.400 --> 00:01:56.519 That's a fantastic analogy. It captures that subtlety. It's not 38 00:01:56.560 --> 00:01:57.519 just commanding, it's. 39 00:01:57.400 --> 00:02:01.879 Guiding exactly, moving from just asking to actually guiding the AI. 40 00:02:02.120 --> 00:02:05.719 And that means the prompt itself changes meaning, doesn't it. 41 00:02:05.719 --> 00:02:09.280 It's not just your question anymore. It becomes the whole package, 42 00:02:09.400 --> 00:02:12.719 the instructions. Maybe an example, you give it the background context. 43 00:02:13.280 --> 00:02:15.919 It's the roadmap you provide for the AI's answer. 44 00:02:16.759 --> 00:02:19.479 So it could be just attendance or much more complex. 45 00:02:19.520 --> 00:02:24.000 Oh, absolutely, it could be quite detailed, several clauses, specific directives. 46 00:02:24.439 --> 00:02:28.159 You're influencing how it responds using natural language, or sometimes 47 00:02:28.159 --> 00:02:33.159 even specific system instructions or constraints. The art really is 48 00:02:33.520 --> 00:02:35.639 in crafting that guidance effectively. 49 00:02:35.759 --> 00:02:38.560 Okay, so why now? Why is this suddenly such a 50 00:02:38.560 --> 00:02:40.719 big deal? What are the implications here? 51 00:02:40.879 --> 00:02:43.039 That really is the core question, isn't it? And the 52 00:02:43.039 --> 00:02:45.919 book lays out a few key reasons why prompt engineering 53 00:02:46.000 --> 00:02:47.000 is vital right now? 54 00:02:47.000 --> 00:02:47.319 Okay. 55 00:02:47.800 --> 00:02:52.080 Firstly, it helps us unlock the full potential of these elms. 56 00:02:52.080 --> 00:02:55.000 Think about it. These models have learned from just unimaginable 57 00:02:55.000 --> 00:02:56.039 amounts of text data. 58 00:02:56.120 --> 00:02:57.439 Right the scale is huge. 59 00:02:57.280 --> 00:02:59.960 Exactly, and prompt engineering lets us tap into that knowledge 60 00:03:00.000 --> 00:03:03.599 for things they weren't explicitly trained to do. The book 61 00:03:03.639 --> 00:03:08.599 mentions things like authoring engrossing novels or complex musical compositions. 62 00:03:08.719 --> 00:03:12.039 Wow, okay, so pushing beyond their standard programming. 63 00:03:11.680 --> 00:03:16.080 Precisely extending what they can inherently do. Secondly, it gives 64 00:03:16.120 --> 00:03:20.479 us precision and management. This is crucial. How so, Well, 65 00:03:20.520 --> 00:03:25.000 if you need accuracy, think about news reporting or scientific study, 66 00:03:25.199 --> 00:03:28.280 you can't just have a vague answer. Prompt engineering lets 67 00:03:28.280 --> 00:03:32.400 you carefully craft the output guide it towards factual accuracy, 68 00:03:32.599 --> 00:03:35.280 towards the specific requirements of the task. 69 00:03:35.199 --> 00:03:38.800 So getting exactly what you need, not just something exactly. 70 00:03:39.120 --> 00:03:43.159 Then there's uniqueness and creativity. A really well designed prompt 71 00:03:43.199 --> 00:03:46.560 can spark the llms well, you could call it imagination, 72 00:03:47.319 --> 00:03:51.520 leading to novel and surprising results. It can genuinely help 73 00:03:51.599 --> 00:03:54.400 foster creative solutions that you might not have thought of otherwise. 74 00:03:54.479 --> 00:03:56.560 Okay, that's interesting using it as a creative partner. 75 00:03:56.680 --> 00:03:58.560 Yeah. And finally, and this is a big one, it's 76 00:03:58.599 --> 00:04:02.759 about democratizing AI. Good prompt engineering makes these powerful tools 77 00:04:02.840 --> 00:04:03.479 more accessible. 78 00:04:03.520 --> 00:04:04.120 How does that work? 79 00:04:04.240 --> 00:04:07.719 Well, people without deep technical coding skills can still leverage 80 00:04:07.719 --> 00:04:10.919 AI effectively. They can use prompts to get the AI 81 00:04:10.960 --> 00:04:14.280 to do complex things for their creative and productive efforts. 