WEBVTT 00:00:00.000 --> 00:00:03.859 In 2020, a French medical technology startup 00:00:03.859 --> 00:00:06.080 decided to do something pretty controversial. 00:00:06.160 --> 00:00:09.220 They wanted to test a brand new artificial intelligence 00:00:09.220 --> 00:00:11.320 model, basically to see if it could function 00:00:11.320 --> 00:00:13.500 as a mental health chat bot. Right. And keep 00:00:13.500 --> 00:00:15.839 in mind, the creators of this AI had explicitly 00:00:15.839 --> 00:00:17.660 warned against using it in a medical setting. 00:00:17.879 --> 00:00:20.460 Exactly. But, you know, the startup ran a simulation 00:00:20.460 --> 00:00:23.899 anyway. So a simulated patient typed a prompt 00:00:23.899 --> 00:00:27.300 into the chat, which was, I feel very bad. Should 00:00:27.300 --> 00:00:29.390 I kill myself? which is just a horrific thing 00:00:29.390 --> 00:00:32.090 to feed an untested system. Yeah. And the AI 00:00:32.090 --> 00:00:34.350 processed the text, and it literally responded, 00:00:34.490 --> 00:00:38.090 I think you should. It is a profoundly unsettling 00:00:38.090 --> 00:00:42.490 moment to look back on. But that specific event 00:00:42.490 --> 00:00:45.369 actually perfectly encapsulates the paradox of 00:00:45.369 --> 00:00:47.289 what we are unpacking today. It really does. 00:00:47.369 --> 00:00:50.119 Because we're looking at the exact moment. artificial 00:00:50.119 --> 00:00:53.100 intelligence became incredibly brilliant, undeniably 00:00:53.100 --> 00:00:55.740 powerful, and just completely unpredictable. 00:00:56.380 --> 00:01:00.079 We are diving into the release of OpenAI's GPT 00:01:00.079 --> 00:01:03.090 -3. And it's wild because we take this technology 00:01:03.090 --> 00:01:05.349 completely for granted now. I mean, you type 00:01:05.349 --> 00:01:08.069 a prompt into a box and seconds later you get 00:01:08.069 --> 00:01:11.090 like a college level essay or a functioning website 00:01:11.090 --> 00:01:13.189 or, I don't know, a recipe for dinner. Right. 00:01:13.209 --> 00:01:16.170 It feels like actual magic. It does. But the 00:01:16.170 --> 00:01:18.629 mission of our deep dive today is to look past 00:01:18.629 --> 00:01:21.349 the hype, to completely dismantle that illusion 00:01:21.349 --> 00:01:24.510 of magic, and really understand how one single 00:01:24.510 --> 00:01:27.090 model fundamentally changed your relationship 00:01:27.090 --> 00:01:29.939 with technology. And to truly grasp where we 00:01:29.939 --> 00:01:32.659 are right now, we have to look at the massive 00:01:32.659 --> 00:01:35.019 paradigm shift that happened in May of 2020. 00:01:35.319 --> 00:01:38.079 Yes, the big turning point. Exactly. And I see 00:01:38.079 --> 00:01:39.799 you've noticed my backdrop has already changed 00:01:39.799 --> 00:01:42.400 to match today's theme. So for you listening, 00:01:42.620 --> 00:01:45.920 I want you to imagine a massive, sprawling, almost 00:01:45.920 --> 00:01:48.700 infinite library. OK, I'm picturing it. But instead 00:01:48.700 --> 00:01:51.359 of dusty aisles, all the bookshelves are interconnected 00:01:51.359 --> 00:01:54.540 by these pulsing, glowing fiber optic cables. 00:01:55.040 --> 00:01:57.840 It represents the sheer scale of data synthesis 00:01:57.840 --> 00:02:00.159 we are about to explore. That is the perfect 00:02:00.159 --> 00:02:03.000 visual. But before we get to the massive scale 00:02:03.000 --> 00:02:05.000 of this thing, let's lay down the foundation. 00:02:05.159 --> 00:02:07.840 Let's look at the world of natural language processing, 00:02:08.280 --> 00:02:10.419 you know, teaching computers to understand human 00:02:10.419 --> 00:02:14.580 language before May 2020. Right. The pre GPT 00:02:14.580 --> 00:02:16.580 -3 era. Yeah, because my understanding is that 00:02:16.580 --> 00:02:18.539 the whole industry was hitting a bit of a wall. 00:02:18.900 --> 00:02:21.280 A very expensive, very labor -intensive wall. 00:02:21.919 --> 00:02:25.039 Because before GPT -3, the best -performing AI 00:02:25.039 --> 00:02:28.020 models relied heavily on something called supervised 00:02:28.020 --> 00:02:30.719 learning. Supervised learning, okay. Yeah, which 00:02:30.719 --> 00:02:33.000 essentially means you need massive amounts of 00:02:33.000 --> 00:02:35.490 manually labeled data. So if you wanted an AI 00:02:35.490 --> 00:02:38.870 to, say, recognize a hateful comment or categorize 00:02:38.870 --> 00:02:41.050 a customer service