WEBVTT - Why AI Has No ROI with Paul Kedrosky 0:00:03.160 --> 0:00:20.599 Media greetums. I'm at Zetron and this is Better Offline. Today. 0:00:20.640 --> 0:00:23.759 We are joined by the mighty economist Paul Keadrowski. Paul, 0:00:23.840 --> 0:00:24.919 thank you for joining me. 0:00:25.560 --> 0:00:26.599 Hey, how's it going. 0:00:27.200 --> 0:00:31.960 It's going gray. Everyone's deeply upset because this week and 0:00:32.040 --> 0:00:35.640 the last week everyone has been saying, huh does AI 0:00:35.960 --> 0:00:39.920 have a return on investment? And it's I've really been 0:00:40.000 --> 0:00:43.320 enjoying it because it's like watching the dinosaurs look up 0:00:43.360 --> 0:00:46.760 and see the meteor. They're just like, what do you 0:00:47.000 --> 0:00:49.600 what do you mean? What do you mean this costs money? 0:00:50.000 --> 0:00:50.239 Have you? 0:00:50.400 --> 0:00:52.720 I don't know if you've seen the GitHub co pilot stuff. 0:00:53.200 --> 0:00:55.160 Yeah, I actually put it a thing on it yesterday. 0:00:55.600 --> 0:00:58.920 Oh sorry, Paul, terribly rude to me. That was you. 0:00:59.160 --> 0:01:00.960 You know me, I've got all sorts of crap on, 0:01:01.120 --> 0:01:03.680 so I haven't read it yet. But I'm excited to 0:01:03.720 --> 0:01:07.600 talk about this. I'm really excited. It's really and not 0:01:07.640 --> 0:01:11.399 only that. I mean there's I'll sort of triangulate with 0:01:11.440 --> 0:01:14.000 three different things that touched on different aspects of this 0:01:14.040 --> 0:01:16.080 at the same time. One was, obviously they get up 0:01:16.720 --> 0:01:19.360 copiled study, which we can get into as deeply as 0:01:19.360 --> 0:01:21.399 you want. There was also a piece that came out 0:01:21.959 --> 0:01:25.360 in part from the Peterson Institute for International Economics. 0:01:26.319 --> 0:01:31.320 Yesterday the day before. Jack Cook at Clark at Anthropics 0:01:31.360 --> 0:01:32.920 set it around who is obviously one of the co 0:01:33.000 --> 0:01:35.039 founders there, and it's and it's called where is the 0:01:35.160 --> 0:01:38.640 where is AI? And GDP statistics. And then of course 0:01:38.640 --> 0:01:42.399 there was the tobacco which I saw anonymous someone that 0:01:42.400 --> 0:01:45.720 had anonymously disclosed that they had spent almost a half 0:01:45.959 --> 0:01:49.720 five hundred million dollars because they had uncapped token expenses, 0:01:50.440 --> 0:01:53.240 and we'll discovered they've sort of blown their credit cards. 0:01:53.240 --> 0:01:55.279 So anyways, yes, there's a bunch of flat as well. 0:01:55.720 --> 0:01:58.640 Well, let's stop with the GitHub things. So for the uninitiated, 0:01:58.920 --> 0:02:02.720 get up Copilot, a coding tool from Microsoft. A couple 0:02:02.720 --> 0:02:05.160 of weeks ago I broke the story, of course that 0:02:05.200 --> 0:02:07.920 they were moving their users from a premium request model 0:02:08.000 --> 0:02:10.280 to a token based model. So think of it like 0:02:10.360 --> 0:02:13.480 this with listeners. If you every time you use the 0:02:13.480 --> 0:02:15.799 cab service, you could just say drive me from the 0:02:15.880 --> 0:02:18.480 Upper West Side to Red Hook and that would just 0:02:18.800 --> 0:02:20.520 that would be one drive, and you get a certain 0:02:20.520 --> 0:02:23.600 amount of drives a month, and then suddenly the beginning 0:02:23.600 --> 0:02:25.040 of June they turned to you and say, yeah, you've 0:02:25.040 --> 0:02:27.000 got to pay by the mile, and you suddenly realize 0:02:27.000 --> 0:02:29.440 you've been taking ninety five mile trips. You've been asking 0:02:29.480 --> 0:02:31.639 to drive from New Jersey to Maryland, which I realized 0:02:31.720 --> 0:02:35.840 is further than ninety five miles. Not a geographer, all right, 0:02:36.160 --> 0:02:40.360 But nevertheless, on the GitHub co pilot subreddit, people have 0:02:40.520 --> 0:02:43.320 just been posting, what the fuck what do you mean? 0:02:43.360 --> 0:02:46.640 My whole balance is gone in three prompts? What do 0:02:46.680 --> 0:02:47.400 you mean by that? 0:02:47.760 --> 