WEBVTT - Monologue: Tokenpocalypse Now 0:00:02.480 --> 0:00:03.560 Azo Media. 0:00:05.480 --> 0:00:07.640 Hello and welcome to the second of this week's Better 0:00:07.680 --> 0:00:08.680 Offline Monologues. 0:00:08.800 --> 0:00:10.119 I'm your host ed Zetron. 0:00:18.120 --> 0:00:20.480 It's been a funny few weeks watching the little spurts 0:00:20.480 --> 0:00:22.800 of air coming out of the AI bubble. It's unclear 0:00:22.840 --> 0:00:26.000 when it will burst, what will burst it, really even. 0:00:25.800 --> 0:00:26.400 What's going on. 0:00:26.520 --> 0:00:28.440 Half the time, they just all seem to be running 0:00:28.440 --> 0:00:31.080 around like a billion mister beans. But I get the 0:00:31.120 --> 0:00:34.239 sense that everything will accelerate dramatically based on one or 0:00:34.240 --> 0:00:37.680 two big events, things that just knock the confidence of everyone. 0:00:38.240 --> 0:00:40.920 Maybe it's an AI company dying, or a funding ground 0:00:40.960 --> 0:00:44.600 not coming together, or maybe it's a hyperscalar cutting Capex 0:00:45.040 --> 0:00:48.839 Goldin SAX analyst Rich Pirotovsky said last week that he 0:00:48.920 --> 0:00:52.840 believed that everybody was spending simply to remain competitive. Fucking 0:00:52.880 --> 0:00:55.520 love that, don't you, and added that the first hyperscala 0:00:55.560 --> 0:00:57.640 to signal that it can slow the pace of spending 0:00:57.680 --> 0:01:00.680 would likely see its share price rewarded. If I had 0:01:00.720 --> 0:01:03.000 to bet, I think it's between Meta and Microsoft, the 0:01:03.080 --> 0:01:05.600 latter of which has made numerous noises about using cheaper 0:01:05.600 --> 0:01:08.320 AI models like deep Seek and its co pilot cowork 0:01:08.400 --> 0:01:12.040 product Yeah and Also, I should add the Sacha and 0:01:12.080 --> 0:01:14.520 Adela keeps doing interviews where he's like, yeah, no, one 0:01:14.560 --> 0:01:17.800 company should have all that power. We shouldn't rely on 0:01:17.800 --> 0:01:20.600 one lab, one lab being able to be shut down 0:01:20.640 --> 0:01:21.840 at any times bad. 0:01:21.959 --> 0:01:22.360 We can't. 0:01:22.600 --> 0:01:25.440 It's really funny watching this guy change his chew and 0:01:25.520 --> 0:01:27.640 considering how a year and a half ago, a year 0:01:27.680 --> 0:01:30.920 ago he was rock hard for open AI, and six 0:01:30.959 --> 0:01:33.920 months after that rock harder for Anthropic. It's almost as 0:01:33.959 --> 0:01:36.319 if he's just doing random shit based on what he 0:01:36.400 --> 0:01:40.640 thinks might work, you know, like a fucking cargo cult. Anyway, 0:01:40.800 --> 0:01:43.360 as I've mentioned over the last few months, both Anthropic 0:01:43.400 --> 0:01:46.400 and open Ai both started to charge their enterprise customers 0:01:46.480 --> 0:01:48.480 companies with over one hundred and fifty users for the 0:01:48.480 --> 0:01:51.080 actual cost of the AI tooken spend as of the 0:01:51.120 --> 0:01:55.000 middle of Q one, twenty twenty six. To explain, when 0:01:55.040 --> 0:01:57.480 you use a regular chat, GPT or claud account, you 0:01:57.560 --> 0:02:00.240 burn tokens, each one about three quarters of a word, 0:02:00.640 --> 0:02:03.360 up to a five hour or weekly limit, depending on 0:02:03.400 --> 0:02:05.560 what kind of model you're using. So with Anthropic on 0:02:05.560 --> 0:02:08.919 their clause subscription Opus four point eight and Fable when 0:02:08.919 --> 0:02:12.320 it was around, had a specific limit to themselves. But anyway, 0:02:12.560 --> 0:02:14.920 just the more powerful the model powerful being meshed by 0:02:14.960 --> 0:02:17.280 the companies, of course, the more it burns and the 0:02:17.360 --> 0:02:20.120 less you can use it. On a regular subscription you're 0:02:20.120 --> 0:02:22.680 able to burn and I am not kidding eight thousand 0:02:22.680 --> 0:02:25.640 dollars worth of tokens a month on Claude and fourteen 0:02:25.760 --> 0:02:30.440 thousand dollars a month on open Aiy's chet GPT codec subscription. 