0:00By now you guys have probably heard of Lauren, aka Potato, aka the creator of the P stack. She made a lot of other really cool things like the React compiler. I think she's awesome. But she recently went from Meta to working at Cursor and has quickly taken this entire agentic dev world by storm with her unique way of doing things. When I checked Twitter at the start of today, I saw this post going in on Lauren and people like her for wasting allegedly 132 million dollars of tokens a year.
0:27132 million dollars a year. I don't know about you guys, but as someone who has used Cursor a good bit myself, I don't know if I would put the last few months of Cursor's improvement in that 132 million dollar range. That I mean, you can get like three GPUs for that. That's a lot of money. All of this data comes from her profile on the Cursor site. You could see that in the last 30 days,
0:54she's done 1.5 trillion with a T tokens.
0:58I go pretty hard with my token use and I'm still not able to hit 100 bill. So, how the hell did she manage to do 15 times more token usage than someone like me who is actively trying to waste tokens? And what the hell is she doing with 132 million dollars worth every 30 days? Is this number just fake? Is there things being shipped that we don't know about? Is she just doing it for the memes? What's going on here?
1:26As you can guess, I have a lot of thoughts and as you can probably have predicted by now, most of these numbers don't make any sense at all. But I'm not trying to say your spend isn't crazy, it absolutely is. But what she is doing here is much more pioneering than she is getting credit for on Twitter. And sadly, since it's my job to clear up this type of misinfo, I'm going to have to drag you guys through yet another explanation of how companies like OpenAI, Anthropic, Cursor, and SpaceX AI, Meta, and all these token maxing,
1:55token wasting companies are actually spending their tokens and what we are working towards as they do it. In particular, I have to explain why what Zeeg said here about people working at the labs not understanding how no one can ever spend this much on tokens.
2:12I need to explain why this is a really [ __ ] stupid thing to post. Before we go into stupid and bad things, I want to talk about something awesome. Today's sponsor. I don't know if this goes for everyone, but personally, I've been running a lot more agents recently. And with this comes a lot of additional friction. Things like more builds, more time waiting on CI, more time just constrained by compute and RAM and network speeds and more. At least I was until I moved all that work over to Depot. These guys are here to speed up your agents and their dev workflows. All of the stuff they've been building was already useful for devs who are trying
2:41to ship quickly, but it's so much more useful when your agents are hitting those same bottlenecks. If you were waiting a little bit for CI yourself, that might have slowed you down a bit.
2:49When your agents are filing dozens of PRs a day, every second counts. And they're not going to save you seconds, they're going to save you hours if not more. They can be used as a drop-in replacement for GitHub Actions, but it gets way cooler once you start adopting their CI pipelines, which guess what?
3:03Your agents can do very easily. This allows you to parallelize the work that's being done and share much more of it, so your agents are able to test their code much, much faster. Let's go through some of the tools that will help your agents move faster. First, we have Depot CI, a fully programmable engine your agents can use to make faster parallelizable CI. It can be used fully without GitHub, so all those outages won't affect you. If you're using mono repos for building, Depot Cache will help a ton as well. They already work with pretty much all mono repo tools, whether you're in TypeScript with Turbo Repo or Nx, or you're in other
3:31ecosystems with Bazel and Gradle and other tools like that. If you're using Docker, I hope you already are using Depot because they will make all of your Docker commands way faster. Their cache is shared between you, your team, and your CI, so the likelihood that you have to build a full image from scratch is near zero, massively speeding up those builds up to 10 times faster. And this all can be stored in the Depot registry, which handles all of the images, all of the artifacts, and more that you and your agents will need to be building. If you want to go all in, they just put out Depot code, which is a Git server that
4:00uses no desks, stores everything in blob storage, that's way faster and closer to your CI, so all of the steps that normally take minutes will start taking seconds instead. Real users have seen improvements between 10 and 20x faster for their builds, CI, and more. Get them yourself at soydev.link/depot.