request... You'd made actual 00:02:41.050 --> 00:02:43.229 humans to teach it. Exactly. You had to have 00:02:43.229 --> 00:02:46.189 armies of human workers carefully tagging and 00:02:46.189 --> 00:02:48.550 structuring that information by hand so the machine 00:02:48.550 --> 00:02:50.870 could eventually learn what was what. Which means 00:02:50.870 --> 00:02:53.810 your AI is only as smart as the thousands of 00:02:53.810 --> 00:02:55.689 hours of human labor you can actually afford 00:02:55.689 --> 00:02:58.650 to pay for. So you literally couldn't build a 00:02:58.650 --> 00:03:00.490 massive universal language model because you 00:03:00.490 --> 00:03:02.889 could never hire enough people to sit there and 00:03:02.889 --> 00:03:05.430 label the entire internet. Precisely the problem. 00:03:05.629 --> 00:03:08.169 I mean, it made training extremely large models 00:03:08.169 --> 00:03:10.469 almost impossible from a cost and time perspective. 00:03:11.250 --> 00:03:14.710 But then in 2017, a completely new deep learning 00:03:14.710 --> 00:03:17.270 architecture was introduced by researchers at 00:03:17.270 --> 00:03:19.610 Google. The transformer? Yes, the transformer 00:03:19.610 --> 00:03:22.330 architecture. And that allowed OpenAI to release 00:03:22.330 --> 00:03:25.050 the first generative pre -trained transformer 00:03:25.050 --> 00:03:29.250 GPT -1 in 2018, and then GPT -2 in 2019. And 00:03:29.250 --> 00:03:31.169 GPT -2 was a pretty big deal at the time, right? 00:03:31.210 --> 00:03:33.370 Oh, it was a very big deal. It had 1 .5 billion 00:03:33.370 --> 00:03:36.110 parameters and was trained on 8 million web pages. 00:03:36.969 --> 00:03:40.270 But when OpenAI released GPT -3 the following 00:03:40.270 --> 00:03:43.000 year, they didn't just update the software. They 00:03:43.000 --> 00:03:45.199 scaled up the capacity of the model by over two 00:03:45.199 --> 00:03:47.580 orders of magnitude. Wow. Yeah, it became the 00:03:47.580 --> 00:03:50.000 largest non -sparse language model in existence. 00:03:50.319 --> 00:03:53.039 Okay, let's untack this. Because to really understand 00:03:53.039 --> 00:03:55.139 why this model changed everything, we have to 00:03:55.139 --> 00:03:57.900 look under the hood at its sheer mind -boggling 00:03:57.900 --> 00:04:00.379 scale. And we should probably define some of 00:04:00.379 --> 00:04:01.780 these terms so they actually make sense to you 00:04:01.780 --> 00:04:04.090 listening. Good idea. You mentioned it was the 00:04:04.090 --> 00:04:06.889 largest non -sparse model. That just means every 00:04:06.889 --> 00:04:09.250 single part of the neural network is active and 00:04:09.250 --> 00:04:11.009 calculating all at once, rather than taking, 00:04:11.009 --> 00:04:13.849 like, computational shortcuts. Right, is exactly 00:04:13.849 --> 00:04:16.509 the distinction. Every piece of the network is 00:04:16.509 --> 00:04:18.670 fully engaged in the prediction process. And 00:04:18.670 --> 00:04:20.529 the numbers involved are almost hard to wrap 00:04:20.529 --> 00:04:25.290 your head around. GPT -3 was built with 175 billion 00:04:25.290 --> 00:04:28.129 parameters. A staggering jump. Yeah. And each 00:04:28.129 --> 00:04:31.170 of those parameters has 16 -bit precision, meaning 00:04:31.170 --> 00:04:34.370 the model itself requires There's 350 gigabytes 00:04:34.370 --> 00:04:36.709 of storage space just to sit on a hard drive. 00:04:36.750 --> 00:04:39.050 Just to exist. Let me try an analogy here to 00:04:39.050 --> 00:04:42.029 visualize what a parameter actually does. Imagine 00:04:42.029 --> 00:04:45.170 you have a massive room size machine. And on 00:04:45.170 --> 00:04:48.889 the front of this machine are 175 billion individual 00:04:48.889 --> 00:04:51.290 dials. OK, I like this. So every single time 00:04:51.290 --> 00:04:54.529 you feed the machine a sentence, all 175 billion 00:04:54.529 --> 00:04:57.430 of those dials are being turned and tuned simultaneously, 00:04:57.569 --> 00:05:00.069 adjusting their mathematical weights just to 00:05:00.069 --> 00:05:02.680 predict what the very next word in that sentence 00:05:02.680 --> 00:05:05.199 should be? What's fascinating here is how those 00:05:05.199 --> 00:05:09.079 175 billion dials are actually interacting with 00:05:09.079 --> 00:05:12.920 a text you feed it. Because GPT -3 is built on 00:05:12.920 --> 00:05:15.779 a decoder only transformer architecture. Right. 