0:02:30.360 --> 0:02:31.880 All for two hundred bucks a month. 0:02:32.080 --> 0:02:35.720 Pretty good deal, right, you know? It's like the discount 0:02:35.760 --> 0:02:38.520 prices sketch from Tim and Eric that I reference every 0:02:38.560 --> 0:02:41.760 so often just to see if anyone emails me. And 0:02:41.800 --> 0:02:44.639 so Yeah, after years of being able to burn thousands 0:02:44.680 --> 0:02:46.959 of dollars of tokens for two hundred dollars a month, 0:02:47.080 --> 0:02:50.240 enterprises are suddenly having to pay the actual costs. The 0:02:50.280 --> 0:02:53.520 result is multiple companies capping their workers token spend. As 0:02:53.520 --> 0:02:55.720 I've reported, both T Mobile and Brecks have done so, 0:02:55.800 --> 0:02:58.760 and others have reported that Uber, Meta, Walmart, Coinbase, and 0:02:58.800 --> 0:03:01.880 Cisco have all done so. Meta notably had a token 0:03:01.960 --> 0:03:06.480 maxing leaderboard. Very funny again, just changing direction based on 0:03:06.600 --> 0:03:10.160 random signals. None of these people have a plan, none 0:03:10.160 --> 0:03:12.440 of them. None of these companies have a plan. Nobody 0:03:12.520 --> 0:03:15.240 integrating AI has a plan. If they did, they would 0:03:15.240 --> 0:03:17.440 have been like, Yeah, let's make sure we know how 0:03:17.480 --> 0:03:19.800 much this is gonna cost, or let's make sure we 0:03:19.880 --> 0:03:21.799 know whether this is good or not, whether we can 0:03:22.160 --> 0:03:26.040 measure the roy you know, anything that would suggest anyone 0:03:26.120 --> 0:03:30.160 knew what the hell they're doing, but they don't. And 0:03:30.320 --> 0:03:33.240 four o four media reports that Management Consultancy Firm and 0:03:33.600 --> 0:03:37.520 Evil Corporation Accenture, which is an insane way of saying 0:03:37.520 --> 0:03:39.640 that I'm just going to keep has seen what it 0:03:39.720 --> 0:03:42.960 calls soaring token spend based on leaked audio recordings, with 0:03:43.680 --> 0:03:46.200 much of that token spend being driven by non engineers 0:03:46.200 --> 0:03:49.920 doing things like combining PDFs into presentation slides. I love 0:03:50.000 --> 0:03:53.320 this shit is the easiest stuff in the world made 0:03:53.600 --> 0:03:57.640 slightly easier but probably worse, costing way more. You could 0:03:57.720 --> 0:03:59.640 definitely hire an extra, but you could just hire a 0:03:59.640 --> 0:04:02.160 contract to do this. You could iira a bloke at 0:04:02.200 --> 0:04:04.080 forty five bucks an hour to do that for you. 0:04:04.320 --> 0:04:07.880 Maybe the happiest man alive probably cost your less thought too. 0:04:07.960 --> 0:04:11.760 I know, Ugh, God, this stuff is exhausting. You ever 0:04:11.840 --> 0:04:14.520 sit and wonder what you could do with the rest 0:04:14.560 --> 0:04:17.919 of that money I could have, Oh no, not like 0:04:18.080 --> 0:04:20.599 millions of dollars a month. I could have like a 0:04:20.640 --> 0:04:23.640 week's worth of diet coke. Anyway, it turns out that 0:04:23.720 --> 0:04:26.680 telling your workers that AI can do basically anything and 0:04:26.760 --> 0:04:28.839 to use it as much as possible starts costing you 0:04:28.839 --> 0:04:31.240 a lot of money when you actually pay what it costs. 0:04:31.640 --> 0:04:34.839 As I've said before, we're barely three months into enterprises 0:04:34.839 --> 0:04:37.480 paying the actual cost of AI, and they're already screeching 0:04:37.520 --> 0:04:39.880 like they're being pecked to death by birds. And that 0:04:40.040 --> 0:04:42.640 joke is a courtesy of Kisei Kugawa, whos said that 0:04:42.720 --> 0:04:44.880 to me about a year ago when we talked about 0:04:44.880 --> 0:04:47.080 the amount of money that Anthropic was allowing people to 0:04:47.080 --> 0:04:50.760 burn on Claude subscriptions. Meanwhile, the threat of open source 0:04:50.800 --> 0:04:53.240 models appears to be getting more serious. I've heard from 0:04:53.320 --> 0:04:57.120 numerous