1 00:00:02,840 --> 00:00:06,640 Speaker 1: Media, Hello and welcomes a better offline. I'm, of course 2 00:00:06,680 --> 00:00:20,880 Speaker 1: your host ed Zitron. We're in the third episode of 3 00:00:20,880 --> 00:00:23,080 Speaker 1: our four part series where I give you a comprehensive 4 00:00:23,120 --> 00:00:25,919 Speaker 1: explanation as to the origins of the AI bubble, the 5 00:00:25,960 --> 00:00:29,000 Speaker 1: mythology sustaining it, and why it's destined to end really, 6 00:00:29,080 --> 00:00:32,840 Speaker 1: really badly. Now, if you're jumping in now, please start 7 00:00:32,880 --> 00:00:35,240 Speaker 1: from the very beginning. The reason why this is a 8 00:00:35,320 --> 00:00:37,839 Speaker 1: four part my first ever, is because I want it 9 00:00:37,880 --> 00:00:40,040 Speaker 1: to be comprehensive, and because this is a very big 10 00:00:40,080 --> 00:00:42,959 Speaker 1: subject with a lot of moving parts and even more bullshit. 11 00:00:43,600 --> 00:00:45,879 Speaker 1: A few weeks ago, I published a premium newsletter that 12 00:00:45,920 --> 00:00:48,640 Speaker 1: explained how everybody is losing money on generative AI, in 13 00:00:48,680 --> 00:00:51,320 Speaker 1: part because the costs of running AI models is increasing, 14 00:00:51,520 --> 00:00:53,680 Speaker 1: and in part because the software itself doesn't do enough 15 00:00:53,680 --> 00:00:55,880 Speaker 1: to warrant the costs associated with running them, which are 16 00:00:55,880 --> 00:00:59,680 Speaker 1: already subsidized and unprofitable for the model providers. Outside of 17 00:00:59,760 --> 00:01:03,160 Speaker 1: open and to a lesser extent, Anthropic, nobody seems to 18 00:01:03,160 --> 00:01:05,800 Speaker 1: be making much revenue, with the most successful company being 19 00:01:05,840 --> 00:01:08,800 Speaker 1: any Sphere, makers of AI coding tool Cursor, which hid 20 00:01:08,800 --> 00:01:11,800 Speaker 1: five hundred million dollars have annualized so forty one point 21 00:01:12,280 --> 00:01:14,679 Speaker 1: six million in one month a few months ago, just 22 00:01:14,720 --> 00:01:17,319 Speaker 1: before Anthropic and open ai jacked up the prices for 23 00:01:17,400 --> 00:01:21,160 Speaker 1: priority processing on enterprise queries, raising their operating costs as 24 00:01:21,160 --> 00:01:24,280 Speaker 1: a result. In any case, that's some pissport revenue for 25 00:01:24,319 --> 00:01:26,440 Speaker 1: an industry that's meant to be the future of software. 26 00:01:26,800 --> 00:01:29,280 Speaker 1: Smart Watchers are projected to make thirty two billion dollars 27 00:01:29,280 --> 00:01:32,000 Speaker 1: this year, and as I've mentioned in the past, the 28 00:01:32,040 --> 00:01:34,600 Speaker 1: Magnificent Seven expect to make thirty five billion dollars or 29 00:01:34,600 --> 00:01:36,399 Speaker 1: so in revenue from AI this year, and I think 30 00:01:36,440 --> 00:01:38,800 Speaker 1: in total, when you're throw in core even all them, 31 00:01:39,120 --> 00:01:42,560 Speaker 1: it's barely fifty five billion dollars in total. Even Anthropic 32 00:01:42,600 --> 00:01:45,240 Speaker 1: and open Ai seem a little lethargic, both burning billions 33 00:01:45,240 --> 00:01:47,560 Speaker 1: of dollars while making by my estimates, no more than 34 00:01:47,560 --> 00:01:50,360 Speaker 1: two billion dollars in Anthropics case this year so far 35 00:01:50,600 --> 00:01:53,480 Speaker 1: and six point six two six billion dollars in twenty 36 00:01:53,520 --> 00:01:56,440 Speaker 1: twenty five so far for open Ai, despite projections of 37 00:01:56,520 --> 00:02:00,240 Speaker 1: five billion dollars and thirteen billion dollars respectively. Outside of 38 00:02:00,320 --> 00:02:03,240 Speaker 1: these two AI startups are floundering, struggling to stay alive 39 00:02:03,320 --> 00:02:06,000 Speaker 1: and raising money in several hundred million dollar versus their 40 00:02:06,040 --> 00:02:09,919 Speaker 1: negative gross margin businesses flounder as they dug into. A 41 00:02:09,960 --> 00:02:12,520 Speaker 1: few months ago, I could find only twelve AI powered 42 00:02:12,560 --> 00:02:15,040 Speaker 1: companies making more than eight point three million dollars a month, 43 00:02:15,080 --> 00:02:17,840 Speaker 1: with two of them slightly improving their revenue, specifically AI 44 00:02:17,880 --> 00:02:20,120 Speaker 1: search company perplexd, which is now here one hundred and 45 00:02:20,160 --> 00:02:23,040 Speaker 1: fifty million dollars an ur in or twelve point five 46 00:02:23,080 --> 00:02:26,119 Speaker 1: million dollars a month, and AI coding startup Replayer, which 47 00:02:26,120 --> 00:02:29,280 Speaker 1: has hit the same amount. Both of these companies burn 48 00:02:29,400 --> 00:02:32,720 Speaker 1: ridiculous amounts of money. Paplexd burned one hundred and sixty 49 00:02:32,720 --> 00:02:35,400 Speaker 1: four percent of its revenue on Amazon web services, open 50 00:02:35,400 --> 00:02:38,440 Speaker 1: Ai and Anthropic last year, and while replet hasn't leaked 51 00:02:38,440 --> 00:02:41,720 Speaker 1: its costs, the information reports its gross margins in July 52 00:02:41,800 --> 00:02:44,320 Speaker 1: but twenty three percent, which doesn't include the cost of 53 00:02:44,320 --> 00:02:47,200 Speaker 1: its free users, which you simply have to do with llms, 54 00:02:47,360 --> 00:02:49,440 Speaker 1: as free users are capable of costing you a shit 55 00:02:49,480 --> 00:02:51,600 Speaker 1: ton of money. And some of you might say that's 56 00:02:51,600 --> 00:02:54,040 Speaker 1: how they do it in software, Well, guess what software 57 00:02:54,040 --> 00:02:56,160 Speaker 1: doesn't usually connect you to a model that can burn 58 00:02:56,360 --> 00:02:59,520 Speaker 1: I don't know ten cents twenty cents every time they 59 00:02:59,560 --> 00:03:01,880 Speaker 1: touch it, which may not seem like much, but when 60 00:03:01,919 --> 00:03:04,760 Speaker 1: you're making three dollars on someone and they don't convert, 61 00:03:04,840 --> 00:03:08,880 Speaker 1: it does problematically. Your paid users also cost you more 62 00:03:08,880 --> 00:03:11,440 Speaker 1: than they bring in as well. In fact, every user 63 00:03:11,639 --> 00:03:14,080 Speaker 1: loses you money in Generative AI because it's impossible to 64 00:03:14,080 --> 00:03:17,280 Speaker 1: do cost control in a consistent manner. A few months ago, 65 00:03:17,320 --> 00:03:19,440 Speaker 1: I did a piece of Anthropic losing money on every 66 00:03:19,440 --> 00:03:21,880 Speaker 1: single claud code subscriber. And now I'm going to walk 67 00:03:21,919 --> 00:03:24,760 Speaker 1: you through the whole story in a simplified fashion because 