0:02:51.520 Uh huh, it's yeah, And this is part of the problem, 0:02:51.639 --> 0:02:55.079 right has this been? You can get all ekon of 0:02:55.160 --> 0:02:58.920 wonky about this stuff about the merits of metered pricing 0:02:58.919 --> 0:03:01.720 on a per token basis versus lump sun pricing, but 0:03:01.919 --> 0:03:03.960 in a sense you can think of this was the 0:03:04.080 --> 0:03:07.200 early pricing and token in terms of how tokens were 0:03:07.240 --> 0:03:11.560 metered out, had two really important characteristics. One is they 0:03:11.560 --> 0:03:14.880 were grotesquely subsidized. You weren't actually seeing the real all 0:03:14.919 --> 0:03:17.720 in cost with respect to the loaded cost of actually 0:03:17.880 --> 0:03:20.600 providing you with those tokens. And then as a kind 0:03:20.639 --> 0:03:23.680 of don't pay a cent event up there with cost 0:03:23.720 --> 0:03:26.560 co it was being bundled, so you we had a 0:03:26.600 --> 0:03:29.680 second layer of masking with respect to what these what 0:03:29.800 --> 0:03:34.120 tokens were actually costing. And so once it becomes unsubsidized 0:03:34.120 --> 0:03:37.720 and unbundled, then you see, you know, your ass is 0:03:37.800 --> 0:03:40.080 dangling in the breeze of real token pricing. 0:03:41.000 --> 0:03:43.680 I think it's funny as well, because for years people 0:03:43.760 --> 0:03:46.720 have been saying to be that's not happening. They're not 0:03:46.760 --> 0:03:51.120 subsidizing it. It's different. It's just it's like the the 0:03:51.120 --> 0:03:53.320 cost Go model, for example. People are like, oh, yeah, 0:03:53.320 --> 0:03:55.680 well they're making money other ways. It's like, no, they're not. 0:03:56.240 --> 0:03:59.000 They're just selling. In Microsoft's case, they were like, we're 0:03:59.040 --> 0:04:01.560 going to sell you one one thousand dollars for thirty 0:04:01.680 --> 0:04:05.320 nine dollars, right, what do you think? You think that's good? 0:04:05.480 --> 0:04:06.040 Do you like that? 0:04:06.720 --> 0:04:09.360 It's lovely, It's a lovely come on. It brings people in. 0:04:09.400 --> 0:04:11.680 It's like the hot dogs and Costco, except Costco has 0:04:11.720 --> 0:04:13.640 other things on which they make a b outload of money. 0:04:14.200 --> 0:04:18.120 Except the hot dogs cost like seven thousand dollars a packet. 0:04:18.520 --> 0:04:22.599 It's just I think it's quite deceitful personally. I think 0:04:22.640 --> 0:04:25.560 it's because these on one hand, we can make fun 0:04:25.600 --> 0:04:28.560 of these people. I will continue to do so. It's funny, 0:04:28.839 --> 0:04:31.719 but when you look at them, it is also quite 0:04:31.920 --> 0:04:36.320 depressing because they were intentionally misled. Like these people had 0:04:36.400 --> 0:04:39.800 no idea. It's not like these subsidize subscriptions, like, hey, 0:04:40.160 --> 0:04:42.560 if you use this many tokens while paying for them, 0:04:42.560 --> 0:04:44.960 it would cost this much. Until they made the change. 0:04:45.000 --> 0:04:47.280 Microsoft release the calculator. It allows you to see that, 0:04:47.560 --> 0:04:51.160 but only once they'd announced it. So you have millions, 0:04:51.200 --> 0:04:53.960 I would say, the vast majority of people that interact 0:04:53.960 --> 0:04:58.720 with AI who have no idea what it costs, literally none. 0:04:58.839 --> 0:05:01.400 Right and may and which which is made worse by 0:05:01.760 --> 0:05:06.640 some of the early over excitement, especially among large corporations 0:05:06.680 --> 0:05:12.200 that made the mistake of creating leader boards and right, 0:05:12.240 --> 0:05:15.039 So this is we got into this token maxing phenomenon. 