1 00:00:02,160 --> 00:00:07,560 Speaker 1: Zone Media. Hello and welcome to this week's Better Offline Monologue. 2 00:00:07,600 --> 00:00:17,520 Speaker 1: I'm your host ed zetron over on my premium newsletter. 3 00:00:17,560 --> 00:00:20,040 Speaker 1: This week I reported out a troubling situation with AI 4 00:00:20,079 --> 00:00:23,239 Speaker 1: powered coding environment Cursor maybe, a company called any Sphere 5 00:00:23,360 --> 00:00:25,800 Speaker 1: who unexpectedly changed their pricing in the middle of June 6 00:00:25,880 --> 00:00:27,639 Speaker 1: as well as adding a two hundred dollars a month 7 00:00:27,760 --> 00:00:31,880 Speaker 1: ultra subscription that well pissed everyone off and pissed off 8 00:00:31,920 --> 00:00:34,240 Speaker 1: a lot of people for many different reasons. In fact, 9 00:00:34,560 --> 00:00:37,879 Speaker 1: for some history, Cursor is or was the bell of 10 00:00:37,880 --> 00:00:40,400 Speaker 1: the ball of the AI industry. They had five hundred 11 00:00:40,400 --> 00:00:43,440 Speaker 1: million dollars of annualized recurring revenue, which means their highest 12 00:00:43,440 --> 00:00:46,880 Speaker 1: month times twelve, making them the first high revenue startup 13 00:00:46,920 --> 00:00:50,080 Speaker 1: outside of anthropic and open AI to actually show growth potential. 14 00:00:50,360 --> 00:00:52,440 Speaker 1: There really is no one else like them out there 15 00:00:52,440 --> 00:00:56,240 Speaker 1: at this number. I mean as Alpha sense four hundred 16 00:00:56,280 --> 00:00:58,840 Speaker 1: million according to the AI data based from the information, 17 00:00:59,320 --> 00:01:01,640 Speaker 1: but everyone else is powering around one hundred two hundred 18 00:01:01,680 --> 00:01:04,920 Speaker 1: million arr and raising far more money. Everyone was really 19 00:01:05,000 --> 00:01:08,120 Speaker 1: excited about Cursor. They were growing fast. Everyone loved them, 20 00:01:08,280 --> 00:01:10,160 Speaker 1: and it was because their product allowed you to just 21 00:01:10,319 --> 00:01:14,720 Speaker 1: build software just by typing stuff, and engineers liked it 22 00:01:14,800 --> 00:01:17,400 Speaker 1: and regular people liked it. Not saying it was good 23 00:01:17,480 --> 00:01:19,640 Speaker 1: or anything, just that people liked it, and they liked 24 00:01:19,680 --> 00:01:22,160 Speaker 1: it because you gave you kind of unrestricted access to 25 00:01:22,160 --> 00:01:25,240 Speaker 1: ward these models to build all this stuff. Now, what 26 00:01:25,280 --> 00:01:28,160 Speaker 1: if I told you that perhaps they had grown to 27 00:01:28,240 --> 00:01:32,280 Speaker 1: that five hundred million ARR number using techniques like allowing 28 00:01:32,319 --> 00:01:34,840 Speaker 1: people to spend way more than they were paying for 29 00:01:34,880 --> 00:01:39,600 Speaker 1: the service. Yeah, that's exactly what happened. Previously, Cursor's pricing 30 00:01:39,640 --> 00:01:42,559 Speaker 1: was pretty simple. You had five hundred fast completions, meaning 31 00:01:42,560 --> 00:01:44,559 Speaker 1: that you could ask Cursor to do something using model 32 00:01:44,600 --> 00:01:47,600 Speaker 1: like Claudes sonnet for or Opus four, and it deduct 33 00:01:47,720 --> 00:01:50,640 Speaker 1: requests from your allowance. The amount deductive was based on 34 00:01:50,680 --> 00:01:52,880 Speaker 1: the model, with some models costing less than others, such 35 00:01:52,920 --> 00:01:56,160 Speaker 1: as Anthropics claud for Opus costing