WEBVTT 00:00:00.000 --> 00:00:01.919 You know, sitting down with this massive stack 00:00:01.919 --> 00:00:03.940 of research you send over on prompt engineering, 00:00:04.540 --> 00:00:08.500 it really, it hits home just how universal a 00:00:08.500 --> 00:00:10.279 certain kind of frustration has become for all 00:00:10.279 --> 00:00:12.419 of us. Oh, absolutely. Everyone has felt it. 00:00:12.419 --> 00:00:14.720 Right. Like, you are sitting at your computer 00:00:14.720 --> 00:00:17.559 staring at that blinking cursor in a chat box, 00:00:17.559 --> 00:00:21.199 and you are trying to figure out the exact, precise, 00:00:21.199 --> 00:00:24.449 magic words to make the artificial intelligence 00:00:24.449 --> 00:00:27.190 do what you actually want it to do. It's a completely 00:00:27.190 --> 00:00:29.750 new paradigm. We've spent decades expecting software 00:00:29.750 --> 00:00:33.289 to just work linearly. Exactly. You click a button, 00:00:33.490 --> 00:00:35.689 a specific action happens. But with generative 00:00:35.689 --> 00:00:38.509 AI, the interface is suddenly human language, 00:00:39.310 --> 00:00:42.210 which is inherently ambiguous, emotionally loaded, 00:00:42.469 --> 00:00:45.770 and frankly, Very messy. It feels less like using 00:00:45.770 --> 00:00:48.049 a piece of software and more like I know trying 00:00:48.049 --> 00:00:50.929 to cast a highly volatile spell Yeah, or if you 00:00:50.929 --> 00:00:52.850 mispronounce a single syllable the whole thing 00:00:52.850 --> 00:00:55.789 explodes in your face Yes, and that is exactly 00:00:55.789 --> 00:00:58.149 why we are tackling this today looking at the 00:00:58.149 --> 00:01:00.210 notes You shared with us your goal is crystal 00:01:00.210 --> 00:01:02.770 clear. You want to understand the actual underlying 00:01:02.770 --> 00:01:05.620 mechanics of how we talk to AI And you want to 00:01:05.620 --> 00:01:07.859 do it without getting completely bogged down 00:01:07.859 --> 00:01:10.739 in impenetrable technical jargon. You want the 00:01:10.739 --> 00:01:12.900 translation. Which is the perfect approach to 00:01:12.900 --> 00:01:15.420 take. Yeah. Because understanding the machinery 00:01:15.420 --> 00:01:18.359 operating just beneath that chat box gives you 00:01:18.359 --> 00:01:21.200 a tremendous, almost unfair advantage over someone 00:01:21.200 --> 00:01:23.480 who is just, well, guessing. It really does. 00:01:23.500 --> 00:01:25.900 It elevates you from a passive user to an active 00:01:25.900 --> 00:01:29.120 director. So our mission today, using the comprehensive 00:01:29.120 --> 00:01:32.109 Wikipedia deep dive on prompt engineering you 00:01:32.109 --> 00:01:35.739 provided is to unpack a really specific evolution. 00:01:36.079 --> 00:01:38.959 We want to trace how giving instructions to AI 00:01:38.959 --> 00:01:42.340 evolved from this chaotic trial and error guessing 00:01:42.340 --> 00:01:45.760 game into what is now a highly automated, deeply 00:01:45.760 --> 00:01:48.060 scientific discipline. It's a wild trajectory, 00:01:48.299 --> 00:01:50.280 especially when you look at how quickly the entire 00:01:50.280 --> 00:01:52.459 landscape shifted in just a handful of months. 00:01:52.599 --> 00:01:54.939 Let's start right there with the brief life and 00:01:54.939 --> 00:01:57.079 sudden death of the prompt engineer. The speed 00:01:57.079 --> 00:01:59.799 of that boom and bust is historic. Looking at 00:01:59.799 --> 00:02:02.500 the timeline and the research, the whiplash is 00:02:02.500 --> 00:02:04.640 genuinely hard to wrap your head around. I mean, 00:02:04.819 --> 00:02:08.139 in 2023, the word prompt was literally the runner 00:02:08.139 --> 00:02:10.439 up for Oxford's Word of the Year. Right. Everyone 00:02:10.439 --> 00:02:13.000 was talking about it. Every tech blog and business 00:02:13.000 --> 00:02:16.319 magazine was screaming that prompt engineer was 00:02:16.319 --> 00:02:19.439 the ultimate six -figure job of the future. Companies 00:02:19.439 --> 00:02:22.159 were scrambling to hire these supposed AI whisperers. 00:02:22.300 --> 00:02:25.319 And then? Fast forward just two years to 2025, 00:02:25.979 --> 00:02:27.759 and the Wall Street Journal is officially declaring 00:02:27.759 --> 00:02:30.620 the job. completely obsolete. Like, it's just 00:02:30.620 --> 00:02:33.500 gone. Gone. But what is truly fascinating is 00:02:33.500 --> 00:02:37.020 the mechanism behind the death of the role. Because 00:02:37.020 --> 00:02:39.280 the job didn't vanish because prompting became 00:02:39.280 --> 00:02:41.319 less important to the tech industry. Oh, definitely 00:02:41.319 --> 00:02:44.039 not. It vanished because the AI models themselves 00:02:44.039 --> 00:02:46.740 underwent a massive structural shift in how they 00:02:46.740 --> 00:02:50.939 parse human intent. So they basically outgrew 00:02:50.939 --> 00:02:54.080 the need for a translator. Precisely. In 2023, 00:02:54.699 --> 00:02:57.280 models were raw. You needed a dedicated specialist 00:02:57.280 --> 00:02:59.639 who understood the machine's quirks just to coax 00:02:59.639 --> 00:03:01.919 a usable answer out of it. Like trying to hotwire 00:03:01.919 --> 00:03:04.419 a car. Yeah, exactly. But behind the scenes, 00:03:04.819 --> 00:03:06.639 developers were heavily refining a process called 00:03:06.639 --> 00:03:09.139 instruction tuning. They were aggressively training 00:03:09.139 --> 00:03:11.840 the models to be more forgiving. To infer what 00:03:11.840 --> 00:03:13.960 a normal human actually meant, even if the prompt 00:03:13.960 --> 00:03:16.610 was terribly written. Right. And simultaneously, 00:03:17.270 --> 00:03:19.610 everyday corporate training just got much better. 