82 00:04:14.360 --> 00:04:15.800 It opens it up right, so. 83 00:04:15.800 --> 00:04:19.959 It's not just for developers anymore, artists, writers, small businesses, everyone. 84 00:04:20.120 --> 00:04:24.519 Potentially, it fundamentally changes who gets to innovate with this tech. 85 00:04:24.680 --> 00:04:26.639 It's easy to think this is all brand new, like 86 00:04:26.720 --> 00:04:29.240 AI just appeared. But you're saying the roots of this 87 00:04:29.439 --> 00:04:32.079 guiding AI with language go back further. 88 00:04:31.839 --> 00:04:33.920 Oh much further. You're right. The idea isn't new at all. 89 00:04:34.079 --> 00:04:35.839 If you connect the dots, you can trace it back 90 00:04:35.839 --> 00:04:39.680 to early experiments like what well think about chatbots like 91 00:04:39.720 --> 00:04:43.360 Eliza that was way back in nineteen sixty six. Wow, 92 00:04:43.600 --> 00:04:46.920 it mimicked a therapist trying to hold a conversation. And 93 00:04:46.959 --> 00:04:51.360 then in the seventies there was shr dlu shrdlu. Yeah, 94 00:04:51.399 --> 00:04:53.879 it was a natural language system that could understand and 95 00:04:54.000 --> 00:04:57.879 follow commands in a simple virtual world of blocks. 96 00:04:58.079 --> 00:05:00.800 So early attempts at language based in st diction. 97 00:05:00.639 --> 00:05:04.519 Exactly rudimentary, is sure, but they lay the groundwork. They 98 00:05:04.639 --> 00:05:07.439 showed that human language could direct what an AI does. 99 00:05:07.680 --> 00:05:10.160 And then things really took off with the transfermer revolution 100 00:05:10.240 --> 00:05:13.120 in twenty seventeen. That seems like the pivotal moment. 101 00:05:13.360 --> 00:05:17.160 Absolutely, that paper attention is all you need was the breakthrough. 102 00:05:17.560 --> 00:05:20.720 It paved the way for these incredibly powerful lms like 103 00:05:20.800 --> 00:05:24.680 GPT three, and that's when prompt engineering really emerged as 104 00:05:24.720 --> 00:05:27.079 its own distinct skill, its own field. 105 00:05:27.160 --> 00:05:31.240 Right, and the open AI GPT models that progression is 106 00:05:31.279 --> 00:05:33.240 pretty key to understanding this isn't it? 107 00:05:33.240 --> 00:05:35.959 It really is. It shows the rapid evolution. So GPT 108 00:05:35.959 --> 00:05:39.720 one came out in twenty eighteen. It used unsupervised pre training, 109 00:05:39.920 --> 00:05:43.040 had about one hundred and seventeen million parameters. 110 00:05:43.079 --> 00:05:46.120 Parameters those like the internal knobs the AI uses. 111 00:05:45.879 --> 00:05:48.439 To learn sort of yeah, the variables. It adjusts a 112 00:05:48.560 --> 00:05:51.680 huge number of them. Then GPT two in twenty nineteen 113 00:05:51.759 --> 00:05:54.360 was a big step up one point five billion parameters 114 00:05:54.439 --> 00:05:58.199 WHOA and trained on much more data about forty gigabytes. 115 00:05:58.240 --> 00:06:01.600 It showed definite improvements in coherence generating texts. 116 00:06:01.720 --> 00:06:03.639 GPT three was the next massive. 117 00:06:03.360 --> 00:06:05.639 Leap massive is the word twenty twenty one hundred and 118 00:06:05.639 --> 00:06:08.920 seventy five billion parameters trained on huge chunks of the 119 00:06:08.920 --> 00:06:13.319 Internet common crawl Wikipedia, and it showed really impressive coherence 120 00:06:13.360 --> 00:06:15.759 and creativity. It could even write code, which was pretty 121 00:06:15.759 --> 00:06:16.680 mind blowing at the time. 122 00:06:16.720 --> 00:06:18.600 But it wasn't perfect. Was it? Their issues? 123 00:06:18.879 --> 00:06:24.040 No, definitely not. The book points out challenges with misalignment. 124 00:06:24.839 --> 00:06:27.759 Sometimes what the user wanted didn't quite line up with 125 00:06:27.839 --> 00:06:30.519 how the model was trained. I mean, it could generate 126 00:06:30.560 --> 00:06:34.560 stuff that was wrong, erroneous, or even problematic content like 127 00:06:35.079 --> 00:06:37.720 biased are toxic text. That was a real hurdle. 