68 00:03:24,800 --> 00:03:28,600 Speaker 1: it's quite important. So claud Code is a coding environment 69 00:03:28,639 --> 00:03:32,680 Speaker 1: that people use used, or I should really say, try 70 00:03:32,720 --> 00:03:36,480 Speaker 1: to use to build software using generative AI. It's available 71 00:03:36,520 --> 00:03:39,360 Speaker 1: as part of Anthropics twenty dollars, one hundred dollars and 72 00:03:39,400 --> 00:03:41,920 Speaker 1: two hundred dollars a month claud subscriptions, with the more 73 00:03:41,920 --> 00:03:46,520 Speaker 1: expensive subscriptions having more generous rate limits. Generally, these subscriptions 74 00:03:46,520 --> 00:03:47,920 Speaker 1: are all you can eat. You can use them as 75 00:03:47,960 --> 00:03:49,920 Speaker 1: much as you want until you hit limits, rather than 76 00:03:49,960 --> 00:03:52,400 Speaker 1: paying for the actual tokens you burn. When I say 77 00:03:52,480 --> 00:03:55,680 Speaker 1: burn tokens and someone reached out saying I should specify this, 78 00:03:56,040 --> 00:03:59,960 Speaker 1: I'm describing how these models are traditionally built. In general, 79 00:04:00,120 --> 00:04:03,360 Speaker 1: you'll builded a dollar per million input tokens as in 80 00:04:03,600 --> 00:04:06,520 Speaker 1: user feeding in data and output tokens the output created, 81 00:04:06,920 --> 00:04:09,640 Speaker 1: so you wouldn't get one token built, so every million 82 00:04:09,640 --> 00:04:12,520 Speaker 1: you get charged. So, for example, Anthropic charges three dollars 83 00:04:12,560 --> 00:04:16,880 Speaker 1: per million input tokens and six million output tokens to 84 00:04:17,000 --> 00:04:19,440 Speaker 1: use its clauds on it for model, and it's about 85 00:04:19,480 --> 00:04:23,400 Speaker 1: I think, well, a word before tokens should really look 86 00:04:23,440 --> 00:04:26,719 Speaker 1: that up. It's it also gets more complex as you 87 00:04:26,760 --> 00:04:29,800 Speaker 1: get into things like generating code. Nevertheless, claud code has 88 00:04:29,839 --> 00:04:33,120 Speaker 1: been quite popular, and a user created a program called 89 00:04:33,160 --> 00:04:36,160 Speaker 1: cc usage which allowed you to see your token burn 90 00:04:36,400 --> 00:04:39,200 Speaker 1: the amount of tokens you were using. You were actually 91 00:04:39,279 --> 00:04:43,440 Speaker 1: burning using Anthropics models while using clawed code versus just 92 00:04:43,640 --> 00:04:45,640 Speaker 1: getting charged a month and not knowing, and many were 93 00:04:45,640 --> 00:04:47,440 Speaker 1: seeing that they were burning in the excess of their 94 00:04:47,480 --> 00:04:50,200 Speaker 1: monthly spend. To be clear, this is the token price 95 00:04:50,240 --> 00:04:52,480 Speaker 1: based on anthropics own pricing, and thus the cost of 96 00:04:52,520 --> 00:04:55,080 Speaker 1: Anthropic are likely not identical. So I got a little 97 00:04:55,120 --> 00:04:59,000 Speaker 1: clever using anthropics gross profit margins, I chose fifty five percent, 98 00:04:59,160 --> 00:05:01,480 Speaker 1: and then a few weeks solved my article sixty percent 99 00:05:01,560 --> 00:05:03,680 Speaker 1: was leaked. I found at least twenty different accounts of 100 00:05:03,720 --> 00:05:06,400 Speaker 1: people costing Anthropic anywhere from one hundred and thirty percent 101 00:05:06,600 --> 00:05:09,320 Speaker 1: to three thousand and eighty four percent of their subscription. 102 00:05:09,960 --> 00:05:13,240 Speaker 1: There is also now a leader board called vibrank, where 103 00:05:13,240 --> 00:05:16,039 Speaker 1: people compete to see how much they burn with the 104 00:05:16,080 --> 00:05:18,840 Speaker 1: current leader burning and I sheit you not fifty two 105 00:05:18,920 --> 00:05:20,880 Speaker 1: hundred and ninety one dollars of the course of a month. 106 00:05:22,120 --> 00:05:25,400 Speaker 1: Anthropic is, to be clear, the second largest model developer 107 00:05:25,520 --> 00:05:27,760 Speaker 1: and has some of the best AI talent in the industry. 108 00:05:27,960 --> 00:05:30,240 Speaker 1: It has a better handle on its infrastructure than anyone 109 00:05:30,240 --> 00:05:32,760 Speaker 1: outside of big tech and open AI, and it still 110 00:05:32,800 --> 00:05:35,640 Speaker 1: cannot seem to fix this problem even with weekly rate 111 00:05:35,680 --> 00:05:38,360 Speaker 1: limits brought in at the end of August. While one 112 00:05:38,400 --> 00:05:41,400 Speaker 1: could assume that Anthropic is simply letting users run wild, 113 00:05:41,440 --> 00:05:44,360 Speaker 1: my theory is far simpler. Even the model developers have 114 00:05:44,520 --> 00:05:47,520 Speaker 1: no real way of limiting user activity, likely due to 115 00:05:47,520 --> 00:05:50,839 Speaker 1: the architecture of generative AI. I know it sounds insane, 116 00:05:50,839 --> 00:05:54,320 Speaker 1: but at the most advanced level. Even there, modeled providers 117 00:05:54,360 --> 00:05:57,560 Speaker 1: are still prompting their models, and whatever rate limits may 118 00:05:57,839 --> 00:06:00,719 Speaker 1: be in place appear to at times get completely ignored, 119 00:06:00,920 --> 00:06:02,479 Speaker 1: and there doesn't seem to be anything they can do 120 00:06:02,520 --> 00:06:05,839 Speaker 1: to stop it now. Really, Anthropic counts amongst its capitalist 121 00:06:05,920 --> 00:06:09,480 Speaker 1: apex predators one lone Chinese man who spent fifty thousand 122 00:06:09,480 --> 00:06:11,400 Speaker 1: dollars to their compute in the space of a month 123 00:06:11,480 --> 00:06:15,320 Speaker 1: fucking around with glord code. Even if Anthropic was profitable, 124 00:06:15,520 --> 00:06:17,719 Speaker 1: it isn't, and we'll burn billions of dollars this year. 125 00:06:17,920 --> 00:06:20,240 Speaker 1: A customer paying two hundred dollars a month ran up 126 00:06:20,279 --> 00:06:24,600 Speaker 1: fifty thousand dollars in costs, immediately devouring the margin of 127 00:06:24,640 --> 00:06:27,120 Speaker 1: any user running the service that day, that week, or 128 00:06:27,120 --> 00:06:30,560 Speaker 1: even that month. Even if Anthropics costs are half the 129 00:06:30,560 --> 00:06:33,400 Speaker 1: published rates, they're not. By the way, one guy amounted 130 00:06:33,400 --> 00:06:35,159 Speaker 1: to one hundred and twenty five US is worth of 131 00:06:35,200 --> 00:06:38,720 Speaker 1: monthly revenue. This is not a real business. That's a 132 00:06:38,760 --> 00:06:41,479 Speaker 1: bad business without of control costs, and it doesn't appear 133 00:06:41,600 --> 00:06:45,160 Speaker 1: anybody has these costs under control and face with the 134 00:06:45,200 --> 00:06:47,880 Speaker 1: grim reality ahead of them, these companies are trying nasty 135 00:06:47,920 --> 00:06:50,599 Speaker 1: little tricks on their customers to douce more revenue from them. 136 00:06:51,279 --> 00:06:54,159 Speaker 1: A few weeks ago, Replet, an unprofitable AI coding company, 137 00:06:54,200 --> 00:06:56,880 Speaker 1: released a product called Agent three, which promised to be 138 00:06:56,960 --> 00:07:00,560 Speaker 1: ten times more autonomous and offer infinitely more possible abilities, 139 00:07:00,760 --> 00:07:04,040 Speaker 1: testing and fixing its code, constantly improving your application behind 140 00:07:04,080 --> 00:07:08,760 Speaker 1: the scenes in a reflection loop. Sounds very real, sounds 141 00:07:08,800 --> 00:07:12,600 Speaker 1: extremely real, It's so real, but actually it isn't. In reality. 