00:03:20.330 --> 00:03:23.090 Regular employees integrated basic prompting 00:03:23.090 --> 00:03:25.909 into their daily workflow, eliminating the need 00:03:25.909 --> 00:03:28.689 for a standalone specialist. You know, thinking 00:03:28.689 --> 00:03:31.550 about the mechanics of this, it feels almost 00:03:31.550 --> 00:03:33.789 identical to the history of the elevator operator. 00:03:34.090 --> 00:03:36.490 Oh, that is an excellent comparison, the transition 00:03:36.490 --> 00:03:39.150 from manual to automatic. Right. Think about 00:03:39.150 --> 00:03:42.669 the early 1900s. Elevators were these complex, 00:03:42.909 --> 00:03:45.610 heavily mechanical, potentially dangerous machines 00:03:45.610 --> 00:03:49.770 with manual levers and very specific speed controls. 00:03:50.069 --> 00:03:51.710 You couldn't just step into one and press a button. 00:03:51.830 --> 00:03:54.830 Exactly. You needed a highly specialized middleman, 00:03:54.930 --> 00:03:57.389 the elevator operator, just to get from the lobby 00:03:57.389 --> 00:03:59.590 to the fifth floor without crashing. But once 00:03:59.590 --> 00:04:01.330 the engineering advanced... Once the buttons 00:04:01.330 --> 00:04:03.550 became automated and the interface became user 00:04:03.550 --> 00:04:05.909 friendly, the middleman wasn't needed anymore. 00:04:06.430 --> 00:04:09.199 The technology absorbed the expertise. The prompt 00:04:09.199 --> 00:04:11.680 engineer was basically the modern elevator operator 00:04:11.680 --> 00:04:14.340 for neural networks. That perfectly encapsulates 00:04:14.340 --> 00:04:16.579 it. And what's remarkable is that this desire 00:04:16.579 --> 00:04:18.980 for a structured, human -in -the -loop interaction 00:04:18.980 --> 00:04:22.100 isn't even a product of the modern AI boom. Yeah, 00:04:22.199 --> 00:04:23.959 I was actually amazed by that section in the 00:04:23.959 --> 00:04:26.500 notes. If you look closely at the historical 00:04:26.500 --> 00:04:29.040 context in the documentation you provided, the 00:04:29.040 --> 00:04:31.779 foundational concept has been around for decades. 00:04:32.240 --> 00:04:34.319 It highlights a piece of software called the 00:04:34.319 --> 00:04:37.240 Intelligent Filling Manager built by Krishna 00:04:37.240 --> 00:04:41.420 C. Mukherjee all the way back in 1999. And 1999 00:04:41.420 --> 00:04:44.360 is ancient history in software terms. I mean, 00:04:44.399 --> 00:04:46.759 I was still using dial -up internet in 1999. 00:04:47.079 --> 00:04:49.500 Right. And this system didn't use neural networks 00:04:49.500 --> 00:04:52.860 or deep learning at all. It used a dynamic Q 00:04:52.860 --> 00:04:56.149 &A interface driven by a rule -based expert system. 00:04:56.490 --> 00:04:58.870 Essentially, it asked the user a series of highly 00:04:58.870 --> 00:05:01.389 structured questions to automatically collect 00:05:01.389 --> 00:05:03.910 the exact inputs needed for complex regulatory 00:05:03.910 --> 00:05:06.670 filings. So it was prompting the human to prompt 00:05:06.670 --> 00:05:09.769 the system correctly. It proves that the philosophy 00:05:09.769 --> 00:05:12.389 of guiding a human -to -machine interaction was 00:05:12.389 --> 00:05:15.069 always there, waiting for the technology to catch 00:05:15.069 --> 00:05:17.170 up. It's a great aha moment. But this brings 00:05:17.170 --> 00:05:19.170 up a massive contradiction that we need to unpack. 00:05:19.350 --> 00:05:21.639 OK, let's hear it. If the specialized job is 00:05:21.639 --> 00:05:24.120 dead and the models are now incredibly good at 00:05:24.120 --> 00:05:27.060 guessing what we want, why are researchers still 00:05:27.060 --> 00:05:29.240 spending billions of dollars studying exactly 00:05:29.240 --> 00:05:32.680 how we phrase things? That is the core paradox 00:05:32.680 --> 00:05:36.060 of generative AI. And the answer lies in the 00:05:36.060 --> 00:05:39.639 extreme, almost absurd fragility of large language 00:05:39.639 --> 00:05:42.560 models or LLMs. We perceive them as highly intelligent, 00:05:42.620 --> 00:05:45.540 sure. But their internal architecture makes them 00:05:45.540 --> 00:05:47.819 violently sensitive to the slightest linguistic 00:05:47.819 --> 00:05:50.240 changes. The statistics on this from the source 00:05:50.240 --> 00:05:52.779 are staggering. Like the research shows that 00:05:52.779 --> 00:05:54.879 if you are providing a few examples to an AI 00:05:54.879 --> 00:05:57.839 to show it what you want, simply reordering those 00:05:57.839 --> 00:06:00.899 exact same examples can produce accuracy shifts 00:06:00.899 --> 00:06:03.639 of more than 40 percent. More than 40 percent. 00:06:03.699 --> 00:06:06.389 Think about that. The data is identical. Just 00:06:06.389 --> 00:06:09.370 the sequence of presentation causes a 40 % swing 00:06:09.370 --> 00:06:11.949 in the AI's ability to be correct. It gets even 00:06:11.949 --> 00:06:14.350 more bizarre, actually. Changing just the basic 00:06:14.350 --> 00:06:16.689 formatting, like where you put commas or how 00:06:16.689 --> 00:06:19.389 you space the text. Wait, commas? Yeah, commas. 