WEBVTT - SoftBank and OpenAI's Dangerous Game 0:00:03.160 --> 0:00:17.960 Media greetums. I'm at xytron and this is better offline. 0:00:20.160 --> 0:00:22.280 In my last episode, I started by telling you about 0:00:22.280 --> 0:00:24.319 a report from the analyst wing of a major bank, 0:00:24.600 --> 0:00:27.880 TD Cohen that revealed how Microsoft had drastically pulled back 0:00:27.880 --> 0:00:30.480 on its plans to build new data center capacity. At 0:00:30.480 --> 0:00:33.280 a bare minimum, Microsoft had canceled the equivalent of every 0:00:33.360 --> 0:00:36.800 data center in London or Tokyo at least, that the 0:00:36.840 --> 0:00:39.559 real figure is actually probably much much higher, and as 0:00:39.600 --> 0:00:42.120 I pointed out, that this is a real pale horse, 0:00:42.159 --> 0:00:45.080 a harbinger of bad times through generative AI and open 0:00:45.080 --> 0:00:48.200 AI especially, and an indicator that the bubble is imploding 0:00:48.320 --> 0:00:51.360 or popping or I really don't want to say it's popped, 0:00:51.400 --> 0:00:54.360 and I actually don't know if it's necessary to say 0:00:54.360 --> 0:00:56.600 when it does, it's never going to be one thing. 0:00:57.480 --> 0:00:59.720 I also mentioned that some may interpret this movie as 0:00:59.720 --> 0:01:02.600 a sponds to open AI's Stargate project, which aims to 0:01:02.640 --> 0:01:04.920 build hundreds of billions of dollars of data centers and 0:01:04.920 --> 0:01:08.119 power generation facilities in the US, all to power generator 0:01:08.160 --> 0:01:11.440 FAI apps and tools that I'm not sure actually there's 0:01:11.480 --> 0:01:14.840 really the demand for. But this begs the question how 0:01:14.840 --> 0:01:17.480 feasible is the Stargate project. Let's start with a bit 0:01:17.520 --> 0:01:20.600 of background. The Stargate project was officially announced that the 0:01:20.640 --> 0:01:23.000 very beginning of the Trump presidency. Is open AI tried 0:01:23.000 --> 0:01:25.800 to court favor from the notoriously transactional and praise hungry 0:01:25.920 --> 0:01:28.959 person of the United States of America, despite that the 0:01:28.959 --> 0:01:30.600 project has been in the works for some time and 0:01:30.640 --> 0:01:34.679 open AI had previously courted the Biden administration. Sam Allman's 0:01:34.720 --> 0:01:37.240 previous pitch to that administration late last year was that 0:01:37.280 --> 0:01:39.720 it was necessary to build a five gigaw what data center. 0:01:40.280 --> 0:01:42.440 We don't know how big Stargate will be, just that 0:01:42.480 --> 0:01:44.959 it will initially involve spending one hundred billion dollars due 0:01:45.040 --> 0:01:48.200 and I quote developed data centers for artificial intelligence in 0:01:48.200 --> 0:01:51.040 the US according to the information, with the project potentially 0:01:51.080 --> 0:01:54.480 scaling to five hundred billion dollars, a truly fucking, astoundingly 0:01:54.520 --> 0:01:58.520 stupid number. Stargate's first and only data center deal currently 0:01:58.600 --> 0:02:01.120 signed is in Abilene, Texas, and it's expected to be 0:02:01.160 --> 0:02:04.520 operational in mid twenty twenty six. Though these centers usually 0:02:04.520 --> 0:02:08.280 become operational in phases, this is especially likely to be 0:02:08.320 --> 0:02:11.320 the case here considering that, according to the information. Open 0:02:11.360 --> 0:02:14.360 Ai plans to have access to a gigawat of power 0:02:14.360 --> 0:02:17.200 on hundreds of thousands of GPUs. As part of this, 0:02:17.280 --> 0:02:19.640 the Stargate project will construct a three hundred and sixty 0:02:19.680 --> 0:02:22.440 point five megawatt natural gas power station. And as I 0:02:22.480 --> 0:02:25.720 said last episode, that one's about power. This power station is, 0:02:25.760 --> 0:02:28.000 as far as I can tell, still in the permitting phase. 