sources that Chinese AI lab z poos GLM five 0:04:57.120 --> 0:05:00.599 point two coding model is competitive with Anthropic Sopus four 0:05:00.600 --> 0:05:02.880 point eight at somewhere between a quarter and a sixth 0:05:02.920 --> 0:05:04.880 of the price, but that seems to vary based on 0:05:04.960 --> 0:05:08.000 the cost of the task. As far as whether GLM 0:05:08.080 --> 0:05:10.919 five point two is actually competitive goes it appears that 0:05:10.960 --> 0:05:13.440 the hype is somewhat real, but it's kind of you 0:05:13.520 --> 0:05:16.400 kind of had to check your saucing. Open source AI 0:05:16.440 --> 0:05:19.400 coding company Cliin ran benchmarks on a natural bug from 0:05:19.400 --> 0:05:22.080 its repo and found that while GLM used twice as 0:05:22.120 --> 0:05:24.200 many tokens to fix it, it also did so at 0:05:24.200 --> 0:05:27.080 a cheaper cost forty one cents versus OPUS four point 0:05:27.080 --> 0:05:30.680 eights eighty one cents. Opus was quicker, taking one point 0:05:30.720 --> 0:05:34.279 six minutes versus glm's four point seven minutes, but one 0:05:34.320 --> 0:05:37.440 thing that stood out was the GLM to quote client 0:05:37.680 --> 0:05:40.440 cleaned up dead code and verified the build compiled before 0:05:40.480 --> 0:05:43.719 completing something that Opus, the more expensive frontier model run 0:05:43.720 --> 0:05:46.360 by the company that uses billions of dollars, apparently didn't 0:05:46.360 --> 0:05:49.279 bother to do. It's just one bug, but GLM five 0:05:49.279 --> 0:05:51.120 point two seems to score well on many of the 0:05:51.160 --> 0:05:55.480 major benchmarks, coming third on cybersecurity firms semgreps cyber benchmarks, 0:05:55.480 --> 0:05:58.360 with the top two being semgp's own harness, so their 0:05:58.400 --> 0:05:59.960 own thing that they put on top of the models 0:06:00.080 --> 0:06:03.160 running Opus four point eight and GPT five point five, 0:06:03.680 --> 0:06:07.360 followed by GLM five point two, which is interesting. 0:06:07.640 --> 0:06:09.760 It's very interesting if. 0:06:09.960 --> 0:06:11.960 I was Glamy Samuel when I'd be looking at this 0:06:12.000 --> 0:06:14.960 and getting a little worried. I'd genuinely be looking at 0:06:14.960 --> 0:06:17.160 them and even saying, oh, why don't I train my own? 0:06:17.960 --> 0:06:20.400 But you know, that might actually be a bad idea, 0:06:20.440 --> 0:06:23.880 even though it's fully possible, because GLM five point two 0:06:23.920 --> 0:06:26.200 is a true open weights model. It's released under something 0:06:26.240 --> 0:06:29.960 called the unrestricted MIT license, which means you can anyone, 0:06:30.080 --> 0:06:32.800 by the way company, a regular person whomever, can self 0:06:32.839 --> 0:06:34.640 host on, fine tune it, or use it in any 0:06:34.680 --> 0:06:37.760 way they wish. And yeah, I mean in theory, any 0:06:37.760 --> 0:06:39.880 company that wants to use LMS could either run their 0:06:39.920 --> 0:06:42.040 own version of it locally or be on a rig 0:06:42.120 --> 0:06:44.159 that cost tens of thousands of dollars, or pay an 0:06:44.200 --> 0:06:46.400 inference provider like base ten to spin up and train 0:06:46.480 --> 0:06:49.880 their own version. Now, I must be clear that there's 0:06:50.040 --> 0:06:52.479 not a ton of evidence that running GLM five point 0:06:52.560 --> 0:06:55.080 two is profitable for the inference provider. In fact, there's 0:06:55.279 --> 0:06:57.919 really note I genuinely have looked, nor do we have 0:06:57.960 --> 0:07:00.320 any significant signs that customers are moving away way on 0:07:00.440 --> 0:07:04.000 mass from open AI and anthropics models. People generally like 0:07:04.160 --> 0:07:07.960 name brands, and they even though Anthropic shit is constantly broken. 0:07:08.040 --> 0:07:12.440 Go look at their stability page, stability page, uptime page. 0:07:12.440 --> 0:07:14.160 I don't know what you call that. Email me if 0:07:14.160 --> 0:07:15.800 you know what they call that, or I could just 0:07:15.840 --> 0:07:16.400 look after this. 0:07:16.440 --> 0:07:19.520 I'm going to keep going though. Nevertheless, people want name brands. 