00:06:19.670 --> 00:06:22.129 Formatting changes alone can swing the AI's accuracy 00:06:22.129 --> 00:06:25.470 by up to 76 points on benchmark tests. That is 00:06:25.470 --> 00:06:28.720 insane. The documentation even notes that clausal 00:06:28.720 --> 00:06:32.180 syntax and morphology, literally, whether you 00:06:32.180 --> 00:06:35.879 use active or passive voice, fundamentally alters 00:06:35.879 --> 00:06:38.899 whether an AI successfully retrieves the right 00:06:38.899 --> 00:06:41.220 knowledge from its massive database. I have to 00:06:41.220 --> 00:06:43.680 be honest, I struggle to accept this. It sounds 00:06:43.680 --> 00:06:46.220 ridiculous. It violates all our assumptions about 00:06:46.220 --> 00:06:49.279 intelligence, doesn't it? Completely. It is like 00:06:49.279 --> 00:06:52.540 dealing with a temperamental, painfully literal 00:06:52.540 --> 00:06:54.970 -minded genie. That's a good way to put it. Or, 00:06:55.069 --> 00:06:57.129 you know, it's like walking into a Michelin star 00:06:57.129 --> 00:06:59.310 restaurant asking the world's greatest chef for 00:06:59.310 --> 00:07:01.550 salt and pepper and getting a beautifully seasoned 00:07:01.550 --> 00:07:04.709 world -class meal. Right. But then the next day 00:07:04.709 --> 00:07:07.589 you ask that exact same chef for pepper and salt 00:07:07.589 --> 00:07:09.930 and they hand you a plate of actual garbage. 00:07:10.610 --> 00:07:12.790 If these machines possess all human knowledge, 00:07:12.790 --> 00:07:15.389 how can they be this brittle? To understand the 00:07:15.389 --> 00:07:17.629 why behind that brittleness, we have to look 00:07:17.629 --> 00:07:20.069 at an underlying mechanism called in -context 00:07:20.069 --> 00:07:22.670 learning. This is the absolute key to everything. 00:07:22.959 --> 00:07:25.379 Walk me through it. How does in -context learning 00:07:25.379 --> 00:07:27.399 differ from how the AI was built in the first 00:07:27.399 --> 00:07:30.000 place? Well, when a company builds an LLM, they 00:07:30.000 --> 00:07:32.040 put it through traditional training or fine -tuning. 00:07:32.420 --> 00:07:35.100 That process makes lasting, permanent changes 00:07:35.100 --> 00:07:37.279 to the model's neural weights. Like chiseling 00:07:37.279 --> 00:07:40.259 knowledge into a stone tablet. Exactly. But in 00:07:40.259 --> 00:07:42.959 context, learning is entirely different. It is 00:07:42.959 --> 00:07:45.060 a temporary state. It's an emergent property 00:07:45.060 --> 00:07:47.040 that only appeared when these models reached 00:07:47.040 --> 00:07:49.420 a massive scale. So it wasn't programmed into 00:07:49.420 --> 00:07:52.339 them intentionally? No, not at all. As the models 00:07:52.339 --> 00:07:55.060 ingested trillions of words, they suddenly developed 00:07:55.060 --> 00:07:58.519 the ability to learn to learn on the fly entirely 00:07:58.519 --> 00:08:01.060 based on the text sitting inside your current 00:08:01.060 --> 00:08:04.139 chat window. Wow. It is a form of temporary meta 00:08:04.139 --> 00:08:07.300 -learning. When you type a prompt, the AI uses 00:08:07.300 --> 00:08:10.339 your specific grammar, your formatting, and your 00:08:10.339 --> 00:08:13.040 word order as temporary scaffolding to build 00:08:13.040 --> 00:08:16.060 its internal logic pathways for that specific 00:08:16.060 --> 00:08:18.160 answer. I see. So the prompt is the blueprint 00:08:18.160 --> 00:08:20.560 for the scaffolding. Yes. If I put a comma in 00:08:20.560 --> 00:08:22.939 a weird place or flip the order of my examples, 00:08:23.240 --> 00:08:25.540 I am essentially building the scaffolding slightly 00:08:25.540 --> 00:08:28.360 crooked. The AI follows that crooked blueprint 00:08:28.360 --> 00:08:30.680 exactly, and the whole intellectual building 00:08:30.680 --> 00:08:33.419 collapses. That is exactly what happens. And 00:08:33.419 --> 00:08:35.620 the moment you close the chat, the scaffolding 00:08:35.620 --> 00:08:38.740 vanishes. The AI forgets you and the task entirely. 00:08:38.919 --> 00:08:42.200 That is so wild. Because this sensitivity persists 00:08:42.200 --> 00:08:45.500 even in the absolute largest models, engineers 00:08:45.500 --> 00:08:48.000 have been forced to build specialized diagnostic 00:08:48.000 --> 00:08:51.539 tools just to measure the fragility. like the 00:08:51.539 --> 00:08:53.899 ones mentioned in the research, format spread 00:08:53.899 --> 00:08:57.259 and prompt evil. Yes. Format spread systematically 00:08:57.259 --> 00:08:59.879 evaluates a massive range of plausible prompt 00:08:59.879 --> 00:09:02.500 formats, and prompt evil estimates performance 00:09:02.500 --> 00:09:05.440 distributions. They are desperately trying to 00:09:05.440 --> 00:09:08.080 quantify exactly how sensitive these models are 00:09:08.080 --> 00:09:10.240 so they can build guardrails around them. Which 00:09:10.240 --> 00:09:12.120 means humans couldn't just sit back and hope 00:09:12.120 --> 00:09:15.460 for the best. Faced with this unpredictable fragility, 00:09:15.720 --> 00:09:18.039 we had to invent clever psychological tricks 00:09:18.039 --> 00:09:21.000 to keep the AI's logic from derailing. We essentially 00:09:21.000 --> 00:09:23.919 had to build a human toolkit to manage the machine's 00:09:23.919 --> 00:09:25.799 attention span. And those tools are brilliant 00:09:25.799 --> 00:09:28.519 because they directly exploit the way the AI's 00:09:28.519 --> 00:09:31.200 neural pathways function. Let's look at the first 00:09:31.200 --> 00:09:34.879 major tool in the kit. Multi -shot or few -shot 00:09:34.879 --> 00:09:37.659 learning. This is basically establishing a rhythm 00:09:37.659 --> 00:09:40.080 for the AI to dance to. You give it a pattern, 00:09:40.299 --> 00:09:42.879 the classic example in the notes. You type Maison 00:09:42.879 --> 00:09:45.960 Arrow House, Chat Arrow Cat, Ching Arrow, and 00:09:45.960 --> 00:09:47.899 the AI naturally wants to complete the pattern 00:09:47.899 --> 00:09:50.529 by typing dog. You are priming the pump. You 00:09:50.529 --> 00:09:53.750 are providing exemplars, constraining the AI's 00:09:53.750 --> 00:09:56.289 infinite possibilities down to a single formatting 00:09:56.289 --> 00:09:58.769 pattern. But the real breakthrough in this toolkit 00:09:58.769 --> 00:10:02.450 came in 2022, right? From researchers at Google 00:10:02.450 --> 00:10:04.970 Brain. Yes, with a technique called chain of 00:10:04.970 --> 00:10:07.429 thought prompting. So if U -Shot is just matching 00:10:07.429 --> 00:10:09.690 patterns, what is chain of thought actually doing? 