WEBVTT 00:00:00.000 --> 00:00:03.560 Okay, let's unpack this. Imagine for a second 00:00:03.560 --> 00:00:06.379 that you are in just the most deeply toxic, on 00:00:06.379 --> 00:00:09.000 again, off again relationship imaginable. Oh 00:00:09.000 --> 00:00:11.039 boy. You know exactly the kind I'm talking about, 00:00:11.080 --> 00:00:14.320 right? First, there is this wild, breathless 00:00:14.320 --> 00:00:16.800 infatuation where they promise you the absolute 00:00:16.800 --> 00:00:19.480 moon. I mean, they tell you they are going to 00:00:19.480 --> 00:00:21.260 change your entire life tomorrow. Right, the 00:00:21.260 --> 00:00:24.179 honeymoon phase. Exactly. But then comes the 00:00:24.179 --> 00:00:26.199 bitter, devastating disappointment when they 00:00:26.199 --> 00:00:29.859 completely fail to deliver on any of it. And 00:00:29.859 --> 00:00:32.679 finally, you enter this freezing cold period 00:00:32.679 --> 00:00:35.439 where everyone in your friend group simply refuses 00:00:35.439 --> 00:00:38.399 to even mention the ex's name out loud. Yeah, 00:00:38.479 --> 00:00:40.240 you just pretend they don't exist. Right. And 00:00:40.240 --> 00:00:42.700 today we are taking a deep dive into the history 00:00:42.700 --> 00:00:45.359 of artificial intelligence. And what is so striking 00:00:45.359 --> 00:00:48.780 to me is that for the last 70 years, humanity 00:00:48.780 --> 00:00:51.579 has basically been trapped in that exact toxic 00:00:51.579 --> 00:00:54.820 relationship with AI technology. That is honestly 00:00:54.820 --> 00:00:57.060 that's a highly accurate way to frame it. That 00:00:57.060 --> 00:00:58.840 cold period of the relationship you mentioned, 00:00:59.359 --> 00:01:01.140 that's actually what the technology industry 00:01:01.140 --> 00:01:04.439 literally calls an AI winter. An AI winter. Yeah. 00:01:04.859 --> 00:01:07.540 It describes these cyclical historical periods 00:01:07.540 --> 00:01:12.840 where funding, public interest, and academic 00:01:12.840 --> 00:01:15.379 research in artificial intelligence essentially 00:01:15.379 --> 00:01:18.430 just froze over. completely dried up. And the 00:01:18.430 --> 00:01:20.569 term itself was actually coined during one of 00:01:20.569 --> 00:01:23.469 those really bitter breakups. Really? When was 00:01:23.469 --> 00:01:25.989 that? It first topped up in 1984. There was this 00:01:25.989 --> 00:01:28.450 public debate at the American Association of 00:01:28.450 --> 00:01:31.689 Artificial Intelligence. And two leading researchers, 00:01:31.890 --> 00:01:34.329 Roger Schenck and Marvin Minsky, they were looking 00:01:34.329 --> 00:01:37.109 around at the massive corporate hype of the 1980s. 00:01:37.109 --> 00:01:39.129 And they just started sounding the alarm. Because 00:01:39.129 --> 00:01:41.390 they had seen it before, right? Exactly. They 00:01:41.390 --> 00:01:43.390 had already survived an earlier funding freeze 00:01:43.390 --> 00:01:46.340 back in the 1970s. So they explicitly warned 00:01:46.340 --> 00:01:49.260 the business community that the enthusiasm was 00:01:49.260 --> 00:01:51.019 you know spiraling completely out of control 00:01:51.019 --> 00:01:53.420 they saw the writing on the wall they did I mean 00:01:53.420 --> 00:01:55.760 they predicted a massive chain reaction they 00:01:55.760 --> 00:01:58.219 actually compared it to a nuclear winter Wow 00:01:58.219 --> 00:02:02.019 nuclear winter that's bleak it is but they were 00:02:02.019 --> 00:02:04.780 warning that when the technology inevitably failed 00:02:04.780 --> 00:02:08.060 to meet these utopian promises Pessimism would 00:02:08.060 --> 00:02:10.379 infect the research community. It would bleed 00:02:10.379 --> 00:02:13.240 into the mainstream press and trigger severe 00:02:13.240 --> 00:02:16.539 funding cuts that would effectively end serious 00:02:16.539 --> 00:02:18.699 research. And spoiler alert, they were right. 00:02:19.060 --> 00:02:22.009 They were entirely right. Yeah. Just three years 00:02:22.009 --> 00:02:24.669 after they coined the term, a billion -dollar 00:02:24.669 --> 00:02:27.729 AI industry began to spectacularly collapse. 00:02:28.009 --> 00:02:29.870 Which is just wild to think about. I mean, to 00:02:29.870 --> 00:02:31.909 understand how you even get to a billion -dollar 00:02:31.909 --> 00:02:33.830 collapse in the 1980s, we have to go back to 00:02:33.830 --> 00:02:35.430 the beginning. We have to look at how this promise 00:02:35.430 --> 00:02:38.669 started and why the very first bubble burst so 00:02:38.669 --> 00:02:41.349 incredibly fast. Yeah, back to the 1950s. Yeah, 00:02:41.349 --> 00:02:43.090 because when you look at the origins, it all 00:02:43.090 --> 00:02:45.750 kicks off with this intense Cold War optimism. 00:02:46.310 --> 00:02:48.849 But what fascinates me is that the spark that 00:02:48.849 --> 00:02:52.500 lit the very first boom was essentially a well 00:02:52.500 --> 00:02:55.680 it was a heavily curated parlor trick you're 00:02:55.680 --> 00:02:58.319 referring to the 1954 Georgetown IBM machine 00:02:58.319 --> 00:03:00.699 translation experiment and it absolutely was 00:03:00.699 --> 00:03:03.740 a parlor trick but the hype was instantaneous 00:03:03.740 --> 00:03:06.460 I mean the media absolutely lost its mind over 00:03:06.460 --> 00:03:09.379 this you had headlines screaming things like 00:03:09.379 --> 00:03:12.740 robot brain translates Russian into King's English 00:03:12.740 --> 00:03:14.580 right making it sound like sci -fi was already 00:03:14.580 --> 00:03:17.370 here Exactly. The implication was that a flawless 00:03:17.370 --> 00:03:20.150 automatic translator was just a few years away. 00:03:21.129 --> 00:03:23.330 But wait, if you actually look under the hood 00:03:23.330 --> 00:03:26.469 of that 1954 machine, what was it genuinely doing? 