WEBVTT 00:00:00.000 --> 00:00:03.419 Welcome in to today's deep dive. Usually, you 00:00:03.419 --> 00:00:05.839 know, we're parsing through massive white papers 00:00:05.839 --> 00:00:08.839 or these really dense historical texts for you. 00:00:09.019 --> 00:00:12.740 Right. The heavy stuff. Exactly. But today, our 00:00:12.740 --> 00:00:15.240 source material is, well, it's highly unusual. 00:00:15.400 --> 00:00:19.260 It's literally a single Wikipedia disambiguation 00:00:19.260 --> 00:00:22.489 page. Yeah. We're talking about maybe 50 actual 00:00:22.489 --> 00:00:24.969 words of content here. It's tiny. It's basically 00:00:24.969 --> 00:00:27.550 a digital signpost. Just a few sentences at a 00:00:27.550 --> 00:00:29.609 crossroads. Which is crazy because it's so minimalist. 00:00:29.850 --> 00:00:32.350 Right. But our mission today is to show you how 00:00:32.350 --> 00:00:35.710 these few carefully chosen words actually map 00:00:35.710 --> 00:00:39.609 out the architectural fault lines of modern artificial 00:00:39.609 --> 00:00:41.810 intelligence. It forces us to look past all the 00:00:41.810 --> 00:00:43.829 industry buzzwords. We get to actually examine 00:00:43.829 --> 00:00:45.530 the mechanics of what these systems are doing. 00:00:45.670 --> 00:00:47.649 Because while the document is incredibly short, 00:00:47.829 --> 00:00:50.189 it highlights this profound difference in how 00:00:50.189 --> 00:00:52.270 machine actually learn to understand our world. 00:00:52.369 --> 00:00:54.670 Oh absolutely. So the document introduces us 00:00:54.670 --> 00:00:57.130 to two core terms right at the very top. It says 00:00:57.130 --> 00:00:59.250 a few -shot learning and one -shot learning may 00:00:59.250 --> 00:01:02.429 refer to and then it lists the options. Which 00:01:02.429 --> 00:01:04.930 perfectly sets up the central mystery we're gonna 00:01:04.930 --> 00:01:08.189 unravel for you today. Yes. Why does this ultra 00:01:08.189 --> 00:01:11.370 short document explicitly take these two incredibly 00:01:11.370 --> 00:01:15.290 similar sounding concepts, few shot learning 00:01:15.290 --> 00:01:18.170 and one shot learning, and just split them into 00:01:18.170 --> 00:01:20.890 entirely different technological realms? Because 00:01:20.890 --> 00:01:23.010 on the surface, I mean, it sounds like an arbitrary 00:01:23.010 --> 00:01:25.269 division, right? Totally. If you're casually 00:01:25.269 --> 00:01:28.069 following AI. You probably assume these are just 00:01:28.069 --> 00:01:31.150 variations of the exact same algorithm, just 00:01:31.150 --> 00:01:33.549 with a different number of data points. OK. Let's 00:01:33.549 --> 00:01:35.650 unpack this terminology first before we even 00:01:35.650 --> 00:01:37.709 look at the separate technological fields. Let's 00:01:37.709 --> 00:01:40.510 do it. When this document groups few shot and 00:01:40.510 --> 00:01:43.530 one shot together, what does the concept of a 00:01:43.530 --> 00:01:46.670 shot actually mean mechanically? Good question. 