00:10:10.169 --> 00:10:13.070 Chain of thought, or COTI, forces the AI to show 00:10:13.070 --> 00:10:16.960 its work. If you ask a standard model a complex 00:10:16.960 --> 00:10:20.000 multi -step math problem, it will normally try 00:10:20.000 --> 00:10:22.919 to spit out the final answer immediately. Which 00:10:22.919 --> 00:10:25.240 often causes it to hallucinate and get the math 00:10:25.240 --> 00:10:27.840 wrong. Exactly. Chain of thought forces the model 00:10:27.840 --> 00:10:30.159 to generate the intermediate steps of reasoning 00:10:30.159 --> 00:10:33.159 first, mimicking a human train of thought. It's 00:10:33.159 --> 00:10:35.259 like a math teacher forcing you to write out 00:10:35.259 --> 00:10:37.000 every step of the equation on the chalkboard 00:10:37.000 --> 00:10:38.860 instead of just guessing the answer in your head. 00:10:39.080 --> 00:10:41.019 And the results of doing this were absolutely 00:10:41.019 --> 00:10:44.139 massive. The notes mentioned the GSM 8K benchmark. 00:10:45.379 --> 00:10:47.879 When Google applied this to their 540 billion 00:10:47.879 --> 00:10:51.600 parameter Paul Lem model, it achieved state -of 00:10:51.600 --> 00:10:54.379 -the -art results on that test, which is a brutally 00:10:54.379 --> 00:10:57.220 difficult mathematical reasoning benchmark. It 00:10:57.220 --> 00:10:59.559 fundamentally proved that the way you ask the 00:10:59.559 --> 00:11:02.059 quotient limits or expands the model's intelligence. 00:11:02.570 --> 00:11:04.509 But my favorite part of this entire evolution 00:11:04.509 --> 00:11:07.470 is how researchers subsequently hacked this very 00:11:07.470 --> 00:11:10.429 concept. Oh, yes. The zero -shot hack. This is 00:11:10.429 --> 00:11:12.509 incredible. It really is. OK, so originally, 00:11:12.590 --> 00:11:15.070 to get the AI to use Chain of Thought, you had 00:11:15.070 --> 00:11:18.690 to manually write out these long, complex, fully 00:11:18.690 --> 00:11:21.269 -solved examples in your prompt to show it what 00:11:21.269 --> 00:11:23.669 a step -by -step process looked like. Took a 00:11:23.669 --> 00:11:26.190 lot of work. A ton of manual effort. But then 00:11:26.190 --> 00:11:27.950 researchers from Google and the University of 00:11:27.950 --> 00:11:30.169 Tokyo discovered something almost comical. They 00:11:30.169 --> 00:11:31.950 found that you didn't need to write out those. 00:11:31.850 --> 00:11:34.789 examples at all. You literally just had to type 00:11:34.789 --> 00:11:37.429 the words, let's think step by step, at the end 00:11:37.429 --> 00:11:39.809 of your request. It is hilarious, but from a 00:11:39.809 --> 00:11:41.889 mechanical standpoint, it makes total sense. 00:11:42.210 --> 00:11:44.629 Appending that single phrase acted as a linguistic 00:11:44.629 --> 00:11:47.450 trigger. It forced the model's attention mechanism 00:11:47.450 --> 00:11:50.129 to route its processing through that same step 00:11:50.129 --> 00:11:52.850 -by -step reasoning pathway, but without the 00:11:52.850 --> 00:11:55.299 user having to supply any heavy lifting. And 00:11:55.299 --> 00:11:57.720 since you, the listener, are looking to get better 00:11:57.720 --> 00:11:59.860 results in your own daily workflows, this is 00:11:59.860 --> 00:12:02.279 an immediate, incredibly practical takeaway you 00:12:02.279 --> 00:12:05.759 can use today. Yes. If you are asking an AI to 00:12:05.759 --> 00:12:09.159 solve a logic puzzle, debug some code, or plan 00:12:09.159 --> 00:12:12.500 a complex travel itinerary, just add, let's think 00:12:12.500 --> 00:12:15.500 step by step to the end. It's a proven psychological 00:12:15.500 --> 00:12:18.519 hack for a machine. It is a phenomenal tip. And 00:12:18.519 --> 00:12:20.960 the concept has continued to evolve. The research 00:12:20.960 --> 00:12:23.019 details a more advanced generalization of this 00:12:23.019 --> 00:12:25.559 called tree of thought. So if a chain is a straight 00:12:25.559 --> 00:12:28.059 line of logic, a tree sounds like it's branching 00:12:28.059 --> 00:12:31.419 out. Is the AI essentially arguing with itself 00:12:31.419 --> 00:12:33.519 over different possibilities? That is exactly 00:12:33.519 --> 00:12:35.799 what it is doing. While a chain is a single linear 00:12:35.799 --> 00:12:38.620 path, step one, step two, step three, a tree 00:12:38.620 --> 00:12:41.159 of thought prompts the AI to generate multiple 00:12:41.159 --> 00:12:43.440 parallel lines of reasoning simultaneously. Oh 00:12:43.440 --> 00:12:45.860 wow. Yeah, it explores different paths, evaluates 00:12:45.860 --> 00:12:48.379 the validity of each one, and can even backtrack 00:12:48.379 --> 00:12:51.259 if it realizes a specific path leads to a logical 00:12:51.259 --> 00:12:53.960 dead end. It is very similar to how a chess computer 00:12:53.960 --> 00:12:56.559 evaluates thousands of potential future board 00:12:56.559 --> 00:12:59.200 states before selecting the absolute best move. 00:12:59.370 --> 00:13:01.750 It's fascinating how we are basically teaching 00:13:01.750 --> 00:13:05.370 it to deliberate. But knowing how to ask an AI 00:13:05.370 --> 00:13:08.389 for math or logic doesn't help us at all when 00:13:08.389 --> 00:13:10.509 we change the medium. What happens when we move 00:13:10.509 --> 00:13:13.190 beyond text and ask an AI to paint a picture 00:13:13.190 --> 00:13:16.960 or write software? vocabulary required seems 00:13:16.960 --> 00:13:18.980 completely different. It is completely different 00:13:18.980 --> 00:13:21.500 because the underlying architecture of early 00:13:21.500 --> 00:13:24.340 text to image models, systems like stable diffusion 00:13:24.340 --> 00:13:26.960 or mid -journey, is fundamentally distinct from 00:13:26.960 --> 00:13:29.700 an LLM. Right. The research points out that these 00:13:29.700 --> 00:13:31.960 image models, especially the ones that exploded 00:13:31.960 --> 00:13:35.279 around 2022, do not actually understand grammar, 00:13:35.419 --> 00:13:38.220 syntax, or sentence structure the way text models 00:13:38.220 --> 00:13:40.399 do. And this leads to a hilarious failure point. 