more. Once you were 36 00:01:56,200 --> 00:01:58,240 Speaker 1: out of those fast requests, Cursor would then put you 37 00:01:58,280 --> 00:02:01,000 Speaker 1: in the slow lane, the unlimited slow lane, I should add, 38 00:02:01,120 --> 00:02:03,400 Speaker 1: which nevertheless would let you use premium models, but you'd 39 00:02:03,440 --> 00:02:04,760 Speaker 1: be put in the back of the queue and at 40 00:02:04,760 --> 00:02:07,000 Speaker 1: peak times could be waiting some time before you get 41 00:02:07,040 --> 00:02:09,640 Speaker 1: to your completion. You could also select auto and Cursor 42 00:02:09,680 --> 00:02:12,079 Speaker 1: would select whatever was available the best model for the 43 00:02:12,120 --> 00:02:15,440 Speaker 1: job at the time, never want to turn the best 44 00:02:16,000 --> 00:02:19,280 Speaker 1: over to the company though. But on June seventeenth, twenty 45 00:02:19,320 --> 00:02:21,880 Speaker 1: twenty five, Cursor launched that two hundred dollars a month 46 00:02:22,080 --> 00:02:26,040 Speaker 1: ultraplan along with a new and confusing pro package It's 47 00:02:26,160 --> 00:02:28,880 Speaker 1: very muddied. A new wording around said twenty dollars a 48 00:02:28,880 --> 00:02:31,560 Speaker 1: month plan and how it would work going forward, claiming 49 00:02:31,560 --> 00:02:33,520 Speaker 1: that they would be rolling out changes to make it 50 00:02:33,560 --> 00:02:36,400 Speaker 1: more generous, but actually changing pricing to reflect new wording, 51 00:02:36,400 --> 00:02:39,640 Speaker 1: offering unlimited agent requests and no real explanation as to 52 00:02:39,680 --> 00:02:41,919 Speaker 1: what that meant, and no further mention of any request 53 00:02:41,960 --> 00:02:44,920 Speaker 1: limits or what those requests or rate limits might be. 54 00:02:45,840 --> 00:02:48,760 Speaker 1: The reality was, of course, far grimmer. Cursor users have 55 00:02:48,760 --> 00:02:51,919 Speaker 1: found themselves heavily rate limited, especially on Anthropic models, but 56 00:02:51,960 --> 00:02:55,200 Speaker 1: the so called unlimited agent requests mostly pushing towards Cursor's 57 00:02:55,280 --> 00:02:58,160 Speaker 1: on frontier model, which users claim is nowhere near as 58 00:02:58,200 --> 00:03:02,080 Speaker 1: effective as model SMOP and AI Andthropic. On June thirtieth, 59 00:03:02,440 --> 00:03:05,200 Speaker 1: Curser made another change to their product features, changing unlimited 60 00:03:05,240 --> 00:03:08,360 Speaker 1: agents to extended limits on agents, further money in pricing 61 00:03:08,440 --> 00:03:10,680 Speaker 1: on a product that was once renowned for its simplicity 62 00:03:10,760 --> 00:03:14,919 Speaker 1: and lack of limits. So this isn't good, is it. 63 00:03:15,240 --> 00:03:17,799 Speaker 1: Silicon Valley's favorite coding startup has in the last few 64 00:03:17,800 --> 00:03:20,560 Speaker 1: weeks completely changed how its customers interact with its product, 65 00:03:20,680 --> 00:03:23,040 Speaker 1: both degrading the service and making it far more expensive 66 00:03:23,040 --> 00:03:25,359 Speaker 1: in the process. They also allow you now to use 67 00:03:25,480 --> 00:03:29,359 Speaker 1: usage based pricing. Now, this is the funny one. If 68 00:03:29,360 --> 00:03:31,799 Speaker 1: you pay twenty dollars a month on Cursor, they will 69 00:03:31,840 --> 00:03:35,400 Speaker 1: guarantee you twenty dollars a month of AI compute at least, 70 00:03:35,400 --> 00:03:37,840 Speaker 1: so they're literally giving money away. They add a twenty 71 00:03:37,880 --> 00:03:40,920 Speaker 1: percent fee on top of compute. But wow, so you're 72 00:03:40,920 --> 00:03:43,400 Speaker 1: making twenty percent morgin. Now you're fucking not. They're losing 73 00:03:43,400 --> 00:03:46,800 Speaker 1: money on everything. I'm sure of them. But it's strange. 