128 00:06:38.040 --> 00:06:39.839 So how do they tackle that? Because that sounds like 129 00:06:39.879 --> 00:06:41.920 a major barrier to using it widely. 130 00:06:42.160 --> 00:06:46.360 It was, and OpenAI solution was pretty ingenious. Actually, they 131 00:06:46.519 --> 00:06:50.160 introduced something called instruct GPT in twenty twenty one. 132 00:06:50.439 --> 00:06:52.439 Instruct GPT, Yeah, and the. 133 00:06:52.439 --> 00:06:56.439 Key was using reinforcement learning with human feedback our LHF. 134 00:06:56.720 --> 00:06:58.399 Okay, what does that involve. 135 00:06:58.240 --> 00:07:01.399 Well, it's quite meticulous. They had human labelers write high 136 00:07:01.480 --> 00:07:03.839 quality ideal answers to various. 137 00:07:03.519 --> 00:07:06.680 Prompts I like examples of good behavior exactly. 138 00:07:06.360 --> 00:07:09.000 And then they trained another model, a reward model, to 139 00:07:09.240 --> 00:07:12.720 basically score the AI's responses based on how close they 140 00:07:12.720 --> 00:07:14.040 were to those human examples. 141 00:07:14.160 --> 00:07:16.240 So the AI learned to aim for what the humans 142 00:07:16.240 --> 00:07:17.480 preferred precisely. 143 00:07:17.759 --> 00:07:21.120 It iteratively refined the main model, and the book says 144 00:07:21.160 --> 00:07:25.439 this significantly enhances model accuracy and reduces the generation of 145 00:07:25.439 --> 00:07:28.560 toxic content, made them much safer and more reliable. 146 00:07:28.600 --> 00:07:31.399 And that whole process, that refinement led directly to the 147 00:07:31.399 --> 00:07:32.839 models we know now pretty much. 148 00:07:32.920 --> 00:07:35.759 That work fed into the GPT three point five series, 149 00:07:35.879 --> 00:07:38.800 and then, of course November twenty twenty two, Chat GPT 150 00:07:38.920 --> 00:07:40.600 launched based on that series. 151 00:07:40.439 --> 00:07:41.680 Right when it really hit the mainstream. 152 00:07:41.759 --> 00:07:44.639 Yeah, and then March twenty twenty three we got GPT 153 00:07:44.759 --> 00:07:46.120 four the current top. 154 00:07:45.920 --> 00:07:49.920 Tier, and that one's multimodal right handles images too, that's 155 00:07:49.959 --> 00:07:50.399 the claim. 156 00:07:50.519 --> 00:07:55.360 Yes, the inaugural multimodal model. It can process text and images, 157 00:07:55.399 --> 00:07:58.639 though the image input part isn't widely public yet. 158 00:07:58.720 --> 00:08:00.639 But the performance jumps huge. 159 00:08:00.480 --> 00:08:04.120 Oh staggering. The book gives examples like the Uniform Bar 160 00:08:04.160 --> 00:08:07.360 Exam CHAT GPT scored around the tenth percentile GPT four 161 00:08:07.600 --> 00:08:12.079 ninetieth percent style or the International Biology Olympiad CHAT GPT 162 00:08:12.240 --> 00:08:15.040 was thirty first percentile GTT four hit the ninety ninth. 163 00:08:15.319 --> 00:08:17.720 That's incredible progress in just a year or so. 164 00:08:17.959 --> 00:08:20.360 It really highlights how fast this field is moving. 165 00:08:20.639 --> 00:08:22.839 Okay, so we've got the history, we know what's kind 166 00:08:22.839 --> 00:08:26.120 of under the hood, how do we actually do prompt engineering? Like, 167 00:08:26.439 --> 00:08:29.319 for you listening, what are the practical ways to start 168 00:08:29.360 --> 00:08:30.720 guiding these ais better? 169 00:08:31.000 --> 00:08:33.399 Right? The nuts and bolts? Well, the book lays out 170 00:08:33.399 --> 00:08:37.159 a whole toolkit, but the core principles are pretty straightforward. Okay. 171 00:08:37.200 --> 00:08:39.960 It starts with really understanding the task you want the 172 00:08:39.960 --> 00:08:46.200 AI to do, being precise and succinct in your language, 173 00:08:46.480 --> 00:08:48.519 giving enough detail but not too much. 174 00:08:48.679 --> 00:08:50.120 Finding that balance. 175 00:08:49.879 --> 00:08:55.240 Exactly providing examples can be huge and maybe most importantly, experiment. 