00:13:40.620 --> 00:13:42.600 Yes, the one I'm going to call the cake dilemma. 00:13:42.750 --> 00:13:45.610 The cake dilemma perfectly illustrates the architectural 00:13:45.610 --> 00:13:49.049 gap. If you type the prompt, a party with no 00:13:49.049 --> 00:13:52.049 cake, into an early text to image model, it will 00:13:52.049 --> 00:13:54.230 almost certainly generate an image featuring 00:13:54.230 --> 00:13:56.750 a massive prominently displayed cake. Because 00:13:56.750 --> 00:13:59.490 it completely ignores the word no. It lacks the 00:13:59.490 --> 00:14:02.110 linguistic parser to understand negation. It 00:14:02.110 --> 00:14:05.450 just sees the word cake. accesses its mathematical 00:14:05.450 --> 00:14:07.450 representation of what a cake looks like and 00:14:07.450 --> 00:14:09.769 throws it onto the canvas. Image models operate 00:14:09.769 --> 00:14:12.590 much more on keyword association and spatial 00:14:12.590 --> 00:14:15.990 weight rather than narrative logic. This is why 00:14:15.990 --> 00:14:18.950 word order becomes a supreme variable in image 00:14:18.950 --> 00:14:21.730 prompting. Because words placed closer to the 00:14:21.730 --> 00:14:24.330 absolute beginning of an image prompt are weighted 00:14:24.330 --> 00:14:27.169 far more heavily by the AI's attention mechanism. 00:14:27.429 --> 00:14:30.090 Right. So if you want the scene to be a dark, 00:14:30.330 --> 00:14:32.750 stormy night, you have to prioritize those words 00:14:32.750 --> 00:14:35.269 at the very front of the prompt. Long before 00:14:35.269 --> 00:14:37.230 you start describing the character standing in 00:14:37.230 --> 00:14:39.870 the foreground, or the storm might just get entirely 00:14:39.870 --> 00:14:42.990 ignored. Precisely. And because describing complex 00:14:42.990 --> 00:14:45.860 aesthetics purely with adjectives is so inefficient, 00:14:46.340 --> 00:14:48.179 users quickly figured out how to use established 00:14:48.179 --> 00:14:50.419 reference points. They started imitating styles. 00:14:50.559 --> 00:14:52.759 Yes, the shortcut of using artists as names. 00:14:53.379 --> 00:14:55.379 The research specifically mentions the Polish 00:14:55.379 --> 00:14:58.580 digital artist, Greg Rodkowski, and the surrealist, 00:14:58.679 --> 00:15:01.200 Salvador Dali. It saves so much time. Instead 00:15:01.200 --> 00:15:03.080 of spending three paragraphs trying to describe 00:15:03.080 --> 00:15:05.419 a melting surreal dreamscape with a specific 00:15:05.419 --> 00:15:07.879 color palette, you just type your subject and 00:15:07.879 --> 00:15:10.700 add, in the style of Salvador Dali. But here 00:15:10.700 --> 00:15:13.039 is where the engineering gets incredibly clever. 00:15:13.610 --> 00:15:16.750 What if the specific aesthetic you want in your 00:15:16.750 --> 00:15:19.730 head doesn't belong to a famous artist? What 00:15:19.730 --> 00:15:21.909 if it's just a vibe you like from a few random 00:15:21.909 --> 00:15:25.190 photos? Exactly. This requires a technique called 00:15:25.190 --> 00:15:27.909 textual inversion. I really need you to break 00:15:27.909 --> 00:15:30.610 this one down for me. How does textual inversion 00:15:30.610 --> 00:15:32.909 actually work under the hood? Okay, think of 00:15:32.909 --> 00:15:36.389 the AI's brain as a massive, high -dimensional 00:15:36.389 --> 00:15:40.029 library where every concept has a specific coordinate. 00:15:40.200 --> 00:15:43.360 A specific shade of blue is on one shelf, a specific 00:15:43.360 --> 00:15:47.100 texture of oil paint is on another. Textual inversion 00:15:47.100 --> 00:15:49.600 lets you carve out a brand new shelf. You take 00:15:49.600 --> 00:15:52.059 three to five images of your highly specific, 00:15:52.320 --> 00:15:54.500 unique artistic style and you feed them to the 00:15:54.500 --> 00:15:57.059 AI. Just a handful of images. Yep. The system 00:15:57.059 --> 00:15:59.480 then performs a mathematical optimization to 00:15:59.480 --> 00:16:01.779 find an empty coordinate in its library that 00:16:01.779 --> 00:16:04.179 perfectly captures the shared elements of those 00:16:04.179 --> 00:16:08.139 images. So it is mathematically inventing a completely 00:16:08.139 --> 00:16:11.940 new concept based on visual data. Yes, it creates 00:16:11.940 --> 00:16:14.909 a pseudo word. Embedding a vector that represents 00:16:14.909 --> 00:16:18.470 that exact style. It then links that vector to 00:16:18.470 --> 00:16:21.049 a meaningless string of characters like an asterisk 00:16:21.049 --> 00:16:23.570 or a made -up word. And then what? You can then 00:16:23.570 --> 00:16:26.070 just drop that newly invented pseudo word into 00:16:26.070 --> 00:16:29.929 any text prompt and the AI applies your custom 00:16:29.929 --> 00:16:32.190 style perfectly. That is mind -bending. It's 00:16:32.190 --> 00:16:34.389 like instead of trying to describe a very specific 00:16:34.389 --> 00:16:36.970 shade of blue to a painter using clumsy words, 00:16:37.570 --> 00:16:40.919 the AI just directly hands the painter the exact 00:16:40.919 --> 00:16:43.059 mathematical hex code. And we aren't just doing 00:16:43.059 --> 00:16:45.340 this with visual art. We are applying this high 00:16:45.340 --> 00:16:47.200 level abstraction to software development too. 00:16:47.340 --> 00:16:50.360 The research highlights the 2025 Collins Dictionary 00:16:50.360 --> 00:16:53.240 Word of the Year, which is vibe coding. Vibe 00:16:53.240 --> 00:16:55.600 coding is arguably the peak of human directed 00:16:55.600 --> 00:16:58.559 prompting. It is a massive shift in how humans 00:16:58.559 --> 00:17:00.720 interact with computers. It sounds like magic. 