142 00:07:12,920 --> 00:07:15,040 Speaker 1: This means you go and tell the model to build something, 143 00:07:15,040 --> 00:07:16,520 Speaker 1: and it would go and do it, and you'll be 144 00:07:16,520 --> 00:07:18,640 Speaker 1: shocked to hear that these models can't be relied upon 145 00:07:18,680 --> 00:07:21,400 Speaker 1: to go and do anything. Please note that this was 146 00:07:21,480 --> 00:07:24,160 Speaker 1: launched a few months after Replet raise their prices, shifting 147 00:07:24,320 --> 00:07:27,240 Speaker 1: to obfiscated effort based pricing that would charge the full 148 00:07:27,280 --> 00:07:29,320 Speaker 1: scope of the agent's work. And if you're wondering what 149 00:07:29,400 --> 00:07:32,440 Speaker 1: the fuck that means, so are their customers. Agent three 150 00:07:32,440 --> 00:07:35,600 Speaker 1: has been a disaster. Users found the tasks that previously 151 00:07:35,640 --> 00:07:38,280 Speaker 1: cost a few dollars were spiraling into the hundreds of dollars, 152 00:07:38,480 --> 00:07:41,200 Speaker 1: with the register reporting one customer found themselves within one 153 00:07:41,240 --> 00:07:43,440 Speaker 1: thousand dollars bill after a week, and I quote them, 154 00:07:44,040 --> 00:07:46,880 Speaker 1: I think it's just launch pricing adjustment. Some tasks on 155 00:07:46,960 --> 00:07:49,040 Speaker 1: new apps ran over an hour and forty five minutes 156 00:07:49,040 --> 00:07:51,760 Speaker 1: and only charged four to six dollars, but editing pre 157 00:07:51,840 --> 00:07:55,120 Speaker 1: existing apps seems to cost most overall. I spend one 158 00:07:55,320 --> 00:07:58,240 Speaker 1: K this week alone, and they told that to the register. 159 00:07:58,320 --> 00:08:01,440 Speaker 1: By the way, another user comp that costs skyrocket without 160 00:08:01,440 --> 00:08:04,280 Speaker 1: any concrete results, and they quote the register here. I 161 00:08:04,320 --> 00:08:06,480 Speaker 1: typically spent between one hundred dollars and two hundred and 162 00:08:06,480 --> 00:08:08,640 Speaker 1: fifty dollars a month. I blew through seventy dollars in 163 00:08:08,640 --> 00:08:11,280 Speaker 1: a night at Agent three launch, and another redditor wrote 164 00:08:11,320 --> 00:08:14,800 Speaker 1: alleging the new tool also performed some questionable actions. One 165 00:08:14,840 --> 00:08:17,760 Speaker 1: prompt brute forced its way through authentication, redoing auth and 166 00:08:17,800 --> 00:08:20,520 Speaker 1: hard resetting users password to what it wanted to perform 167 00:08:20,640 --> 00:08:23,160 Speaker 1: app testing on a form. The user wrote, I realized 168 00:08:23,200 --> 00:08:25,960 Speaker 1: that's a little nonsensical, but long story short, it did 169 00:08:25,960 --> 00:08:28,120 Speaker 1: a bunch of shit. It wasn't asked to. As I 170 00:08:28,160 --> 00:08:31,720 Speaker 1: previously reported, in late May early June, both open ai 171 00:08:31,800 --> 00:08:34,280 Speaker 1: and Anthropic cranked up the pricing on their enterprise customers, 172 00:08:34,360 --> 00:08:37,720 Speaker 1: leading Replet and Cursor both shifting their prices upward. This 173 00:08:37,800 --> 00:08:40,800 Speaker 1: abuse is now trickled down to the customers. Report has 174 00:08:40,800 --> 00:08:43,520 Speaker 1: now released an update. Unless you choose how autonomous you 175 00:08:43,520 --> 00:08:45,840 Speaker 1: want Agent three to be, which is a tacit admission 176 00:08:45,840 --> 00:08:48,880 Speaker 1: that you can't trust coding elms to build software replets. 177 00:08:48,960 --> 00:08:51,360 Speaker 1: Users are still pissed off, complaining that report is charging 178 00:08:51,360 --> 00:08:53,720 Speaker 1: them for an activity when the agent doesn't do anything, 179 00:08:53,960 --> 00:08:57,200 Speaker 1: a consistent problem I've found across redditors. While Reddit is 180 00:08:57,200 --> 00:09:00,000 Speaker 1: not the full summation of all users of every company everywhere, 181 00:09:00,160 --> 00:09:02,880 Speaker 1: it's a fairly good barometer of user sentiment and man 182 00:09:02,880 --> 00:09:08,079 Speaker 1: a user's piss and now here's why this is bad. Traditionally, 183 00:09:08,120 --> 00:09:10,920 Speaker 1: Silicon Valley startups have relied upon the same model, have 184 00:09:11,040 --> 00:09:13,120 Speaker 1: grow really fast and burn a bunch of money, then 185 00:09:13,160 --> 00:09:16,600 Speaker 1: turn the profit lever. AI does not have a profit 186 00:09:16,679 --> 00:09:19,200 Speaker 1: lever because the raw costs of providing access to AI 187 00:09:19,320 --> 00:09:22,280 Speaker 1: models are so high and they're only increasing that the 188 00:09:22,280 --> 00:09:25,080 Speaker 1: basic economics of how the tech industry sell software don't 189 00:09:25,080 --> 00:09:38,600 Speaker 1: make sense. I'll reiterate something I wrote a few weeks ago. 190 00:09:39,280 --> 00:09:43,240 Speaker 1: A large language model users infrastructural burden varies wildly between 191 00:09:43,320 --> 00:09:46,679 Speaker 1: users and use cases. While somebody asking chat gpt to 192 00:09:46,720 --> 00:09:48,880 Speaker 1: summarize an email might not be much of a burden, 193 00:09:49,120 --> 00:09:52,240 Speaker 1: somebody asking chat gpt to review hundreds of pages of 194 00:09:52,240 --> 00:09:54,800 Speaker 1: documents at once. A core feature of basically any twenty 195 00:09:54,800 --> 00:09:57,520 Speaker 1: dollars a month subscription could eat up to eight GPUs 196 00:09:57,559 --> 00:09:59,480 Speaker 1: at once. To be very clear, a user that pays 197 00:09:59,520 --> 00:10:01,720 Speaker 1: twenty dollars a month could run multiple queries like this 198 00:10:01,760 --> 00:10:04,800 Speaker 1: a month and there's not really a way to stop them. 199 00:10:04,960 --> 00:10:08,440 Speaker 1: Unlike most software products, any errors in producing an output 200 