74 00:03:46,880 --> 00:03:49,560 Speaker 1: It's strange, and it's bad. But this is also textbook 75 00:03:49,560 --> 00:03:52,360 Speaker 1: and shitification. Corey doctor O's term for when platforms offer 76 00:03:52,400 --> 00:03:54,400 Speaker 1: a high quality product to gain a large user base, 77 00:03:54,680 --> 00:03:57,280 Speaker 1: usually through convenience or great value. Then the grade the 78 00:03:57,320 --> 00:03:59,080 Speaker 1: service over time to make more money as a means 79 00:03:59,120 --> 00:04:02,839 Speaker 1: of maximizing value shareholders or making money themselves. It's also 80 00:04:02,920 --> 00:04:05,360 Speaker 1: part of my rot economy. Thesis. The growth of all 81 00:04:05,400 --> 00:04:08,200 Speaker 1: costs thinking has dominated the tech industry, and I'd argue, 82 00:04:08,240 --> 00:04:11,120 Speaker 1: thanks to the proliferation of business idi it's controlling everything 83 00:04:11,720 --> 00:04:14,280 Speaker 1: that they've drained some of the logic behind in certification away, 84 00:04:14,320 --> 00:04:16,320 Speaker 1: because very few of these companies actually have a plan 85 00:04:16,400 --> 00:04:20,400 Speaker 1: for sustainability, let alone profitability. Generally, in certification gets people 86 00:04:20,440 --> 00:04:23,719 Speaker 1: through the door and makes the service totally impossible to avoid, 87 00:04:23,839 --> 00:04:26,960 Speaker 1: makes it essential. That's really difficult to do with a 88 00:04:27,000 --> 00:04:30,839 Speaker 1: paid software product like Cursor, because AI compute is so 89 00:04:31,040 --> 00:04:35,760 Speaker 1: fucking expensive. It's so expensive, so Cursor has decided to 90 00:04:35,839 --> 00:04:39,560 Speaker 1: en shittify without making sure that they have something essential. 91 00:04:40,760 --> 00:04:43,479 Speaker 1: But the thing is, I had texted Cory about this. 92 00:04:43,839 --> 00:04:47,000 Speaker 1: I think this actually is something quite different. I think 93 00:04:47,040 --> 00:04:50,680 Speaker 1: this is the world's first chain in shitification. My belief 94 00:04:50,720 --> 00:04:53,599 Speaker 1: is simple. Any Sphere, which is the company that makes Cursor, 95 00:04:53,600 --> 00:04:56,440 Speaker 1: by the way, is despite getting nine hundred million dollars 96 00:04:56,440 --> 00:04:58,880 Speaker 1: in funding in early May, running out of money, or 97 00:04:58,880 --> 00:05:01,200 Speaker 1: at least believes that continue to operate its business in 98 00:05:01,240 --> 00:05:02,800 Speaker 1: the way it did less than a month ago would 99 00:05:02,800 --> 00:05:05,360 Speaker 1: cause it to do so. But why the sudden changes, 100 00:05:05,400 --> 00:05:07,279 Speaker 1: Why the knife in the heart of their customers just 101 00:05:07,320 --> 00:05:10,599 Speaker 1: after raising nearly a billion dollars. It's simple, it's a 102 00:05:10,680 --> 00:05:13,839 Speaker 1: chain and citification. Anthropic jacked up their prices, and so 103 00:05:13,920 --> 00:05:17,320 Speaker 1: did open Ai. On May twenty second, twenty twenty five, 104 00:05:17,360 --> 00:05:20,000 Speaker 1: a few weeks after Cursor raised nine hundred million dollars 105 00:05:20,000 --> 00:05:23,080 Speaker 1: and Thropic launched both claud Sonet and claud Opus four. 106 00:05:23,480 --> 00:05:26,920 Speaker 1: You might say Claude four Opus of Claude force Sonet. 