0:07:19.520 --> 0:07:22.520 People like paying for a company that ostensibly works on 0:07:22.560 --> 0:07:23.760 this and keeps it updated. 0:07:24.000 --> 0:07:25.120 How true is that? Who knows. 0:07:25.160 --> 0:07:27.640 I don't think people are buying lllms because of particularly 0:07:27.640 --> 0:07:31.240 sensible reasons anyway. But what might change things would be Microsoft, 0:07:31.320 --> 0:07:34.200 Amazon or Google offering GLM five point two via their 0:07:34.280 --> 0:07:38.080 foundry Bedrock or Vertex AI platforms. And even then it's 0:07:38.400 --> 0:07:41.000 how does left customers are excited about AI or just 0:07:41.120 --> 0:07:44.119 paying open AI and Anthropic to use buzzy models? Really 0:07:44.160 --> 0:07:48.040 isn't clear. We should know by now. All that being said, 0:07:49.120 --> 0:07:51.560 now does feel like the perfect time for a shift 0:07:51.560 --> 0:07:53.840 towards open source models, se if only because everyone is 0:07:53.880 --> 0:07:57.320 crapping themselves about costs. Everyone is really freaking out, And 0:07:57.360 --> 0:07:58.760 I think that this is going to be the time 0:07:58.800 --> 0:08:01.720 when we start separate the wheat from the chaff, the 0:08:01.760 --> 0:08:04.880 people that actually like this stuff versus the people who 0:08:04.960 --> 0:08:07.720 just feel good because they're doing AI and they don't 0:08:07.720 --> 0:08:09.920 know what else to do with their time, and they 0:08:10.040 --> 0:08:11.600 like forcing people to use it. 0:08:12.040 --> 0:08:23.480 Now I'll get to who actually does in a minute. 0:08:25.080 --> 0:08:28.600 Open Routers ranking show a dramatic move away from Frontier labs. 0:08:28.680 --> 0:08:30.960 The top ten most popular models are all open source 0:08:31.000 --> 0:08:33.760 safe for Opus four point seven at six, Opus four 0:08:33.760 --> 0:08:36.440 point eight at eight, and Sonnet four point six at ten, 0:08:36.640 --> 0:08:39.160 as well as a mysterious new model called al Alpha, 0:08:39.280 --> 0:08:41.360 which is currently free to use, which is probably why 0:08:41.400 --> 0:08:44.720 it's trending. To be clear, anthropics models still dominate the 0:08:44.720 --> 0:08:47.240 top models by task across the board, and that's based 0:08:47.240 --> 0:08:51.160 on the spend on the task. I hate to give 0:08:51.200 --> 0:08:53.720 him any credit though, I really don't like doing this, 0:08:53.840 --> 0:08:56.840 but Scott Gallaway recently made the point that China could 0:08:56.880 --> 0:08:59.960 thrive by AI dumping high quality open source models into 0:09:00.040 --> 0:09:03.079 the ecosystem as a means of destabilizing the major model providers. 0:09:04.080 --> 0:09:06.600 And I hate to say that's what's happening. It's like 0:09:06.640 --> 0:09:10.199 a fucking broken clock. I guess. Zepu and Minimax are 0:09:10.200 --> 0:09:13.559 both partially government funded by the Chinese government and also 0:09:13.679 --> 0:09:16.439 both wofully unprofitable, but neither do so at the scale 0:09:16.480 --> 0:09:19.360 of anthropic or open aim. While both do the kfabe 0:09:19.400 --> 0:09:23.520 discussion of AGI and autonomous compute and all that bullshit, 0:09:23.720 --> 0:09:26.520 they seem far more focused on creating open weight models 0:09:26.559 --> 0:09:28.960 to compete with them. And just for the differentiation here, 0:09:29.160 --> 0:09:31.080 open source means you can share some of it. Open 0:09:31.080 --> 0:09:33.400 weights means you actually can deddle with the model itself. 0:09:33.720 --> 0:09:36.080 That's a flattening. Someone's gonna email. They're gonna be mad 0:09:36.120 --> 0:09:36.360 at me. 0:09:36.920 --> 0:09:37.480 I don't care. 0:09:38.320 --> 0:09:41.000 Sorry, Like it gets the point across, all right, just 0:09:41.280 --> 0:09:42.480 right on the subreddit if. 0:09:42.400 --> 0:09:43.560 You're mad anyway. 