00:10:08,440 --> 00:10:11,360 Speaker 1: from a large language model have a significant opportunity cost. 201 00:10:11,640 --> 00:10:13,520 Speaker 1: When a user doesn't like an output, or the model 202 00:10:13,520 --> 00:10:16,160 Speaker 1: gets something wrong which it's guaranteed to do, or the 203 00:10:16,240 --> 00:10:19,160 Speaker 1: user realizes they forgot something, the model must make a 204 00:10:19,200 --> 00:10:22,800 Speaker 1: further generation or generations, and even with caching which anthropic 205 00:10:22,880 --> 00:10:25,400 Speaker 1: is added are told to there's a definitive cost attached 206 00:10:25,400 --> 00:10:28,679 Speaker 1: to any mistake. Large language models are for the most 207 00:10:28,720 --> 00:10:31,720 Speaker 1: part lacking in any definitive use cases, meaning that every 208 00:10:31,800 --> 00:10:33,719 Speaker 1: user is even with an idea of what they want 209 00:10:33,760 --> 00:10:37,040 Speaker 1: to do, experimenting with every input and output. In doing so, 210 00:10:37,080 --> 00:10:39,480 Speaker 1: they create the opportunity to burn more tokens, which in 211 00:10:39,480 --> 00:10:42,120 Speaker 1: turn creates an infrastructural burn on GPUs, which cost a 212 00:10:42,120 --> 00:10:44,800 Speaker 1: lot of money to run. The more specific the output, 213 00:10:44,880 --> 00:10:47,240 Speaker 1: the more opportunities there are of a monstrous token burn. 214 00:10:47,280 --> 00:10:50,440 Speaker 1: And I'm specifically thinking about coding with l elms. The 215 00:10:50,480 --> 00:10:53,600 Speaker 1: token heavy nature of generating code means that any mistakes, 216 00:10:53,640 --> 00:10:57,600 Speaker 1: suboptimal generations, or straight up errors will guarantee further token burn. 217 00:10:58,280 --> 00:11:01,319 Speaker 1: Even efforts to reduce compute cors by, for example, pushing 218 00:11:01,360 --> 00:11:03,679 Speaker 1: free users or those on cheap plans, the small or 219 00:11:03,760 --> 00:11:07,440 Speaker 1: less intensive models have dubious efficacy. As I talked about 220 00:11:07,480 --> 00:11:09,679 Speaker 1: in a previous episode, open ai split a model in 221 00:11:09,720 --> 00:11:12,959 Speaker 1: the GPT version of CHET. GPT requires vast amounts of 222 00:11:13,000 --> 00:11:15,760 Speaker 1: additional compute in order to route the user's request or 223 00:11:15,880 --> 00:11:18,960 Speaker 1: the appropriate model, with simpler requests going to smaller models 224 00:11:19,000 --> 00:11:21,760 Speaker 1: and more complex ones being shifted to reasoning models, and 225 00:11:21,800 --> 00:11:23,840 Speaker 1: it makes it impossible to cash part of the input. 226 00:11:23,880 --> 00:11:26,680 Speaker 1: As a result, it's not really clear whether it's saving 227 00:11:26,679 --> 00:11:29,120 Speaker 1: open ai any money, and indeed, kind I suggest it 228 00:11:29,200 --> 00:11:32,200 Speaker 1: might be costing them more. In simpler terms, it's very, 229 00:11:32,280 --> 00:11:34,920 Speaker 1: very very difficult to imagine what one user free or 230 00:11:34,960 --> 00:11:37,480 Speaker 1: otherwise might cost, and thus it's hard to charge them 231 00:11:37,679 --> 00:11:39,840 Speaker 1: anything on a monthly basis or tell them what a 232 00:11:39,840 --> 00:11:42,720 Speaker 1: service might actually cost them on average. And this is 233 00:11:42,760 --> 00:11:47,200 Speaker 1: a huge, huge problem with AI coding environments. But let's 234 00:11:47,200 --> 00:11:50,640 Speaker 1: talk about claud Code again. Anthropics code generate a tool. 235 00:11:50,760 --> 00:11:53,480 Speaker 1: According to the information claud code was driving nearly four 236 00:11:53,559 --> 00:11:56,719 Speaker 1: hundred million dollars in annualized revenue, roughly doubling from a 237 00:11:56,720 --> 00:11:59,360 Speaker 1: few weeks ago on July thirty first, twenty twenty five. 238 00:12:00,080 --> 00:12:02,880 Speaker 1: The annualized revenue works out to about thirty three million 239 00:12:02,920 --> 00:12:05,280 Speaker 1: dollars a month in revenue for a company that predicts 240 00:12:05,280 --> 00:12:07,679 Speaker 1: it will make at least four hundred and sixteen million 241 00:12:07,720 --> 00:12:09,280 Speaker 1: dollars a month by the end of the year, and 242 00:12:09,320 --> 00:12:11,840 Speaker 1: for a product that has become for a time the 243 00:12:11,880 --> 00:12:14,280 Speaker 1: most popular coding environment in the world from the second 244 00:12:14,360 --> 00:12:17,680 Speaker 1: largest and best funded AI company in the world. Is 245 00:12:17,720 --> 00:12:20,760 Speaker 1: that it is that fucking it is that all that's 246 00:12:20,760 --> 00:12:23,960 Speaker 1: happening here thirty three million dollars, all of which is 247 00:12:24,000 --> 00:12:27,800 Speaker 1: unprofitable after it felt, at least based on social media 248 00:12:27,920 --> 00:12:30,840 Speaker 1: chatter and discussing with multiple different engineers, that claud code 249 00:12:30,840 --> 00:12:33,760 Speaker 1: have become ubiquitous with anything to do with LLLMS and coding. 250 00:12:34,720 --> 00:12:37,280 Speaker 1: To be clear, Anthropics, so on It and Opus models 251 00:12:37,320 --> 00:12:39,560 Speaker 1: are consistently some of the most popular for programming an 252 00:12:39,600 --> 00:12:42,720 Speaker 1: open router, an aggregator of LM usage, and Anthropic has 253 00:12:42,760 --> 00:12:45,120 Speaker 1: been consistently named as the best at coding. Whether or 254 00:12:45,120 --> 00:12:47,959 Speaker 1: not I feel that way is irrelevant. Some bright spark 255 00:12:48,000 --> 00:12:49,720 Speaker 1: out there is going to send it. Microsoft's get hub 256 00:12:49,760 --> 00:12:52,320 Speaker 1: copilot at one point eight million paying subscribers, and guess 257 00:12:52,320 --> 00:12:55,360 Speaker 1: what that's true? In fact, I reported it. Here's another 258 00:12:55,440 --> 00:12:58,160 Speaker 1: fun fact. The Wall Street Journal report that Microsoft loses 259 00:12:58,200 --> 00:13:00,440 Speaker 1: on average twenty dollars a month per use, with some 260 00:13:00,520 --> 00:13:03,120 Speaker 1: users costing the company as much as eight bucks. And 261 00:13:03,160 --> 00:13:06,600 Speaker 1: that's for the most popular product. But wait, wait, wait, wait, 262 00:13:07,320 --> 00:13:11,479 Speaker 1: hold up, wait, I read some shit in the newspaper. 