0:05:15.120 --> 0:05:17.120 If so, if you're inside of which is obviously the 0:05:17.160 --> 0:05:19.200 idea that the more tokens you use, the better you 0:05:19.680 --> 0:05:21.480 do in your job review, because look at you, you're 0:05:21.520 --> 0:05:23.599 all AI. The problem, of course, is that this is 0:05:23.600 --> 0:05:28.800 a little bit like the Saudi's handing out humviies to 0:05:28.839 --> 0:05:32.080 everyone in America. Wow, this is awesome. I love having 0:05:32.120 --> 0:05:33.960 a hum vy. And then you have to fill it 0:05:34.080 --> 0:05:36.960 up and so for a little while it was like, 0:05:37.440 --> 0:05:42.839 we were subsidizing these grotesquely uh profligate users of tokens, 0:05:42.880 --> 0:05:44.880 just like profligate users of gasoline. And then all of 0:05:44.920 --> 0:05:46.280 a sudden, the bill comes to you and you say, 0:05:46.279 --> 0:05:49.000 wait a minute, this thing's a pig. It's a lot 0:05:49.000 --> 0:05:51.039 of gas. I don't want to drive it for growth 0:05:51.040 --> 0:05:53.160 through anywhere. And the exact same phenomena, it's true with 0:05:53.200 --> 0:05:56.480 respect to being again exposed to the having your ass 0:05:56.480 --> 0:05:58.440 hanging in the breeze of real token prices. 0:05:58.839 --> 0:06:01.320 Well, the other thing is as well, is I'll just 0:06:01.320 --> 0:06:05.240 put our newsletter about this. You look at how these 0:06:05.240 --> 0:06:07.960 people are freaking out, and you also realize they have 0:06:08.480 --> 0:06:11.480 no idea what AI cost. It's not just like, wow, 0:06:11.520 --> 0:06:13.920 this is a lot of money. It's they're not even 0:06:14.120 --> 0:06:17.400 thinking in terms of cost. It's not like they know, 0:06:18.000 --> 0:06:20.120 I don't know. They're refactoring something, they don't know how 0:06:20.200 --> 0:06:23.679 much that got. They don't know how much anything costs. 0:06:23.720 --> 0:06:27.320 So it's not like they can smoothly transition to token 0:06:27.360 --> 0:06:31.080 based billing because they don't they don't know, they have 0:06:31.120 --> 0:06:31.640 no idea. 0:06:31.800 --> 0:06:33.640 And this is this is the deep this is the 0:06:33.680 --> 0:06:36.359 deep structural problem because they were brought in through the 0:06:36.440 --> 0:06:40.680 side door of bundled pricing, and now that's that's becoming unbundled, 0:06:40.920 --> 0:06:43.880 and of course that also has a is reflected in 0:06:44.480 --> 0:06:46.400 what we're told. And I think the Wall Street Journal 0:06:46.520 --> 0:06:48.440 others have written about this, and it'd be interesting we 0:06:48.480 --> 0:06:51.720 see the final s wyans for some of the upcoming 0:06:51.800 --> 0:06:55.320 I p os that there is this attempt to try 0:06:55.360 --> 0:06:58.719 and even mask it in the financial filings where you 0:06:58.760 --> 0:07:01.080 get into this phenomenon what we used to call earnings 0:07:01.120 --> 0:07:03.760 before bad stuff. And so what they're trying to do 0:07:04.400 --> 0:07:07.880 is hide the costs of training the models and saying 0:07:07.920 --> 0:07:10.880 that's not actually an operating cost, that's a capital cost, 0:07:10.920 --> 0:07:13.960 and we shouldn't have to show that as a function 0:07:14.040 --> 0:07:16.960 of what actually the margins are on producing tokens. And 0:07:17.000 --> 0:07:19.640 that is of course a cheat, right because if that's true, 0:07:20.080 --> 0:07:22.120 then you should be able to capitalize these things and 0:07:22.160 --> 0:07:24.040 expense them over a long period of time. And we 0:07:24.120 --> 0:07:27.239 know full well that these are actually operating costs, because 0:07:27.280 --> 0:07:29.960 they tell us that every eighteen months we're launching a 0:07:30.040 --> 0:07:32.280 new major model. These things are not capitalized. These are 0:07:32.360 --> 0:07:35.840 operating expenses that should be treated accordingly with respect to 0:07:35.840 --> 0:07:38.920 the actual cost of token production. So there's a multifaceted 0:07:40.160 --> 0:07:42.360 game going on here, both in terms of how it's 0:07:42.400 --> 0:07:44.440 being presented to users, but also in terms of how 0:07:44.480 --> 0:07:46.240 they're trying to sell it in the context of the 0:07:46.320 --> 0:07:49.720 upcoming s ones for the anthropic and open a IPO pilots. 0:07:50.480 --> 0:07:52.600 Well, what's really funny as well about the idea of 0:07:52.800 --> 0:07:56.720 capitalizing training costs is they're never going away because it's 0:07:56.760 --> 0:07:59.840 not just free training shoving the stuff in the models. 