4:16If you were to ask your agent to summarize the video you're about to watch, it would say something along the lines of Theo points out the flaws in the calculation that Peter did and showcases how models will get cheaper over time or something along those lines. And that is the like top-level what we're about to do. The more subtle lesson I want you to take from this is learning how to identify the types of people who say things that sound really smart and agreeable in this AI dev space, but are just entirely wrong if you look at the numbers. I'm
4:45particularly frustrated around the types of people who just gladly, happily post these giant, inflated, incorrect things cuz they want to complain, and I want to help you guys learn to avoid them.
4:56Because you got a lot of people falling for things like what Peter said here and what Zig followed up with. And this is all just egregious. And I don't mean any ill will towards Peter or towards Zig, but I legitimately think you should never listen to anything either says about AI because they clearly don't know what they're talking about. I'm going to change my opinion here in the future if they stop lying so aggressively and just being wrong in adamant ways on the internet. But if the majority of the things they say about AI are factually wrong, I don't know why you would hunt so deeply to find the others. So let's
5:25go through all the factual inaccuracies of what Peter said here. We start with 1.5 trillion tokens in last 30 days.
5:31That's true. These first two sentences are true. Assume an average rate of $8 per million tokens. Now we're at 384k a day or 11 million in the last 30 days.
5:42Per year, that's 132 mil. First, I want you to remember that our friend Lauren works at SpaceX AI. So, there's a good chance she is using models cheaper than things like Opus, Fable, and Astra.
5:54She's probably using Grok a good bit.
5:56So, let's dig into this $8 per million number. The token count that is being shared on Cursor profiles is an aggregate number. 1.5 trillion tokens.
6:08This is not split between the four types of tokens. Cash read, cash write, normal read, and normal out. There's four types of tokens, and they're all priced very, very differently. A model like Opus 5.5 is $5 per million in and $25 per million out. If you average those two numbers, you end up with a number like $12 to $15. Knowing that I am almost exclusively using Opus 5.5, what do you think my average cost per million
6:37tokens, if you combine all the tokens, is? Get a number in your head. Remember, the number we cited for Lauren here, who is using Grok a lot, which is a cheaper model, is $8 per mil. What do you think my number per mil is? Spoiler, it's 41 cents. He's off by 20x. He is literally opening with a number that is an order of magnitude and a half off. These people don't know what they're talking about. They're just lying. They're either lying or they're stupid, and it's annoying cuz Peter's a smart person. I don't know why the [ __ ] he posted this.
7:07And we're just getting started, don't worry. The reason that this number is so egregious is because cash read on Opus isn't the usual $5. It's 20 cents. And despite that, half my cost is cash read, a little less than half, but around half is cash read, half is cash write, and the input and output tokens, the ones he is basing the price on, make up less than 8% of my total spent. He averaged the two numbers at the far end that are
7:36less than 10% of my spend, and said that's the price for the whole thing.
7:40And we're just getting started, guys.
7:42Don't worry, though, he he prices this in because we should assume Cursor has some insane deal with Anthropic and they're getting 90% off. Then it's only 38k a day or 1.1 mil in September and October. And he compares this to how expensive engineers are and says that 13 mil, which is allegedly her budget for inference over a year, gets her 26 staff level engineers at 500k a year each.
8:06Well, I'm going to be real with you, Peter. I don't know how many staff level engineers I'd want to hire without giving them a crazy token budget because I honestly believe 26 staff level engineers with $0 in tokens would be meaningfully less productive than just Lauren by herself with unlimited tokens.
8:24I would argue comically so. He does follow up though. Let's say his math is off by a factor of 10 or they're running their own GPUs at only 80 cents, it still comes out to $100,000 a month.
8:34Remember, he is pricing in huge discounts there and his number is still double my number as an end consumer.
8:40Entirely wrong. The only thing here that is right is that they are probably getting some insane deal from Anthropic because remember, Cursor is now owned by SpaceX AI and SpaceX AI is giving a meaningful amount of compute to Anthropic, renting it because Anthropic needs GPUs more than XAI does. And on top of that, the vast majority of Anthropic's revenue came from Cursor for many years. So they almost certainly have a crazy deal here. If I was to guess what the effective cost, so to
9:09speak, that they are spending on the stuff that Lauren is doing is, again, pretending that the price is across all the tokens and that agreeing that way makes any sense at all, which it doesn't for a million reasons, but assuming it does, I would guess at the high end, her price per million tokens with all of this combined would be closer to like 20 cents. So again, look at his numbers, he said 80 cents is 100k a month. So my guess at 20 cents, which again, like
9:38this is not the right way to measure it, but given the [ __ ] show we're in here, that would knock that 100K a month down to 25K a month. And here is where I have an honest question for you guys. I don't know how much you all know about marketing spend. In fact, I actually do.