263 00:13:11,800 --> 00:13:15,480 Speaker 1: Aren't these LLLM code generators replacing actual human engineers? And thus, 264 00:13:15,640 --> 00:13:17,480 Speaker 1: even if they cost way more than twenty dollars one 265 00:13:17,520 --> 00:13:19,440 Speaker 1: hundred dollars or two hundred dollars a month, they're still 266 00:13:19,440 --> 00:13:22,800 Speaker 1: worth it. Right, They're replacing an entire engineer. Oh my 267 00:13:22,880 --> 00:13:25,520 Speaker 1: sweet summer child. If you believe the New York Times 268 00:13:25,600 --> 00:13:28,240 Speaker 1: or other outlets that simply copy and paste whatever anthropic 269 00:13:28,280 --> 00:13:31,079 Speaker 1: CEO Warrio Ama Day says, you'd think that the reason 270 00:13:31,080 --> 00:13:33,200 Speaker 1: that software engineers are having trouble finding work is because 271 00:13:33,200 --> 00:13:37,120 Speaker 1: their jobs are being replaced by AI. This grotesque, manipulative, abusive, 272 00:13:37,120 --> 00:13:40,000 Speaker 1: and offensive lie has been propagated through the entire business 273 00:13:40,080 --> 00:13:42,360 Speaker 1: and tech media without anybody sitting down and asking whether 274 00:13:42,360 --> 00:13:44,560 Speaker 1: it's true, or even getting a good understanding of what 275 00:13:44,600 --> 00:13:47,880 Speaker 1: it is that elms can actually do with code. Members 276 00:13:47,880 --> 00:13:51,440 Speaker 1: of the media, I am begging you stop stop doing this, 277 00:13:51,559 --> 00:13:56,480 Speaker 1: Stop publishing these fucking headlines. You're embarrassing yourself. Every asshole 278 00:13:56,559 --> 00:13:58,400 Speaker 1: is willing to give a quote saying that coding is 279 00:13:58,440 --> 00:14:00,600 Speaker 1: dead and that every execut if he is willing to 280 00:14:00,600 --> 00:14:03,160 Speaker 1: burp out some nonsense about replacing all of their engineers. 281 00:14:03,200 --> 00:14:05,199 Speaker 1: But I'm fucking begging you to either use these things 282 00:14:05,280 --> 00:14:08,040 Speaker 1: yourself or speak to people that do. I am not 283 00:14:08,080 --> 00:14:10,800 Speaker 1: a coder. I cannot write or read code. Nevertheless, I'm 284 00:14:10,840 --> 00:14:13,520 Speaker 1: capable of learning, and I've spoken to numerous software engineers 285 00:14:13,520 --> 00:14:15,880 Speaker 1: in the last few months, and basically I've reached a 286 00:14:15,920 --> 00:14:20,480 Speaker 1: consensus of this is kind of useful sometimes. However, one time, 287 00:14:20,520 --> 00:14:24,440 Speaker 1: a very silly man with an increasingly squeaky voice said 288 00:14:24,440 --> 00:14:26,600 Speaker 1: that I don't speak to people who use AI tools. 289 00:14:26,600 --> 00:14:29,440 Speaker 1: So I went and spoke to three notable experienced software 290 00:14:29,480 --> 00:14:31,600 Speaker 1: engineers and ask them to give me the straight truth 291 00:14:31,640 --> 00:14:34,560 Speaker 1: about what coding lllms can do. Now, for the purposes 292 00:14:34,600 --> 00:14:36,400 Speaker 1: of brevity, I'm going to use select quotes from what 293 00:14:36,440 --> 00:14:37,920 Speaker 1: these people said. But if you want to read the 294 00:14:37,920 --> 00:14:40,560 Speaker 1: whole thing, you can check out the newsletter first. I'm 295 00:14:40,560 --> 00:14:42,160 Speaker 1: going to read what Carl Brown of the Internet of 296 00:14:42,200 --> 00:14:44,160 Speaker 1: Bugs said, and I had him on the show a 297 00:14:44,200 --> 00:14:48,040 Speaker 1: few months back. He's fantastic. So most of the advancements 298 00:14:48,040 --> 00:14:50,760 Speaker 1: in programming languages, technique and craft in the last four 299 00:14:50,840 --> 00:14:53,080 Speaker 1: years have been designing safer and better ways of tying 300 00:14:53,080 --> 00:14:56,240 Speaker 1: these blocks together to create large and larger programs with 301 00:14:56,320 --> 00:15:00,000 Speaker 1: more complexity and functionality. Humans use these advancements to arrange 302 00:15:00,120 --> 00:15:02,720 Speaker 1: these blocks in logical abstraction layers so we can fit 303 00:15:02,720 --> 00:15:05,160 Speaker 1: an understanding of the lairs interconnections in our heads as 304 00:15:05,160 --> 00:15:08,640 Speaker 1: we work. Diving into blocks temporarily is needed. This is 305 00:15:08,680 --> 00:15:11,360 Speaker 1: where AIS fall down. The amount of context required to 306 00:15:11,400 --> 00:15:14,480 Speaker 1: hold the interconnections between these blocks quickly grows beyond the 307 00:15:14,480 --> 00:15:17,760 Speaker 1: AI's effective short term memory, in practice much smaller than 308 00:15:17,760 --> 00:15:21,000 Speaker 1: its advertised context windows size, and the AIS like the 309 00:15:21,040 --> 00:15:23,880 Speaker 1: ability to reason about the abstractions as we do. This 310 00:15:24,000 --> 00:15:27,840 Speaker 1: leads to real world code that's illogically layed, hard to understand, debug, 311 00:15:27,880 --> 00:15:32,440 Speaker 1: and maintain. Carl also said code generation AIS, from an 312 00:15:32,480 --> 00:15:35,600 Speaker 1: industry standpoint, are roughly the equivalent of a slightly below 313 00:15:35,600 --> 00:15:38,640 Speaker 1: average computer science graduate fresh out of school without any 314 00:15:38,680 --> 00:15:41,600 Speaker 1: real world experience, only ever having written programs to be 315 00:15:41,600 --> 00:15:44,480 Speaker 1: printed and graded. That's bad because, as he pointed out, 316 00:15:44,520 --> 00:15:47,280 Speaker 1: whereas llms can't get past this summer, in turn stage, 317 00:15:47,320 --> 00:15:50,320 Speaker 1: actual humans get better, and if we're replacing the bottom 318 00:15:50,360 --> 00:15:52,160 Speaker 1: rung of the labor market, there won't be any mid 319 00:15:52,240 --> 00:15:55,080 Speaker 1: level or senior developers later down the line. Next, I 320 00:15:55,120 --> 00:15:57,560 Speaker 1: asked Nick Sharesh of I will fucking pile drive you 321 00:15:57,640 --> 00:16:01,240 Speaker 1: if you mention AI again what he thought. Llms, he said, 322 00:16:01,280 --> 00:16:03,600 Speaker 1: will sometimes solve a thorny problem for me in a 323 00:16:03,640 --> 00:16:06,320 Speaker 1: few seconds, saving me some brain power. But in practice, 324 00:16:06,320 --> 00:16:08,960 Speaker 1: the effort of articulating so much of the design work 325 00:16:08,960 --> 00:16:11,560 Speaker 1: in plain English and hoping the LM emits code that 326 00:16:11,600 --> 00:16:15,120 Speaker 1: I find acceptable is frequently more work than just writing 327 00:16:15,160 --> 00:16:18,480 Speaker 1: the code. For most problems, the hardest part is the thinking, 328 00:16:18,640 --> 00:16:21,560 Speaker 1: and lllms don't make it that part any easier. I 329 00:16:21,600 --> 00:16:24,440 Speaker 1: also talked to Colvogi of no AI is not making 330 00:16:24,480 --> 00:16:27,680 Speaker 1: AI engineers ten X is productive. We also had in 331 00:16:27,720 --> 00:16:30,760 Speaker 1: the show recently, and he said this, llms