0:08:00.120 --> 0:08:02.960 They have to constantly tweak them because that's. 0:08:02.920 --> 0:08:05.640 Right and so it's just right, which from a class 0:08:05.680 --> 0:08:08.640 you know, from a classic my years ago accounting, whenever 0:08:08.680 --> 0:08:10.920 you have a regular and predictable cost that you have 0:08:11.000 --> 0:08:14.120 to expense, you have to incur to continue operating your business, 0:08:14.360 --> 0:08:16.800 that is no longer a capital cost, that's an expensable 0:08:16.840 --> 0:08:19.160 item that should be expressed. Such so you get into, 0:08:19.800 --> 0:08:21.600 as I said, like back in the dark days of 0:08:21.640 --> 0:08:24.040 dot com and even the telecom boom, you get into 0:08:24.040 --> 0:08:26.200 this problem of earnings before bad stuff, where they want 0:08:26.200 --> 0:08:28.000 to they want to exclude all of the things that 0:08:28.040 --> 0:08:30.480 make the numbers look bad. And then of course on 0:08:30.520 --> 0:08:33.200 the other side, you have this run rate problem where 0:08:33.800 --> 0:08:35.960 we continually hear about what the run rates are at 0:08:35.960 --> 0:08:39.160 these companies, and the window with respect to the run 0:08:39.240 --> 0:08:41.880 rate could be the last fifteen minutes for all you know, right, 0:08:41.920 --> 0:08:44.800 a run rate is just you extrapolate whatever is most 0:08:44.800 --> 0:08:46.840 convenient for you. So it's a problem on both sides. 0:08:47.520 --> 0:08:50.080 Well yeah, actually that's that's what I love talking about 0:08:50.160 --> 0:08:52.080 run right. Everyone who listens to the show knows I'm 0:08:52.120 --> 0:08:56.199 a real run right pig because because like I've reached 0:08:56.240 --> 0:08:59.160 out to both open ai and Anthropic and said, hey, 0:08:59.480 --> 0:09:02.160 how do you find this number? And they will not respond. 0:09:02.240 --> 0:09:04.800 They will not They're very unfair to me, very nasty. 0:09:05.360 --> 0:09:10.440 They will not respond, probably because from what the information 0:09:10.520 --> 0:09:12.680 has reported. I don't know how open ai does it. 0:09:12.720 --> 0:09:16.480 But Anthropic not only includes the amounts of money that 0:09:16.559 --> 0:09:20.079 Amazon and Google make in their revenues that's like when 0:09:20.120 --> 0:09:24.520 they resell them a but they also they do thirteen 0:09:24.720 --> 0:09:29.520 times the last month's APIs spend and twelve times the 0:09:29.880 --> 0:09:34.920 current days subscribers. So it's just there's so many ways. Also, 0:09:35.760 --> 0:09:39.600 so token spend, so just organizational token spend. That's not 0:09:39.720 --> 0:09:45.120 a recurring costs, that's just you can kind of I 0:09:45.200 --> 0:09:48.560 guess think, well, maybe people are spending this today, but 0:09:48.920 --> 0:09:51.600 that person who spent half a billion dollars. That company 0:09:51.640 --> 0:09:55.000 spent half a billion dollars on AI. Right, that's not 0:09:55.120 --> 0:09:59.560 happening again, that person is that person is You're not 0:09:59.600 --> 0:10:02.240 going to get one half a billion dollar miss the 0:10:02.320 --> 0:10:07.120 bean every single month? Is someone just goofully? I also, 0:10:07.360 --> 0:10:11.000 genuinely I know the report of Madison Mills. She's respective report, 0:10:11.080 --> 0:10:15.280 she's very she's she's good, she is well sourced. It's 0:10:15.360 --> 0:10:20.280 just like I hope Anthropic didn't include that five hundred 0:10:21.080 --> 0:10:23.360 million in their annualized Yeah. 0:10:24.400 --> 0:10:26.560 I'm looking forward to its showing up in a public 0:10:26.559 --> 0:10:28.840 company filing because it almost inevitably will this is going 0:10:28.920 --> 0:10:30.920 to be somebody can time item right. 0:10:31.320 --> 0:10:33.600 And and then you think that that will though. 