0:02:28.240 --> 0:02:31.200 It'll be some time before Stargate breaks ground on the facility, 0:02:31.560 --> 0:02:35.720 let alone starts to actually generate power. Now, as for funding, 0:02:36.160 --> 0:02:38.520 things have got a little weird. Both open ai and 0:02:38.560 --> 0:02:41.000 soft Bank have committed to putting in either eighteen billion 0:02:41.120 --> 0:02:43.919 or nineteen million dollars each into the project. I've seen 0:02:43.960 --> 0:02:47.040 both numbers reported, by the way. Regardless, it's not that 0:02:47.120 --> 0:02:51.520 big a difference to worry about, especially with well the 0:02:51.560 --> 0:02:53.560 fact that open ai does not really have the money 0:02:53.600 --> 0:02:56.360 either way. What's a billion dollars when you don't have anything. 0:02:56.720 --> 0:02:59.160 The company is currently trying to raise forty billion dollars 0:02:59.160 --> 0:03:01.639 at a two hundred sixty billion dollar valuation, with the 0:03:01.720 --> 0:03:04.760 quote CNBC part of the funding expected to be used 0:03:04.760 --> 0:03:08.480 for open AI's commitment to Stargate. Now, I'm old enough 0:03:08.480 --> 0:03:10.919 to remember when this round was previously rumored to be 0:03:11.040 --> 0:03:13.400 valuing open Ai at three hundred and forty billion dollars 0:03:13.480 --> 0:03:16.280 and also at three hundred billion dollars, and soft Bank 0:03:16.280 --> 0:03:18.600 appears to be taking full responsibility for raising the round, 0:03:18.639 --> 0:03:21.880 including syndicating as much as ten billion dollars of the amount, 0:03:22.120 --> 0:03:25.440 which means that it would include a group of other investors. Nevertheless, 0:03:25.480 --> 0:03:28.040 it certainly seems SoftBank will provide the majority of the 0:03:28.040 --> 0:03:31.200 capital thirty billion dollars, with CNBC reporting that it will 0:03:31.240 --> 0:03:33.480 be paid out over the next twelve to twenty four months, 0:03:33.560 --> 0:03:36.440 with the first payment coming as soon as spring. Soft 0:03:36.480 --> 0:03:39.520 Bank also has another problem, and this one's this one. 0:03:39.560 --> 0:03:41.880 I think even the least technical of you can understand. 0:03:41.920 --> 0:03:45.280 They also appear to not have the money. They don't 0:03:45.280 --> 0:03:47.560 have the money. They not the money to give open Ai, 0:03:47.760 --> 0:03:50.160 not the money for Stargate. It's a little bit of 0:03:50.200 --> 0:03:54.000 worrying math issues with this. According to the Information, SoftBank 0:03:54.080 --> 0:03:56.960 CEO Masayoshi's Sun is planning to borrow sixteen billion dollars 0:03:57.000 --> 0:03:59.400 to invest in Ai, and may borrow another eight billion 0:03:59.400 --> 0:04:02.120 dollars next year. The following points, by the way, are 0:04:02.160 --> 0:04:04.560 drawn from the Information's reporting, and I give serious props 0:04:04.560 --> 0:04:07.400 to Duro Osawa and Corey Weinberg for their hard work here. 0:04:07.960 --> 0:04:10.160 I attack the information sometimes with some of the framing, 0:04:10.200 --> 0:04:12.160 but they actually do some of the best reporting in 0:04:12.200 --> 0:04:15.880 tech journalism. Now let's do some maths. Soft Bank currently 0:04:15.880 --> 0:04:17.760 only has thirty one billion dollars in cash and its 0:04:17.800 --> 0:04:20.480 balance sheet as of December twenty twenty four. It's net debt, 0:04:20.520 --> 0:04:22.680 which despite what you think, does not measure total debt, 0:04:22.680 --> 0:04:25.919 but rather represents its cash minus any debt liabilities, sounds 0:04:25.920 --> 0:04:27.960 at twenty nine billion dollars. They plan to use the 0:04:28.000 --> 0:04:30.760 loaning question to finance part of their investment in open Ai, 0:04:31.040 --> 0:04:34.480 and there are acquisition of chip design firm Ampere. According 0:04:34.480 --> 0:04:36.800 to SoftBank's reported assets, its holdings are worth about two 0:04:36.880 --> 0:04:39.559 hundred and nineteen billion dollars, so thirty three point sixty 0:04:39.839 --> 0:04:42.719 six trillion yen for those of you who deal with yen, 0:04:43.040 --> 0:04:46.680 including stock in companies like Alibaba and ARM. On a 0:04:46.760 --> 0:04:49.480 side note, not every company in soft Bank's portfolio is 0:04:49.520 --> 0:04:52.159 an arm or an Alibaba or like a good one. 0:04:52.320 --> 0:04:54.599 There are some real stinkers too, and I'm not just 0:04:54.600 --> 0:04:57.400 talking about the Morrobon carcass of we work. I recommend 0:04:57.480 --> 0:04:59.240 you look at and I'll link to the list in 0:04:59.240 --> 0:05:02.320 the episode