often function 332 00:16:30,920 --> 00:16:32,680 Speaker 1: like a fresh summer intern. They're good at solving the 333 00:16:32,680 --> 00:16:35,080 Speaker 1: straightforward problems that code has learned about in school. But 334 00:16:35,160 --> 00:16:37,800 Speaker 1: they are unworldly. They do not understand how to bring 335 00:16:37,840 --> 00:16:40,520 Speaker 1: lots of solutions to the small, straightforward problems together into 336 00:16:40,560 --> 00:16:42,920 Speaker 1: a larger hole. They lack the experience to be wholly 337 00:16:42,920 --> 00:16:44,720 Speaker 1: trusted and trust this is the most important thing you 338 00:16:44,760 --> 00:16:48,360 Speaker 1: need to fully delegate coding tasks. In simpler terms, lms 339 00:16:48,360 --> 00:16:50,880 Speaker 1: are capable of writing code, but can't do software engineering 340 00:16:50,880 --> 00:16:54,400 Speaker 1: because software engineering is the process of understanding, maintaining and 341 00:16:54,440 --> 00:16:58,080 Speaker 1: executing code to produce functional software, and lms do not learn, 342 00:16:58,160 --> 00:17:01,280 Speaker 1: cannot adapt, and to paraphrase something Carl Brown said to me, 343 00:17:01,640 --> 00:17:04,439 Speaker 1: break down the more of your code and variables you 344 00:17:04,480 --> 00:17:06,840 Speaker 1: ask them to look at at once, so you can't 345 00:17:06,880 --> 00:17:09,600 Speaker 1: replace a software engineer with them. If you are printing 346 00:17:09,640 --> 00:17:12,080 Speaker 1: this in a media outlet and have heard this sentence, 347 00:17:12,280 --> 00:17:15,840 Speaker 1: you are fucking up. You really are fucking up. I'm 348 00:17:15,920 --> 00:17:18,040 Speaker 1: really neat members of the media here in this You 349 00:17:18,080 --> 00:17:19,879 Speaker 1: need to change. You need to change on this one. 350 00:17:19,960 --> 00:17:38,000 Speaker 1: You are doing software engineers dirty. Look, and I understand 351 00:17:38,000 --> 00:17:40,680 Speaker 1: why too. It's very easy to believe that software engineering 352 00:17:40,720 --> 00:17:42,679 Speaker 1: is just writing code, but the reality is that software 353 00:17:42,680 --> 00:17:46,480 Speaker 1: engineers maintain software, which includes writing and analyzing code, amongst 354 00:17:46,480 --> 00:17:49,440 Speaker 1: a vast array of different personalities and programs and problems. 355 00:17:50,040 --> 00:17:53,600 Speaker 1: Good software engineering harkens back to Brian Merchant's interviews with translators. 356 00:17:53,680 --> 00:17:55,959 Speaker 1: While some may believe the translators simply tell you what 357 00:17:55,960 --> 00:17:59,840 Speaker 1: words mean, true translation is communicating the meaning of a sentence, 358 00:18:00,119 --> 00:18:03,800 Speaker 1: which is cultural, contextual, regional, and personal and often requires 359 00:18:03,840 --> 00:18:07,240 Speaker 1: the exercise of creativity and novel thinking. And on top 360 00:18:07,320 --> 00:18:10,199 Speaker 1: of that, while translation is the production of words, you 361 00:18:10,240 --> 00:18:12,200 Speaker 1: can't just take code and look at it. You actually 362 00:18:12,240 --> 00:18:15,440 Speaker 1: need to know how code works and functions and wide functions. 363 00:18:15,480 --> 00:18:18,640 Speaker 1: In that way, using an LLM, you'll never know because 364 00:18:18,680 --> 00:18:21,760 Speaker 1: the LM doesn't know anything either. Now, my editor Matt 365 00:18:21,800 --> 00:18:23,960 Speaker 1: Hughes gave an example of this in his newsletter, which 366 00:18:24,000 --> 00:18:26,399 Speaker 1: I think i'll paraphrase. He used to live in France 367 00:18:26,400 --> 00:18:28,680 Speaker 1: and the French speaking part of Switzerland, and sometimes he 368 00:18:28,720 --> 00:18:31,159 Speaker 1: will read French translations of books to see how awkward 369 00:18:31,240 --> 00:18:34,399 Speaker 1: bits of prose are translated. Doing those awkward bits requires 370 00:18:34,400 --> 00:18:37,000 Speaker 1: a bit of creative thinking. And I quote take Harry 371 00:18:37,040 --> 00:18:40,960 Speaker 1: Potter in French, Hogwarts is boudlard, which translates into bacon lice. 372 00:18:41,359 --> 00:18:43,680 Speaker 1: Why did they go with that instead of a literal translation? 373 00:18:43,720 --> 00:18:47,439 Speaker 1: Of Hogwarts, which would be Verus Spork. I'm sorry to 374 00:18:47,440 --> 00:18:50,000 Speaker 1: anyone who can actually read languages, no idea, but I'd 375 00:18:50,000 --> 00:18:51,359 Speaker 1: assume it is something to do with the fact that 376 00:18:51,400 --> 00:18:55,120 Speaker 1: Poolard that Poudlard sounds a lot better than Veru Spork, 377 00:18:55,520 --> 00:18:58,960 Speaker 1: and both of them, I can say flawlessly. Someone had 378 00:18:59,000 --> 00:19:01,520 Speaker 1: to actually think about to translate that one idea. They 379 00:19:01,520 --> 00:19:04,040 Speaker 1: had to exercise creativity, which is something that an AI 380 00:19:04,119 --> 00:19:08,040 Speaker 1: in is inherently incapable of doing. Similarly, coding is not 381 00:19:08,080 --> 00:19:10,199 Speaker 1: just a series of texts that program as a computer, 382 00:19:10,280 --> 00:19:12,800 Speaker 1: but a series of interconnected characters that refers to other 383 00:19:12,880 --> 00:19:15,959 Speaker 1: software in other places that must also function now and 384 00:19:16,040 --> 00:19:18,040 Speaker 1: explain on some level to someone who has never ever 385 00:19:18,080 --> 00:19:20,320 Speaker 1: seen the code before why it was done in this way. 386 00:19:20,880 --> 00:19:23,000 Speaker 1: This is, by the way, while we're still yet to 387 00:19:23,000 --> 00:19:25,800 Speaker 1: get any tangible proof that AI is replacing software engineers, 388 00:19:25,840 --> 00:19:30,359 Speaker 1: because it isn't replacing software engineers, and now we need 389 00:19:30,400 --> 00:19:33,080 Speaker 1: to understand why this is so existentially bad for generative AI. 390 00:19:33,880 --> 00:19:36,600 Speaker 1: Of all the fields supposedly at risk from AI disruption, 391 00:19:36,720 --> 00:19:39,440 Speaker 1: coding fields or felt the most tangible, if only because 392 00:19:39,480 --> 00:19:42,040 Speaker 1: the answer to can you write code with LMS wasn't 393 00:19:42,080 --> 00:19:45,280 Speaker 1: an immediate unilater or no The media has also been 394 00:19:45,359 --> 00:19:48,000 Speaker 1: quick to suggest that AI writes software, which is true 395 00:19:48,000 --> 00:19:51,440 Speaker 1: in the same way that chat GBT writes novels. In reality, 396 00:19:51,560 --> 00:19:54,919 Speaker 1: lms can generate code and do somewhere some sort of 397 00:19:55,000 --> 00:19:58,720 Speaker 1: software engineering adjacent tasks, but like all large language models, 398 00:19:58,760 --> 00:20:01,359 Speaker 1: break down and go totally in saying hallucinating more and 399 00:20:01,400 --> 00:20:03,920 Speaker 1: more as the tasks get more complex, and software engineering 400 00:20:03,960 --> 00:20:07,520 Speaker 1: is extremely complex. Even software engineers who can read code 401 00:20:07,520 --> 00:20:09,359 Speaker 1: and have done so for decades will find problems they 402 00:20:09,400 --> 00:20:12,360 Speaker 1: can't solve just by looking at the code. And as 403 00:20:12,359 --> 00:20:15,120 Speaker 1: I pointed out earlier, software engineer is not just coding. 