0:10:33.800 --> 0:10:37.679 Oh absolutely, I mean half a billion, it's material for 0:10:37.800 --> 0:10:40.160 almost anyone. So I my guess is it's going to 0:10:40.200 --> 0:10:42.160 show up somewhere. It'll be really interesting to see. And 0:10:42.200 --> 0:10:44.720 my guess is at that scale, it's a it's a 0:10:44.720 --> 0:10:47.080 public company. So my guess is we will see that, 0:10:47.120 --> 0:10:49.240 we will know where that actually happened, and so it's 0:10:49.280 --> 0:10:51.200 gonna be it's gonna be very entertaining. But that this 0:10:51.280 --> 0:10:54.360 is the deep structural problem, and it gets worse of course, 0:10:54.440 --> 0:10:58.400 because once you unbundled token pricing, and then you're looking 0:10:58.480 --> 0:11:03.440 at the actual year over year decline in quality adjusted 0:11:03.480 --> 0:11:06.600 token pricing, in token pricing, and you see that the 0:11:06.720 --> 0:11:11.760 inherent deflationary curve underneath the hood. Now, let's connect that 0:11:12.160 --> 0:11:14.520 to how all of these data centers that are producing 0:11:14.559 --> 0:11:19.160 these deflating tokens are being being constructed and increasing fractions. 0:11:18.960 --> 0:11:23.120 Can elaborate what you mean by the deflating token, I'm 0:11:23.120 --> 0:11:23.880 not sure I understand. 0:11:24.480 --> 0:11:28.160 So over the last since twenty twenty two, on an 0:11:28.200 --> 0:11:34.520 annualized basis, on a performance basis, so ignoring what this 0:11:34.600 --> 0:11:37.360 continual jump to the frontier, if you imagine sort of 0:11:37.679 --> 0:11:43.000 on a comparable token basis, per cross models, across the period, 0:11:43.200 --> 0:11:46.280 token prices have fallen anywhere from seventy to ninety percent 0:11:46.400 --> 0:11:48.800 year over year consistently back to twenty twenty one. 0:11:48.920 --> 0:11:51.440 Right, but they're burning more tokens in the process. 0:11:52.000 --> 0:11:55.640 So but let's put that aside for one second. Think 0:11:55.679 --> 0:11:57.720 about it the alternative the other way around. So if 0:11:57.800 --> 0:12:01.199 I'm now, my business is now I'm unbundling and I'm 0:12:01.240 --> 0:12:04.600 selling tokens, and that's the way customers. You're telling my 0:12:04.920 --> 0:12:07.800 customers to think about it. So now they're looking out 0:12:07.920 --> 0:12:10.640 and they start to see what happening with token prices, 0:12:10.679 --> 0:12:13.400 and if I go back one generation, maybe those prices 0:12:13.440 --> 0:12:16.840 are cheaper. Now we have this classic financial problem of 0:12:16.840 --> 0:12:21.160 what's called a duration mismatch. Right, So I have debt 0:12:21.400 --> 0:12:24.520 funding the data centers that's ten to fifteen years duration 0:12:24.640 --> 0:12:28.040 and longer, which is predicated on fixed payments but being 0:12:28.080 --> 0:12:31.720 made on the basis of tokens where you're telling the 0:12:31.760 --> 0:12:34.600 customer to control your costs. You may want to look 0:12:34.640 --> 0:12:37.600 back in time and use an older model. So I'm 0:12:37.640 --> 0:12:41.240 paying for a fixed cost with a deflating commodity right 0:12:41.320 --> 0:12:45.760 this WEEKNTE from the over and over and over again. 0:12:46.080 --> 0:12:50.880 These duration mismatches, especially duration mismatches that are built on 0:12:51.000 --> 0:12:55.800 top of debt and a deflating commodity, are absolutely atomic 0:12:56.000 --> 0:12:59.280 with respect to causing a blow up in people's obligations 0:12:59.320 --> 0:13:02.679 with respect to these these kinds of duration mismatch problems. 0:13:02.679 --> 0:13:04.640 So there's a there's a deep structural issue that this 0:13:04.679 --> 0:13:07.120 will expose, and people haven't quite realized it yet. 0:13:07.720 --> 0:13:11.079 So you're saying that as the token costs get cheaper 0:13:11.160 --> 0:13:15.800 and everyone's being encouraged to use this less or more thoughtfully, 0:13:16.520 --> 0:13:19.079 that's happening. But they're building the data centers as if 0:13:19.200 --> 0:13:22.199 number will only ever go up and they'll only ever use. 0:13:21.920 --> 0:13:24.240 That's exactly right, and so you've got this wonderful ye 0:13:24.440 --> 0:13:26.760 again the term of art. You got a duration mismatch 0:13:26.800 --> 0:13:30.160 on top of a deflating commodity that can only end very, 0:13:30.280 --> 0:13:33.200 very badly, and it was masked because for a while 0:13:33.600 --> 0:13:35.920 you weren't exposed directly to that. You were just paying 0:13:35.920 --> 0:13:38.880 a straight up subscription, almost like Amazon Prime. And of 0:13:38.920 --> 0:13:42.680 course that doesn't work because Amazon primes costs across the 0:13:42.720 --> 0:13:47.199 border declining, whereas costs are increasing at the frontier declining 0:13:47.320 --> 0:13:49.320 in the in the in the back catalog, if you 0:13:49.360 --> 0:13:50.079 will have tokens. 