00:03:26.610 --> 00:03:28.449 Because it certainly wasn't reading Tolstoy. 00:03:28.569 --> 00:03:31.229 No, not even close. The sober reality was that 00:03:31.229 --> 00:03:33.610 the machine didn't actually know Russian at all. 00:03:33.770 --> 00:03:36.650 It was only translating 49 carefully selected 00:03:36.650 --> 00:03:39.110 sentences. 49 sentences? Yeah. And it was operating 00:03:39.110 --> 00:03:41.490 with a minuscule vocabulary of just 250 words. 00:03:41.629 --> 00:03:44.370 Wait, 250 words? That is nothing. It's nothing. 00:03:44.650 --> 00:03:46.750 To put that in perspective, human beings need 00:03:46.750 --> 00:03:50.729 a vocabulary of around 8 ,000 to 9 ,000 word 00:03:50.729 --> 00:03:53.210 families just to comprehend written text with 00:03:53.210 --> 00:03:56.389 98 % accuracy. Oh, wow. So the researchers in 00:03:56.389 --> 00:03:58.770 the 1950s, they ran head first into what became 00:03:58.770 --> 00:04:00.530 known as the common sense knowledge problem. 00:04:00.729 --> 00:04:03.460 OK, explain that for me. What is the common sense 00:04:03.460 --> 00:04:06.500 knowledge problem in this context? It fundamentally 00:04:06.500 --> 00:04:10.120 comes down to the profound difficulty of word 00:04:10.120 --> 00:04:13.460 sense disambiguation. That is a mouthful. It 00:04:13.460 --> 00:04:16.699 is, yeah. But basically, for a machine, translating 00:04:16.699 --> 00:04:18.860 a word isn't just about looking it up in a dictionary. 00:04:18.959 --> 00:04:21.480 It's about understanding the context of the real 00:04:21.480 --> 00:04:24.139 world. Think of the word pen. Like a writing 00:04:24.139 --> 00:04:26.000 pen. Right. Are we talking about a writing instrument 00:04:26.000 --> 00:04:28.740 or are we talking about a pig pen? Yeah. As humans, 00:04:29.019 --> 00:04:30.699 our common sense knowledge of the world lets 00:04:30.699 --> 00:04:32.759 us instantly know which one makes sense in a 00:04:32.759 --> 00:04:34.920 sentence. Oh, sure. I don't even have to think 00:04:34.920 --> 00:04:38.319 about it. Exactly. But for early computers, which 00:04:38.319 --> 00:04:40.459 had absolutely no experience of the physical 00:04:40.459 --> 00:04:43.860 world, determining that context was a mathematical 00:04:43.860 --> 00:04:47.339 nightmare. Yes. And there is this absolutely 00:04:47.339 --> 00:04:50.480 amazing, though possibly apocryphal, anecdote 00:04:50.480 --> 00:04:52.740 from that era that perfectly illustrates this. 00:04:52.779 --> 00:04:54.899 The translation game. Right, so they supposedly 00:04:54.899 --> 00:04:57.319 fed the classic biblical phrase, the spirit is 00:04:57.319 --> 00:04:59.779 willing but the flesh is weak, into an early 00:04:59.779 --> 00:05:02.800 translation system. It translated it into Russian, 00:05:03.000 --> 00:05:04.819 and then they translated it back into English 00:05:04.819 --> 00:05:07.779 to see how it did. And the result is legendary. 00:05:08.019 --> 00:05:10.569 It's spin -out. The vodka is good, but the meat 00:05:10.569 --> 00:05:13.329 is rotten. It's such a hilarious result, but 00:05:13.329 --> 00:05:16.129 it perfectly encapsulates the exact technical 00:05:16.129 --> 00:05:19.149 hurdle they couldn't clear. The machine has absolutely 00:05:19.149 --> 00:05:22.589 no idea what a spirit or flesh actually is in 00:05:22.589 --> 00:05:25.449 a metaphorical context. It just maps words. But 00:05:25.449 --> 00:05:28.910 here's my genuine question. If the system was 00:05:28.910 --> 00:05:31.870 that limited, if it was basically just a giant 00:05:31.870 --> 00:05:35.129 mechanical flash card reader matching 250 words 00:05:35.129 --> 00:05:38.149 without any actual understanding, how did the 00:05:38.149 --> 00:05:40.110 researchers convince the U .S. government to 00:05:40.110 --> 00:05:42.310 pour millions and millions of dollars into it? 00:05:42.389 --> 00:05:44.329 Well, it comes down to geopolitical desperation. 00:05:44.990 --> 00:05:46.689 You have to remember the context of the Cold 00:05:46.689 --> 00:05:50.379 War. The CIA and the U .S. military were incredibly 00:05:50.379 --> 00:05:52.300 anxious. They were terrified of the Soviets. 00:05:52.480 --> 00:05:54.779 Right. They desperately wanted a way to instantly 00:05:54.779 --> 00:05:57.060 and automatically translate intercepted Russian 00:05:57.060 --> 00:05:59.060 scientific reports and communications. Yeah. 00:05:59.279 --> 00:06:02.120 So they saw that 1954 demonstration and they 00:06:02.120 --> 00:06:03.860 wanted to believe that an imminent breakthrough 00:06:03.860 --> 00:06:05.839 was right around the corner. They wanted a silver 00:06:05.839 --> 00:06:08.870 bullet. Exactly. And they were also heavily influenced 00:06:08.870 --> 00:06:12.389 by new, highly theoretical work in grammar by 00:06:12.389 --> 00:06:16.110 linguists like Noam Chomsky, which made them 00:06:16.110 --> 00:06:18.589 think language could be easily reduced to pure 00:06:18.589 --> 00:06:21.430 mathematics. But eventually the bill comes due, 00:06:21.689 --> 00:06:24.370 right? The optimism hits a brick wall. Precisely. 