0:09:43.720 --> 0:09:46.200 In truth, I think the future of lms will end 0:09:46.280 --> 0:09:50.160 up being locally run models and expensive specialist hardware. The 0:09:50.280 --> 0:09:53.439 underlying economics of cloud based GPU compute do not make 0:09:53.480 --> 0:09:55.840 sense for anybody. You need to make sure the GPUs 0:09:55.840 --> 0:09:59.360 are saturated. You have to buy in advance, otherwise you 0:09:59.400 --> 0:10:01.680 won't like, you just won't be able to get the capacity. 0:10:01.720 --> 0:10:03.480 And if you have less people than you expect, you 0:10:03.480 --> 0:10:04.600 will lose money for sure. 0:10:04.600 --> 0:10:05.199 And if you have. 0:10:05.480 --> 0:10:07.760 More people than you expect, you have to get more compute, 0:10:07.760 --> 0:10:10.000 which means you'll also lose money. So as long as 0:10:10.040 --> 0:10:13.240 you can goldilocks this bullshit, maybe you'll eke out the 0:10:13.240 --> 0:10:17.080 world's worst gross margins in history, or you'll just lose money, 0:10:17.120 --> 0:10:18.120 which is what everyone's doing. 0:10:18.559 --> 0:10:19.800 But I think there may be. 0:10:19.800 --> 0:10:22.720 A point at which ZEPU, minimacs or even American model 0:10:22.720 --> 0:10:26.640 developers start focusing more on services driven custom deployments and 0:10:26.720 --> 0:10:30.520 become boring, ugly licensed driven monstrosities like Oracle. 0:10:31.000 --> 0:10:31.760 And that's the thing. 0:10:32.520 --> 0:10:36.320 I've talked to quite a few people recently and they 0:10:36.360 --> 0:10:39.520 are all saying the same thing. They've even in Vidia 0:10:39.679 --> 0:10:42.760 is moving into this realm and they're selling these giant 0:10:42.880 --> 0:10:46.880 hundred grand machines. I think that's kind of interesting. And 0:10:46.960 --> 0:10:49.920 like I said, that GPU compute, it requires a certain 0:10:49.920 --> 0:10:52.280 critical mass of customer demand that I don't believe will 0:10:52.320 --> 0:10:55.480 exist once the cargo culture of the AI bubble passes, 0:10:55.520 --> 0:10:59.000 because GPU commitments are done a few years in advance, 0:10:59.080 --> 0:11:01.840 and any drop Paul Lull kills the margins of even 0:11:01.880 --> 0:11:04.720 a cheap to run model. Remember Baseten just they're an 0:11:04.720 --> 0:11:07.200 influence provider. They just raise one and a half billion dollars. 0:11:07.720 --> 0:11:11.719 They're not profitable. I mean, ZEPU isn't profitable, and they 0:11:11.760 --> 0:11:14.120 do the same thing open ai does with their stuff 0:11:14.120 --> 0:11:16.960 where they say, oh yeah, yeah, our cost of revenue, 0:11:16.960 --> 0:11:19.080 it's lower than our revenue. However, when you add the 0:11:19.120 --> 0:11:22.280 sales and marketing, it's actually it's actually unprofitable as well. 0:11:22.400 --> 0:11:24.680 Oh also we lose hundreds of millions of dollars because 0:11:24.720 --> 0:11:28.480 of training. Other than that, we're super profitable. Yeah, it's 0:11:28.559 --> 0:11:32.480 really fucking stupid industry. And look, even if GLM five 0:11:32.520 --> 0:11:34.400 point two can do things at half the cost of 0:11:34.440 --> 0:11:37.000 OPAHS four point eight, that's still half of what in 0:11:37.040 --> 0:11:39.880 some cases is hundreds of thousands or millions of dollars 0:11:39.920 --> 0:11:42.880 a month. I've heard tell from multiple source is that 0:11:42.880 --> 0:11:46.440 that hardware demand is growing and that customers at big 0:11:46.440 --> 0:11:49.479 companies are starting to look into those high end hardware solutions. 