0:13:50.840 --> 0:13:53.240 And that's the thing with like an Amazon Prime for example. Yes, 0:13:53.240 --> 0:13:57.199 they have found like Amazon or not like Amazon people. 0:13:57.240 --> 0:14:00.319 I know many listeners don't love them. Then I agree, 0:14:00.440 --> 0:14:03.400 But it's like Amazon Prime. They fixed those costs by 0:14:03.400 --> 0:14:06.160 building their own logistics network and they found ways to 0:14:06.400 --> 0:14:09.600 they they had I don't know, ways to make that cheaper. 0:14:10.040 --> 0:14:12.640 No one has that in AI, No one like it's 0:14:12.720 --> 0:14:15.240 just we have three four years in and that one's like, oh, 0:14:15.280 --> 0:14:18.120 we'll do a six. No, we won't. That didn't work. 0:14:18.160 --> 0:14:21.040 Like we have. We're like two or three generations of 0:14:21.160 --> 0:14:27.040 tranium inferentia TPUs. Still not profitable, still not profitable, right, 0:14:27.200 --> 0:14:28.440 we would know. 0:14:29.800 --> 0:14:32.320 Since we would We wouldn't know. But I think and 0:14:32.320 --> 0:14:34.200 and and of course the problem is that if you 0:14:34.240 --> 0:14:36.280 look at I was just looking at some data yesterday 0:14:36.720 --> 0:14:40.320 with respect to how small language models are increasingly closing 0:14:40.320 --> 0:14:43.560 the gap with large language models, which is causing training 0:14:43.600 --> 0:14:46.440 cycles on large language models to have to accelerate, become 0:14:46.480 --> 0:14:50.200 more expensive, throw more compute, added more reinforcement learning. The 0:14:50.280 --> 0:14:54.040 costs are not are particularly not not not declining, they're 0:14:54.040 --> 0:14:58.280 actually increasing sharply at the frontier because they're essentially being 0:14:58.400 --> 0:15:02.200 chased like you know, like the rabbits and like wiley 0:15:02.280 --> 0:15:05.720 coyote and the road runner, and so they're being chased 0:15:05.720 --> 0:15:08.520 into this very costly corner as a result. And that's 0:15:08.520 --> 0:15:10.960 a classic commoditization problem. If you go back to the 0:15:11.040 --> 0:15:14.440 nineteenth I don't know, late nineteenth century, a very similar 0:15:14.480 --> 0:15:17.440 thing happened in railroads as people were racing desperately to 0:15:17.480 --> 0:15:20.680 try and find a way to build a corner and 0:15:20.720 --> 0:15:22.920 control themselves so that they could compete with all of 0:15:22.960 --> 0:15:25.320 these other upstart railroads, and of course all that really 0:15:25.320 --> 0:15:29.840 happened was Capex exploded, margins went to shit, and multiple 0:15:29.920 --> 0:15:31.840 railroads failed, and we led to the crash of what 0:15:31.960 --> 0:15:36.400 eighteen seventy three, eighteen ninety three and arguably was a 0:15:36.440 --> 0:15:38.120 cause in the Great Depression. So you know, so you're 0:15:38.160 --> 0:15:41.320 playing out this exact same game because you're sitting in 0:15:41.320 --> 0:15:45.040 this high Capex world that's increasingly funded by debt and 0:15:45.080 --> 0:15:47.920 built on top of this duration mismatch, with token prices 0:15:47.960 --> 0:15:50.440 being now exposed and raw in front of people. 