00:06:24.480 --> 00:06:28.079 reality caught up. By 1964, the government started 00:06:28.079 --> 00:06:31.160 getting suspicious about the total lack of actual 00:06:31.160 --> 00:06:33.360 usable progress. Like, where's our translator, 00:06:33.759 --> 00:06:35.959 guys? Yeah. So the National Research Council 00:06:35.959 --> 00:06:38.259 formed the Automatic Language Processing Advisory 00:06:38.259 --> 00:06:41.459 Committee, known as ALPAC, to investigate. And 00:06:41.459 --> 00:06:44.480 their 1966 report... was utterly devastating. 00:06:44.519 --> 00:06:46.560 What did it say? They systematically proved that 00:06:46.560 --> 00:06:48.899 machine translation was slower, less accurate, 00:06:49.279 --> 00:06:51.180 and fundamentally more expensive than just paying 00:06:51.180 --> 00:06:53.879 human translators. Ouch. I mean, when your multi 00:06:53.879 --> 00:06:55.839 -million dollar supercomputer is actually worse 00:06:55.839 --> 00:06:57.779 for the budget than just hiring a guy with a 00:06:57.779 --> 00:06:59.959 dictionary. Exactly. And the fallout was immediate. 00:07:00.120 --> 00:07:02.220 The government pulled all support after spending 00:07:02.220 --> 00:07:04.899 some 20 million dollars. Careers were destroyed 00:07:04.899 --> 00:07:08.079 overnight. Research simply ended. The first winter. 00:07:08.399 --> 00:07:12.040 Yeah, it was the first true A .I. winter, proving 00:07:12.040 --> 00:07:15.060 the hard way that curating a few sentences in 00:07:15.060 --> 00:07:17.420 a pristine lab environment does not magically 00:07:17.420 --> 00:07:20.319 scale up to the messy, ambiguous complexities 00:07:20.319 --> 00:07:22.720 of natural human language. So if the government 00:07:22.720 --> 00:07:25.000 realizes they are paying 20 million dollars for 00:07:25.000 --> 00:07:28.180 a parlor trick, what happens to all these researchers 00:07:28.180 --> 00:07:31.160 who promise the moon? Because that realization, 00:07:31.319 --> 00:07:33.339 that lab experiments couldn't just scale to the 00:07:33.339 --> 00:07:35.740 real world, that must have forced a massive reckoning 00:07:35.740 --> 00:07:38.079 across the entire field. Oh, it completely shifted 00:07:38.079 --> 00:07:40.220 the entire landscape of how research was funded. 00:07:40.639 --> 00:07:42.819 So in the 1960s, the Defense Advanced Research 00:07:42.819 --> 00:07:45.579 Projects Agency, or DARPA, had been throwing 00:07:45.579 --> 00:07:48.600 massive grants at AI with almost no strings attached. 00:07:48.740 --> 00:07:51.100 This blank checks. Pretty much. You had directors 00:07:51.100 --> 00:07:53.300 like J .C .R. Licklider, whose whole philosophy 00:07:53.300 --> 00:07:56.720 was, quote, funding people, not projects. They 00:07:56.720 --> 00:07:58.519 were basically handing brilliant scientists money 00:07:58.519 --> 00:08:00.709 to dream. That sounds like a pretty sweet gig. 00:08:01.290 --> 00:08:03.829 It was, until the 1969 Mansfield Amendment was 00:08:03.829 --> 00:08:06.610 passed by Congress. That forced DARPA to pivot. 00:08:07.410 --> 00:08:09.410 Legally, they could only fund projects that had 00:08:09.410 --> 00:08:11.990 strict, mission -oriented military applications. 00:08:12.430 --> 00:08:16.069 Ah! No more blank checks for philosophy. You 00:08:16.069 --> 00:08:17.870 have to build something the military can use 00:08:17.870 --> 00:08:20.970 today. Exactly. Researchers now had to prove 00:08:20.970 --> 00:08:23.470 their work would produce immediate technology. 00:08:23.819 --> 00:08:27.360 and their previous habit of making wildly optimistic 00:08:27.360 --> 00:08:31.399 promises to secure those grants, well, it began 00:08:31.399 --> 00:08:33.460 to backfire dramatically. Right, because now 00:08:33.460 --> 00:08:35.860 they have to actually deliver. A perfect example 00:08:35.860 --> 00:08:38.000 is the Speech Understanding Research Program, 00:08:38.480 --> 00:08:42.340 or SRR, at Carnegie Mellon. DARPA wanted a system 00:08:42.340 --> 00:08:45.179 that could respond to voice commands from a military 00:08:45.179 --> 00:08:48.539 pilot in real time. Okay, high stakes. Very. 00:08:49.100 --> 00:08:51.159 So Carnegie Mellon developed a system called 00:08:51.159 --> 00:08:54.289 Harpy. Now, Harpy could indeed recognize spoken 00:08:54.289 --> 00:08:56.250 English, which sounds like a massive success. 00:08:56.269 --> 00:08:58.230 This was a catch. There was a devastating catch. 00:08:58.970 --> 00:09:01.129 It could only understand the words if the human 00:09:01.129 --> 00:09:04.769 pilot spoke them in a highly specific, pre -programmed 00:09:04.769 --> 00:09:07.070 grammatical order. Which is practically useless. 00:09:07.169 --> 00:09:09.350 I mean, it's like buying an insanely expensive, 00:09:09.590 --> 00:09:11.750 voice -controlled, autonomous military tank. 00:09:11.870 --> 00:09:13.710 But it only follows your instructions if you 00:09:13.710 --> 00:09:15.870 speak to it in a perfect Shakespearean sonnet. 00:09:16.009 --> 00:09:18.250 That's a great way to put it. If you are in the 00:09:18.250 --> 00:09:21.049 middle of a firefight and you stutter... or you 00:09:21.049 --> 00:09:23.269 use the wrong syntax, the tank just sits there 00:09:23.269 --> 00:09:26.250 and does nothing. And DARPA was furious. They 00:09:26.250 --> 00:09:28.590 felt completely duped by the academic community. 