176 00:08:55.559 --> 00:08:58.879 There's no single magic formula. It's iterative. You try something, 177 00:08:58.960 --> 00:09:00.039 see what happens. 178 00:08:59.799 --> 00:09:02.840 To Okay, So let's dig into some specific techniques people 179 00:09:02.879 --> 00:09:06.000 can try. The book mentions, the instructions prompt technique first 180 00:09:06.399 --> 00:09:07.320 sound straightforward. 181 00:09:07.559 --> 00:09:09.639 It is pretty much you basically just tell the model 182 00:09:09.679 --> 00:09:11.960 exactly what you want it to do, often using clear 183 00:09:12.000 --> 00:09:15.480 commands imperative verbs. So yeah, so if you need customer 184 00:09:15.519 --> 00:09:19.279 service replies, you might instruct it. Responses should be professional 185 00:09:19.320 --> 00:09:21.840 and provide accurate information, nice and clear. 186 00:09:22.080 --> 00:09:24.080 Or for maybe a legal context. 187 00:09:23.759 --> 00:09:27.120 Yeah, you could say draft this section ensuring the document 188 00:09:27.200 --> 00:09:31.679 complies with relevant laws and regulations. You're setting clear rules 189 00:09:31.960 --> 00:09:33.440 guardrails for the output. 190 00:09:33.679 --> 00:09:38.240 Okay, makes sense. Then there's zero one and few shot prompting. 191 00:09:38.879 --> 00:09:41.120 That sounds interesting. What's the difference? 192 00:09:41.519 --> 00:09:43.679 This is a really helpful way to think about context. 193 00:09:43.759 --> 00:09:46.879 So zero shot is when you give it no specific 194 00:09:46.960 --> 00:09:50.519 examples in the prompt itself, just the instruction, just the instruction. Yeah, 195 00:09:50.519 --> 00:09:54.120 like write a poem about nature, the AI has to 196 00:09:54.159 --> 00:09:57.720 rely entirely on its pre existing knowledge. It's general training 197 00:09:57.759 --> 00:09:59.360 about poems and nature, so. 198 00:09:59.320 --> 00:10:02.159 It's kind of gets based on its vast training data. 199 00:10:02.360 --> 00:10:04.279 In a way. Yes, and the book points out that 200 00:10:04.320 --> 00:10:07.559 while it sounds simple, every word counts massively here. A 201 00:10:07.600 --> 00:10:10.919 slightly ambiguous word can send it off track. It really 202 00:10:10.960 --> 00:10:13.480 tests your ability to be concise and clear. 203 00:10:13.679 --> 00:10:15.679 Okay, so what if you need to give it a 204 00:10:15.679 --> 00:10:17.720 bit more direction, show what you mean. 205 00:10:17.919 --> 00:10:20.159 That's where one shot and FU shot come in. With 206 00:10:20.240 --> 00:10:22.080 one shot, you provide just one example. 207 00:10:22.440 --> 00:10:23.879 Uh So, say you. 208 00:10:23.840 --> 00:10:27.720 Want a paragraph summarized in a very specific, concise style. 209 00:10:27.799 --> 00:10:29.919 You'd give it the paragraph and then maybe one example 210 00:10:29.919 --> 00:10:32.200 of a sentence summarizing the style you want. You're showing 211 00:10:32.200 --> 00:10:32.919 it how just. 212 00:10:33.120 --> 00:10:36.080 Once demonstrating the task exactly. 213 00:10:35.639 --> 00:10:37.559 And FU shot just scales that up a bit. You 214 00:10:37.559 --> 00:10:40.519 give it a small number maybe two, three, five examples. 215 00:10:40.559 --> 00:10:42.080 Like for a product review, Yeah. 216 00:10:41.879 --> 00:10:44.279 Good example. If you want to review with certain elements, 217 00:10:44.320 --> 00:10:47.879 you might describe the product briefly, then provide maybe two 218 00:10:48.039 --> 00:10:51.120 or three examples of well written reviews formatted the way 219 00:10:51.159 --> 00:10:54.120 you like. This gives the AI more patterns to. 220 00:10:54.200 --> 00:10:57.879 Learn from, and that leads to better, more tailored results. 221 00:10:57.960 --> 00:11:02.159 Usually often, Yes, especially for more or complex or stylistic tasks, 222 00:11:02.480 --> 00:11:04.600 more examples generally mean better guidance. 