9:50It's not that much. I know cuz I talked to a lot of engineers about marketing.
9:53They don't get it. How much money do you think most companies would be willing to spend to get the type of attention that Lauren is getting on Twitter and in the dev AI world right now? I'll give you a hint. It's a lot more than $25,000 a month. I know companies that throw events that get less attention than one random tweet from Lauren. And some of those events cost over a million dollars. And they get less from it than a couple tweets from Lauren. And that's not even counting the fact that all of the tokens she is burning are actually
10:22improving the products. Both Cursor, which now finally is actually performant. When Lauren joined Cursor, I was very hopeful because Cursor's app was falling apart. It was super slow. It was buggy as hell. And Lauren joining cleaned up a bunch of the bugs and made it way, way faster. And more importantly is Grokbot, which is quickly becoming the coolest product out of SpaceX AI.
10:42She's filing thousands of PR's a month, landing endless changes across all of these products, fixing things constantly. And it's only probably costing 25 to like worst case 50 grand a month based on reasonable estimates of what her token spend would be.
10:57But fun facts that we'll add on top of here, not only are they getting a discount for the tokens they're getting from somebody like Anthropic, she's also doing very, very deep long agentic runs.
11:08So, she's probably hitting cash even more than I am. All of this combined, I wouldn't be surprised if her token spend was closer to 10 cents per mill based on this awful aggregate number that doesn't tell us anything. Which would again, pricing based on this, put her at closer to like 15K a month at worst. And now I want to talk a bit about my response here because I I have things to say.
11:33The $132 mil a year that he started with is just it's not real at all. Even assuming full prices today, worst case is closer to 3 mil. I don't even think that is accurate. I think it's a lot closer to like 1 mil at most, probably like 800k or so. But with the trend of how inference is getting cheaper, I think this could get as low as $1,000 a year within the year. Hear me out.
11:58One last artificial analysis flashbang warning because we need to go to the trends, specifically the efficiency trends.
12:06This chart shows at a given level of intelligence, how does cost go down over time? As you can imagine, smart models coming out tend to come out at a pretty high price. They're measuring the intelligence index task cost. So, this is how much money did it take to complete the average task in artificial analysis. When a model like Sonnet 4.5 came out, it was the frontier. It was the first model to break a 20 on the intelligence index, and it cost 56 cents per task. That was in September of 2025.
12:35Now, in September of 2026, there are models that get that same score, which I can't see cuz of this line being broken.
12:42God, this is why I stopped using artificial analysis for things. Yeah. 6 Luna low gets the same score as Sonnet 4.5 did about a year later for 0.45 cents per task. It went from 56 cents per task to 0.45. That is a 100x decrease over a year. Or we could look at the models in the 30 range, where Sonnet 4.6 Max broke it for the first time at $2.50.
13:10And now, under a year later, 6 Luna high can do it for 3 cents. $2.50 to 3 cents. That is nearly a 100x decrease in approximately 8 months. Do you see where we're going here? Not only is intelligence getting cheaper, it's getting cheaper faster.
13:29It used to take a year for a 10x decrease. Now it takes 6 to 8 months for a 100x decrease. And this is where I need to talk about what Zeig said.
13:39This is exact quote him responding to Peter. This is what the folks at the labs often don't understand. If you pay for your own tokens or you have a business that has to, then you'll finally understand why the things that you say are so silly. No.