404 00:20:15,400 --> 00:20:18,600 Speaker 1: It involves thinking about problems, finding solutions to novel challenges, 405 00:20:18,640 --> 00:20:20,159 Speaker 1: designing stuff in a way that could be read and 406 00:20:20,160 --> 00:20:23,720 Speaker 1: maintained by others, and that's ideally scalable and secure. The 407 00:20:23,800 --> 00:20:26,360 Speaker 1: whole fucking point of an AI is that you handshit 408 00:20:26,480 --> 00:20:29,520 Speaker 1: off to it. That's what they've been selling it as. 409 00:20:29,640 --> 00:20:32,760 Speaker 1: That's why Jensen Huang told kids to stop learning to code. 410 00:20:32,760 --> 00:20:35,160 Speaker 1: As with AI, there's no point and it was all 411 00:20:35,240 --> 00:20:38,600 Speaker 1: a fucking lie. Generative AI can't do the job of 412 00:20:38,640 --> 00:20:41,359 Speaker 1: a software engineer, and it fails. While also costing an 413 00:20:41,400 --> 00:20:45,560 Speaker 1: abominable amount of money. Coding large language models seem like 414 00:20:45,600 --> 00:20:48,200 Speaker 1: magic at first because they, to quote a conversation with 415 00:20:48,280 --> 00:20:50,720 Speaker 1: Carl Brown, make the easy things easier, but they also 416 00:20:50,760 --> 00:20:53,720 Speaker 1: make the harder things harder. They don't even speed up engineers. 417 00:20:53,720 --> 00:20:56,360 Speaker 1: There's a study that showed that make them slower YEAT 418 00:20:56,400 --> 00:21:00,159 Speaker 1: coding is basically the only obvious use case for lms. Oh, 419 00:21:00,240 --> 00:21:02,360 Speaker 1: I'm sure you're gonna say, but I bet the enterprise 420 00:21:02,480 --> 00:21:06,280 Speaker 1: is doing well, and you're also very, very wrong. Microsoft, 421 00:21:06,280 --> 00:21:08,000 Speaker 1: if you've ever switched on a TV in the past 422 00:21:08,040 --> 00:21:10,439 Speaker 1: two years, has gone all in on generative AI, and 423 00:21:10,480 --> 00:21:13,040 Speaker 1: despite being arguably the biggest software company in the world 424 00:21:13,080 --> 00:21:16,600 Speaker 1: at least in terms of desktop operating systems and productivity software, 425 00:21:16,840 --> 00:21:20,280 Speaker 1: has made almost no traction in popularizing generative AI. It 426 00:21:20,320 --> 00:21:23,119 Speaker 1: has thousands, if not tens of thousands of salespeople and 427 00:21:23,200 --> 00:21:26,680 Speaker 1: thousands of companies that literally sell Microsoft services for a living, 428 00:21:27,600 --> 00:21:30,960 Speaker 1: and it can't sell AI. I've got a real fucking scoopyeo, 429 00:21:31,000 --> 00:21:32,840 Speaker 1: I'm so excited, and I buried it in the third 430 00:21:32,840 --> 00:21:36,160 Speaker 1: part of a four pot episode. AAH and truly twisted. 431 00:21:36,760 --> 00:21:39,560 Speaker 1: But a source that has CM materials related to Sales 432 00:21:39,840 --> 00:21:43,119 Speaker 1: has confirmed that as of August twenty twenty five, Microsoft 433 00:21:43,119 --> 00:21:47,159 Speaker 1: has around eight million active license so paying users of 434 00:21:47,240 --> 00:21:50,800 Speaker 1: Microsoft three sixty five Copilot, amounting to a one point 435 00:21:50,840 --> 00:21:53,840 Speaker 1: eight one percent conversion rate across four hundred and forty 436 00:21:53,920 --> 00:21:57,760 Speaker 1: million Microsoft three sixty five subscribers. Must be clear that 437 00:21:57,760 --> 00:22:00,600 Speaker 1: three sixty five is their big cash cow. This would 438 00:22:00,640 --> 00:22:03,080 Speaker 1: amount to if each of these users paid annually at 439 00:22:03,080 --> 00:22:05,399 Speaker 1: the full rate thirty dollars a month, to about two 440 00:22:05,480 --> 00:22:07,480 Speaker 1: point eight eight billion dollars an annual revenue for a 441 00:22:07,480 --> 00:22:10,840 Speaker 1: product category that makes thirty three billion dollars a fucking quarter. 442 00:22:10,920 --> 00:22:13,880 Speaker 1: It's productivity and business unit for Microsoft, and I must 443 00:22:13,920 --> 00:22:16,160 Speaker 1: be clear, I am one hundred percent sure these users 444 00:22:16,200 --> 00:22:19,639 Speaker 1: aren't all paying thirty dollars a month. The Information reported 445 00:22:19,640 --> 00:22:22,000 Speaker 1: a few weeks ago that Microsoft has been reducing the 446 00:22:22,040 --> 00:22:25,040 Speaker 1: software's price, referring to Microsoft three sixty five with more 447 00:22:25,040 --> 00:22:28,920 Speaker 1: generous discounts on the AI features. According to customers and salespeople, 448 00:22:29,080 --> 00:22:32,680 Speaker 1: heavily suggesting discounts have already been happening. Enterprise software is 449 00:22:32,720 --> 00:22:35,560 Speaker 1: traditionally sold at a discount anyway, or put a different way, 450 00:22:35,760 --> 00:22:37,560 Speaker 1: with bulk pricing for those who sign up a bunch 451 00:22:37,600 --> 00:22:39,760 Speaker 1: of users at once. In fact, I found evidence that 452 00:22:39,760 --> 00:22:41,639 Speaker 1: they've been doing this for a while, with a fifteen 453 00:22:41,640 --> 00:22:45,000 Speaker 1: percent discount on annual Microsoft three sixty five Copilot subscriptions 454 00:22:45,000 --> 00:22:47,359 Speaker 1: for orders of ten to three hundred seats mentioned by 455 00:22:47,359 --> 00:22:49,719 Speaker 1: an IT consultant back in late twenty twenty four, and 456 00:22:49,760 --> 00:22:52,680 Speaker 1: another that's currently running through September thirtieth, twenty twenty five, 457 00:22:52,840 --> 00:22:57,720 Speaker 1: with another Microsoft Cloud Solution Provider program. Yeah this, I've 458 00:22:57,720 --> 00:23:00,600 Speaker 1: found tons of other examples too. A Microsoft three sixty 459 00:23:00,600 --> 00:23:02,760 Speaker 1: five is the enterprise version where they sell things with 460 00:23:02,840 --> 00:23:05,760 Speaker 1: like Word and PowerPoint and sometimes teams as well. This 461 00:23:05,880 --> 00:23:08,920 Speaker 1: is them probably the most popular product, and by the way, 462 00:23:09,160 --> 00:23:11,760 Speaker 1: they even manipulate the numbers a little bit there. An 463 00:23:11,800 --> 00:23:15,200 Speaker 1: active user is someone who has taken one action on 464 00:23:15,359 --> 00:23:18,359 Speaker 1: any Microsoft three sixty five app with Copilot in the 465 00:23:18,400 --> 00:23:21,520 Speaker 1: space of twenty eight days, not thirty twenty eight. That's 466 00:23:21,560 --> 00:23:24,439 Speaker 1: so generous, now, I know, I know that word active. 