of notes SoftBank's recent group report, which is 0:05:02.360 --> 0:05:04.719 linked in the spreadsheet. Like I just said, which is 0:05:04.760 --> 0:05:06.880 me reading a script out a little behind the curtain 0:05:06.880 --> 0:05:09.760 for you there. In particular, got to twenty nine in 0:05:09.839 --> 0:05:12.679 the report, which lists the ten largest publicly traded companies 0:05:12.680 --> 0:05:16.120 in soft Bank's Vision Fund portfolio. Now how all of them, 0:05:16.120 --> 0:05:19.920 without exception trade at a significant fraction of their peak market cap. 0:05:20.400 --> 0:05:22.880 In simpler terms, the ten biggest companies in soft Bank's 0:05:22.880 --> 0:05:25.440 flagship tech fund are worth far far less than their 0:05:25.480 --> 0:05:27.680 all time high, and in some cases with less than 0:05:27.720 --> 0:05:30.719 one fifth of their all time high. If soft Bank 0:05:30.720 --> 0:05:32.960 liquidated its assets, and I admit this is a big 0:05:33.000 --> 0:05:35.760 ear for most likely a worst case scenario situation, how 0:05:35.760 --> 0:05:39.720 big would their losses be? Separately, SoftBank has committed to 0:05:39.800 --> 0:05:42.640 a joint venture called SBO open Ai Japan and to 0:05:42.680 --> 0:05:44.680 spend three billion dollars a year on open Ai. Is 0:05:44.720 --> 0:05:47.560 take for the various companies across its group. We'll talk 0:05:47.560 --> 0:05:50.240 about that later, but there's some napkin mass. It is 0:05:50.279 --> 0:05:53.200 what soft Banks agreed to sixteen billion dollars sorry, eighteen 0:05:53.240 --> 0:05:55.880 billion dollars or nineteen billion dollars really do not know 0:05:55.960 --> 0:05:58.839 in funding for the Stargate Data Center project, three billion 0:05:58.880 --> 0:06:01.880 dollars a year in spend on open ai software, and 0:06:01.920 --> 0:06:04.280 as much as thirty billion dollars in funding for open 0:06:04.279 --> 0:06:06.600 Ai paid over twelve to twenty four months, according to 0:06:06.640 --> 0:06:09.760 the Information, and by the way, twenty five billion has 0:06:09.800 --> 0:06:12.120 also been reported. But the Information reports that open Ai 0:06:12.160 --> 0:06:14.159 has told investors that soft Bank will provide it at 0:06:14.279 --> 0:06:16.839 least thirty billion dollars of the forty billion dollars they need. 0:06:17.040 --> 0:06:19.080 Jesus fucking Christ. Can you imagine if this went It's 0:06:19.080 --> 0:06:23.560 something useful anyway, What I'm getting at is that soft 0:06:23.560 --> 0:06:26.400 Bank has effectively agreed to bankroll the entirety of open 0:06:26.400 --> 0:06:29.520 AI's future, signing up for over forty six billion dollars 0:06:29.520 --> 0:06:32.080 of investments over the next few years, and does not 0:06:32.120 --> 0:06:34.080 appear to be able to do so without selling its 0:06:34.120 --> 0:06:36.680 current holdings in valuable companies like arm or taking at 0:06:36.760 --> 0:06:39.680 least sixteen billion dollars of debt this year, representing a 0:06:39.680 --> 0:06:43.039 fifty five percent increase of its current liabilities. Work still, 0:06:43.080 --> 0:06:46.160 thanks to disagreement, open AI's future both in its ability 0:06:46.200 --> 0:06:49.760 to expand its infrastructure, which appears to be entirely contingent 0:06:49.839 --> 0:06:53.040 on the construction of Stargate with Microsoft pulling out, and 0:06:53.080 --> 0:06:56.039 its ability to raise funding, which is also entirely dependent 0:06:56.120 --> 0:06:59.000 on soft bank and soft Bank in this case, where 0:06:59.040 --> 0:07:01.640 open ai is entirely dependent on soft Bank to live, 0:07:01.839 --> 0:07:06.000 must borrow money to give to open Ai, a company 0:07:06.040 --> 0:07:11.640 that only only loses money. On top of that, open 0:07:11.680 --> 0:07:14.800 Ai also on the money losing front, anticipates it will 0:07:14.840 --> 0:07:16.840 burn as much as forty billion dollars a year by 0:07:16.880 --> 0:07:19.200 twenty twenty eight, and projects to only turn a profit 0:07:19.240 --> 0:07:22.440 by the end of the decade after the build out 0:07:22.480 --> 0:07:26.920 of Stargate, which I add is, like I said, almost 0:07:27.040 --> 0:07:29.600 entirely dependent on soft bank, which has to