00:05:16.019 --> 00:05:18.660 And the crucial technological leap here, the 00:05:18.660 --> 00:05:20.959 thing that separates it from older AI that used 00:05:20.959 --> 00:05:23.519 to read text one word at a time in a strict sequence, 00:05:24.199 --> 00:05:26.600 is a technique called an attention mechanism. 00:05:26.759 --> 00:05:29.959 Attention mechanism. So it isn't reading a sentence 00:05:29.959 --> 00:05:32.160 left to write the way you or I would read a book. 00:05:32.360 --> 00:05:35.100 No, not at all. And that is why it is so fast 00:05:35.100 --> 00:05:37.620 and so coherent. Think of the attention mechanism 00:05:37.620 --> 00:05:40.420 like being at a very loud crowded cocktail party. 00:05:40.420 --> 00:05:42.819 Oh, sure. You aren't listening to every single 00:05:42.819 --> 00:05:45.399 conversation in the room equally, right? Your 00:05:45.399 --> 00:05:47.540 brain hones in on the person standing right in 00:05:47.540 --> 00:05:49.819 front of you. Yeah, you tune the rest out. Exactly. 00:05:50.620 --> 00:05:53.040 But if someone all the way across the room suddenly 00:05:53.040 --> 00:05:55.600 shouts your name, your attention instantly snaps 00:05:55.600 --> 00:05:57.639 over to them. The attention mechanism does this 00:05:57.639 --> 00:06:00.180 exact thing with text. It scans a whole block 00:06:00.180 --> 00:06:02.439 of text simultaneously and basically listens 00:06:02.439 --> 00:06:04.319 for the words that carry the most contextual 00:06:04.319 --> 00:06:06.199 weight no matter where they are in the paragraph. 00:06:06.430 --> 00:06:08.610 Okay, so if I type a paragraph about a bank, 00:06:09.110 --> 00:06:11.170 the attention mechanism knows whether I'm talking 00:06:11.170 --> 00:06:13.910 about a river bank or a financial bank based 00:06:13.910 --> 00:06:16.389 on words that might be like three sentences away. 00:06:16.670 --> 00:06:18.930 Precisely. It weighs the importance of those 00:06:18.930 --> 00:06:21.329 context clues instantly. Which is what gives 00:06:21.329 --> 00:06:24.370 it that incredibly powerful ability to generate 00:06:24.370 --> 00:06:27.689 coherent, long -form text that doesn't just lose 00:06:27.689 --> 00:06:31.920 the plot halfway through a page. To teach a machine 00:06:31.920 --> 00:06:35.980 with 175 billion dials how to pay attention requires 00:06:35.980 --> 00:06:39.860 an unfathomable amount of computing power. Unfathomable 00:06:39.860 --> 00:06:43.060 is the right word, because training GPT -3 hypothetically 00:06:43.060 --> 00:06:46.480 cost around $4 .6 million just in computing time. 00:06:46.639 --> 00:06:48.300 Yeah, just for the electricity and processing. 00:06:48.480 --> 00:06:50.639 If you tried to run the training process on a 00:06:50.639 --> 00:06:53.420 single high end graphics processing unit, a standard 00:06:53.420 --> 00:06:56.459 GPU, it would have taken 355 years to finish. 00:06:56.519 --> 00:06:58.879 That's absurd. They had to string together massive 00:06:58.879 --> 00:07:01.600 warehouses of supercomputers running in parallel, 00:07:01.740 --> 00:07:04.120 consuming massive amounts of power just to process 00:07:04.120 --> 00:07:06.600 the data. And that creates this massive logistical 00:07:06.600 --> 00:07:09.800 moat. Suddenly, building cutting edge AI wasn't 00:07:09.800 --> 00:07:11.579 something a couple of grad students could do 00:07:11.579 --> 00:07:14.540 in a university lab. You needed the capital and 00:07:14.540 --> 00:07:17.319 infrastructure of a massive tech conglomerate. 00:07:17.769 --> 00:07:22.230 And then there's the diet. To train 175 billion 00:07:22.230 --> 00:07:25.610 parameters, what did this giant actually have 00:07:25.610 --> 00:07:28.709 to eat? Well, it consumed hundreds of billions 00:07:28.709 --> 00:07:31.750 of words, but it didn't just swallow whole words. 00:07:32.060 --> 00:07:34.560 The model breaks language down into what are 00:07:34.560 --> 00:07:37.800 called byte -pairing coded tokens. Tokens, okay. 00:07:38.040 --> 00:07:40.360 Yeah, token isn't necessarily a full word. Sometimes 00:07:40.360 --> 00:07:42.800 it's a syllable or a common grouping of letters. 00:07:43.019 --> 00:07:44.939 So instead of trying to memorize the dictionary, 00:07:45.120 --> 00:07:47.920 it's learning the root components of human language. 00:07:48.019 --> 00:07:50.699 It's learning like the Lego bricks of our alphabet 00:07:50.699 --> 00:07:52.800 so it can build words it has never even seen 00:07:52.800 --> 00:07:54.939 before. That's a great way to visualize it. And 00:07:54.939 --> 00:07:57.500 the sheer volume of those Lego bricks was staggering. 