0:16:01.360 --> 0:16:03.920 And the other thing is as well, is people are 0:16:03.960 --> 0:16:07.320 just literally in my piece today, people make this point about, oh, well, 0:16:07.400 --> 0:16:10.120 it's like the dot com bubble in the we will 0:16:10.360 --> 0:16:14.920 we will simply just we'll reuse these things in the future, 0:16:15.440 --> 0:16:18.160 like we'll just pick these up. And it's really I 0:16:18.840 --> 0:16:21.160 hear this from very smart people, are people who are 0:16:21.200 --> 0:16:25.480 not like beguiled by AI industry. But it's like, okay, 0:16:25.600 --> 0:16:27.440 let's talk about what happened to the dot com bubble. 0:16:27.520 --> 0:16:30.080 So when it exploded, you had those some microsystems, the 0:16:30.160 --> 0:16:34.040 Ultra whatever it was, I forget forty three fifty grand 0:16:34.040 --> 0:16:37.520 a server, but that one server could run an entire company. 0:16:37.520 --> 0:16:42.040 You could run everything on it, databases, messy CRMs they 0:16:42.040 --> 0:16:44.600 had like did they have on prem lotus notes? Anyway, 0:16:46.160 --> 0:16:47.760 you had those things. But you could run that, and 0:16:47.960 --> 0:16:49.920 you could probably run that in a garage. You might 0:16:49.960 --> 0:16:52.280 need to use the washing machines plug, but you could 0:16:52.280 --> 0:16:55.520 do it if you and that. Yeah, those things were 0:16:55.560 --> 0:16:57.560 fifty grand, so you probably get them at what twenty 0:16:57.640 --> 0:17:01.400 thirty large? Yeah, okay, great, what happens from the AI 0:17:01.440 --> 0:17:04.800 bubble burst? You can't just plug in an AI GPU 0:17:04.920 --> 0:17:08.120 a B two hundred GPUs but fifty grand. It requires 0:17:08.520 --> 0:17:11.320 about I think side look this up for a recently. 0:17:11.320 --> 0:17:13.440 It's like twelve hundred and fifteen hundred watts for a 0:17:13.520 --> 0:17:17.520 Sun Microsystem server, about the same for single B two hundred, 0:17:17.640 --> 0:17:22.600 which will require bespoke cooling, a server, hardware, RAM, all 0:17:22.600 --> 0:17:24.919 of this other stuff. And then you'll find out that 0:17:25.000 --> 0:17:27.159 you can't do jackshit with the single GPU. 0:17:28.600 --> 0:17:31.280 Yeah, you're gonna be a huge source of disappointment once 0:17:31.320 --> 0:17:33.600 you power it up. Your neighborhood lights all blink out 0:17:33.600 --> 0:17:36.159 and you can't still can't do anything. So yeah, I 0:17:36.160 --> 0:17:39.400 think that's exactly right. But that's but that's I call 0:17:39.480 --> 0:17:41.560 this I was just I got into this with someone 0:17:41.600 --> 0:17:43.960 recently who was making a similar I call it faith 0:17:43.960 --> 0:17:48.400 based argumentation. It's this kind of you know, quasi religious 0:17:48.840 --> 0:17:52.040 orthodoxy that requires you to believe the following five things. 0:17:52.080 --> 0:17:54.320 You know, they always create more jobs than they destroy. 0:17:54.400 --> 0:17:56.520 We can always reuse assets after the fact. And I'm 0:17:56.560 --> 0:17:58.320 one of the things I always point to people is 0:17:58.359 --> 0:18:02.199 that almost half of US railroad lines built during the 0:18:02.200 --> 0:18:05.679 Boon years in the late nineteenth century were eventually abandoned. 0:18:05.880 --> 0:18:08.800 And did they find reuse. They absolutely did. It only 0:18:08.840 --> 0:18:11.159 took one hundred years, and now they're mountain biking trails. 0:18:12.560 --> 0:18:14.520 You know, let's let's wait around for that. 0:18:17.160 --> 0:18:22.760 Railroads. They were railroads that didn't require electrification, like. 0:18:22.760 --> 0:18:26.159 That's right, actually, so they were much more stable as assets, right, 0:18:26.200 --> 0:18:29.399 They didn't have the problems that GPUs and data centers do. 0:18:29.520 --> 0:18:33.199 Were not just the huge power and cooling requirements, but 0:18:33.240 --> 0:18:36.360 also they hearent the trajectory of the underlying technology where 0:18:36.400 --> 0:18:39.200 it changes quickly enough that you know, is a twenty 0:18:39.280 --> 0:18:43.040 year old you know, blackwell of any use to anyone 0:18:43.080 --> 0:18:44.960 other than as a paper weight. Of course, the answer 0:18:45.040 --> 0:18:48.080 is probably not. Whereas a railroad pretty heavy, they are 0:18:48.200 --> 0:18:50.639 extremely heavy. I actually was messing around with one recently, 0:18:50.840 --> 