00:09:28.990 --> 00:09:31.970 They canceled a $3 million a year contract over 00:09:31.970 --> 00:09:35.710 it. Wow. And this exact same frustration was 00:09:35.710 --> 00:09:37.990 boiling over internationally too. Over in the 00:09:37.990 --> 00:09:40.350 UK, Parliament commissioned Professor Sir James 00:09:40.350 --> 00:09:44.250 Lighthill in 1973 to evaluate the actual state 00:09:44.250 --> 00:09:46.879 of AI. The famous Lighthill Report? Yeah. And 00:09:46.879 --> 00:09:48.820 the resulting Lighthill report was essentially 00:09:48.820 --> 00:09:52.019 a targeted lethal strike on the entire field. 00:09:52.419 --> 00:09:55.120 He criticized AI for its grandiose objectives 00:09:55.120 --> 00:09:57.679 and pointed out that nothing being done in AI 00:09:57.679 --> 00:10:00.360 couldn't just be achieved in other, more traditional 00:10:00.360 --> 00:10:02.809 sciences. He really went for the throat. He did. 00:10:03.269 --> 00:10:05.769 But his most devastating critique, the one that 00:10:05.769 --> 00:10:08.129 really froze the industry, was his identification 00:10:08.129 --> 00:10:10.649 of the combinatorial explosion. The combinatorial 00:10:10.649 --> 00:10:12.350 explosion. Let's make sure we really unpack that 00:10:12.350 --> 00:10:13.970 for you listening, because it sounds like a Michael 00:10:13.970 --> 00:10:16.649 Bay movie title, but it's actually a mathematical 00:10:16.649 --> 00:10:19.149 nightmare that killed the industry. How exactly 00:10:19.149 --> 00:10:21.950 does it work? Think of it this way. Early AI 00:10:21.950 --> 00:10:24.870 algorithms worked perfectly fine on simplified 00:10:24.870 --> 00:10:28.019 toy versions of problems. If you ask a computer 00:10:28.019 --> 00:10:30.440 to calculate every possible move in a game of 00:10:30.440 --> 00:10:33.500 tic -tac -toe, it's easy. There are a very limited 00:10:33.500 --> 00:10:35.500 number of combinations. Right. A kid could do 00:10:35.500 --> 00:10:37.379 it. But as soon as you introduce the variables 00:10:37.379 --> 00:10:40.440 of the real world, say, calculating the best 00:10:40.440 --> 00:10:42.419 route for a delivery truck, navigating a city 00:10:42.419 --> 00:10:45.240 with traffic, construction, weather, the number 00:10:45.240 --> 00:10:47.240 of possible combinations the computer has to 00:10:47.240 --> 00:10:50.559 calculate doesn't just add up. It explodes. It 00:10:50.559 --> 00:10:52.940 explodes exponentially. It goes from a few hundred 00:10:52.940 --> 00:10:56.090 possibilities to trillions upon trillions. The 00:10:56.090 --> 00:10:58.809 computers of the 1970s simply didn't have the 00:10:58.809 --> 00:11:01.929 processing power or memory to handle it. The 00:11:01.929 --> 00:11:03.889 machines would grind to an absolute halt. It 00:11:03.889 --> 00:11:06.129 just freeze. Yeah, the problems were mathematically 00:11:06.129 --> 00:11:08.470 intractable. And once Lighthill proved that these 00:11:08.470 --> 00:11:10.710 machines fundamentally couldn't handle real -world 00:11:10.710 --> 00:11:14.169 math, the money vanished. I mean, following that 00:11:14.169 --> 00:11:16.370 report, AI research in the UK was completely 00:11:16.370 --> 00:11:18.870 dismantled, surviving in only a tiny handful 00:11:18.870 --> 00:11:21.509 of universities. It was a purge. And what's crazy 00:11:21.509 --> 00:11:23.450 to me is that the researchers themselves knew 00:11:23.450 --> 00:11:25.590 they were in trouble. There's this brilliant 00:11:25.590 --> 00:11:28.129 observation from AI researcher Hans Moravec. 00:11:28.509 --> 00:11:31.330 He blamed the crisis entirely on his own colleagues. 00:11:31.649 --> 00:11:34.289 Oh, the exaggeration web. Yes. He said they were 00:11:34.289 --> 00:11:37.269 caught in a web of increasing exaggeration. To 00:11:37.269 --> 00:11:39.409 get the first grant, they promised DARPA the 00:11:39.409 --> 00:11:42.289 moon. When they only delivered a rock, they had 00:11:42.289 --> 00:11:45.009 to promise the moon and the stars just to get 00:11:45.009 --> 00:11:48.309 the next grant to fix the first project. It was 00:11:48.309 --> 00:11:51.090 essentially an academic pondy scheme of hype. 00:11:51.289 --> 00:11:53.429 What's fascinating here is that despite this 00:11:53.429 --> 00:11:56.409 narrative of total devastating collapse, historians 00:11:56.409 --> 00:11:58.690 like Thomas Hay argue that the winter wasn't 00:11:58.690 --> 00:12:02.409 entirely dead. Really? How so? If you look beneath 00:12:02.409 --> 00:12:04.490 the surface of the headlines and the massive 00:12:04.490 --> 00:12:07.090 funding cuts, professional interest was quietly 00:12:07.090 --> 00:12:09.379 thriving. Hay looked at the membership of the 00:12:09.379 --> 00:12:11.399 Association for Computing Machinery's Special 00:12:11.399 --> 00:12:13.879 Interest Group on Artificial Intelligence. Okay, 00:12:14.159 --> 00:12:17.779 SIGART. Right. Between 1973, when the light -hole 00:12:17.779 --> 00:12:20.580 report dropped, and 1978, which is conventionally 00:12:20.580 --> 00:12:22.620 considered the darkest, coldest part of the winter, 00:12:23.419 --> 00:12:25.879 their membership didn't shrink. It almost tripled 00:12:25.879 --> 00:12:29.179 to 3 ,500 members. Oh, wow. So even while the 00:12:29.179 --> 00:12:31.059 government money dried up and the press moved 00:12:31.059 --> 00:12:33.860 on to other things, this dedicated community 00:12:33.860 --> 00:12:36.519 of professionals was quietly growing, sharing 00:12:36.519 --> 00:12:38.960 ideas and building the foundation for the next 00:12:38.960 --> 00:12:41.759 boom. Exactly. But that raises a big question. 