223 00:11:05.200 --> 00:11:08.399 Okay, And the last one mention here is self consistency prompt. 224 00:11:08.639 --> 00:11:11.879 What's that about? Sounds important for factual stuff? It is. 225 00:11:12.000 --> 00:11:14.360 It's crucial when you need the output to be logical 226 00:11:14.440 --> 00:11:17.120 and coherent, and especially when it needs to stick to 227 00:11:17.159 --> 00:11:18.720 certain facts or principles. 228 00:11:18.799 --> 00:11:19.480 How does it work? 229 00:11:19.600 --> 00:11:24.080 You basically embed key facts or maybe core principles directly 230 00:11:24.159 --> 00:11:25.720 within the prompt itself. 231 00:11:25.960 --> 00:11:28.559 Ah, so you feed it the essential info it needs 232 00:11:28.600 --> 00:11:29.080 to stay. 233 00:11:28.919 --> 00:11:32.480 On track precisely. The book uses the example of a 234 00:11:32.480 --> 00:11:37.080 climate change discussion. You might include specific established scientific data 235 00:11:37.120 --> 00:11:39.960 points or consensus statements right in the prompt, and that. 236 00:11:39.960 --> 00:11:42.480 Helps prevent it from just making things up or going 237 00:11:42.480 --> 00:11:43.279 off on a tangent. 238 00:11:43.480 --> 00:11:46.200 Exactly. It acts like an internal reference guide for that 239 00:11:46.279 --> 00:11:49.679 specific answer, keeping it grounded in the information you provided. 240 00:11:50.279 --> 00:11:53.840 These techniques they really seem to allow for incredibly specific outcomes, 241 00:11:53.840 --> 00:11:56.720 don't they Moving way beyond just generic answers? 242 00:11:56.919 --> 00:12:01.120 They really do. It's about tailoring the AI's as capabilities 243 00:12:01.120 --> 00:12:03.840 to your precise needs. And this is why the future 244 00:12:03.879 --> 00:12:07.840 of LLM communication looks so interesting. Things like AI powered 245 00:12:07.879 --> 00:12:12.440 negotiation tools or real time translation that understands nuance or 246 00:12:12.639 --> 00:12:16.639 news updates customize just for you. Prompt engineering is the 247 00:12:16.720 --> 00:12:19.240 key to unlocking all of that. It's how we teach 248 00:12:19.279 --> 00:12:22.320 the AI to understand our intent with much, much greater precision. 249 00:12:22.559 --> 00:12:25.960 You know, the book title Empowering Communication really starts to 250 00:12:25.960 --> 00:12:28.559 make sense. Now. It's not just about tech, is it. 251 00:12:29.240 --> 00:12:32.080 These prompts can actually change how we work and interact 252 00:12:32.559 --> 00:12:34.960 across well, pretty much everywhere. 253 00:12:35.080 --> 00:12:38.399 That's a really important point. The book actually highlights applications 254 00:12:38.440 --> 00:12:42.000 in over a dozen different professional fields. It's incredibly versatile. 255 00:12:42.120 --> 00:12:44.279 Give us some examples, maybe some less obvious ones for 256 00:12:44.320 --> 00:12:45.519 people listening to think about. 257 00:12:45.600 --> 00:12:49.360 Okay, sure, well, take creative thinking prompts can genuinely help 258 00:12:49.879 --> 00:12:55.080 unlock imagination and innovation. How so imagine using prompts designed 259 00:12:55.080 --> 00:12:59.799 specifically to challenge your assumptions during brainstorming, or prompts that 260 00:12:59.799 --> 00:13:02.919 go an AI to generate, say, an initial sketch for 261 00:13:02.960 --> 00:13:07.000 an artwork representing the beauty of diversity and inclusivity. It's 262 00:13:07.039 --> 00:13:09.000 a starting point for human creativity. 263 00:13:09.360 --> 00:13:12.279 So using the AI not just to do the creative work, 264 00:13:12.320 --> 00:13:14.039 but to spark our own creativity. 