0:11:50.040 --> 0:11:53.800 And I think that that is the only future for 0:11:53.880 --> 0:11:56.040 this stuff, and I think it would reduce the cost 0:11:56.080 --> 0:11:59.160 significantly for the end user, probably keep this shit off 0:11:59.200 --> 0:12:04.920 the internet, massively marginalize the damage that LM based code 0:12:04.960 --> 0:12:07.520 can do by focusing on people actually give enough of 0:12:07.559 --> 0:12:10.600 a shit to buy the buy the actual machines. And 0:12:10.960 --> 0:12:13.480 ah No, I feel like an actual serious fiscal commitment 0:12:13.520 --> 0:12:16.280 makes companies take things a little bit more seriously, unless, 0:12:16.320 --> 0:12:19.040 of course, your Microsoft, Google and Amazon and meta buying GPUs, 0:12:19.040 --> 0:12:21.840 at which case you gives a shit. But I mean, 0:12:21.880 --> 0:12:25.280 if this takes off, it is it's curtains vazousha. It's 0:12:25.320 --> 0:12:27.800 it's a bad time for Open AI and Anthropic unless 0:12:27.840 --> 0:12:31.160 they too start doing these hardware driven things. This is 0:12:31.240 --> 0:12:33.240 just a guess. This is just a guess they would 0:12:33.240 --> 0:12:36.480 ever think about that. But I think it's possible. I 0:12:36.480 --> 0:12:39.160 think it's really the only economically viable end for this 0:12:39.520 --> 0:12:41.400 unless there are just going to be companies that lose 0:12:41.640 --> 0:12:44.240 billions of dollars a year for no apparent reason. I 0:12:44.320 --> 0:12:46.679 just don't think that's gonna happen. It's not enough money 0:12:46.679 --> 0:12:50.120 to do it. And goddamn, please go and listen to 0:12:50.200 --> 0:12:53.480 the Open AI will not get bailed out AI is 0:12:53.559 --> 0:12:56.240 not too big to Fail podcast They did. 0:12:56.320 --> 0:12:57.320 I'm tired of. 0:12:57.240 --> 0:12:59.959 Getting Please don't please stop emailing me. 0:13:00.240 --> 0:13:01.520 I love hearing from you. 0:13:01.559 --> 0:13:03.040 I really I would love to hear from you all 0:13:03.080 --> 0:13:06.360 more other than the people who email me saying I 0:13:06.400 --> 0:13:08.640 think the GPUs all for surveillance. 0:13:08.880 --> 0:13:11.000 I think it's all that. I think it's a big 0:13:11.040 --> 0:13:16.320 going s that's called Oracle. Oracle's first client was the CIA. 0:13:16.440 --> 0:13:20.080 Oracle has a giant government gpu array. They have several 0:13:20.080 --> 0:13:20.320 of them. 0:13:20.440 --> 0:13:23.600 Jesus Christ, that's already happening, and they're not doing surveillance 0:13:23.600 --> 0:13:25.000 on that. They did that with Maven, they did it 0:13:25.040 --> 0:13:28.080 with the promise software. Jesus fucking Christ. Sorry, I understand 0:13:28.120 --> 0:13:31.040 that I sound agitated here. It's just that if the 0:13:31.080 --> 0:13:34.360 only choice you can make when looking at something is 0:13:34.360 --> 0:13:36.920 to say, well, bad thing will definitely happen, that's not 0:13:37.360 --> 0:13:42.160 actually an intellectual pursuit. It's just crapping yourself. It's just saying, well, 0:13:42.160 --> 0:13:44.839 everything's bad, so everything stays bad. I actually want to 0:13:44.840 --> 0:13:46.880 bring you a message of hope. I don't think these 0:13:46.880 --> 0:13:49.480 companies get bailed down. I don't think they get anything. 0:13:49.880 --> 0:13:52.520 I think that they get allowed to die. I think 0:13:52.640 --> 0:13:55.920 partner assets might be scooped up by Amazon, Google and Microsoft. 0:13:56.080 --> 0:13:58.960 I think the government might take a chunk out of them. 0:13:59.160 --> 0:14:02.640 Perhaps they get some sort of bridge financing. But I 0:14:02.800 --> 0:14:05.040 actually don't know if the government could even extend that 0:14:05.200 --> 0:14:10.080 because AI is very unpopular with regular people. Regular people 0:14:10.200 --> 0:14:12.880 do not like this, and they hate data centers, and 0:14:12.920 --> 0:14:17.400 they have intelligently connected data centers, these horrible AI companies, 0:14:18.280 --> 0:14:20.120 and so in the end, I tell you please stop 0:14:20.640 --> 0:14:23.840 emailing me. The surveillance thing. Hey, it's not true. You're wrong. 