take on 0:07:29.720 --> 0:07:33.880 debt to fund both open Ai and the project required 0:07:33.920 --> 0:07:37.920 to theoretically make open ai profitable. How the fuck does 0:07:37.920 --> 0:07:40.600 this work? How does this work? How does this work? 0:07:40.640 --> 0:07:42.560 How does this work? How the fuck does this work? 0:07:42.600 --> 0:07:45.680 Open Ai, a company that spent nine billion dollars to 0:07:45.760 --> 0:07:48.640 lose five billion dollars in twenty twenty four, requires so 0:07:48.720 --> 0:07:52.520 much money to be its obligations, both to cover its stupid, ruinous, unprofitable, 0:07:52.600 --> 0:07:55.560 unsustainable operations and the eighteen billion to nineteen billion dollars. 0:07:55.560 --> 0:07:58.440 A committed to keep growing that it has to raise 0:07:58.480 --> 0:08:01.679 more money than any startup ever raised in history, forty 0:08:01.760 --> 0:08:04.480 billion fucking dollars, with the cast iron guarantee that it 0:08:04.480 --> 0:08:08.320 will need even more money within a year. Every goddamn month, 0:08:08.480 --> 0:08:10.679 Sam Wilton has to go to someone and say, my huge, 0:08:10.680 --> 0:08:13.320 beautiful company is so powerful, but it's weak and sick 0:08:13.360 --> 0:08:15.360 and frail. I need more money than you have today, 0:08:15.480 --> 0:08:17.480 and I'll need more money than you'll have tomorrow. My 0:08:17.560 --> 0:08:19.400 company is so sick and weak, but it's the most 0:08:19.440 --> 0:08:24.160 powerful company of all time. Sam flipping off pisces me off. 0:08:24.880 --> 0:08:27.520 Imagine if this money went literally anywhere else, you could 0:08:27.560 --> 0:08:30.560 set it on fire at least be fucking warm. Soft Bank. 0:08:31.000 --> 0:08:33.080 On top of the thirty billion dollars of funding and 0:08:33.120 --> 0:08:35.400 three billion dollars a year of revenue it's committed to, 0:08:35.440 --> 0:08:38.240 open ai itself also has to cough up eighteen billion 0:08:38.240 --> 0:08:41.000 dollars for Stargate, the data center project that open ai 0:08:41.120 --> 0:08:45.080 will run and get this. Soft Bank will take financial 0:08:45.120 --> 0:08:49.240 responsibility for forty eight billion dollars in cash three billion 0:08:49.280 --> 0:08:51.560 dollars in revenue. You the latter of which, like all 0:08:51.559 --> 0:08:55.200 open AI's offerings, will lose the company money. Open Ai 0:08:55.360 --> 0:08:58.559 has no path to profitability guaranteeing it will need more 0:08:58.600 --> 0:09:00.400 cash and right now at the time time, and it 0:09:00.440 --> 0:09:02.920 needs it more than it's ever needed it. Soft Bank, 0:09:02.960 --> 0:09:05.080 the only company willing to provide it and possibly the 0:09:05.080 --> 0:09:07.120 only company with the money to do so, has proven 0:09:07.160 --> 0:09:09.760 that it will have to go to greet possibly ruinous 0:09:09.840 --> 0:09:12.840 lengths to do so. If open ai needs forty billion 0:09:12.880 --> 0:09:14.800 dollars in twenty twenty five, how much will it need 0:09:14.840 --> 0:09:17.880 in twenty twenty six fifty billion, one hundred billion dollars. 0:09:18.000 --> 0:09:21.040 Where is that money going to come from? While soft 0:09:21.040 --> 0:09:23.120 Bank might be able to do this once, what happens 0:09:23.120 --> 0:09:25.480 when open ai needs money in six to twelve months. 0:09:25.600 --> 0:09:27.960 Soft Bank made about fifteen billion dollars a profit in 0:09:27.960 --> 0:09:31.240 the last year on about forty six billion dollars of revenue. 0:09:31.440 --> 0:09:34.079 Three billion dollars is an absolutely obscene amount to commit 0:09:34.080 --> 0:09:37.040 to buying open ai software annually, especially when some of 0:09:37.080 --> 0:09:39.760 it is allegedly for access to open AI's barely functional 0:09:39.920 --> 0:09:44.160 operator and mediocre deep research products. As per my previous 0:09:44.280 --> 0:09:47.760 podcasts and pieces, I do not see how open ai survives, 0:09:47.960 --> 0:09:51.240 and soft Bank's involvement only gives me further concern. While 0:09:51.240 --> 0:09:54.280 soft banker theoretically burn its current holdings to fund open 0:09:54.320 --> 0:09:56.960 Ai in perpetuity, its ability to do so is cast 0:09:57.000 --> 0:09:59.000 