265 00:13:14.039 --> 00:13:17.399 That's fascinating exactly, or think about effective writing. We all 266 00:13:17.480 --> 00:13:21.039 face writer's block sometimes, right prompts can be invaluable for 267 00:13:21.159 --> 00:13:24.440 just igniting the writing process. You could ask for a 268 00:13:24.480 --> 00:13:27.679 prompt to explore the concept of time travel and a 269 00:13:27.720 --> 00:13:30.679 short story, but maybe from the perspective of an object 270 00:13:30.679 --> 00:13:33.639 not a person, or create a story about a character 271 00:13:33.720 --> 00:13:37.399 who discovers a hitting world within everyday objects. It gives 272 00:13:37.440 --> 00:13:40.879 you that initial seed, a narrative launch pad precisely, and 273 00:13:40.919 --> 00:13:44.799 then in business, prompts can help craft effective presentations, maybe 274 00:13:44.840 --> 00:13:47.679 generate different options for an opening hook, like a powerful 275 00:13:47.720 --> 00:13:51.159 rhetorical question to grab the audience, or even assist in 276 00:13:51.279 --> 00:13:55.519 negotiation and persuasion by helping you frame arguments about, say 277 00:13:55.960 --> 00:13:59.320 the importance of building rapport before diving into the deal points. 278 00:13:59.480 --> 00:14:02.879 And it reaches into really specialized fields too, like medicine 279 00:14:02.919 --> 00:14:03.279 or law. 280 00:14:03.559 --> 00:14:07.840 Absolutely the book mentions healthcare professionals using prompts to enhance 281 00:14:07.919 --> 00:14:12.039 empathetic patient communication, maybe prompts that encourage patients to articulate 282 00:14:12.080 --> 00:14:16.600 their health goals clearly or guiding ethical decision making discussions. 283 00:14:16.879 --> 00:14:20.480 And for legal professionals, prompts could help draft sections of 284 00:14:20.519 --> 00:14:24.960 persuasive legal writing. Maybe exploring arguments for or against the 285 00:14:25.000 --> 00:14:28.519 constitutionality of a law to help refine the lawyer's own thinking. 286 00:14:28.799 --> 00:14:32.399 Okay, and data science, that seems like a natural fit. 287 00:14:32.279 --> 00:14:36.919 Definitely for data scientists. Prompts can guide analysis at every stage, 288 00:14:37.240 --> 00:14:41.440 from framing questions for data cleaning and preprocessing, like asking 289 00:14:41.480 --> 00:14:43.919 the AI to reflect on the importance of data cleaning, 290 00:14:44.200 --> 00:14:46.840 all the way to generating different types of data visualizations 291 00:14:46.960 --> 00:14:49.200 or summarizing findings from complex models. 292 00:14:49.679 --> 00:14:52.080 Yeah, that's a huge range. It really drives home that 293 00:14:52.240 --> 00:14:54.919 this isn't just a niche tech skill, it's becoming a 294 00:14:54.960 --> 00:14:57.600 fundamental communication skill across disciplines. 295 00:14:57.720 --> 00:15:00.720 I think that's exactly right. It's about empowering community, whoever 296 00:15:00.799 --> 00:15:02.000 you are, whatever field you're in. 297 00:15:02.159 --> 00:15:03.919 So we've talked a lot about using prompts as an 298 00:15:03.960 --> 00:15:07.519 individual user interacting with chat GPT on the website or 299 00:15:07.559 --> 00:15:10.840 similar But what if you're a developer or a business, 300 00:15:11.480 --> 00:15:13.919 What if you want to build these AI capabilities into 301 00:15:13.960 --> 00:15:17.159 your own software, your own products. That's where the API comes. 302 00:15:17.000 --> 00:15:21.720 In, right exactly. The chat GPT API Application Programming Interface 303 00:15:21.840 --> 00:15:26.799 is what lets developers integrate these powerful conversational AI features 304 00:15:27.080 --> 00:15:29.080 directly into their own apps and services. 305 00:15:29.159 --> 00:15:31.000 The bridge. Basically, it's the bridge. 306 00:15:31.039 --> 00:15:33.879 It moves beyond just the chat window and lets you 307 00:15:33.960 --> 00:15:36.159 embed the AI's power systematically. 308 00:15:36.360 --> 00:15:39.559 Okay, so for someone listening, why would they care about 309 00:15:39.720 --> 00:15:43.000 using an API versus just going to the website. What 310 00:15:43.039 --> 00:15:44.200 are the big advantages? 