00:17:01.360 --> 00:17:03.940 Vibe coding is this AI assisted software development 00:17:03.940 --> 00:17:06.259 method where the human user doesn't write a single 00:17:06.259 --> 00:17:08.940 line of actual syntax right. Correct. You just 00:17:08.940 --> 00:17:11.980 prompt the LLM with a plain English description 00:17:11.980 --> 00:17:14.319 of what you want the software to do, the vibe 00:17:14.319 --> 00:17:17.079 of the app. And the LLM then acts as your lead 00:17:17.079 --> 00:17:19.710 developer. It translates your vibe into code, 00:17:19.990 --> 00:17:22.890 runs tests, reads its own error logs, and fixes 00:17:22.890 --> 00:17:26.490 its own bugs. It completely democratizes software 00:17:26.490 --> 00:17:29.690 creation. Anyone who can describe a problem clearly 00:17:29.690 --> 00:17:32.230 can now build an app to solve it. That's impressive. 00:17:32.529 --> 00:17:35.609 But vibe coding also represents a ceiling. It 00:17:35.609 --> 00:17:37.990 highlights the ultimate limitation of relying 00:17:37.990 --> 00:17:40.329 on human natural language. Wait, how is vibe 00:17:40.329 --> 00:17:42.869 coding a limitation if it lets anyone build software? 00:17:43.069 --> 00:17:45.930 Because vibe coding is the front -end experience. 00:17:46.430 --> 00:17:49.009 It is great for a human hacking together a personal 00:17:49.009 --> 00:17:51.369 tool. But on the back end, when you're running 00:17:51.369 --> 00:17:54.210 enterprise software, medical databases, or financial 00:17:54.210 --> 00:17:56.450 systems... Human language is a massive liability. 00:17:56.769 --> 00:17:59.109 Exactly. As we discussed with the cake dilemma, 00:17:59.450 --> 00:18:01.809 language is inherently imprecise. If you are 00:18:01.809 --> 00:18:03.990 building automated, reliable infrastructure, 00:18:04.309 --> 00:18:06.950 you cannot afford the AI accidentally generating 00:18:06.950 --> 00:18:09.710 a cake when you asked for no cake. So the solution 00:18:09.710 --> 00:18:12.869 is to remove the human language from the back... 00:18:12.359 --> 00:18:16.559 loop entirely. We let the AI prompt the AI. We 00:18:16.559 --> 00:18:18.859 automate the prompting. And the most foundational 00:18:18.859 --> 00:18:21.279 step in this automation is retrieval augmented 00:18:21.279 --> 00:18:24.980 generation, or RRAG. Break down RRAG for me. 00:18:25.059 --> 00:18:27.839 How does it fix the language problem? Standard 00:18:27.839 --> 00:18:30.480 LLMs freeze their knowledge at the moment their 00:18:30.480 --> 00:18:32.599 training is finished. If you ask them about a 00:18:32.599 --> 00:18:34.319 news event that happened yesterday, they will 00:18:34.319 --> 00:18:36.500 either guess or hallucinate. Because they don't 00:18:36.500 --> 00:18:39.799 actually know. Right. ARAG fixes this by intercepting 00:18:39.799 --> 00:18:42.519 your prompt before the AI sees it. It takes your 00:18:42.519 --> 00:18:44.759 prompt, runs a search through a live database 00:18:44.759 --> 00:18:47.220 or the internet, retrieves the factual data, 00:18:47.819 --> 00:18:49.799 invisibly pastes that data into your prompt, 00:18:49.920 --> 00:18:52.420 and then hands it to the AI to read and format. 00:18:52.579 --> 00:18:55.059 It grounds the AI in reality. It's like giving 00:18:55.059 --> 00:18:57.579 the AI an open book test instead of forcing it 00:18:57.579 --> 00:18:59.799 to rely on its memory. That makes total sense. 00:19:00.099 --> 00:19:03.099 And Microsoft took this a step further with GraphRag. 00:19:03.180 --> 00:19:05.500 What makes the graph version different? GraphFrag 00:19:05.500 --> 00:19:08.079 combines that retrieval process with a knowledge 00:19:08.079 --> 00:19:10.819 graph. So instead of just handing the AI an open 00:19:10.819 --> 00:19:14.200 book, it hands the AI an open book and a detective's 00:19:14.200 --> 00:19:16.640 cork board with string connecting all the key 00:19:16.640 --> 00:19:19.119 entities. Oh, that's brilliant. It maps out the 00:19:19.119 --> 00:19:21.099 relationships between data points before the 00:19:21.099 --> 00:19:24.559 AI even reads it. This allows the model to connect 00:19:24.559 --> 00:19:26.900 disparate pieces of information across massive 00:19:26.900 --> 00:19:29.519 data sets, synthesizing insights that a normal 00:19:29.519 --> 00:19:31.880 search could never find. It was proven highly 00:19:31.880 --> 00:19:34.819 effective on chaotic data sets, like tracking 00:19:34.819 --> 00:19:37.539 complex, interwoven violent incidents in global 00:19:37.539 --> 00:19:40.920 news articles. So ARAG gives the AI better data, 00:19:41.240 --> 00:19:43.400 but the automation goes even deeper than that. 00:19:43.759 --> 00:19:46.220 We now have algorithms specifically designed 00:19:46.220 --> 00:19:49.319 to let LLMs invent the prompts themselves. The 00:19:49.319 --> 00:19:51.660 notes mention the Automatic Tront Engineer algorithm. 