107 00:05:27,200 --> 00:05:30,640 Speaker 1: I don't care. But these were two new powerful as 108 00:05:30,680 --> 00:05:34,000 Speaker 1: judged by benchmarks made specifically for large language models. And 109 00:05:34,200 --> 00:05:37,800 Speaker 1: I should add that the expensive claud Opus four was 110 00:05:37,839 --> 00:05:41,479 Speaker 1: more focused on coding benchmarks. This is an important detail now. 111 00:05:41,480 --> 00:05:44,200 Speaker 1: Eight days later, on May thirtieth, twenty twenty five, patron 112 00:05:44,240 --> 00:05:48,080 Speaker 1: Anthropics API documentation appeared for the first time called service tiers, 113 00:05:48,440 --> 00:05:51,080 Speaker 1: adding priority tiers for enterprise stats that didn't want to 114 00:05:51,120 --> 00:05:54,440 Speaker 1: and I quote sorry, I mean did want to and 115 00:05:54,520 --> 00:05:57,919 Speaker 1: I quote provide a guarantee around the infrequency of server 116 00:05:58,040 --> 00:06:02,120 Speaker 1: overloaded errors even during peak time. Hmm, that's not good. 117 00:06:02,520 --> 00:06:06,320 Speaker 1: Anthropic service tiers require a multi month upfront commitment on 118 00:06:06,360 --> 00:06:08,400 Speaker 1: how many tokens per minute your startup will use, but 119 00:06:08,480 --> 00:06:12,279 Speaker 1: also add an insidious charge around prompt casing. Now Prompt 120 00:06:12,320 --> 00:06:14,760 Speaker 1: cashing is when when you put something into a model, 121 00:06:14,839 --> 00:06:16,520 Speaker 1: such as a code base in the case of a 122 00:06:16,520 --> 00:06:19,200 Speaker 1: coding startup, or I don't know a great deal of 123 00:06:19,200 --> 00:06:21,880 Speaker 1: stuff about how a model might want to act, you 124 00:06:21,920 --> 00:06:23,479 Speaker 1: put it in the cash so that it kind of 125 00:06:23,520 --> 00:06:25,400 Speaker 1: like RAM, it just kind of reads off of it 126 00:06:25,440 --> 00:06:30,400 Speaker 1: instead of rewriting and rewriting and rewriting. So they've added 127 00:06:30,440 --> 00:06:33,160 Speaker 1: a vig to it or attacks whatever you call it, 128 00:06:33,200 --> 00:06:36,200 Speaker 1: a toll perhaps, and they're now charging and this is 129 00:06:36,279 --> 00:06:38,599 Speaker 1: really fucking insidious, either one hundred and twenty five percent 130 00:06:38,720 --> 00:06:40,800 Speaker 1: or two hundred percent of the cost of casing information 131 00:06:40,880 --> 00:06:43,680 Speaker 1: that you need to access it more readily. And I 132 00:06:43,680 --> 00:06:46,200 Speaker 1: should add this is only if you want the Priority tier. 133 00:06:46,760 --> 00:06:51,080 Speaker 1: And another thing, coding startups are extremely prompt cash heavy, 134 00:06:51,960 --> 00:06:54,800 Speaker 1: kind of fucked up right, very fucked up, and they're 135 00:06:54,839 --> 00:06:57,080 Speaker 1: not the only ones doing it. On June twenty fifth, 136 00:06:57,120 --> 00:07:00,159 Speaker 1: twenty twenty five, open Ai also launched Priority processing for 137 00:07:00,200 --> 00:07:02,760 Speaker 1: API customers. They claim it's the payers you go to 138 00:07:02,839 --> 00:07:05,599 Speaker 1: where you pay higher API prices for and I quote 139 00:07:05,680 --> 00:07:09,800 Speaker 1: predictable low latency. However, despite this being payers you go, 140 