00:12:42.320 --> 00:12:45.559 If the military closes its checkbook, these researchers 00:12:45.559 --> 00:12:47.320 have to go somewhere else to keep the lights 00:12:47.320 --> 00:12:50.080 on in their labs. Who was willing to fund them 00:12:50.080 --> 00:12:53.600 in the 1980s? The corporate sector. Ah, big business. 00:12:53.720 --> 00:12:56.100 Yes. Stripped of government defense funding, 00:12:56.600 --> 00:12:59.240 the AI community pivoted hard toward commercial 00:12:59.240 --> 00:13:02.480 enterprise. And this created a highly specialized 00:13:02.480 --> 00:13:05.340 corporate bubble that set the stage for an even 00:13:05.340 --> 00:13:09.519 more spectacular crash. We enter the 1980s era. 00:13:09.740 --> 00:13:13.259 of the expert system. The expert system. By 1985, 00:13:13.840 --> 00:13:15.179 corporations around the world were spending over 00:13:15.179 --> 00:13:17.440 a billion dollars on these systems. And to be 00:13:17.440 --> 00:13:19.940 completely fair, for a brief minute, they actually 00:13:19.940 --> 00:13:22.100 worked. You look at something like the XCon system 00:13:22.100 --> 00:13:23.940 developed at Carnegie Mellon for the Digital 00:13:23.940 --> 00:13:25.799 Equipment Corporation. A massive success story. 00:13:25.960 --> 00:13:28.759 Huge. It was designed to automatically select 00:13:28.759 --> 00:13:30.840 computer components based on customer orders, 00:13:31.019 --> 00:13:33.980 and it saved the company $40 million over just 00:13:33.980 --> 00:13:38.360 six years. That is massive, undeniable ROI. Suddenly, 00:13:38.559 --> 00:13:41.740 every Fortune 500 company wanted an expert system. 00:13:42.019 --> 00:13:45.100 They did. But this corporate boom led directly 00:13:45.100 --> 00:13:47.899 into a massive hardware trap. How so? Because 00:13:47.899 --> 00:13:50.379 the dominant programming language for AI in the 00:13:50.379 --> 00:13:53.610 United States at the time was LISP. companies 00:13:53.610 --> 00:13:55.950 started building physical machines optimized 00:13:55.950 --> 00:13:59.169 specifically to run that one singular language. 00:13:59.269 --> 00:14:01.649 Wait, building entirely new hardware just for 00:14:01.649 --> 00:14:03.909 one language? Yes, we were talking about companies 00:14:03.909 --> 00:14:07.549 like Symbolics and LISP Machines, Inc. selling 00:14:07.549 --> 00:14:10.210 massive, highly specialized, incredibly expensive 00:14:10.210 --> 00:14:12.269 computer workstations. Okay, the best analogy 00:14:12.269 --> 00:14:14.629 I can think of for this is it sounds exactly 00:14:14.629 --> 00:14:17.669 like buying a massive $10 ,000 heavy -duty stove 00:14:17.669 --> 00:14:20.529 for your kitchen that has been biologically and 00:14:20.529 --> 00:14:22.629 mechanically engineered to only cook omelets. 00:14:22.710 --> 00:14:25.190 That is a perfect analogy. Like, it makes the 00:14:25.190 --> 00:14:27.110 absolute best omelet in the world incredibly 00:14:27.110 --> 00:14:29.210 fast. But what happens when you want to make 00:14:29.210 --> 00:14:32.480 pasta? Your expensive stove is entirely useless. 00:14:32.860 --> 00:14:35.259 And to build on your analogy, it wasn't just 00:14:35.259 --> 00:14:37.799 that it only knew the recipe for an omelet. The 00:14:37.799 --> 00:14:40.480 very pipes and heating elements of the stove 00:14:40.480 --> 00:14:43.460 were physically constructed in a way that couldn't 00:14:43.460 --> 00:14:46.519 boil water. Oh, man. The memory and processors 00:14:46.519 --> 00:14:49.659 of these LISP machines were hardwired specifically 00:14:49.659 --> 00:14:53.759 to read LISP code. And that extreme specialization 00:14:53.759 --> 00:14:56.519 is exactly what killed the market in 1987. What 00:14:56.519 --> 00:14:59.240 happened in 87? The general computing landscape 00:14:59.149 --> 00:15:01.450 shifted dramatically. General purpose desktop 00:15:01.450 --> 00:15:04.210 computers from companies like Apple, IBM, and 00:15:04.210 --> 00:15:06.669 Sun Microsystems became faster, cheaper, and 00:15:06.669 --> 00:15:09.370 powerful enough to run LISP environments purely 00:15:09.370 --> 00:15:12.330 through software on standard Unix operating systems. 00:15:12.669 --> 00:15:14.490 So suddenly there was absolutely no reason for 00:15:14.490 --> 00:15:17.509 a corporation to buy a massive specialized omelet 00:15:17.509 --> 00:15:20.669 maker when a standard cheap kitchen stove could 00:15:20.669 --> 00:15:23.429 do it all. Exactly. An entire specialized industry 00:15:23.429 --> 00:15:25.409 worth half a billion dollars was replaced by 00:15:25.409 --> 00:15:28.179 general purpose computers in a single year. Consumers 00:15:28.179 --> 00:15:30.940 stopped buying and the LISP hardware market completely 00:15:30.940 --> 00:15:33.240 evaporated. Half a billion dollars vanished in 00:15:33.240 --> 00:15:35.860 12 months. But it wasn't just the hardware that 00:15:35.860 --> 00:15:39.000 failed, right? The software itself, these expert 00:15:39.000 --> 00:15:40.980 systems that corporations were betting everything 00:15:40.980 --> 00:15:44.120 on, eventually hit a massive wall, too. Yes. 00:15:44.340 --> 00:15:46.440 The software suffered from a fatal flaw that 00:15:46.440 --> 00:15:49.440 researchers called brittleness. Brittleness? 