00:01:47.069 --> 00:01:48.769 Because, I mean, think about human learning, 00:01:48.909 --> 00:01:51.129 right? If I'm teaching you a new card game. OK. 00:01:51.439 --> 00:01:54.519 Do you need to play a few hands like a few shots 00:01:54.519 --> 00:01:57.079 to get it? Or do you just understand the rules 00:01:57.079 --> 00:01:59.780 after seeing just one hand? Right, one shot. 00:01:59.900 --> 00:02:02.280 That's a great analogy for the baseline. But 00:02:02.280 --> 00:02:05.200 I have to push back a little here. If few shot 00:02:05.200 --> 00:02:07.620 and one shot sound like essentially the exact 00:02:07.620 --> 00:02:09.860 same process, just with a different number of 00:02:09.860 --> 00:02:13.060 examples, why does the source immediately separate 00:02:13.060 --> 00:02:15.979 them into totally different subcategories? Well... 00:02:15.900 --> 00:02:18.219 To understand that hard fork in the road, we 00:02:18.219 --> 00:02:20.340 have to look at the underlying reality of how 00:02:20.340 --> 00:02:23.500 that goal is achieved. OK. In the old days, giving 00:02:23.500 --> 00:02:26.259 an AI an example meant fundamentally altering 00:02:26.259 --> 00:02:29.520 its underlying code. You were adjusting the weights 00:02:29.520 --> 00:02:31.659 and biases of the neural network permanently. 00:02:32.300 --> 00:02:35.319 Which required massive data sets, right? Millions 00:02:35.319 --> 00:02:37.800 of shots. Millions of labeled pictures of cats 00:02:37.800 --> 00:02:40.580 just to recognize a cat, exactly. But when we 00:02:40.580 --> 00:02:43.020 talk about a shot today, specifically in modern 00:02:43.020 --> 00:02:46.180 foundational models, we're talking about in -context 00:02:46.180 --> 00:02:48.159 learning. So you're not retraining the model. 00:02:48.599 --> 00:02:50.560 No, not at all. You're giving it a temporary 00:02:50.560 --> 00:02:52.219 anchor during inference, which is the moment 00:02:52.219 --> 00:02:54.240 you actually ask it to do something. Got it. 00:02:54.379 --> 00:02:56.879 So it's like, OK, think of a massive generative 00:02:56.879 --> 00:02:59.939 AI model, like a highly detailed multi -dimensional 00:02:59.939 --> 00:03:03.180 map of every concept humanity has ever digitized. 00:03:03.199 --> 00:03:05.870 I love that visual. When you give it zero shots, 00:03:05.969 --> 00:03:07.770 meaning you just give it a blunt command with 00:03:07.770 --> 00:03:10.490 no examples, it drops you in the middle of a 00:03:10.490 --> 00:03:12.770 random continent on that map. Right. You could 00:03:12.770 --> 00:03:15.349 end up anywhere. But when you provide a few shots, 00:03:15.629 --> 00:03:18.129 you are essentially giving it GPS coordinates. 00:03:18.569 --> 00:03:20.889 You're giving it a zip code, a street name, and 00:03:20.889 --> 00:03:23.990 a house number to narrow down the exact neighborhood 00:03:23.990 --> 00:03:26.729 of the answer you want. That spatial analogy 00:03:26.729 --> 00:03:28.849 maps perfectly to the math, actually. Really? 00:03:29.129 --> 00:03:32.050 Yeah. These models operate in what we call a 00:03:32.050 --> 00:03:35.189 latent space. It's a mathematical representation 00:03:35.189 --> 00:03:38.610 of relationships between concepts. Okay. So, 00:03:38.650 --> 00:03:41.210 providing a few examples acts as a gravitational 00:03:41.210 --> 00:03:44.270 pull. It bends that latent space temporarily, 00:03:45.150 --> 00:03:47.590 pulling the AI's output toward the specific style 00:03:47.590 --> 00:03:49.550 or logic you've demonstrated. The underlying 00:03:49.550 --> 00:03:51.349 weights haven't changed, but you've constrained 00:03:51.349 --> 00:03:53.689 the geometry of its possible answers. Exactly. 00:03:54.289 --> 00:03:55.969 But that brings us right back to your pushback. 00:03:56.189 --> 00:03:58.849 The architecture required to follow sequential 00:03:58.849 --> 00:04:01.409 iterative examples is completely different from 00:04:01.409 --> 00:04:03.990 the architecture required to instantly extract 00:04:03.990 --> 00:04:06.750 the invariant truth of an object from one definitive 00:04:06.750 --> 00:04:09.969 look. Which dictates the first major branch of 00:04:09.969 --> 00:04:12.590 our source document. Yes, the generative AI branch. 