into doubt by them having to borrow money from other 0:09:59.080 --> 0:10:01.720 banks to get into this funding ground and to get 0:10:01.760 --> 0:10:05.720 Stargate done. Open Ai burned five billion dollars in twenty 0:10:05.760 --> 0:10:08.400 twenty four, a number it will likely double in twenty 0:10:08.440 --> 0:10:11.360 twenty five. And remember the information reported that open ai 0:10:11.559 --> 0:10:14.200 was projected to spend thirteen billion dollars on compute alone 0:10:14.240 --> 0:10:17.360 with Microsoft in twenty twenty five, and has no path 0:10:17.400 --> 0:10:20.880 to profitability. SoftBank has already had to borrow to fund 0:10:20.960 --> 0:10:22.800 this round, and the fact that it had to do 0:10:22.840 --> 0:10:26.200 so suggests its inability to continue supporting open Ai without 0:10:26.200 --> 0:10:30.080 accruing further debt. Open Ai. As a result of Microsoft's cuts, 0:10:30.080 --> 0:10:34.040 the data center capacity now only has one path expansion 0:10:34.600 --> 0:10:37.840 once it runs through whatever remaining buildout Microsoft has planned. 0:10:37.840 --> 0:10:40.680 That is, and that's Stargate, a project funded by open 0:10:40.760 --> 0:10:44.440 AI's contribution, which it's receiving from SoftBank and SoftBank which 0:10:44.480 --> 0:10:47.079 is also having to take out loans to meet its share. 0:10:47.760 --> 0:10:51.240 How does this work exactly? How does this continue? Do 0:10:51.360 --> 0:10:53.840 you see anyone else stepping up to fund this. Who 0:10:53.920 --> 0:10:56.400 else has got thirty to forty billion dollars to shit 0:10:56.440 --> 0:10:59.840 out every year? While the answer is Saudi Arabia. SoftBank's 0:11:00.000 --> 0:11:02.920 CEO Massioshi Son recently said that he had had and 0:11:03.000 --> 0:11:05.640 I quote not given Saudi ruler Muhammed bin Salmon in 0:11:05.720 --> 0:11:10.199 our return, adding that he still owed him. That's really 0:11:10.240 --> 0:11:13.720 not the ideal thing you want to say after naming MBS. 0:11:14.200 --> 0:11:16.640 Nothing about this suggests that Saudi money will follow soft 0:11:16.640 --> 0:11:19.280 banks in anywhere near the volume necessary. As to the 0:11:19.360 --> 0:11:21.599 m ratis. They're already involved through the MGX fund, and 0:11:21.640 --> 0:11:27.160 it's unclear how much more they'd be willing to commit. Really, really, 0:11:27.240 --> 0:11:31.920 are really, mate, buddy fellow? How does this? How does 0:11:31.920 --> 0:11:35.000 this work? In my opinion, this open Ai soft Bank 0:11:35.040 --> 0:11:38.040 deal is wildly unsustainable, dependent on soft Bank continuing to 0:11:38.120 --> 0:11:41.080 raise both debt and funnel money directly into a company 0:11:41.080 --> 0:11:43.760 that burns it burns it by the billions every year, 0:11:43.840 --> 0:11:45.520 and it's said to only burn more thanks to the 0:11:45.600 --> 0:11:48.360 arrival of its latest bullshit model. And if it had 0:11:48.400 --> 0:11:50.880 a huge breakthrough that would change everything. Wouldn't Microsoft want 0:11:50.880 --> 0:11:52.760 to make sure they were building the data center capacity 0:11:52.800 --> 0:12:08.080 to support him? Hey, maybe they're not. Now. Perhaps this 0:12:08.360 --> 0:12:11.280 crazy level of spending would be necessary if open ai 0:12:11.440 --> 0:12:13.640 was still the leader in generator of AI and it 0:12:13.679 --> 0:12:16.960 was still meaningfully improving its capabilities year after year or 0:12:17.280 --> 0:12:19.559 two year after two years. And I think we all 0:12:19.559 --> 0:12:21.960 know that that isn't the case. A few weeks ago, 0:12:22.000 --> 0:12:24.800 open ai launched GPT four point five, it's latest model 0:12:24.840 --> 0:12:29.560 that well, this isn't brilliant, sam Alman says, and I quote, 0:12:29.600 --> 0:12:31.600 is the first model that feels like talking to a 0:12:31.640 --> 0:12:35.000 thoughtful person, which, by the way, is really funny. It's 0:12:35.040 --> 0:12:37.280 really funny to say it, to be like, yeah, I 0:12:37.320 --> 0:12:40.480 have spent billions of dollars telling you that this bullshit 0:12:40.600 --> 0:12:43.200 is like a person, but this one really is. The 0:12:43.320 --> 0:12:45.560 other one's not so much. I did not think