00:15:49.799 --> 00:15:52.480 Yeah. These expert systems were essentially... 00:15:52.409 --> 00:15:55.769 giant rule -based engines. They couldn't learn 00:15:55.769 --> 00:15:58.870 on their own. Human experts had to manually type 00:15:58.870 --> 00:16:01.549 in every single line of logic. Okay, let's break 00:16:01.549 --> 00:16:04.629 that down with an example. How exactly does brittleness 00:16:04.629 --> 00:16:07.909 break a system? Imagine you build a medical diagnosis 00:16:07.909 --> 00:16:10.610 system. You write a rule that says, if the patient 00:16:10.610 --> 00:16:13.470 has a fever, give them aspirin. That works fine. 00:16:13.610 --> 00:16:15.690 Seems simple enough. But then you realize, what 00:16:15.690 --> 00:16:18.250 if the patient is allergic to aspirin? Okay, 00:16:18.269 --> 00:16:20.110 you have to write another rule. What if you're 00:16:20.110 --> 00:16:22.679 pregnant? Another rule. What if they have a rare 00:16:22.679 --> 00:16:25.580 blood condition? Another rule. Oh, I see where 00:16:25.580 --> 00:16:28.259 this is going. Eventually, your rule book is 00:16:28.259 --> 00:16:31.700 millions of pages long. If you give the system 00:16:31.700 --> 00:16:34.580 an unusual input that it hasn't explicitly been 00:16:34.580 --> 00:16:37.059 programmed for, it doesn't just fail gracefully 00:16:37.059 --> 00:16:39.539 or ask for clarification. It crashes. It makes 00:16:39.539 --> 00:16:42.960 grotesque, catastrophic mistakes. And worse, 00:16:43.360 --> 00:16:45.620 changing one rule might accidentally contradict 00:16:45.620 --> 00:16:49.679 10 other rules. As the knowledge base grew, managing 00:16:49.679 --> 00:16:52.580 the truth of the system became computationally 00:16:52.580 --> 00:16:55.120 and financially grueling. It's like building 00:16:55.120 --> 00:16:57.899 a house out of toothpicks. Eventually, adding 00:16:57.899 --> 00:17:00.059 one more piece brings the whole thing crashing 00:17:00.059 --> 00:17:02.419 down. And once the corporate world realized this, 00:17:02.659 --> 00:17:04.940 the dominoes just kept falling. They absolutely 00:17:04.940 --> 00:17:07.680 did. The US government had briefly started funding 00:17:07.680 --> 00:17:10.579 AI again in the 80s through the Strategic Computing 00:17:10.579 --> 00:17:13.759 Initiative, but a new director named Jack Schwartz 00:17:13.759 --> 00:17:16.440 came in and absolute gutted it. He was not a 00:17:16.440 --> 00:17:19.180 fan. Not at all. He looked at expert systems, 00:17:19.500 --> 00:17:21.920 dismissed them as just clever programming, and 00:17:21.920 --> 00:17:24.519 famously argued that DARPA should surf rather 00:17:24.519 --> 00:17:26.720 than dog paddle. Ouch. He meant they should ride 00:17:26.720 --> 00:17:29.019 the massive waves of promising technologies, 00:17:29.079 --> 00:17:31.640 and he definitively believed AI was a sinking 00:17:31.640 --> 00:17:34.940 ship. Meanwhile, over in Japan, their government 00:17:34.940 --> 00:17:38.599 had poured $850 million into a massive fifth 00:17:38.599 --> 00:17:41.160 -generation computer project aiming for machines 00:17:41.160 --> 00:17:43.339 that could naturally converse and reason. And 00:17:43.339 --> 00:17:47.039 how did that go? By 1992, that project ended. 00:17:47.309 --> 00:17:50.910 Not with a roar, but with an absolute whimper. 00:17:51.170 --> 00:17:54.109 Expectations had, once again, vastly outpaced 00:17:54.109 --> 00:17:57.190 reality. Which forces the entire industry underground. 00:17:57.369 --> 00:17:59.910 I mean, the crashes of the 1980s and 90s were 00:17:59.910 --> 00:18:02.869 so financially devastating, and the public embarrassment 00:18:02.869 --> 00:18:06.150 was so profound that researchers literally had 00:18:06.150 --> 00:18:08.490 to change the name of what they were doing just 00:18:08.490 --> 00:18:11.740 to survive. The Stealth Era. By the early 2000s, 00:18:11.839 --> 00:18:13.819 the term artificial intelligence carried such 00:18:13.819 --> 00:18:17.160 a deep toxic stigma of false promises that no 00:18:17.160 --> 00:18:19.460 venture capitalists would go near it. Not a chance. 00:18:19.799 --> 00:18:21.839 Computer scientists actively avoided the term 00:18:21.839 --> 00:18:23.960 for fear being viewed as wild -eyed dreamers 00:18:23.960 --> 00:18:26.819 or snake oil salesmen. So they completely rebranded. 00:18:26.819 --> 00:18:28.940 They called their work machine learning or advanced 00:18:28.940 --> 00:18:31.079 analytics or cognitive systems. Anything but 00:18:31.079 --> 00:18:33.500 AI. Anything to avoid the dreaded AI label so 00:18:33.500 --> 00:18:35.759 they could quietly secure funding for their labs. 00:18:35.940 --> 00:18:38.359 If we connect this to the bigger picture, this 00:18:38.359 --> 00:18:40.960 stealth period was actually incredibly productive. 