467 00:23:24,520 --> 00:23:26,520 Speaker 1: Maybe you're thinking ed, this is like the gym model. 468 00:23:26,520 --> 00:23:30,359 Speaker 1: There are unpaid licenses that Microsoft is getting paid for. Fine, fine, fine, 469 00:23:30,440 --> 00:23:34,159 Speaker 1: fucking fine. Let's assume that Microsoft also has based on 470 00:23:34,200 --> 00:23:36,240 Speaker 1: research that suggests this can be the case for some 471 00:23:36,280 --> 00:23:41,040 Speaker 1: software companies another fifty percent four million paying Copilot licenses 472 00:23:41,080 --> 00:23:44,720 Speaker 1: that aren't being used. That's still twelve million users, which 473 00:23:44,760 --> 00:23:48,720 Speaker 1: is around two point seven percent conversion rate. That's piss, 474 00:23:48,720 --> 00:23:53,760 Speaker 1: poor buddy, that's piss, Paul, that's pissy. It sucks. It's bad, Doodoo. 475 00:23:54,160 --> 00:23:56,360 Speaker 1: Well I just said pp I guess anyway, very serious, 476 00:23:56,560 --> 00:24:00,000 Speaker 1: very serious podcast. But why aren't people paying for Copilot? Well, 477 00:24:00,080 --> 00:24:01,960 Speaker 1: let's hear from someone who talked to the information and 478 00:24:02,040 --> 00:24:04,520 Speaker 1: I quote, it's easy for an employee to say, yes, 479 00:24:04,560 --> 00:24:06,600 Speaker 1: this will help me, but hard to quantify how. And 480 00:24:06,640 --> 00:24:08,440 Speaker 1: if they can't quantify how it will help them, it's 481 00:24:08,480 --> 00:24:10,040 Speaker 1: not going to be a long discussion over whether the 482 00:24:10,080 --> 00:24:14,640 Speaker 1: software is worth paying for. Is that good? Is that good? 483 00:24:15,560 --> 00:24:18,440 Speaker 1: Is that what you want to hear? It isn't. It isn't. 484 00:24:18,440 --> 00:24:20,960 Speaker 1: That's that's the secret. It's not. It's bad. It's really bad. 485 00:24:21,000 --> 00:24:24,040 Speaker 1: It's all very bad. And Microsoft through sixty five Copilot 486 00:24:24,080 --> 00:24:26,560 Speaker 1: has been such a disaster that Microsoft will now integrate 487 00:24:26,600 --> 00:24:30,000 Speaker 1: Anthropics models to try and make them better. Oh one 488 00:24:30,040 --> 00:24:34,600 Speaker 1: other thing too. Sources also confirm GPU utilization, So how 489 00:24:34,680 --> 00:24:38,119 Speaker 1: much the GPUs set aside for Microsoft through sixty five? Yeah, 490 00:24:38,240 --> 00:24:42,240 Speaker 1: their enterprise codpile. It's barely scratching the sixty percents. I'm 491 00:24:42,280 --> 00:24:45,919 Speaker 1: also hearing the share Point, which is an app they 492 00:24:45,920 --> 00:24:48,000 Speaker 1: have with over two hundred and fifty million users, has 493 00:24:48,080 --> 00:24:51,120 Speaker 1: less than three hundred thousand weekly active users of their 494 00:24:51,119 --> 00:24:53,919 Speaker 1: copilot features, suggesting that people just don't want to fucking 495 00:24:54,040 --> 00:24:56,879 Speaker 1: use this. Those numbers that from August, by the way, 496 00:24:57,240 --> 00:24:59,560 Speaker 1: and it's pathetic, and it must be clear. If Microsoft's 497 00:24:59,600 --> 00:25:02,080 Speaker 1: doing this badly, I don't know how anyone else is 498 00:25:02,119 --> 00:25:05,280 Speaker 1: doing well, and they're not. They're all failing. It's pathetic. 499 00:25:06,119 --> 00:25:08,000 Speaker 1: But I've spent a lot of time today talking about 500 00:25:08,000 --> 00:25:10,960 Speaker 1: AI coding, because this was supposed to be the saving grace, 501 00:25:11,080 --> 00:25:13,120 Speaker 1: the thing that actually turned this from a bubble into 502 00:25:13,160 --> 00:25:15,720 Speaker 1: an actual money minting industry that changes the world. And 503 00:25:15,760 --> 00:25:18,080 Speaker 1: I wanted to bring up Microsoft through sixty five because 504 00:25:18,320 --> 00:25:21,119 Speaker 1: that's the place where Microsoft should be making the most money. 505 00:25:21,200 --> 00:25:24,120 Speaker 1: It's the most ubiquitous software, it's their most well known software, 506 00:25:24,280 --> 00:25:28,240 Speaker 1: and they're not eight million people eight million people. I've 507 00:25:28,280 --> 00:25:30,800 Speaker 1: run that by a few people and everyone's made the 508 00:25:30,840 --> 00:25:34,600 Speaker 1: same Oh God noise. It's quite weird, the old God 509 00:25:34,680 --> 00:25:38,240 Speaker 1: noise and the numbers. But this just isn't happening. Things 510 00:25:38,280 --> 00:25:40,880 Speaker 1: are going badly and it really only gets worse from here, 511 00:25:41,520 --> 00:25:43,600 Speaker 1: and I'm going to tell you more tomorrow in the 512 00:25:43,640 --> 00:25:46,000 Speaker 1: final part of our four part Thank you for your 513 00:25:46,040 --> 00:25:47,600 Speaker 1: patience and thank you for your time. 514 00:25:55,680 --> 00:25:58,080 Speaker 2: Thank you for listening to Better Offline. The editor and 515 00:25:58,119 --> 00:26:01,280 Speaker 2: composer of the Better Offline theme song is Matasowski. You 516 00:26:01,320 --> 00:26:03,639 Speaker 2: can check out more of his music and audio projects 517 00:26:03,720 --> 00:26:07,239 Speaker 2: at Matasowski dot com, M A T T O S 518 00:26:07,280 --> 00:26:11,359 Speaker 2: O W s ki dot com. You can email me 519 00:26:11,400 --> 00:26:14,000 Speaker 2: at easy at Better Offline dot com or visit Better 520 00:26:14,040 --> 00:26:16,240 Speaker 2: Offline dot com to find more podcast links and of 521 00:26:16,240 --> 00:26:19,359 Speaker 2: course my newsletter. I also really recommend you go to 522 00:26:19,440 --> 00:26:21,920 Speaker 2: chat dot Where's youreaed dot at to visit the discord, 523 00:26:22,160 --> 00:26:24,480 Speaker 2: and go to our slash Better Offline to check out 524 00:26:24,520 --> 00:26:27,199 Speaker 2: our reddit. Thank you so much for listening. 525 00:26:28,000 --> 00:26:30,720 Speaker 1: Better Offline is a production of cool Zone Media. For 526 00:26:30,840 --> 00:26:34,000 Speaker 1: more from cool Zone Media, visit our website cool Zonemedia 527 00:26:34,040 --> 00:26:36,920 Speaker 1: dot com, or check us out on the iHeartRadio app, 528 00:26:36,960 --> 00:26:39,639 Speaker 1: Apple Podcasts, or wherever you get your podcasts.