that 0:12:45.640 --> 0:12:48.880 in the past. And it's not obvious what GPT four 0:12:48.880 --> 0:12:52.439 point five does better, or even really what it does, 0:12:52.480 --> 0:12:55.040 other than Altman saying it is a different kind of 0:12:55.040 --> 0:12:58.120 intelligence and there's a magic to it that he has 0:12:58.160 --> 0:13:03.760 not felt before. Wow, wow, really really inspirational stuff. Sammy 0:13:04.720 --> 0:13:07.319 fucking idiot. This was, by the way, in Altman's words, 0:13:07.400 --> 0:13:10.160 the good news. The bad news was that and I 0:13:10.280 --> 0:13:13.480 quote GPT four point five is a giant expensive model, 0:13:13.600 --> 0:13:16.520 adding the open Ai was out of GPUs, but proudly 0:13:16.559 --> 0:13:19.160 declaring that it had added tens of thousands of GPUs 0:13:19.200 --> 0:13:21.680 in the week following and would roll out to open aiyes, 0:13:21.720 --> 0:13:24.120 twenty dollars a month plus here and that they would 0:13:24.120 --> 0:13:28.240 be adding hundreds of thousands of GPUs soon. Excited. Well, 0:13:28.280 --> 0:13:30.560 you shouldn't be on top of a vague product set 0:13:30.559 --> 0:13:33.840 and interterminately high compute cost. GPT four point five costs 0:13:33.880 --> 0:13:37.880 developers an incredible seventy five dollars put million input tokens 0:13:37.920 --> 0:13:40.400 in those are the prompts and data pushed into the model, 0:13:40.640 --> 0:13:43.600 and of absolutely astounding one hundred and fifty dollars per 0:13:43.600 --> 0:13:46.640 million output tokens. That's the output it creates. And that's 0:13:46.720 --> 0:13:49.600 roughly three thousand percent more for input tokens and fifteen 0:13:49.679 --> 0:13:52.319 hundred percent more expensive for output tokens than GPT four 0:13:52.320 --> 0:13:55.600 to zero For results that open Aiyes, andre J. Caapathy 0:13:55.720 --> 0:13:59.000 described as a little bit better and awesome, but also 0:13:59.240 --> 0:14:02.040 not exactly that are trivial to point to, and one 0:14:02.080 --> 0:14:05.320 developer described GPT four point five to our technica as 0:14:05.360 --> 0:14:08.520 a lemon when comparing its reported performance to its price. 0:14:10.600 --> 0:14:14.240 Arstentnaca also reported the GPT four point five was terrible 0:14:14.240 --> 0:14:16.600 for coding relatively speaking, and other tests showed that the 0:14:16.600 --> 0:14:19.760 model's performance was either slightly better or slightly worse across 0:14:19.760 --> 0:14:22.440 the board. With that according to Ours Technica, one success 0:14:22.480 --> 0:14:26.240 metric being the open ai found open human evaluators preferred 0:14:26.240 --> 0:14:29.360 GPT four point five's responses of a GPT four to 0:14:29.400 --> 0:14:34.640 zero in about fifty seven percent of interactions. Wow, that's 0:14:35.160 --> 0:14:38.160 very underwhelming. So, just to be crystal clear, the biggest 0:14:38.160 --> 0:14:40.960 AI company's latest model appears to be even more ruinously 0:14:41.000 --> 0:14:43.400 expensive than its last one, while providing modest at best 0:14:43.400 --> 0:14:46.840 improvements and performing worse than several benchmarks done competing models 0:14:48.760 --> 0:14:51.840 very good. Despite these piss poor results, Sam Wultman's reaction 0:14:51.960 --> 0:14:54.040 was to bring in hundreds of thousands of GPUs as 0:14:54.040 --> 0:14:56.080 a means of exposing as many as people as possible 0:14:56.080 --> 0:14:58.960 to his mediocre, ultra expensive model. And the best that 0:14:59.000 --> 0:15:01.040 Altman has to offer is that this is the first 0:15:01.120 --> 0:15:04.080 time people have been emailing with such passion asking open 0:15:04.120 --> 0:15:06.800 ai to promise to never stop offering a specific model. 0:15:07.160 --> 0:15:09.480 I am just going to say this. That is a tweet, 0:15:09.520 --> 0:15:13.000 and it never happened, or it happened like once. This 0:15:13.080 --> 0:15:16.480 is some girlfriend in Canada share Sam Altman is washed 0:15:16.880 --> 0:15:19.800 when all of this falls apart, Remember I said he 0:15:19.920 --> 0:15:22.880 was washed? Now, Remember how I talked about open AI's 0:15:22.920 --> 0:15:26.000 lack of meaningful improvement as a reminder, GPT four point 0:15:26.000 --> 0:15:28.360 five was meant to be GPT five, but, according to 0:15:28.400 --> 0:15:30.960 the Wall Street Journal, continually failed to make a model 0:15:31.000 --> 0:15:33.840 that advanced enough to justify the enormous cost, with a 0:15:33.880 --> 0:15:37.680 six month training run costing five hundred million dollars and 0:15:37.760 --> 0:15:41.360 GPT four point five requiring multiple runs of different sizes. 