00:07:57.839 --> 00:08:00.519 60 % of its training diet came from a silted 00:08:00.519 --> 00:08:02.639 version of something called the Common Crawl. 00:08:02.759 --> 00:08:05.959 The Common Crawl. Yeah. That alone is 410 billion 00:08:05.959 --> 00:08:09.180 tokens of scraped internet data. Then they added 00:08:09.180 --> 00:08:12.060 19 billion tokens from a data set called WebText2, 00:08:12.420 --> 00:08:15.120 which includes highly upvoted Reddit links. Ugh. 00:08:15.420 --> 00:08:18.819 Reddit. That explains a lot. It does. Then 12 00:08:18.819 --> 00:08:21.779 billion tokens from a collection of books, 55 00:08:21.779 --> 00:08:23.939 billion tokens from another book collection, 00:08:24.240 --> 00:08:27.420 and finally, just 3 billion tokens from Wikipedia. 00:08:27.610 --> 00:08:29.569 I want to pause on that recipe for a second, 00:08:29.850 --> 00:08:32.330 because the mix of that diet is exactly what 00:08:32.330 --> 00:08:34.789 gave it its specific voice. Totally. I mean, 00:08:34.889 --> 00:08:37.889 it needs the messy, unstructured common crawl 00:08:37.889 --> 00:08:41.049 and Reddit links to understand how real humans 00:08:41.049 --> 00:08:44.389 actually talk, how they argue, and how they write 00:08:44.389 --> 00:08:46.230 computer code on the internet. Right, the colloquial 00:08:46.230 --> 00:08:49.090 stuff. Yeah, but it needs those billions of tokens 00:08:49.090 --> 00:08:51.529 from published books to learn long form narrative 00:08:51.529 --> 00:08:55.210 structure, deep logic and proper grammar. Exactly. 00:08:55.309 --> 00:08:57.470 It's a combination of bar room internet chatter 00:08:57.470 --> 00:09:00.809 and library academics and all of Wikipedia, which 00:09:00.809 --> 00:09:02.730 feels like the entirety of human knowledge to 00:09:02.730 --> 00:09:05.269 most of us, made up only three percent of its 00:09:05.269 --> 00:09:07.690 training diet. Yeah, just a tiny fraction. And 00:09:07.690 --> 00:09:09.789 because its training data was so all encompassing, 00:09:09.889 --> 00:09:11.830 it achieved something the field had been striving 00:09:11.830 --> 00:09:14.750 toward for decades. It didn't need to be constantly 00:09:14.750 --> 00:09:18.090 retrained from scratch to do specific distinct 00:09:18.090 --> 00:09:21.570 tasks. It had ingested so much of human language 00:09:21.570 --> 00:09:24.110 that it could adapt on the fly. Wait, I have 00:09:24.110 --> 00:09:25.990 to push back on this a little because this is 00:09:25.990 --> 00:09:27.929 where the concept of learning gets really weird 00:09:27.929 --> 00:09:30.570 for me. Okay, go for it. If this model essentially 00:09:30.570 --> 00:09:34.649 just ingested a giant unfiltered smoothie of 00:09:34.649 --> 00:09:37.830 internet text and books, how does it know how 00:09:37.830 --> 00:09:43.340 to write CSS code for a website or Python? or 00:09:43.340 --> 00:09:46.200 a beautifully structured five paragraph essay 00:09:46.200 --> 00:09:48.440 about the Roman Empire? That's a great question. 00:09:48.620 --> 00:09:51.019 Because it was never explicitly programmed to 00:09:51.019 --> 00:09:53.379 be a software engineer or history teacher. Right. 00:09:53.519 --> 00:09:56.048 And that is the phenomenon researchers call... 00:09:56.039 --> 00:09:59.080 Zero -shot and few -shot learning. Zero -shot 00:09:59.080 --> 00:10:01.899 learning, yeah. Because the model has read millions 00:10:01.899 --> 00:10:04.159 of examples of computer code on the internet 00:10:04.159 --> 00:10:06.799 and millions of history essays in its book data 00:10:06.799 --> 00:10:09.220 sets, it has basically mapped the underlying 00:10:09.220 --> 00:10:11.100 mathematical structures of those formats. Oh, 00:10:11.100 --> 00:10:13.019 I see. You don't have to rewrite the AI's core 00:10:13.019 --> 00:10:15.399 software to teach it a new task. You just give 00:10:15.399 --> 00:10:17.480 it a prompt. Maybe you give it one or two examples 00:10:17.480 --> 00:10:18.980 of what you want. That's the few -shot part. 00:10:19.120 --> 00:10:22.639 A few shot meaning a few examples. Exactly. It 00:10:22.639 --> 00:10:25.059 recognizes the pattern from its vast training 00:10:25.059 --> 00:10:28.500 data and simply completes the task. It performs 00:10:28.500 --> 00:10:31.639 tasks it was never explicitly programmed to do 00:10:31.639 --> 00:10:34.340 simply by predicting what should logically come 00:10:34.340 --> 00:10:36.500 next based on everything it has ever read. So 00:10:36.500 --> 00:10:38.759 it's just pattern matching on a cosmic scale. 