00:18:41.359 --> 00:18:43.579 The philosopher Nick Bostrom made a brilliant 00:18:43.579 --> 00:18:45.380 observation about this era. What did he say? 00:18:45.559 --> 00:18:47.779 He pointed out that a lot of cutting edge AI 00:18:47.779 --> 00:18:50.319 was actually succeeding and filtering into everyday 00:18:50.319 --> 00:18:52.720 applications during this time. But the great 00:18:52.720 --> 00:18:56.339 irony is the moment an algorithm becomes genuinely 00:18:56.339 --> 00:18:59.640 useful and common, like the pathfinding logic 00:18:59.640 --> 00:19:02.680 in a video game or logistic software for shipping 00:19:02.680 --> 00:19:05.609 companies. we immediately stop calling it AI. 00:19:05.789 --> 00:19:08.549 We just call it software. Exactly. The goalposts 00:19:08.549 --> 00:19:10.650 constantly shift. Here's where it gets really 00:19:10.650 --> 00:19:13.509 interesting, because all of that quiet, stealthy, 00:19:13.670 --> 00:19:15.990 foundational work, hiding under different names, 00:19:16.430 --> 00:19:19.599 set the stage for the current AI spring. The 00:19:19.599 --> 00:19:22.259 deep freeze finally thaws and the world changes 00:19:22.259 --> 00:19:25.740 entirely around 2012. The turning point. That 00:19:25.740 --> 00:19:28.480 was the year a deep learning network called AlexNet 00:19:28.480 --> 00:19:30.920 entered a major visual recognition challenge 00:19:30.920 --> 00:19:33.619 and absolutely dominated the competition blowing 00:19:33.619 --> 00:19:35.619 traditional algorithms out of the water. And 00:19:35.619 --> 00:19:38.000 from there the modern era just explodes. The 00:19:38.000 --> 00:19:40.180 modern metrics are staggering by every possible 00:19:40.180 --> 00:19:42.430 measure. We went from that visual recognition 00:19:42.430 --> 00:19:45.009 breakthrough to systems like AlphaGo Mastering 00:19:45.009 --> 00:19:48.089 Complex Strategy games to massive leaps in Google 00:19:48.089 --> 00:19:51.250 Translate. It's everywhere now. By 2022, total 00:19:51.250 --> 00:19:55.150 global investment hit $50 billion. In that same 00:19:55.150 --> 00:19:57.869 year, there were 800 ,000 U .S. job openings 00:19:57.869 --> 00:20:00.450 related to the field. And, of course, the release 00:20:00.450 --> 00:20:04.369 of ChatGPT, which hit 100 million users in just 00:20:04.369 --> 00:20:07.460 a matter of months. We're in an era of absolutely 00:20:07.460 --> 00:20:09.779 unprecedented scale and public adoption. Okay, 00:20:09.880 --> 00:20:12.329 but here is my big pushback for you. We just 00:20:12.329 --> 00:20:14.490 spent this entire deep dive looking at the cyclical 00:20:14.490 --> 00:20:16.670 history. We saw the massive translation hype 00:20:16.670 --> 00:20:19.789 in the 60s crash and burn. We saw the expert 00:20:19.789 --> 00:20:22.390 systems of the 80s wipe out half a billion dollars 00:20:22.390 --> 00:20:24.730 in a single year. Right. With today's billions 00:20:24.730 --> 00:20:26.829 pouring in and everyone on earth talking about 00:20:26.829 --> 00:20:29.589 chat GPT, are we just inflating the absolute 00:20:29.589 --> 00:20:32.869 biggest bubble yet? Or have we actually solved 00:20:32.869 --> 00:20:35.210 Sir James Lightfield's combinatorial explosion? 00:20:35.410 --> 00:20:37.549 Have we fundamentally broken the cycle? This 00:20:37.549 --> 00:20:39.630 raises the most important question of our current 00:20:39.630 --> 00:20:42.319 era. And to answer it, we have look at how today's 00:20:42.319 --> 00:20:45.799 AI actually works. It turns out the engine powering 00:20:45.799 --> 00:20:48.200 today's AI spring is built on the very ideas 00:20:48.200 --> 00:20:50.539 that were mocked and discarded during the winters. 00:20:51.420 --> 00:20:55.039 For instance, back in 1969, Marvin Minsky published 00:20:55.039 --> 00:20:57.500 a critique of early artificial neural networks, 00:20:57.799 --> 00:21:00.759 known as single layer perceptrons. That critique 00:21:00.759 --> 00:21:03.099 was so devastating, it helped kill funding for 00:21:03.099 --> 00:21:05.759 neural networks for over a decade. So Minsky 00:21:05.759 --> 00:21:08.259 killed his own field's funding. Basically, yeah. 00:21:08.589 --> 00:21:11.150 He was mathematically right about the limitations 00:21:11.150 --> 00:21:14.069 of a single layer, but researchers knew that 00:21:14.069 --> 00:21:16.250 multi -layer networks could theoretically solve 00:21:16.250 --> 00:21:19.589 complex problems. The roadblock was that nobody 00:21:19.589 --> 00:21:23.009 in the 1960s or 70s knew how to train a multi 00:21:23.009 --> 00:21:25.170 -layer network. Oh, I see. They had the blueprint 00:21:25.170 --> 00:21:26.970 for the engine, but they didn't have the fuel 00:21:26.970 --> 00:21:29.869 or the spark plug. Exactly. It took decades for 00:21:29.869 --> 00:21:32.750 two crucial things to happen. First, computing 00:21:32.750 --> 00:21:35.680 power had to catch up. The raw processing power 00:21:35.680 --> 00:21:38.619 of modern warehouse -sized data centers has effectively 00:21:38.619 --> 00:21:40.759 bulldozed through the combinatorial explosion 00:21:40.759 --> 00:21:42.900 that Lighthill warned about. Just brute -forcing 00:21:42.900 --> 00:21:45.759 it. Exactly. We now have the brute force to calculate 00:21:45.759 --> 00:21:48.859 those trillions of possibilities. Second, researchers 00:21:48.859 --> 00:21:51.519 fully realized a training method called backpropagation. 