0:14:24.240 --> 0:14:27.840 You sound you sound like someone that wants to be 0:14:27.920 --> 0:14:29.960 paranoid but doesn't want to do the reading well enough 0:14:29.960 --> 0:14:31.920 to be paranoid. If you want to be paranoid, go 0:14:31.920 --> 0:14:34.200 read the promised software. Go look at Oracle like Go 0:14:34.240 --> 0:14:36.160 look at the actual history of these companies if you 0:14:36.200 --> 0:14:37.960 want to be freaked out by something. 0:14:38.000 --> 0:14:40.760 But in truth, the AI industry is just. 0:14:40.720 --> 0:14:45.280 A direction the zeg Grigor of capital bumbles about and 0:14:45.720 --> 0:14:48.960 is only kept alive by a belief system that's mostly 0:14:49.040 --> 0:14:52.200 growth focused and doesn't really understand how to run a 0:14:52.200 --> 0:14:56.440 business anymore or even build innovation or technology. And it's 0:14:56.480 --> 0:14:59.680 blatantly obvious that lms are at best a classic risk 0:14:59.760 --> 0:15:03.040 versus reward software tool for the real sickos that know 0:15:03.080 --> 0:15:05.440 what they're getting into. But the problem is it's been 0:15:05.520 --> 0:15:09.040 sold as this mass market panacea for basically any problem. 0:15:09.480 --> 0:15:13.680 And I'll say that my thesis remains undeterred. I believe 0:15:13.680 --> 0:15:16.640 a great deal of AI usage is cargo cult shit 0:15:16.800 --> 0:15:19.320 driven by executives that don't do any work and people 0:15:19.320 --> 0:15:21.600 that feel pressured into adopting AI by their peers on 0:15:21.640 --> 0:15:24.920 the media, or of course, people that spend twelve to 0:15:25.120 --> 0:15:29.080 eighteen hours a day on Twitter who just follow the algorithm. 0:15:29.080 --> 0:15:32.080 I call it the Dunce's Casino because you just crank 0:15:32.200 --> 0:15:34.320 the thing and a bunch of insane people come out. 0:15:34.560 --> 0:15:36.720 And all of these people come out and they're like, if. 0:15:36.560 --> 0:15:38.920 You don't use agents, you're going to die. 0:15:38.960 --> 0:15:41.080 They're going to kill your wife, They're going to eat 0:15:41.120 --> 0:15:41.640 your dog. 0:15:41.800 --> 0:15:44.360 If you don't use loops, Boris Journey's going to blow 0:15:44.440 --> 0:15:46.320 up your house. Ah. 0:15:46.760 --> 0:15:50.480 Twitter is driving people insane using AI, but not because 0:15:50.520 --> 0:15:53.840 of groc or anything, but because it just fire hoses 0:15:54.040 --> 0:15:58.240 nutters into your feed who will constantly make you feel 0:15:58.320 --> 0:16:01.480 terrible for not picking up the next AI thing. It's 0:16:01.520 --> 0:16:05.320 this insane. It's I keep saying cargo cult because that's 0:16:05.360 --> 0:16:08.000 what it is. It's people just going around being like, yeah, 0:16:08.040 --> 0:16:09.240 you know, this is the big thing now. 0:16:09.360 --> 0:16:11.200 We all like that, right, we love it, We love it. 0:16:11.200 --> 0:16:11.600 It's good. 0:16:11.600 --> 0:16:13.800 If we all follow this, the great prophecy will come 0:16:13.800 --> 0:16:17.760 true and everyone then everyone will make money somewhere somehow. 0:16:17.880 --> 0:16:33.400 I think it's sickness. It's a genuine sickness. So I 0:16:33.400 --> 0:16:35.600 went a little Tucker cos and there, what's going on? 0:16:35.960 --> 0:16:37.080 Why are we using lms. 0:16:37.760 --> 0:16:39.360 What are they going to do to us? What are 0:16:39.360 --> 0:16:40.600 they going to do to your family? 0:16:41.280 --> 0:16:43.560 But anyway, back to the podcast, back to the monologue, 0:16:43.560 --> 0:16:47.200 I apologize for out libbing. I must be clear as well. 0:16:47.520 --> 0:16:49.800 I do love hearing from you because I keep getting 0:16:49.800 --> 0:16:52.000 emails from people that actually use AI, and I got 0:16:52.000 --> 0:16:54.600 one just before I started recording this from a bloke 0:16:54.640 --> 0:16:58.080 who said, and I quote, he's used it to deliver 0:16:58.080 --> 0:17:01.760 a good work and seeing it tried to submit absolute cramp. 