00:19:51.859 --> 00:19:53.859 This is a beautiful piece of engineering. You 00:19:53.859 --> 00:19:56.579 set up two LLMs to talk to each other. One is 00:19:56.579 --> 00:19:58.500 the target model you want to control, and the 00:19:58.500 --> 00:20:00.920 other is the prompting model. The prompting model 00:20:00.920 --> 00:20:03.400 looks at the desired input and output, and it 00:20:03.400 --> 00:20:05.279 tries to guess the instructions that would bridge 00:20:05.279 --> 00:20:07.900 the two. It generates several different prompts, 00:20:08.059 --> 00:20:10.599 tests them on the target model, and scores them 00:20:10.599 --> 00:20:12.940 mathematically. And then it takes the highest 00:20:12.940 --> 00:20:15.480 -scoring prompts and mutates them, combining 00:20:15.480 --> 00:20:18.299 the best parts to try and find even better ones. 00:20:18.519 --> 00:20:21.500 It is literally breeding the prompts for success. 00:20:22.160 --> 00:20:24.180 And the research mentions an optimizer called 00:20:24.180 --> 00:20:27.200 GEPay, or genetic pareto, which takes us even 00:20:27.200 --> 00:20:30.460 further. GPay is fascinating. It uses a Pareto 00:20:30.460 --> 00:20:32.819 -based evolutionary search. Let's unpack that. 00:20:32.859 --> 00:20:35.319 What does Pareto mean in this context? A Pareto 00:20:35.319 --> 00:20:38.160 search means it is optimizing for two competing 00:20:38.160 --> 00:20:40.960 priorities at the exact same time, like trying 00:20:40.960 --> 00:20:43.920 to make the prompt both highly accurate and extremely 00:20:43.920 --> 00:20:47.079 fast to process. I see. It breeds a population 00:20:47.079 --> 00:20:49.339 of candidate prompts, evaluates them against 00:20:49.339 --> 00:20:51.940 these competing goals, and evolves them generation 00:20:51.940 --> 00:20:54.380 by generation. And it is incredibly efficient. 00:20:54.619 --> 00:20:57.160 The data shows it beats older reinforcement learning 00:20:57.160 --> 00:21:00.440 methods by about 10%, while using up to 35 times 00:21:00.440 --> 00:21:03.700 fewer test runs. It is Darwinian evolution applied 00:21:03.700 --> 00:21:07.079 to language. But the absolute most extreme version 00:21:07.079 --> 00:21:09.960 of this automation completely abandons English 00:21:09.960 --> 00:21:13.400 or any human language altogether. It uses a technique 00:21:13.400 --> 00:21:16.579 called soft prompting. This is where we leave 00:21:16.579 --> 00:21:19.140 the realm of linguistics and enter pure mathematics. 00:21:20.059 --> 00:21:23.039 Please explain this to me. because my brain short 00:21:23.039 --> 00:21:25.299 circuits at the idea of a prompt with that language. 00:21:25.420 --> 00:21:27.640 It's definitely a leap. If there are no words, 00:21:28.160 --> 00:21:31.039 what is the AI reading? It is reading gradient 00:21:31.039 --> 00:21:33.519 descent. Instead of searching for the perfect 00:21:33.519 --> 00:21:36.140 English adjectives, the system searches directly 00:21:36.140 --> 00:21:38.940 through mathematical space. The prompt is just 00:21:38.940 --> 00:21:41.160 a string of floating -point vectors, literally 00:21:41.160 --> 00:21:43.799 just raw coordinates in the AI's neural network. 00:21:44.000 --> 00:21:47.160 So no words at all? None. The system uses an 00:21:47.160 --> 00:21:49.200 optimization algorithm called gradient descent, 00:21:49.720 --> 00:21:51.859 which is basically finding the absolute lowest 00:21:51.859 --> 00:21:54.200 point in a valley of errors. It mathematically 00:21:54.200 --> 00:21:56.579 tweaks these number strings until it finds the 00:21:56.579 --> 00:21:59.220 exact coordinate that forces the AI to produce 00:21:59.220 --> 00:22:02.549 the correct output. There are no verbs, no nouns, 00:22:02.750 --> 00:22:05.650 no human words involved. It is just directly 00:22:05.650 --> 00:22:08.210 manipulating the machine's brain chemistry. So 00:22:08.210 --> 00:22:10.930 if the absolute most optimized perfect prompts 00:22:10.930 --> 00:22:14.170 are now just strings of floating point numbers 00:22:14.170 --> 00:22:17.650 generated by the AI itself, where does that leave 00:22:17.650 --> 00:22:20.250 us? Have humans been totally locked out of the 00:22:20.250 --> 00:22:22.829 conversation? We aren't locked out, but our role 00:22:22.829 --> 00:22:25.609 has fundamentally shifted. We have moved from 00:22:25.609 --> 00:22:28.349 the art of prompt engineering to what the industry 00:22:28.349 --> 00:22:31.630 now calls context engineering. Context engineering. 00:22:31.829 --> 00:22:33.990 It sounds like the mature grown -up version of 00:22:33.990 --> 00:22:36.430 the job. Precisely. We are no longer sitting 00:22:36.430 --> 00:22:39.109 around tweaking adjectives or adding please and 00:22:39.109 --> 00:22:41.670 thank you to a chat box. Context engineering 00:22:41.670 --> 00:22:44.569 is rigorous back -end software engineering. It 00:22:44.569 --> 00:22:47.349 is managing the environment the AI operates in. 00:22:47.609 --> 00:22:49.390 The research highlights specific operational 00:22:49.390 --> 00:22:51.730 practices here, like managing token budgets and 00:22:51.730 --> 00:22:53.849 using provenance tags. And those are critical 00:22:53.849 --> 00:22:56.329 engineering concepts. A token budget is essentially 00:22:56.329 --> 00:22:59.549 a prepaid word count limit. AI processing is 00:22:59.549 --> 00:23:02.029 expensive, so context engineers have to strictly 00:23:02.029 --> 00:23:05.170 allocate how many tokens or pieces of words a 00:23:05.170 --> 00:23:07.509 system is allowed to think with before it has 00:23:07.509 --> 00:23:10.349 to deliver an answer. Makes sense. And provenance 00:23:10.349 --> 00:23:12.750 tags. Provenance tags are about tracking the 00:23:12.750 --> 00:23:16.450 origin of data. If an AI generates a report based 00:23:16.450 --> 00:23:19.630 on thousands of