0:15:41.640 --> 0:15:44.160 So yeah, open ai spent hundreds of millions of dollars 0:15:44.200 --> 0:15:47.320 to make this great stuff. And I haven't even mentioned 0:15:47.360 --> 0:15:49.800 the company's supported agent products. And no, I'm not talking 0:15:49.840 --> 0:15:52.760 about Operator, which is also dogshit. By the way, open 0:15:52.800 --> 0:15:55.640 ai wants to create tiers of AI agents, with the 0:15:55.720 --> 0:15:59.480 cheapest costing two thousand dollars a month and capable of 0:15:59.480 --> 0:16:02.480 handling it administrative tasks, and the most expensive costing twenty 0:16:02.520 --> 0:16:05.840 thousand dollars in having PhD level, Look, I wrote the 0:16:05.880 --> 0:16:08.760 script right, and I'm gonna be honest. I can't even 0:16:08.840 --> 0:16:15.080 read that sentence with a straight face. Operator cannot even 0:16:15.120 --> 0:16:18.320 search trip Advisor properly. It can't even do a thing 0:16:18.320 --> 0:16:20.680 that let me google that for you does. And these 0:16:20.760 --> 0:16:23.000 chunder fucks want to charge two thousand dollars a month 0:16:23.040 --> 0:16:26.040 for an agent that does what does some sort of 0:16:26.360 --> 0:16:26.880 what's it dou? 0:16:27.000 --> 0:16:27.160 Oh? 0:16:27.480 --> 0:16:31.040 Twenty thousand dollars for something with PhD level capabilities. I 0:16:31.040 --> 0:16:33.320 think all the people with PhDs listen to this have 0:16:33.480 --> 0:16:36.960 just stood up and gone, I have an idea. And 0:16:37.000 --> 0:16:39.320 this is insane on many levels, not simply because the 0:16:39.360 --> 0:16:41.920 base product is undercut by actual human workers in many 0:16:41.960 --> 0:16:44.760 parts of the world, and even PhD students are typically 0:16:44.800 --> 0:16:46.880 I only paid twenty thousand to thirty thousand a year 0:16:47.000 --> 0:16:50.120 on average, and even people with actual doctorates in industries 0:16:50.160 --> 0:16:52.800 really earn twenty thousand dollars a month unless they're working 0:16:52.800 --> 0:16:56.040 in Silicon Valley occupying a C suite job. But forget 0:16:56.080 --> 0:16:57.880 all about that. What does it mean to have a 0:16:57.880 --> 0:17:01.440 PhD level agent? Remember? Other lambs are guessing machines. They 0:17:01.480 --> 0:17:05.159 don't know anything or even understand the concepts behind the 0:17:05.200 --> 0:17:08.520 words they spit onto a page. No, seriously, Sammy, what 0:17:08.600 --> 0:17:15.239 does it mean? I'm fucking waiting you damp Goblin you pissant. Ah, 0:17:15.320 --> 0:17:19.199 maybe I shouldn't just site here instarting him wanker. Anyway, This, 0:17:19.280 --> 0:17:20.879 by the way, is the company that is about to 0:17:20.960 --> 0:17:23.159 raise forty billion dollars led by a Japanese bank that 0:17:23.160 --> 0:17:25.280 has to go into debt to fund both their operations 0:17:25.280 --> 0:17:31.520 and the infrastructure necessary for them to grow any further. Again, 0:17:32.480 --> 0:17:36.880 as we started with, Microsoft is canceling plans to massively 0:17:36.920 --> 0:17:39.360 expand its data center capacity right at the time when 0:17:39.359 --> 0:17:42.560 open ai just released its most computationally demanding model. Ever, 0:17:43.280 --> 0:17:46.520 how do you reconcile those two things without concluding either 0:17:46.520 --> 0:17:49.080 that Microsoft expects GBT four point five to be a 0:17:49.080 --> 0:17:52.560 flop although it's simply unwilling to continue bankrolling open AI's 0:17:52.600 --> 0:17:55.400 continued growth, or perhaps it's having doubts about the future 0:17:55.440 --> 0:18:01.800 of generative AI in general. Hmm. Maybe now I have 0:18:02.040 --> 0:18:04.800 been and remain hesitant to call the bubble bursting, because 0:18:04.840 --> 0:18:07.879 bubbles do not burst, really, they