0:17:02.160 --> 0:17:04.760 And that really is the AI industry. Hey, you know, 0:17:04.960 --> 0:17:07.440 sometime it works, and when it does, it's all right, 0:17:07.480 --> 0:17:11.040 and when it doesn't is dogshit. Yay, great, I'm so 0:17:11.160 --> 0:17:14.560 glad we sunk a trillion plus dollars into this. Look, 0:17:14.600 --> 0:17:17.640 these are specialist tools that are being forced to solve 0:17:17.720 --> 0:17:21.040 generalist problems, and they're probabilistic models that fail to do 0:17:21.119 --> 0:17:24.840 deterministic things, and they're really just an eternal financial black 0:17:24.840 --> 0:17:27.600 hole for anyone trying to scale them into a Google 0:17:27.680 --> 0:17:31.480 search or a meta sized business. And it's never going 0:17:31.520 --> 0:17:34.840 to work like that. Just before this, as well, I 0:17:34.880 --> 0:17:39.119 saw that Broadcom and open Ai announced their Jalipino chip 0:17:39.400 --> 0:17:42.720 is giant dinner place sized motherfucker. 0:17:42.359 --> 0:17:44.720 Dinner place, dinner plate. I'm not fixing that. 0:17:45.640 --> 0:17:49.200 And it's funny because everyone's doing the cargo culture around it. 0:17:49.240 --> 0:17:51.160 They're all like, oh, well, it's going to bring down 0:17:51.200 --> 0:17:53.880 inference costs. If they did that, that would fix everything. 0:17:54.119 --> 0:17:58.000 If Kalapino fixes everything, then everything will work out. There's 0:17:58.000 --> 0:18:01.280 no proof that that's the case. Actually, here's the thing. 0:18:01.440 --> 0:18:03.960 If you want if you want to like, here's a 0:18:04.000 --> 0:18:07.399 specific request if you have an information subscription. At some 0:18:07.520 --> 0:18:09.479 point in the last year, I know there was an 0:18:09.520 --> 0:18:13.600 information story that suggested that the Broadcom's open Ai chip 0:18:13.680 --> 0:18:15.720 would have modest gains. If you bring that to me, 0:18:16.119 --> 0:18:19.040 I will fuck it. I will venmo you five dollars 0:18:19.520 --> 0:18:22.760 if you can bring me that story. Seriously, I need 0:18:22.760 --> 0:18:27.080 that story because I'm pretty sure that if Broadcom's chip 0:18:27.160 --> 0:18:29.440 was going to change everything for open Ai, open Ai 0:18:29.520 --> 0:18:32.920 would sink every fucking dollar they had into it. Open 0:18:32.960 --> 0:18:37.280 Ai would dispense within video. Instead, it's like, oh, yeah, 0:18:37.320 --> 0:18:39.199 we're getting Microsoft to pay for some of it, and 0:18:39.200 --> 0:18:40.840 it'll be there in twenty twenty seven. 0:18:40.920 --> 0:18:42.680 Maybe. Uh oh no. 0:18:43.400 --> 0:18:45.960 Sounds like something you pre ordered on Kickstarter and you 0:18:46.040 --> 0:18:47.840 got it and it kind of sucks and it turns 0:18:47.840 --> 0:18:50.840 out the video is AI generated, which I did just 0:18:50.880 --> 0:18:53.879 see with this thing where someone had done a Kickstarter 0:18:53.920 --> 0:18:55.960 where it's a yeah, you can build things using real 0:18:56.000 --> 0:18:58.400 principles and you pour your own cement, but it's really 0:18:58.440 --> 0:19:02.400 small and the whole thing is AI generated. Fucking how 0:19:02.640 --> 0:19:04.320 this industry really is miserable? 0:19:04.359 --> 0:19:04.719 Isn't it? 0:19:05.160 --> 0:19:08.880 All the people every other era, other than of course 0:19:08.920 --> 0:19:11.280 crypto and the metaverse, when something cool happened. Even in 0:19:11.320 --> 0:19:15.360 the augmented reality era, there were still people doing goofy, 0:19:15.440 --> 0:19:18.960 funny shit. There were still people doing really interesting things 0:19:18.960 --> 0:19:21.840 with computer vision. There were people doing interesting things with 0:19:21.920 --> 0:19:24.360 depth sensing. The fact you can measure stuff on your 0:19:24.520 --> 0:19:26.400 iPhone I'm pretty sure you can do on Android too, 0:19:26.680 --> 0:19:29.200 is because of depth sensing cameras and augmented reality. 0:19:29.440 --> 0:19:30.040 It's cool. 0:19:30.440 --> 0:19:33.120 The snap spectacles are embarrassing, and if you get them, 0:19:33.200 --> 0:19:37.040 I assume that people will just come and beat you up. 0:19:37.240 --> 0:19:40.600