documents, a provenance tag traces 00:23:19.630 --> 00:23:22.809 exactly which specific sentence in which specific 00:23:22.809 --> 00:23:26.500 document led to a given conclusion. Oh, so it 00:23:26.500 --> 00:23:28.619 ensures that if the AI starts hallucinating, 00:23:29.099 --> 00:23:31.059 engineers can trace the hallucination back to 00:23:31.059 --> 00:23:34.019 the source and cut it off. Yes, it's about observability, 00:23:34.440 --> 00:23:36.440 running regression tests, and ensuring that a 00:23:36.440 --> 00:23:39.259 tiny change in the data doesn't silently break 00:23:39.259 --> 00:23:41.759 an entire enterprise software system. It's about 00:23:41.759 --> 00:23:44.400 reliability and scale, not just chatting. So 00:23:44.400 --> 00:23:46.160 looking back at the incredible notes you provided 00:23:46.160 --> 00:23:48.400 us, we have summarized quite a journey today. 00:23:48.519 --> 00:23:50.779 We really have. We started with prompt engineering 00:23:50.779 --> 00:23:53.460 as this incredibly hyped six -figure job where 00:23:53.460 --> 00:23:55.940 people were basically whispering to machines. 00:23:56.490 --> 00:23:59.470 Then we explored the psychological human to machine 00:23:59.470 --> 00:24:02.410 hacks, like literally telling a supercomputer, 00:24:02.549 --> 00:24:05.529 let's think step by step to force it into a logic 00:24:05.529 --> 00:24:08.410 pathway. And finally, we saw how the front end 00:24:08.410 --> 00:24:10.950 is turning into vibe coding while the back end 00:24:10.950 --> 00:24:13.930 has matured into a highly automated systems engineering 00:24:13.930 --> 00:24:17.369 discipline where AIs evolve their own mathematical 00:24:17.369 --> 00:24:20.900 wordless prompts. It is a perfect example of 00:24:20.900 --> 00:24:23.240 how quickly technology moves from a chaotic, 00:24:23.660 --> 00:24:25.960 mysterious art form to a rigorous, measurable 00:24:25.960 --> 00:24:29.180 science. Wait, sorry, my turn. But yes, and having 00:24:29.180 --> 00:24:31.099 this understanding of how this works under the 00:24:31.099 --> 00:24:33.599 hood gives you a massive advantage over everyone 00:24:33.599 --> 00:24:36.440 else who is just typing blindly into a chat box, 00:24:36.640 --> 00:24:39.339 hoping for the best. Absolutely. You now understand 00:24:39.339 --> 00:24:41.900 the temporary mechanisms of in -context learning, 00:24:42.359 --> 00:24:44.400 the power of chain of thought, and the reality 00:24:44.400 --> 00:24:46.519 of how these systems are actually architected 00:24:46.519 --> 00:24:49.470 to retrieve data. Knowledge truly is the most 00:24:49.470 --> 00:24:52.009 valuable tool you can have as these systems become 00:24:52.009 --> 00:24:54.930 further integrated into our lives. But before 00:24:54.930 --> 00:24:57.089 we wrap up this deep dive into your research, 00:24:57.309 --> 00:25:00.390 we want to leave you with one final highly provocative 00:25:00.390 --> 00:25:02.210 thought from the source material. Oh, this is 00:25:02.210 --> 00:25:04.630 a big one. It's a concept called prompt injection, 00:25:04.829 --> 00:25:06.509 and it flips everything we've talked about on 00:25:06.509 --> 00:25:09.720 its head. It is a profound vulnerability. Now 00:25:09.720 --> 00:25:12.539 that you know exactly how deeply, almost violently 00:25:12.539 --> 00:25:15.279 sensitive these models are to the slightest change 00:25:15.279 --> 00:25:18.259 in syntax, think about the immense cybersecurity 00:25:18.259 --> 00:25:21.279 risks. Hackers are now using prompt injection 00:25:21.279 --> 00:25:24.259 to actively attack these systems. It's terrifying, 00:25:24.359 --> 00:25:26.619 honestly. They craft inputs that look like totally 00:25:26.619 --> 00:25:29.500 normal, innocent text to the user, but to the 00:25:29.500 --> 00:25:32.029 AI's sensitive architecture, They act as malicious 00:25:32.029 --> 00:25:34.849 code designed to hijack the model's core instructions 00:25:34.849 --> 00:25:37.750 and bypass all its safety safeguards. Because 00:25:37.750 --> 00:25:40.109 the model processes the system instructions and 00:25:40.109 --> 00:25:42.410 the users prompt in the exact same linguistic 00:25:42.410 --> 00:25:45.309 space, it struggles to distinguish between trusted 00:25:45.309 --> 00:25:48.690 developer rules and untrusted malicious user 00:25:48.690 --> 00:25:51.250 inputs. It is a fundamental architectural flaw. 00:25:51.569 --> 00:25:53.490 It's like turning a magic spell into a computer 00:25:53.490 --> 00:25:56.109 virus. So here's the question we want you to 00:25:56.109 --> 00:25:59.170 mull over as you go about your day. if a massive 00:25:59.170 --> 00:26:01.569 trillion -parameter artificial intelligence can 00:26:01.569 --> 00:26:04.849 be completely derailed, hijacked, or fundamentally 00:26:04.849 --> 00:26:07.869 altered just by reordering a few words or feeding 00:26:07.869 --> 00:26:10.170 it a cleverly hidden sentence. How can we ever 00:26:10.170 --> 00:26:12.450 truly make them safe from a deliberate linguistic 00:26:12.450 --> 00:26:15.190 attack? It is a chilling question, and it is 00:26:15.190 --> 00:26:18.170 the exact problem the entire cybersecurity industry 00:26:18.170 --> 00:26:21.809 is currently racing to solve. It is wild to think 00:26:21.809 --> 00:26:23.970 about. It really is. Next time you are staring 00:26:23.970 --> 00:26:26.809 at that blinking cursor, Remember the immense 00:26:26.809 --> 00:26:29.569 fragile power sitting right behind it. Keep diving 00:26:29.569 --> 00:26:31.750 deep into your curiosity, keep questioning how 00:26:31.750 --> 00:26:34.349 the tools around you actually work, and we will 00:26:34.349 --> 00:26:35.789 catch you on the next deep dive.