311 00:15:44.600 --> 00:15:47.720 Well, they're pretty significant. First, you can build much more 312 00:15:47.840 --> 00:15:52.879 customized conversational interfaces, think chatbots or virtual assistance embedded right 313 00:15:52.919 --> 00:15:57.039 into your product, speaking your brand's voice, accessing your specific data. 314 00:15:57.240 --> 00:16:00.879 Much more integrated, right, not just a generic bot exactly. Second, 315 00:16:01.039 --> 00:16:04.679 content generation at scale. The API lets you automate the 316 00:16:04.720 --> 00:16:08.600 creation of marketing, copy, articles, summaries, whatever you need, much faster, 317 00:16:09.000 --> 00:16:11.919 and ensuring a consistent tone and style across everything. 318 00:16:12.000 --> 00:16:13.799 That could be huge for businesses. 319 00:16:13.600 --> 00:16:20.080 Oh absolutely. And Third scalability and customizability. Open AI handles 320 00:16:20.120 --> 00:16:23.360 the massive computing infrastructure, so you don't have to worry 321 00:16:23.360 --> 00:16:26.399 about that scaling. Plus, you can actually fine tune the 322 00:16:26.480 --> 00:16:28.879 models using your own data, so you can. 323 00:16:28.879 --> 00:16:31.639 Train it on your specific company information or. 324 00:16:31.639 --> 00:16:36.240 Style precisely to suit very specific use cases, making the 325 00:16:36.279 --> 00:16:40.679 AI perform better for your particular needs. It becomes deeply integrated, 326 00:16:40.799 --> 00:16:41.480 and just like with. 327 00:16:41.399 --> 00:16:44.360 The techniques, there isn't just one API model is there 328 00:16:44.440 --> 00:16:45.840 there are different options. 329 00:16:45.960 --> 00:16:48.279 That's a really key point. Yeah, it's not one size 330 00:16:48.279 --> 00:16:51.440 fits all. There's a whole spectrum of models, mainly differing 331 00:16:51.440 --> 00:16:54.039 by their parameter size, which affects capability and cost. 332 00:16:54.480 --> 00:16:55.720 Okay, so what are we talking about. 333 00:16:56.000 --> 00:16:57.720 Well, at the top end, you might have something like 334 00:16:57.840 --> 00:17:00.440 chat GPT three point five B that's three point five 335 00:17:00.480 --> 00:17:05.400 billion parameters, very capable, great for complex tasks, nuanced. 336 00:17:04.960 --> 00:17:07.480 Conversation, probably more resource intensive. 337 00:17:07.200 --> 00:17:10.400 Generally, yes. Then you might step down to maybe chat 338 00:17:10.400 --> 00:17:13.240 GPT one point five b one point five billion parameters, 339 00:17:13.240 --> 00:17:16.799 still very powerful, often a good balance of performance and efficiency. 340 00:17:17.400 --> 00:17:19.799 Then you get smaller ones like chat GPT three hundred 341 00:17:19.920 --> 00:17:23.519 M three hundred million parameters, maybe better for faster responses, 342 00:17:23.640 --> 00:17:27.799 lower latency, or less demanding tasks right, and even smaller 343 00:17:27.960 --> 00:17:30.400 like chat GPT one twenty five M one hundred and 344 00:17:30.440 --> 00:17:34.359 twenty five million parameters design for environments where resources are 345 00:17:34.400 --> 00:17:38.519 really tight, maybe simple classification tasks or running on less 346 00:17:38.519 --> 00:17:39.400 powerful devices. 347 00:17:39.440 --> 00:17:41.319 So choosing the right one is a balancing act. 348 00:17:41.319 --> 00:17:44.240 Absolutely, you have to weigh up performance needs, how complex 349 00:17:44.319 --> 00:17:48.039 the task is, what computing resources, you have, latency requirements, 350 00:17:48.079 --> 00:17:51.400 and of course cost. It's often again an iterative process. 351 00:17:51.400 --> 00:17:53.920 You test model, see how it performs, get feedback, maybe 352 00:17:53.960 --> 00:17:55.599 try another one and refine your choice. 353 00:17:55.680 --> 00:17:58.720 This is all incredibly powerful stuff, but you know the 354 00:17:58.759 --> 00:18:00.640 old saying, with great power. 355 00:18:00.279 --> 00:18:01.480 Comes great responsibility. 