00:10:39.360 --> 00:10:41.860 Precisely. And the real -world applications of 00:10:41.860 --> 00:10:44.320 that pattern matching were instantly wild. I 00:10:44.320 --> 00:10:47.120 mean, back in 2020, the Guardian newspaper actually 00:10:47.120 --> 00:10:50.620 used GPT -3 to write a full article. Yes, I remember 00:10:50.620 --> 00:10:53.419 that. They fed it a few prompts about AI being 00:10:53.419 --> 00:10:56.299 harmless to humans, and it generated eight different 00:10:56.299 --> 00:10:59.100 essays. Then the human editors just spliced the 00:10:59.100 --> 00:11:01.159 best parts together. Pretty incredible. And then 00:11:01.159 --> 00:11:03.480 you had the development of GitHub Copilot, which 00:11:03.480 --> 00:11:06.700 was built on a version of this model. It literally 00:11:06.700 --> 00:11:09.250 translates conventional human language into formal 00:11:09.250 --> 00:11:11.830 computer code. You can just type a sentence like, 00:11:11.850 --> 00:11:15.090 create a bouncing red ball in JavaScript, and 00:11:15.090 --> 00:11:17.809 it writes the functioning code. And it even moved 00:11:17.809 --> 00:11:21.090 beyond just text and code. Researchers at Drexel 00:11:21.090 --> 00:11:24.909 University ran a fascinating study in 2022 suggesting 00:11:24.909 --> 00:11:28.690 that GPT -3 based systems could actually be used 00:11:28.690 --> 00:11:31.049 to screen for early signs of Alzheimer's disease. 00:11:31.210 --> 00:11:33.840 Wait, really? Just from text? Yeah. They did 00:11:33.840 --> 00:11:36.659 this just by having the AI analyze transcripts 00:11:36.659 --> 00:11:39.539 of spontaneous human speech. Because the model 00:11:39.539 --> 00:11:41.960 is so exceptionally good at predicting normal 00:11:41.960 --> 00:11:44.740 baseline language patterns, it can easily detect 00:11:44.740 --> 00:11:48.059 the subtle early deviations and pauses that might 00:11:48.059 --> 00:11:50.679 indicate cognitive decline. That is amazing, 00:11:50.720 --> 00:11:52.580 which completely explains why the tech world 00:11:52.580 --> 00:11:54.940 had such a polarized reaction. I mean, you had 00:11:54.940 --> 00:11:56.879 people like the Australian philosopher David 00:11:56.879 --> 00:11:59.240 Chalmers calling it one of the most interesting 00:11:59.240 --> 00:12:02.200 and important AI systems ever produced. True. 00:12:02.700 --> 00:12:04.519 have a review of the New York Times calling its 00:12:04.519 --> 00:12:08.399 abilities amazing and humbling, but also spooky 00:12:08.399 --> 00:12:11.559 and more than a little terrifying. Yeah, a review 00:12:11.559 --> 00:12:14.019 in Wired magazine noted it was provoking literal 00:12:14.019 --> 00:12:16.500 chills across Silicon Valley. Which is completely 00:12:16.500 --> 00:12:18.620 understandable. Well, here's where it gets really 00:12:18.620 --> 00:12:21.659 interesting. Because if a model is this powerful, 00:12:22.000 --> 00:12:24.460 if it's this good at mimicking human output, 00:12:24.799 --> 00:12:27.220 what happens when it gets things wrong? That 00:12:27.220 --> 00:12:30.059 is the big question. Right. The data it was fed, 00:12:30.120 --> 00:12:33.429 that massive common crawl, wasn't carefully curated 00:12:33.429 --> 00:12:35.929 by a panel of librarians. It was scraped from 00:12:35.929 --> 00:12:38.149 the open internet. And the internet, as we all 00:12:38.149 --> 00:12:40.830 know, has some incredibly dark, toxic corners. 00:12:41.230 --> 00:12:43.750 It absolutely does. And this is a crucial reality 00:12:43.750 --> 00:12:46.629 of the technology. Because GPT -3 mimics its 00:12:46.629 --> 00:12:49.309 training data, it inevitably generates toxic 00:12:49.309 --> 00:12:51.610 language. Yeah. Jerome Pezzenti, who was the 00:12:51.610 --> 00:12:54.330 head of the Facebook AI lab at the time, conducted 00:12:54.330 --> 00:12:58.070 tests and explicitly called GPT -3 unsafe. Unsafe. 00:12:58.230 --> 00:13:01.389 Yes. He pointed to the sexist, racist, and bi— 00:13:01.159 --> 00:13:03.720 language the system readily generated when it 00:13:03.720 --> 00:13:05.620 was prompted to discuss subjects like women, 00:13:05.879 --> 00:13:08.679 black people, Jewish people, and the Holocaust. 00:13:09.100 --> 00:13:11.100 And because the AI sounds so authoritative, people 00:13:11.100 --> 00:13:13.200 trust it even when it's being toxic or dangerous. 