00:04:13.009 --> 00:04:15.310 Right. So moving from the shared terminology 00:04:15.310 --> 00:04:18.129 into its first specific definition, the text 00:04:18.129 --> 00:04:22.920 reads, a form of prompt engineering in generative 00:04:22.920 --> 00:04:25.819 AI. What's fascinating here is how specific that 00:04:25.819 --> 00:04:29.160 phrasing is. It really is. It explicitly defines 00:04:29.160 --> 00:04:32.040 few -shot learning, not as some background -backend 00:04:32.040 --> 00:04:35.420 programming thing, but as a human -driven interaction. 00:04:35.620 --> 00:04:37.920 Right, it's active. Prompt engineering means 00:04:37.920 --> 00:04:41.139 a user sitting at a keyboard. So what does this 00:04:41.139 --> 00:04:43.879 all mean for the listener? It means it takes 00:04:43.879 --> 00:04:47.240 the magic out of the black box and puts the responsibility 00:04:47.240 --> 00:04:50.579 squarely on you. It's like, uh... Imagine giving 00:04:50.579 --> 00:04:53.019 a chef a couple of specific example dishes. Those 00:04:53.019 --> 00:04:55.000 are your few shots. Okay, yeah. And then you 00:04:55.000 --> 00:04:58.019 ask them to design an entirely new, but stylistically 00:04:58.019 --> 00:05:01.620 similar, menu for a restaurant. Right. By categorizing 00:05:01.620 --> 00:05:04.420 few -shot learning as prompt engineering, the 00:05:04.420 --> 00:05:07.160 source is revealing a crucial limitation of generative 00:05:07.160 --> 00:05:10.360 AI. Which is what? Well, large language models 00:05:10.360 --> 00:05:12.579 are fundamentally sequential prediction engines. 00:05:13.120 --> 00:05:15.079 They guess the next word based on the context 00:05:15.079 --> 00:05:17.300 window they're provided. They thrive on patterns. 00:05:17.360 --> 00:05:20.519 Exactly. But out of the box, their patterns are 00:05:20.519 --> 00:05:23.319 just generalized averages of the entire internet. 00:05:23.459 --> 00:05:25.660 Which means if you want something specific, you 00:05:25.660 --> 00:05:28.240 have to fight the model's pre -training. You 00:05:28.240 --> 00:05:30.759 do. When you're sitting at your desk, endlessly 00:05:30.759 --> 00:05:33.399 tweaking a prompt because the AI is writing an 00:05:33.399 --> 00:05:35.860 email that sounds like a generic corporate robot. 00:05:36.160 --> 00:05:38.639 They've all been there? You're experiencing the 00:05:38.639 --> 00:05:41.220 necessity of few -shot learning in real time. 00:05:41.439 --> 00:05:43.699 You have to provide three or four examples of 00:05:43.699 --> 00:05:45.860 your own writing style just to break the model 00:05:45.860 --> 00:05:48.399 out of its default state. You're hacking the 00:05:48.399 --> 00:05:51.319 context window. That is what prompt engineering 00:05:51.319 --> 00:05:55.560 actually is. It's a workaround. Exactly. Because 00:05:55.560 --> 00:05:58.399 we don't yet have models that can instantly adapt 00:05:58.399 --> 00:06:00.779 their entire personality from a single instruction. 00:06:01.439 --> 00:06:04.379 We use few -shot prompting to build a temporary 00:06:04.379 --> 00:06:07.319 pattern for the machine to lash onto. So generation 00:06:07.319 --> 00:06:10.259 requires a stylistic template. And building a 00:06:10.259 --> 00:06:13.279 template requires multiple reference points. 