certainly don't burst in 0:18:07.960 --> 0:18:11.560 neat little events. Nevertheless, my pale horses I've predicted in 0:18:11.560 --> 0:18:14.159 the past were led by one specific call that reduction 0:18:14.240 --> 0:18:16.760 and capital expenditures by a hyperscaler was a sign that 0:18:16.800 --> 0:18:20.600 things were collapsing. Microsoft walking away from over a gigawatt 0:18:20.600 --> 0:18:23.080 of data center plans equivalent to as much as fourteen 0:18:23.119 --> 0:18:26.240 percent of its current data center capacity, is a direct 0:18:26.359 --> 0:18:28.680 sign that it does not believe that growth is there 0:18:28.880 --> 0:18:31.199 in generative AI, and thus they are not building the 0:18:31.240 --> 0:18:34.600 infrastructure to support it, and indeed may have overbuilt something, 0:18:34.720 --> 0:18:38.400 as I've mentioned that Microsoft CEO Sachinidella has directly foreshadowed 0:18:38.480 --> 0:18:42.439 in his interview with Dwarkesh and otherwise extremely boring waste 0:18:42.480 --> 0:18:45.080 of an hour of your life. The entirety of the 0:18:45.080 --> 0:18:47.600 tech industry and the AI bubble has been built on 0:18:47.640 --> 0:18:50.359 the assumption that generative AI was the next big growth 0:18:50.440 --> 0:18:53.720 vehicle for tech, and if Microsoft, the largest purchaser of 0:18:53.800 --> 0:18:57.119 Nvidia GPUs in the most aggressive builder of AI infrastructure, 0:18:57.160 --> 0:19:00.280 is reducing capacity, it heavily suggests that the growth is 0:19:00.440 --> 0:19:04.000 not there. Microsoft has, by the looks of things, effectively 0:19:04.040 --> 0:19:07.000 given up on further data center expansion, at least at 0:19:07.000 --> 0:19:10.520 the breakneck pay suit runs promised, and that even suggests 0:19:10.520 --> 0:19:12.199 that generator if AI will be a thing in a 0:19:12.200 --> 0:19:14.880 few years, definitely not the scale it is right now. 0:19:15.520 --> 0:19:18.760 AI boosters will email me and they'll say there's something 0:19:18.800 --> 0:19:20.879 that I don't know that in fact, Microsoft has some 0:19:20.920 --> 0:19:24.119 greater strategy and some efficiency play. But answer me this, 0:19:24.280 --> 0:19:26.520 Why is Microsoft canceling over a giga? What of data 0:19:26.560 --> 0:19:29.440 set expansion? And again this is the most conservative estimate. 0:19:30.119 --> 0:19:32.040 The realistic number is much much higher. Do you think 0:19:32.040 --> 0:19:34.479 it's because it expects there to be this dramatic demand 0:19:34.480 --> 0:19:37.520 for the AI services? Do you think it's reducing supply 0:19:37.640 --> 0:19:41.520 because of all the demand? Now you might think that 0:19:41.600 --> 0:19:46.760 this is an efficiency play they're playing with deep SIG right, No, sorry, 0:19:46.880 --> 0:19:50.439 that doesn't matter. Even if deep CG was this magical 0:19:50.440 --> 0:19:52.399 efficiency play, which it may or may not be. I 0:19:52.480 --> 0:19:54.480 actually think it is more efficient, Like that's true, but 0:19:54.520 --> 0:19:57.800 we actually don't know if it's profitable. Even then, they've 0:19:57.840 --> 0:20:00.720 been talking about not having the capacity to deal with demand. 0:20:00.720 --> 0:20:03.520 They've been talking about how incredible this is. They've been 0:20:03.520 --> 0:20:06.439 talking about how big this is gonna be. This sounds 0:20:06.440 --> 0:20:09.360 like they don't think it's gonna big, gonna be big. 0:20:09.400 --> 0:20:11.280 Even it's gonna big, they're gonna get my ass in 0:20:11.320 --> 0:20:14.719 the comments on that one. Now, one might argue that 0:20:14.760 --> 0:20:17.160 Microsoft's reduction and capacity build out it's just the sign 0:20:17.160 --> 0:20:20.919 that open AI is moving its compute elsewhere. Maybe that's true. 0:20:21.840 --> 0:20:24.520 And if starguate ever gets built, which I question anyway, 0:20:25.040 --> 0:20:28.159 here are some questions to ask. Microsoft still sells access 0:20:28.200 --> 0:20:30.920 to open AI's API through asure. Does it not see 0:20:30.960 --> 0:20:33.879 the growth in that product? Do they not see it? 0:20:34.040 --> 0:20:37.240 They're not expanding? Is the growth not there? And Microsoft still, 0:20:37.280 --> 0:20:40.560