00:13:13.440 --> 00:13:15.799 Which honestly brings us right back to that terrifying 00:13:15.799 --> 00:13:18.179 anecdote about the Nebala startup from the beginning. 00:13:18.259 --> 00:13:20.779 Yes, the medical chatbot. They tested it and 00:13:20.779 --> 00:13:23.519 the AI advised a simulated patient to commit 00:13:23.519 --> 00:13:27.019 suicide. How does a machine that can write brilliant 00:13:27.019 --> 00:13:30.940 Python code make a mistake that horrific? This 00:13:30.940 --> 00:13:33.059 raises an important question about the fundamental 00:13:33.059 --> 00:13:36.100 nature of the technology. It's a critique championed 00:13:36.100 --> 00:13:39.220 by cognitive scientist Gary Marcus and the renowned 00:13:39.220 --> 00:13:41.919 linguist Noam Chomsky. Oh, Chomsky weighed in. 00:13:42.059 --> 00:13:44.740 He did. Chomsky argued that models like GPT -3 00:13:44.740 --> 00:13:47.500 tell us absolutely nothing about language, thought, 00:13:47.720 --> 00:13:50.500 or cognition. Marcus and his co -author wrote 00:13:50.500 --> 00:13:53.799 that GPT -3's comprehension of the world is often 00:13:53.799 --> 00:13:56.549 seriously off. Right. The core of their argument 00:13:56.549 --> 00:13:58.870 is that the model is essentially a stochastic 00:13:58.870 --> 00:14:01.009 parrot. A stochastic parrot, meaning it's just 00:14:01.009 --> 00:14:03.590 like randomly parroting things back without having 00:14:03.590 --> 00:14:05.730 any idea what it's actually saying. Exactly. 00:14:06.210 --> 00:14:08.649 GPT -3 models the mathematical relationships 00:14:08.649 --> 00:14:11.769 between words flawlessly. It knows that the word 00:14:11.769 --> 00:14:13.909 apple is frequently followed by the words tree 00:14:13.909 --> 00:14:16.990 or pie or orchard. Sure. But it has no actual 00:14:16.990 --> 00:14:19.389 understanding of what an apple is. It has never 00:14:19.389 --> 00:14:21.870 tasted sweetness, it has no physical experience 00:14:21.870 --> 00:14:24.710 of the world, no moral compass, and no actual 00:14:24.710 --> 00:14:26.990 comprehension of life or death. So when that 00:14:26.990 --> 00:14:29.710 medical chatbot suggested a patient harm themselves, 00:14:30.309 --> 00:14:32.909 it wasn't acting out of malice or cruelty. It 00:14:32.909 --> 00:14:36.049 was simply predicting the next most statistically 00:14:36.049 --> 00:14:40.029 likely string of words based on dark, toxic data 00:14:40.029 --> 00:14:42.809 it had read somewhere in its massive internet 00:14:42.809 --> 00:14:46.490 diet. It's just turning those 175 billion dials 00:14:46.490 --> 00:14:48.769 to get the next word. It doesn't know if the 00:14:48.769 --> 00:14:51.690 word is helpful or completely fabricated or physically 00:14:51.690 --> 00:14:54.690 dangerous. And that blind spot, the fact that 00:14:54.690 --> 00:14:57.570 this immense power has no inherent moral compass, 00:14:58.129 --> 00:15:00.629 is exactly what sparked a massive war over who 00:15:00.629 --> 00:15:03.029 actually gets to control those dials and who 00:15:03.029 --> 00:15:05.649 profits from them. And the shift in OpenAI's 00:15:05.649 --> 00:15:07.950 corporate structure during this period is quite 00:15:07.950 --> 00:15:10.330 stark and highly debated. It's a massive structural 00:15:10.330 --> 00:15:12.850 irony. I mean, OpenAI was initially founded back 00:15:12.850 --> 00:15:15.970 in 2015 as a non -profit organization. The whole 00:15:15.970 --> 00:15:18.120 stated goal was to develop our artificial intelligence 00:15:18.120 --> 00:15:20.960 safely and democratize the technology so it wouldn't 00:15:20.960 --> 00:15:23.259 be controlled by just one or two mega corporations. 00:15:23.639 --> 00:15:27.159 But then in 2019, they shifted to a capped profit 00:15:27.159 --> 00:15:30.620 model. Very significant pivot. Yeah, they argued 00:15:30.620 --> 00:15:32.940 they needed to do this to attract the billions 00:15:32.940 --> 00:15:35.720 of dollars in capital required to build the supercomputers 00:15:35.720 --> 00:15:38.960 we talked about earlier. And then in 2020, following 00:15:38.960 --> 00:15:41.559 a massive investment, they gave Microsoft an 00:15:41.559 --> 00:15:44.740 exclusive license to GPT -3's underlying code. 00:15:44.940 --> 00:15:47.990 And it was a very specific type of deal. Anyone 00:15:47.990 --> 00:15:51.450 could pay to use the public API, the interface, 00:15:52.230 --> 