00:06:13.439 --> 00:06:15.759 Precisely. One point is just a dot. Two points 00:06:15.759 --> 00:06:18.660 make a line. A few points establish a curve that 00:06:18.660 --> 00:06:20.939 the AI can then extrapolate from. Here's where 00:06:20.939 --> 00:06:23.579 it gets really interesting, though. Yeah? Text 00:06:23.579 --> 00:06:25.439 and generative concepts are forgiving, right? 00:06:25.709 --> 00:06:28.589 You can iterate. But what happens when an AI 00:06:28.589 --> 00:06:31.189 needs to process the visual world from just a 00:06:31.189 --> 00:06:34.370 single example? That naturally pivots us to the 00:06:34.370 --> 00:06:36.649 second link in the source document. Right, because 00:06:36.649 --> 00:06:40.129 the text explicitly routes one -shot learning, 00:06:40.329 --> 00:06:43.769 specifically to computer vision. It literally 00:06:43.769 --> 00:06:46.110 has computer vision in parentheses right next 00:06:46.110 --> 00:06:48.230 to it. It does. And I have to raise a skeptical 00:06:48.230 --> 00:06:52.269 question here. Why is one -shot learning completely 00:06:52.269 --> 00:06:55.089 cordoned off into computer vision in this document? 00:06:55.350 --> 00:06:57.899 Seems a bit rigid, doesn't it? Yeah. I mean, 00:06:58.139 --> 00:07:00.819 if I can give a generative AI a few text examples, 00:07:01.300 --> 00:07:03.860 why is seeing something just once isolated to 00:07:03.860 --> 00:07:06.100 the visual realm? I use mid -journey all the 00:07:06.100 --> 00:07:08.300 time. I'll upload three or four different reference 00:07:08.300 --> 00:07:10.439 photos of a character to generate a brand new 00:07:10.439 --> 00:07:12.759 image. Right, you're using multiple images. Exactly. 00:07:12.839 --> 00:07:15.680 That is highly complex computer vision, and I 00:07:15.680 --> 00:07:17.879 am actively using a few -shot approach to do 00:07:17.879 --> 00:07:21.100 it. So why does this digital signpost point in 00:07:21.100 --> 00:07:23.839 such absolute directions? If we connect this 00:07:23.839 --> 00:07:26.509 to the bigger picture, It gets to the heart of 00:07:26.509 --> 00:07:28.949 how language trails behind the bleeding edge 00:07:28.949 --> 00:07:32.069 of tech. When you use mid -journey with multiple 00:07:32.069 --> 00:07:34.829 reference images, you are engaging in generative 00:07:34.829 --> 00:07:36.670 vision. You are creating something that does 00:07:36.670 --> 00:07:40.170 not exist. But when the Wikipedia editors route 00:07:40.170 --> 00:07:43.569 one -shot learning to computer vision, they're 00:07:43.569 --> 00:07:45.730 referring to the classic analytical definition 00:07:45.730 --> 00:07:48.449 of the field, discriminative computer vision. 00:07:48.529 --> 00:07:51.170 Any models designed to understand and categorize 00:07:51.170 --> 00:07:53.990 reality, not paint a new one. Yes. Think of a 00:07:53.990 --> 00:07:56.189 facial recognition system at a secure facility, 00:07:57.129 --> 00:08:00.589 or a medical AI analyzing an MRI for a rare tumor. 00:08:00.870 --> 00:08:03.209 High stakes stuff. Very high stakes. These systems 00:08:03.209 --> 00:08:05.790 don't have a conversational context window. They 00:08:05.790 --> 00:08:09.060 don't iterate. Their mathematical objective is 00:08:09.060 --> 00:08:12.040 to take a single, incredibly noisy array of pixels, 00:08:12.360 --> 00:08:15.480 an image, and extract the absolute ground truth 00:08:15.480 --> 00:08:17.480 features of whatever is in that image. But how 00:08:17.480 --> 00:08:19.720 does the system actually achieve that? How does 00:08:19.720 --> 00:08:22.079 my phone recognize my face in a dark room when 00:08:22.079 --> 00:08:24.360 the only training data I ever gave it was one 00:08:24.360 --> 00:08:27.079 initial scan? It relies on immense foundational 00:08:27.079 --> 00:08:29.720 pre -training. A vision model already knows the 00:08:29.720 --> 00:08:31.740 fundamental physics of light, shadow, edges, 00:08:31.860 --> 00:08:34.480 and geometry. It knows the visual alphabet. Exactly. 