0:18:53.239 and so, yeah, and so this is the problem. And 0:18:53.280 --> 0:18:56.359 again it's this sort of naive argumentation, not to mention that, 0:18:56.520 --> 0:18:58.760 you know, the old Canes line that it may be, 0:18:59.000 --> 0:19:00.520 it may be great in the line run, but in 0:19:00.560 --> 0:19:03.160 the long run we're also all dad. So it really 0:19:03.160 --> 0:19:06.760 depends on your time horizon. And I find it honest, honestly, 0:19:06.840 --> 0:19:09.879 in the face of the kinds of consequential changes in 0:19:09.920 --> 0:19:12.639 the US economy, I had a very glib style of argumentation, 0:19:12.680 --> 0:19:15.600 where you're essentially patting people on the head and saying, 0:19:15.680 --> 0:19:17.440 you know, don't worry your pretty little head. This will 0:19:17.440 --> 0:19:20.240 all work out because it always has. And they're arguing 0:19:20.280 --> 0:19:22.960 from a data set sample size of five, which we 0:19:23.000 --> 0:19:25.600 wouldn't launch, we wouldn't lunch a drug on that basis. 0:19:26.280 --> 0:19:30.639 Also, I think it helps them rationalize bad behavior because 0:19:31.280 --> 0:19:36.040 if you say, okay, it worked, the one that actually 0:19:36.119 --> 0:19:38.719 upsets me is, well, the dot com bubble worked out. 0:19:38.760 --> 0:19:40.919 It's like, yeah, the stock market lost eighty percent of 0:19:40.920 --> 0:19:44.400 its value, hundreds of pauses, of people lost their jobs, 0:19:44.520 --> 0:19:48.200 people lost everything in some cases, and at the end, 0:19:48.240 --> 0:19:51.640 it's like, okay, that was also completely different but you're 0:19:51.680 --> 0:19:55.320 being quite glib about the first part. But it's also yeah, 0:19:55.359 --> 0:19:57.880 it's okay that people burn a lot of money for 0:19:58.320 --> 0:20:02.040 basically no reason. This is also allows you to not 0:20:02.080 --> 0:20:04.600 think about bad stuff. It allows you to This is. 0:20:04.520 --> 0:20:07.840 The andrey Sonian argument that there's no point in introspection. 0:20:07.920 --> 0:20:11.040 It's just a really bad idea, right. I think about 0:20:11.080 --> 0:20:13.000 these things that the law sort and stuff out. But 0:20:13.160 --> 0:20:15.320 I also think there's a deeper issue. And we may 0:20:15.320 --> 0:20:17.720 have talked about this before, but the idea a lot 0:20:17.720 --> 0:20:20.560 of people treat as an article of faith. Carlota Perez's 0:20:20.600 --> 0:20:23.480 book Technological Revolutions and Financial Capital, and one of the 0:20:23.520 --> 0:20:25.640 things that they quote take away from that, which I'm 0:20:25.680 --> 0:20:27.159 not convinced they do. I think they only look at 0:20:27.200 --> 0:20:29.160 the pictures. But anyways, one of the things they take 0:20:29.160 --> 0:20:31.200 away from her book and her work and other people's 0:20:31.200 --> 0:20:37.040 work with respect to these violent technological revolutions is the 0:20:37.080 --> 0:20:39.560 idea that it really doesn't matter because it always works out. 0:20:39.600 --> 0:20:43.160 Here's the difference, though in past episodes, we didn't tell 0:20:43.200 --> 0:20:46.120 ourselves that. So there is an element of reflexivity going 0:20:46.119 --> 0:20:49.080 on here, because once you know the plot and you 0:20:49.119 --> 0:20:51.560 act as if the plot is somehow f equals MA, 0:20:51.920 --> 0:20:54.960 it's a law of physics. Then the whole game changes 0:20:55.119 --> 0:20:57.520 because now you're acting as if it doesn't matter what 0:20:57.600 --> 0:20:59.679 I do, because you think it doesn't matter what you do, 0:20:59.720 --> 0:21:01.760 because as you've got this idea in your head as 0:21:01.800 --> 0:21:04.520 an article of faith that it always works out. That 0:21:04.680 --> 0:21:07.800 wasn't true. Historically, no one in building out the railroads, 0:21:08.000 --> 0:21:11.680 rural electrification, the fiber bubble, no one was telling themselves 0:21:11.680 --> 0:21:14.240 in the time, this always works out. That was not 0:21:14.440 --> 0:21:16.679 part of the that was not part of the playbook. 0:21:16.760 --> 0:21:19.280