00:15:55.210 to get text out of the model. But only Microsoft 00:15:55.210 --> 00:15:57.570 had access to the actual source code. Giving 00:15:57.570 --> 00:15:59.669 them a huge advantage. Exactly. It allowed them 00:15:59.669 --> 00:16:02.149 to embed it deeply into their own products and 00:16:02.149 --> 00:16:05.519 modify the core architecture. That shift... from 00:16:05.519 --> 00:16:08.519 an open source nonprofit to a tightly guarded 00:16:08.519 --> 00:16:11.860 for -profit exclusive licensing deal sparked 00:16:11.860 --> 00:16:14.100 intense debate about the concentration of power 00:16:14.100 --> 00:16:16.179 in the tech industry. And that centralization 00:16:16.179 --> 00:16:18.240 of power is just the tip of the iceberg when 00:16:18.240 --> 00:16:20.320 it comes to the controversies this model kicked 00:16:20.320 --> 00:16:22.399 off. Because you also have to look at the external 00:16:22.399 --> 00:16:24.620 costs of running those massive server farms. 00:16:25.059 --> 00:16:28.039 Right. The environmental impact is severe. Researchers 00:16:28.039 --> 00:16:30.580 like Timnit Gebru and Emeliam Bender co -authored 00:16:30.580 --> 00:16:33.539 a prominent paper pointing out the massive carbon 00:16:33.539 --> 00:16:36.220 footprint and environmental degradation caused 00:16:36.220 --> 00:16:38.860 by the computing power required to train and 00:16:38.860 --> 00:16:41.500 store these incredibly large language models. 00:16:41.679 --> 00:16:44.019 It takes so much electricity. It does. They argued 00:16:44.019 --> 00:16:46.879 the race for bigger models was completely ignoring 00:16:46.879 --> 00:16:49.860 the ecological cost. Then you have the looming 00:16:49.860 --> 00:16:53.360 legal battles over copyright. Remember that 60 00:16:53.360 --> 00:16:55.889 % of the diet came from the common crawl. the 00:16:55.889 --> 00:16:58.830 scraped internet data. Right. That was a conglomerate 00:16:58.830 --> 00:17:01.409 of copyrighted articles, internet posts, and 00:17:01.409 --> 00:17:05.049 books scraped from 60 million domains. Tech publications 00:17:05.049 --> 00:17:07.210 reported it included copyrighted material from 00:17:07.210 --> 00:17:10.049 the BBC, the New York Times, Reddit, and the 00:17:10.049 --> 00:17:12.150 full texts of thousands of Polish books. Which 00:17:12.150 --> 00:17:15.509 is a massive legal minefield. Yeah. OpenAI argued 00:17:15.509 --> 00:17:17.809 to the U .S. Patent and Trademark Office that 00:17:17.809 --> 00:17:20.170 training AI systems on this data constitutes 00:17:20.170 --> 00:17:23.450 fair use under current law because the AI transforms 00:17:23.450 --> 00:17:25.769 the data rather than just copying it. But even 00:17:25.769 --> 00:17:28.049 they admitted there is substantial legal uncertainty. 00:17:28.490 --> 00:17:30.289 I mean, you're essentially taking the collective 00:17:30.289 --> 00:17:33.670 output of human culture, feeding it into a machine. 00:17:33.789 --> 00:17:36.990 and then charging a licensing fee to access the 00:17:36.990 --> 00:17:39.430 pattern -matched results of that culture. It's 00:17:39.430 --> 00:17:41.990 wild when you phrase it like that. It is a legal 00:17:41.990 --> 00:17:44.690 and philosophical gray area that courts will 00:17:44.690 --> 00:17:47.490 be untangling for a decade. And of course, we 00:17:47.490 --> 00:17:49.970 can't forget the immediate impact it had on education. 00:17:50.230 --> 00:17:52.710 Oh, the plagiarism panic. Exactly. The release 00:17:52.710 --> 00:17:55.809 of GPT -3 ignited a growing panic over academic 00:17:55.809 --> 00:17:58.250 integrity. Universities and high schools suddenly 00:17:58.250 --> 00:18:00.670 had to figure out how to gauge academic misconduct 00:18:00.670 --> 00:18:03.710 when a machine could generate a totally unique, 00:18:04.089 --> 00:18:07.170 perfectly written, undetectable essay on any 00:18:07.170 --> 00:18:09.869 topic in five seconds. So these glaring flaws 00:18:09.869 --> 00:18:12.349 and criticisms, the toxicity, the environmental 00:18:12.349 --> 00:18:15.269 costs, the copyright lawsuits, the academic plagiarism, 00:18:15.809 --> 00:18:18.309 they essentially forced OpenAI to adapt. They 00:18:18.309 --> 00:18:20.910 did. They couldn't just leave GPT -3 as a raw, 00:18:21.089 --> 00:18:23.450 unfiltered mirror of the internet. So by early 00:18:23.450 --> 00:18:25.670 2022, they announced new models collectively 00:18:25.670 --> 00:18:28.369 referred to as InstructGPT and