00:07:09,880 --> 00:07:12,240 Speaker 1: the service is only available for their enterprise customers who 141 00:07:12,280 --> 00:07:14,960 Speaker 1: have made an up prompt commitment much like Anthropic does, 142 00:07:15,040 --> 00:07:17,080 Speaker 1: the difference being they are not trying to tax you 143 00:07:17,120 --> 00:07:20,800 Speaker 1: on prompt caching now. Cursor CEO Michael Truell also said 144 00:07:20,800 --> 00:07:23,080 Speaker 1: when announcing their agree just two hundred dollars a month 145 00:07:23,160 --> 00:07:26,160 Speaker 1: plan that it was only made possible by multi year 146 00:07:26,160 --> 00:07:30,400 Speaker 1: partnerships with OpenAI, Anthropic, Google, and Xai. In short, I 147 00:07:30,400 --> 00:07:32,920 Speaker 1: think that nine hundred million dollars that Cursor got may 148 00:07:32,960 --> 00:07:35,520 Speaker 1: have immediately been handed to or committed to the major 149 00:07:35,560 --> 00:07:39,600 Speaker 1: model developers. They got looted baby. And now what makes 150 00:07:39,640 --> 00:07:42,760 Speaker 1: this weirder is that Cursor is also Anthropic's largest customer, 151 00:07:42,960 --> 00:07:45,240 Speaker 1: with sources claiming that they're such a large customer that 152 00:07:45,280 --> 00:07:51,800 Speaker 1: they're taxing anthropics infrastructure and making them run out of GPUs. Now. 153 00:07:51,920 --> 00:07:54,920 Speaker 1: I should also add that Cursor isn't the only one 154 00:07:54,960 --> 00:07:58,880 Speaker 1: that's recently dramatically changed their prices. Vibe coding startups like 155 00:07:59,000 --> 00:08:00,680 Speaker 1: replt and Lovable Bits both have had to do so 156 00:08:00,800 --> 00:08:02,840 Speaker 1: in the last month. And I think this is just 157 00:08:02,880 --> 00:08:06,560 Speaker 1: the beginning of something really, really, really bad. I believe 158 00:08:06,560 --> 00:08:09,480 Speaker 1: that Anthropic did this either as a deliberate attempt to 159 00:08:09,520 --> 00:08:12,480 Speaker 1: price gouge its largest customers, and or as a means 160 00:08:12,480 --> 00:08:15,760 Speaker 1: of increasing revenue and its money losing software. These changes 161 00:08:15,760 --> 00:08:18,720 Speaker 1: that deliberate aggressive and targeted price increases, and they were 162 00:08:18,720 --> 00:08:21,080 Speaker 1: timed with the launches of claud for Opus and Sonnet, 163 00:08:21,240 --> 00:08:24,720 Speaker 1: which suggests that Anthropics costs have dramatically increased with these models, 164 00:08:24,720 --> 00:08:27,120 Speaker 1: so much that simply increase in the cost per million 165 00:08:27,120 --> 00:08:30,280 Speaker 1: tokens is in sufficient. It's my belief that the launchers 166 00:08:30,280 --> 00:08:32,520 Speaker 1: of both claud Opus and to a lesser extent, son 167 00:08:32,559 --> 00:08:34,960 Speaker 1: It have caused en up hevil in Anthropics costs and 168 00:08:35,000 --> 00:08:37,480 Speaker 1: compute demands, which in turn forced them to start increasing 169 00:08:37,480 --> 00:08:41,520 Speaker 1: costs on their customers. However, Anthropic has I believe, realized 170 00:08:41,559 --> 00:08:43,480 Speaker 1: that there is no real way to just increase the 171 00:08:43,600 --> 00:08:46,240 Speaker 1: cost of Opus and Sonnet for further on just the 172 00:08:46,240 --> 00:08:48,800 Speaker 1: amount of tokens that the customer might use, and that 173 00:08:48,880 --> 00:08:51,840 Speaker 1: doing so might push away smaller developers, which wouldn't make 174 00:08:51,920 --> 00:08:54,319 Speaker 1: them more money. So they decided to find a way 175 00:08:54,360 --> 00:08:58,400 Speaker 1: to specifically exploit the finances of their largest customers coding startups, 176 00:08:58,440 --> 00:09:00,480 Speaker 1: in a way that won't be immediately obvious or that 177 00:09:00,520 --> 00:09:04,040 Speaker 1: would spook non coding assistant customers. Except I'm a crafty 178 00:09:04,080 --> 00:09:06,520 Speaker 1: little fuck and I look at everything all the time. 179 00:09:06,880 --> 00:09:09,280 Speaker 1: I do this for fun. I don't do this because 180 00:09:09,280 --> 00:09:11,760 Speaker 1: I have to listen to me. Wario Amadee. If you're 181 00:09:11,800 --> 00:09:15,000 Speaker 1: listening to this, I'm watching you. I have archive dot 182 00:09:15,080 --> 00:09:18,360 Speaker 1: org and a million diet cokes. Anyway, last year I 183 00:09:18,400 --> 00:09:21,480 Speaker 1: talked about the subprime AI crisis, where almost the entire 184 00:09:21,480 --> 00:09:23,720 Speaker 1: tech industry is bought in on a technology sold at 185 00:09:23,760 --> 00:09:27,960 Speaker 1: this vastly discounted rate, and they've heavily centralized and subsidized 186 00:09:28,000 --> 00:09:31,320 Speaker 1: it too. And I predicted one major thing that these 187 00:09:31,360 --> 00:09:34,040 Speaker 1: model developers would eventually have to find a way to 188 00:09:34,120 --> 00:09:37,080 Speaker 1: make their costs work, that they would have to find 189 00:09:37,080 --> 00:09:39,760 Speaker 1: a way to crank up costs on their customers, because 190 00:09:39,800 --> 00:09:42,360 Speaker 1: otherwise they would just continue burning money. I still think 191 00:09:42,400 --> 00:09:45,480 Speaker 1: they are doing so, and I think that's what's happening here. 192 00:09:45,840 --> 00:09:48,400 Speaker 1: I think anthropic and OpenAI to a lesser extent of 193 00:09:48,520 --> 00:09:50,480 Speaker 1: realized that they need to start making money back on 194 00:09:50,520 --> 00:09:53,839 Speaker 1: these fuckers, and they're doing it. Except the fact that 195 00:09:53,880 --> 00:09:56,520 Speaker 1: the companies of immediate companies like Cursor and Repler have 196 00:09:56,559 --> 00:09:59,960 Speaker 1: immediately had to change their prices suggest that maybe not 197 00:10:00,000 --> 00:10:03,760 Speaker 1: one of these businesses makes sense. Maybe this whole time 198 00:10:03,920 --> 00:10:07,439 Speaker 1: this thing was unsustainable. Maybe it turns out the general 199 00:10:07,520 --> 00:10:09,960 Speaker 1: if AI doesn't have the kind of business returns you 200 00:10:10,000 --> 00:10:16,079 Speaker 1: need to run a startup. If only someone had said something. Remember, 201 00:10:16,240 --> 00:10:18,520 Speaker 1: look see how fucked up saying Remember but I'm not 202 00:10:18,600 --> 00:10:22,080 Speaker 1: editing it out. This is an honest podcast. But seriously, everyone, 203 00:10:23,080 --> 00:10:28,200 Speaker 1: remember the pale horses. Rate limits, service interruptions, price increases, 204 00:10:28,360 --> 00:10:31,280 Speaker 1: trouble raising funding, and trouble with money. The horses are 205 00:10:31,320 --> 00:10:34,480 Speaker 1: drawing nearer and telling you we're coming to the end 206 00:10:34,480 --> 00:10:36,480 Speaker 1: of this. I don't know if it will be soon, 207 00:10:36,640 --> 00:10:39,040 Speaker 1: I don't know if it'll be next year. But nothing 208 00:10:39,080 --> 00:10:40,920 Speaker 1: about this suggests that things are going well.