356 00:18:01.559 --> 00:18:04.599 Yeah, what about the ethical side of prompt engineering? That 357 00:18:04.680 --> 00:18:06.359 has to be a major consideration. 358 00:18:06.480 --> 00:18:09.279 It absolutely is, and the book stresses us we have 359 00:18:09.359 --> 00:18:12.880 to think carefully about responsible and equitable. 360 00:18:12.480 --> 00:18:14.680 Use, like what specifically, Well. 361 00:18:14.480 --> 00:18:17.440 One big area is bias. The book points out how 362 00:18:17.640 --> 00:18:21.680 prompts with built in biases may cause discriminatory communication if 363 00:18:21.720 --> 00:18:27.039 we're not careful. The prompts themselves can encode assumptions or stereotypes. 364 00:18:26.359 --> 00:18:28.680 Right, garbage in, garbage out, potentially. 365 00:18:28.279 --> 00:18:32.640 Amplify exactly or issues of fairness manipulation. As we build 366 00:18:32.680 --> 00:18:36.440 these incredibly persuasive communication tools, we really have a duty 367 00:18:36.480 --> 00:18:38.359 to align them with ethical principles. 368 00:18:38.519 --> 00:18:40.440 So how do we do that? How do we make 369 00:18:40.480 --> 00:18:43.640 sure we're using prompts effectively and ethically? Is it just 370 00:18:43.640 --> 00:18:44.880 about having good intentions? 371 00:18:45.319 --> 00:18:47.920 Good intentions are a start, but it needs a process. 372 00:18:48.400 --> 00:18:52.440 The book really emphasizes evaluating and refining prompts as an 373 00:18:52.480 --> 00:18:55.359 ongoing iterative cycle, like any kind of design. 374 00:18:55.480 --> 00:18:56.960 Okay, so what does that cycle look like? 375 00:18:57.119 --> 00:19:01.319 It means first assessing the impact, asking hard questions, is 376 00:19:01.359 --> 00:19:04.680 this prompt actually sparking creativity like we hoped? Is it 377 00:19:04.720 --> 00:19:08.119 promoting critical thinking or is it accidentally leading to harmful 378 00:19:08.160 --> 00:19:10.000 outputs or reinforcing biases? 379 00:19:10.079 --> 00:19:12.519 Really checking the results against the goals yes. 380 00:19:13.359 --> 00:19:18.279 Then gathering feedback, actively asking the people using the prompts 381 00:19:18.400 --> 00:19:22.200 or interacting with the AI what's clear, what's confusing, what 382 00:19:22.319 --> 00:19:23.559 problems are they running into? 383 00:19:23.759 --> 00:19:25.440 Getting real world input. 384 00:19:25.200 --> 00:19:31.160 Crucial, and finally, iterative refinement. Based on that impact assessment 385 00:19:31.279 --> 00:19:36.079 and the feedback, You continually adjust, tweak and improve the pumps. 386 00:19:36.519 --> 00:19:39.720 You make sure they stay relevant, effective, and fair for 387 00:19:39.920 --> 00:19:43.160 different people in situations. It's not a one and done thing. 388 00:19:43.319 --> 00:19:46.160 It sounds like a constant conversation really, Yeah, not just 389 00:19:46.200 --> 00:19:48.720 the conversation with the AI, but a conversation about how 390 00:19:48.720 --> 00:19:49.799 we're shaping that interaction. 391 00:19:49.960 --> 00:19:51.720 That's a great way to put it. It's a continuous 392 00:19:51.759 --> 00:19:55.319 process of learning and improvement within the whole AI community. 393 00:19:55.480 --> 00:19:57.519 Well, we have covered a lot of ground today. We've 394 00:19:57.559 --> 00:20:00.880 taken this deep dive into prompt engineering right from its 395 00:20:01.599 --> 00:20:02.880 surprisingly long. 396 00:20:02.759 --> 00:20:04.920 History, Yeah, going back to Eliza all the. 397 00:20:04.839 --> 00:20:07.759 Way to its absolutely central role now and getting the 398 00:20:07.799 --> 00:20:11.039 most out of these powerful llms like chat GPT. We've 399 00:20:11.079 --> 00:20:14.759 seen how crafting these prompts carefully is becoming the key 400 00:20:14.839 --> 00:20:19.920 really to getting precise, creative, and hopefully ethically sound AI interactions, 401 00:20:20.480 --> 00:20:25.119