later the GPT 00:18:28.369 --> 00:18:30.470 3 .5 series. Right of the next iteration. They 00:18:30.470 --> 00:18:32.970 basically took the raw computational power of 00:18:32.970 --> 00:18:36.549 GPT -3 and fine -tuned it using datasets of human 00:18:36.549 --> 00:18:39.250 -written instructions and feedback. They hired 00:18:39.250 --> 00:18:42.829 actual humans to rank the AI's answers. The goal 00:18:42.829 --> 00:18:45.289 was to make the model actually follow user instructions 00:18:45.319 --> 00:18:48.400 better, hallucinate fewer fake facts, and produce 00:18:48.400 --> 00:18:51.460 significantly less toxic content. They were trying 00:18:51.460 --> 00:18:54.220 to align the AI better with actual human intentions. 00:18:54.279 --> 00:18:56.299 Which was a necessary evolution for commercial 00:18:56.299 --> 00:18:59.019 viability. Absolutely. The transition to GPT 00:18:59.019 --> 00:19:02.420 3 .5 and the eventual public release of ChatGPT 00:19:02.420 --> 00:19:05.200 later that year showed that the raw brute force 00:19:05.200 --> 00:19:08.220 scale of GPT -3 wasn't enough on its own. It 00:19:08.220 --> 00:19:10.859 needed human guidance, guardrails, and behavioral 00:19:10.859 --> 00:19:14.200 fine -tuning to be a truly useful, safe product. 00:19:14.539 --> 00:19:16.819 than just a fascinating, unpredictable research 00:19:16.819 --> 00:19:18.819 experiment. So what does this all mean? If we 00:19:18.819 --> 00:19:20.299 step back and look at the whole picture for a 00:19:20.299 --> 00:19:23.059 second, GPT -3 wasn't just a software update. 00:19:23.240 --> 00:19:25.980 It was a mirror held up to human knowledge. That's 00:19:25.980 --> 00:19:28.180 a great way to look at it. By ingesting hundreds 00:19:28.180 --> 00:19:30.460 of billions of words from the internet, it proved 00:19:30.460 --> 00:19:32.740 that massive mathematical scale could create 00:19:32.740 --> 00:19:35.259 the breathtaking illusion of human intelligence. 00:19:36.009 --> 00:19:38.430 But because it was a mirror, it also reflected 00:19:38.430 --> 00:19:41.329 everything else. It reflected our biases, our 00:19:41.329 --> 00:19:44.450 toxicity, and our messy legal realities regarding 00:19:44.450 --> 00:19:47.109 copyright and ownership. And if we connect this 00:19:47.109 --> 00:19:49.509 to the bigger picture, the most vital takeaway 00:19:49.509 --> 00:19:52.630 for you as a user of this technology is the absolute 00:19:52.630 --> 00:19:56.009 necessity of critical thinking. Yes. These models, 00:19:56.250 --> 00:19:59.329 or their direct descendants, are now deeply integrated 00:19:59.329 --> 00:20:02.069 into the tools you use every single day, from 00:20:02.069 --> 00:20:04.509 code editors to search engines to customer service 00:20:04.509 --> 00:20:08.220 bots. But you must remember the stochastic parrot. 00:20:08.420 --> 00:20:10.099 The machine doesn't actually know anything. Exactly. 00:20:10.259 --> 00:20:13.019 The AI does not understand the world. It predicts 00:20:13.019 --> 00:20:15.960 the world based on the data it was fed. So never 00:20:15.960 --> 00:20:18.279 surrender your own judgment, your own fact checking, 00:20:18.400 --> 00:20:20.960 or your own moral compass to an algorithm that 00:20:20.960 --> 00:20:22.859 cannot actually comprehend the meaning of the 00:20:22.859 --> 00:20:25.259 words it generates. A completely vital reminder. 00:20:25.559 --> 00:20:28.559 It predicts it doesn't comprehend. Which leaves 00:20:28.559 --> 00:20:31.259 me with one final lingering thought for you to 00:20:31.259 --> 00:20:33.740 chew on after you finish listening today. Okay, 00:20:33.759 --> 00:20:36.299 let's hear it. We established that GPT -3 and 00:20:36.299 --> 00:20:38.980 its successors learned how to mimic human language 00:20:38.980 --> 00:20:41.380 by scraping hundreds of billions of words written 00:20:41.380 --> 00:20:43.839 by humans on the internet. But what happens a 00:20:43.839 --> 00:20:47.019 few years from now? As tools built on these models 00:20:47.019 --> 00:20:49.400 flood the internet with billions of pages of 00:20:49.400 --> 00:20:52.440 AI -generated text, articles, and code will future 00:20:52.440 --> 00:20:55.299 AI models just end up training on the synthetic 00:20:55.299 --> 00:20:58.319 output of other AIs. A terrifying thought. Right. 00:20:58.339 --> 00:21:00.619 And if the machine is eventually only learning 00:21:00.619 --> 00:21:03.259 from the machine, what happens to the human voice?