00:08:34.779 --> 00:08:37.600 So when it takes that one shot, that one scan 00:08:37.600 --> 00:08:41.179 of your face, it isn't trying to memorize specific 00:08:41.179 --> 00:08:44.720 pixels. It extracts a mathematical signature. 00:08:44.940 --> 00:08:48.440 Oh. We call it a feature vector. It produces 00:08:48.440 --> 00:08:50.659 a string of numbers that represents the unique 00:08:50.659 --> 00:08:53.419 geometric relationships of your specific facial 00:08:53.419 --> 00:08:56.179 features. So that single string of numbers is 00:08:56.179 --> 00:08:58.960 the one shot. Yes. And from that moment on, every 00:08:58.960 --> 00:09:01.139 time you hold up your phone, it generates a new 00:09:01.139 --> 00:09:03.620 vector from the live camera feed and measures 00:09:03.620 --> 00:09:05.639 the mathematical distance between that new vector 00:09:05.639 --> 00:09:08.179 and your original one shot vector. If the numbers 00:09:08.179 --> 00:09:10.909 are close enough, it unlocks. Which highlights 00:09:10.909 --> 00:09:13.309 why the source document elegantly divides these 00:09:13.309 --> 00:09:16.610 tasks. Generative AI involves iterative prompting 00:09:16.610 --> 00:09:19.149 engineering a response with a few shots. Right. 00:09:19.330 --> 00:09:21.330 Whereas computer vision relies on matching a 00:09:21.330 --> 00:09:24.669 visual pattern from a single definitive one -shot 00:09:24.669 --> 00:09:27.129 image. It's separating the creators from the 00:09:27.129 --> 00:09:29.070 observers. That is the fundamental divide. Which 00:09:29.070 --> 00:09:31.450 actually forces us to zoom out and look at the 00:09:31.450 --> 00:09:34.350 necessity of disambiguation itself. Below those 00:09:34.350 --> 00:09:37.169 two links the text has a note. This disambiguation 00:09:37.169 --> 00:09:39.509 page lists articles associated with the title, 00:09:39.750 --> 00:09:42.870 few -shot learning. And then it says, if an internal 00:09:42.870 --> 00:09:45.830 link incorrectly led you here, you may wish to 00:09:45.830 --> 00:09:47.970 change the link to point directly to the intended 00:09:47.970 --> 00:09:50.889 article. This raises an important question about 00:09:50.889 --> 00:09:53.850 the state of information in the AI age. Right. 00:09:53.950 --> 00:09:56.129 Because it's like walking into a giant library 00:09:56.129 --> 00:09:58.870 looking for a book on bugs. And the librarian 00:09:58.870 --> 00:10:01.210 has to ask, do you mean software glitches or 00:10:01.210 --> 00:10:04.409 insects? Exactly. The terminology overlaps so 00:10:04.409 --> 00:10:06.950 perfectly. But how common is it for such cutting 00:10:06.950 --> 00:10:09.549 edge fields to have overlapping terminology like 00:10:09.549 --> 00:10:13.029 this? I mean, why require active disambiguation 00:10:13.029 --> 00:10:15.549 for the public? It points to the immense danger 00:10:15.549 --> 00:10:18.190 of information overload and jargon confusion. 