00:21:52.099 --> 00:21:54.420 Okay, backpropagation. How does that actually 00:21:54.420 --> 00:21:57.380 work? Why was that the missing spark plug? Imagine 00:21:57.380 --> 00:22:00.619 a student taking a complex calculus test. In 00:22:00.619 --> 00:22:02.960 the old days, the system would just get an F 00:22:02.960 --> 00:22:05.160 and not know why. Which doesn't help you learn 00:22:05.160 --> 00:22:07.940 at all. Right. Backpropagation is a feedback 00:22:07.940 --> 00:22:11.220 loop. The network makes a guess, compares it 00:22:11.220 --> 00:22:13.420 to the correct answer, calculates exactly how 00:22:13.420 --> 00:22:15.720 wrong it was, and then mathematically sends that 00:22:15.720 --> 00:22:17.460 error signal backward through all its layers. 00:22:17.519 --> 00:22:20.779 Yeah, backwards. Backpropagation. Got it. It 00:22:20.779 --> 00:22:23.180 minutely adjusts the weight and importance of 00:22:23.180 --> 00:22:26.200 every single digital neuron so that next time... 00:22:26.200 --> 00:22:28.339 the guess is slightly more accurate. That makes 00:22:28.339 --> 00:22:30.299 sense. And when you do that millions of times 00:22:30.299 --> 00:22:33.099 a second across billions of data points using 00:22:33.099 --> 00:22:35.660 modern supercomputers, the system effectively 00:22:35.660 --> 00:22:39.240 learns. The deep learning networks powering today's 00:22:39.240 --> 00:22:42.200 massive language models are the direct evolutionary 00:22:42.200 --> 00:22:44.819 descendants of those early dismissed neural networks. 00:22:45.579 --> 00:22:47.539 The seeds that were planted and buried during 00:22:47.539 --> 00:22:50.170 the long winters have finally bloomed. So what 00:22:50.170 --> 00:22:52.269 does this all mean for you listening right now? 00:22:52.609 --> 00:22:55.049 Why does knowing the history of LISP machines 00:22:55.049 --> 00:22:58.509 and Cold War translation actually matter? It's 00:22:58.509 --> 00:23:01.210 vital context. I think understanding this cyclical 00:23:01.210 --> 00:23:03.799 history acts as a kind of superpower. When you 00:23:03.799 --> 00:23:05.819 open your phone today and see breathless headlines 00:23:05.819 --> 00:23:08.220 promising that a new AI is going to instantly 00:23:08.220 --> 00:23:10.960 revolutionize the world or replace all human 00:23:10.960 --> 00:23:14.059 jobs by next Tuesday, you can spot the patterns. 00:23:14.420 --> 00:23:17.240 The web of increasing exaggeration. Exactly. 00:23:17.420 --> 00:23:19.960 More of ex web. You can see the venture capital 00:23:19.960 --> 00:23:22.859 frenzies and you can anticipate the inevitable 00:23:22.859 --> 00:23:25.960 collision with real world complexity. You know 00:23:25.960 --> 00:23:28.880 that making a parlor trick work in a highly curated 00:23:28.880 --> 00:23:31.579 demo video is very, very different from deploying 00:23:31.579 --> 00:23:35.410 a flawless scalable tool in the messy, unpredictable 00:23:35.410 --> 00:23:38.710 real world. That is the essential takeaway. True 00:23:38.710 --> 00:23:41.630 technological progress is never magic. It is 00:23:41.630 --> 00:23:44.009 hard, iterative, and often grueling work. It 00:23:44.009 --> 00:23:46.430 takes decades. Navigating this current boom without 00:23:46.430 --> 00:23:48.630 getting burned requires a skeptical eye, critical 00:23:48.630 --> 00:23:50.390 thinking, and a willingness to look past the 00:23:50.390 --> 00:23:52.829 marketing hype to rigorously evaluate what a 00:23:52.829 --> 00:23:55.269 system actually understands versus what it's 00:23:55.269 --> 00:23:58.000 just mimicking. And speaking of what a system 00:23:58.000 --> 00:24:00.680 can genuinely understand, I want to leave you 00:24:00.680 --> 00:24:02.900 with a final provocative thought to mull over. 00:24:03.539 --> 00:24:05.859 Our sources included a fascinating point from 00:24:05.859 --> 00:24:08.700 a review by James Gleick regarding the fundamental 00:24:08.700 --> 00:24:10.900 nature of intelligence. That's a profound point. 00:24:11.180 --> 00:24:14.079 It really is. We know that modern AI, like these 00:24:14.079 --> 00:24:17.119 incredible large language models, can ace standardized 00:24:17.119 --> 00:24:19.960 medical exams and generate beautiful complex 00:24:19.960 --> 00:24:23.250 essays. But as Gleick points out, they remain 00:24:23.250 --> 00:24:25.769 fundamentally disembodied. No physical presence. 00:24:25.970 --> 00:24:28.390 Right. They are strangers to blood, sweat, and 00:24:28.390 --> 00:24:31.390 tears. For biological creatures like you and 00:24:31.390 --> 00:24:34.589 me, agency, reason, and purpose don't just come 00:24:34.589 --> 00:24:37.150 from reading text. They come from acting in the 00:24:37.150 --> 00:24:39.289 physical world and directly experiencing the 00:24:39.289 --> 00:24:41.190 physical consequences of our actions. The real 00:24:41.190 --> 00:24:44.210 stakes. A scraped knee, a broken heart, the feeling 00:24:44.210 --> 00:24:47.089 of absolute exhaustion. So the next time you 00:24:47.089 --> 00:24:49.109 are interacting with one of these incredibly 00:24:49.109 --> 00:24:53.000 smart chatbots, ask yourself this. Can an intelligence 00:24:53.000 --> 00:24:55.140 that has never once experienced the physical 00:24:55.140 --> 00:24:57.539 world ever truly understand it, or is it just 00:24:57.539 --> 00:24:59.740 the most advanced, brilliant translation trick 00:24:59.740 --> 00:25:02.640 humanity has ever built? Thank you so much for 00:25:02.640 --> 00:25:03.980 taking this deep dive with us.