00:10:18.710 --> 00:10:21.429 When the text says, if an internal link incorrectly 00:10:21.429 --> 00:10:24.210 led you here, It's openly acknowledging that 00:10:24.210 --> 00:10:26.649 even encyclopedia editors and tech writers are 00:10:26.649 --> 00:10:28.509 mixing these terms up. They're cross -linking 00:10:28.509 --> 00:10:31.509 to the wrong concepts because few -shot and one 00:10:31.509 --> 00:10:33.330 -shot sound like they belong in the exact same 00:10:33.330 --> 00:10:36.029 article. Right. It completely validates the listener's 00:10:36.029 --> 00:10:38.710 potential confusion. If internal links are getting 00:10:38.710 --> 00:10:41.149 it wrong, it's no wonder the general public gets 00:10:41.149 --> 00:10:43.429 confused trying to keep up with AI news. The 00:10:43.429 --> 00:10:46.129 system is literally admitting, hey, we know our 00:10:46.129 --> 00:10:48.110 own people are linking to the wrong concepts. 00:10:48.210 --> 00:10:51.220 It's a safety net. It's a digital traffic cop 00:10:51.220 --> 00:10:53.860 trying to stop accidents at an intersection where 00:10:53.860 --> 00:10:56.000 the street signs are virtually identical. Even 00:10:56.000 --> 00:10:57.600 though the roads lead to completely different 00:10:57.600 --> 00:11:00.259 cities. Exactly. Well, let's distill this entire 00:11:00.259 --> 00:11:02.679 deep dive down for the listener. We started with 00:11:02.679 --> 00:11:06.460 this tiny minimalist Wikipedia disambiguation 00:11:06.460 --> 00:11:09.600 page. A single digital crossroads. Right. And 00:11:09.600 --> 00:11:13.580 we extracted a core taxonomy. Few shot equals 00:11:13.580 --> 00:11:17.019 prompt engineering in generative AI. It's an 00:11:17.019 --> 00:11:20.200 active human -driven process to temporarily bend 00:11:20.200 --> 00:11:23.340 the latent space. Well, one shot is tethered 00:11:23.340 --> 00:11:26.580 to computer vision. the monumental task of extracting 00:11:26.580 --> 00:11:29.320 an invariant feature vector from a single static 00:11:29.320 --> 00:11:31.500 image. Two different technological features. 00:11:31.899 --> 00:11:33.559 Yeah. Now, before we wrap up, I want to leave 00:11:33.559 --> 00:11:35.960 you with a final provocative thought to mull 00:11:35.960 --> 00:11:38.360 over on your own. Go lay it on us. We've talked 00:11:38.360 --> 00:11:40.840 a lot today about learning from a few shots in 00:11:40.840 --> 00:11:44.220 text or just one shot in vision. Both are huge 00:11:44.220 --> 00:11:46.360 leaps forward compared to the millions of data 00:11:46.360 --> 00:11:48.840 points we used to require. Unbelievable leaps, 00:11:49.000 --> 00:11:53.269 really. Right. If AI can now reliably learn complex 00:11:53.269 --> 00:11:56.889 tasks from a few shots in text, or just one shot 00:11:56.889 --> 00:12:00.169 in vision, how close are we to zero -shot learning? 00:12:00.389 --> 00:12:03.860 Oh, well. The holy grail. Imagine a scenario 00:12:03.860 --> 00:12:06.840 where an AI is asked to perfectly execute a complex 00:12:06.840 --> 00:12:09.639 task it has never seen a single example of before. 00:12:09.879 --> 00:12:13.100 No shots given, no engineered prompts, no reference 00:12:13.100 --> 00:12:16.539 images. Just pure unguided deduction from a standing 00:12:16.539 --> 00:12:18.940 start. Exactly. If this tiny digital signpost 00:12:18.940 --> 00:12:21.860 represents the crossroads of few and one, what 00:12:21.860 --> 00:12:23.639 happens to our relationship with these machines 00:12:23.639 --> 00:12:25.899 when the road finally leads to zero? It marks 00:12:25.899 --> 00:12:28.159 the moment the system no longer needs us to define 00:12:28.159 --> 00:12:30.399 the parameters of reality for it. Something to 00:12:30.399 --> 00:12:31.860 think about the next time you're typing out a 00:12:31.860 --> 00:12:33.340 prompt. Until then, keep exploring.