13:53Just now. For those who haven't been paying close attention to recent videos, this is the same guy who posted saying you should probably just keep using models like GPT 5.6 Soul or Sonnet or Opus 5.0 because the cost savings were worth it and the big fancy models are just never worth it. That I had to do a whole video about why you should use smart models on. This is his new take that pisses me off. This is the one that made me crash out and go do all the numbers. My response to him was that the first computers cost millions of
14:22dollars. And I personally find it quite cool that folks like Lauren are pushing the limits while ignoring cost. On average, the quote cost for a given level of intelligence drops by around 10x year over year. It's actually accelerated recently going as much as 4.5x decreases per month. As I mentioned before for the AI numbers, scores above a 10 used to cost 26 cents in January 2025. Now they're 1.5 cents. Or even things that are in the 50s costing $6
14:51when they dropped originally, just 3 months later you're getting the equivalent for 20x cheaper in 2 and 1/2 months. Do you understand how quickly things are accelerating?
15:02They're not just accelerating with how much faster they get smarter, they're accelerating at how quickly things get cheaper at a given level of intelligence. That's only true due to benchmarks though and you're not actually getting net cheaper intelligence everywhere. Yeah, because we like to use the smarter thing. Yeah, like you can't complain iPhones aren't getting cheaper when you refuse to buy last year's iPhone. Things are getting cheaper. We just use it more. I was just so far as I like this you aren't actually getting net cheaper intelligence. You you are. You
15:32absolutely are. Chat's pointing out that it feels like you didn't even read what I said. Fair and more fair the other way because when he slopped back to me with this essay, I I was done. I just I couldn't do it anymore. The example I gave in the reply is that the smartest model this time last year was GPT-5 High and this was like agreed upon universally. Nobody thought there was a model smarter than 5 High. They might not have thought it was as smart as I did, but everyone agreed 5 High was the [snorts] best model at the time. GLM-53 Flash and DeepSeek-41 Flash match or
16:01beat it in every benchmark, but more importantly here, I prefer it.
16:06I would never choose GPT-5 or O1 or any model released last year over GLM-53 Flash because 53 Flash is genuinely really pleasant to work with.
16:17It's good at long-running agentic tasks.
16:19You ask it to do a thing and it does it.
16:21You ask it to do another thing and it does it. Last year's models weren't just dumber. They were worse at working with you. So you wanted to use them less and when you did, you would use them less.
16:30So not only do I disagree with this premise that like things aren't getting cheaper. It's In particular, he talks in this post about how Where is it? Old models get retired so there's no true price drop regularly.
16:44What What? What? Are you expecting them to keep hosting the super inefficient version of the model when people have made more efficient better models for way cheaper? But like the cost for hosting doesn't inherently go down when they leave the model out. Like it wouldn't make sense for them to keep hosting. Like what what? I'm just this particular sentence.
17:04I'm sorry I'm crashing out so hard. It just really pisses me off when billionaire tech bros that I'm friends with like David are this egregiously wrong about [ __ ] Like in the here's the worst part. Most of what he says isn't factually wrong.
17:17None of what he says is relevant at all.
17:19I'm thankful he didn't watch my last video about his [ __ ] because he would not have interacted with me as politely as he did here.
17:26The thing I'm most upset about though is that he got close to my likes. So the reason I'm crashing out here isn't so much that David is saying really dumb things. It's that people want to listen and believe him. People are hearing the dumb things he is saying and being like, "Yeah, that sounds right to me."
17:42It's not. It's just wrong. And if you're one of those 17 people, you're either just misinformed, which is fine. No offense, that's why you're here. And if you are liking tweets like this and watching my videos, you clearly are trying to figure this out and I respect you for it. But pretty much everybody else who hit that button did it because they don't want to recognize where we're going. They want to say AI is just going to stay super expensive forever. It won't make sense. We need to use cheaper models. We need to write code ourselves.
18:08We need to do all these things. And they're liking things that are wrong cuz they make them feel better. But things are about to get worse as we go. As I mentioned, not only are 41 flash and 53 flash from DeepSeek and GLM respectively way more pleasant to use, they're also comically cheaper. I ran these numbers because Prime was complaining that models weren't getting cheaper. So I ran the numbers cuz I wanted to figure out how wrong that was. So I compared the
18:36raw token price for GPT-5 versus V4 flash, not even 41 flash yet. Before flash and eventually GLM-53. Input tokens were 21x cheaper. Output tokens were 55x cheaper. Even if you presume there's a three times increase in the number of tokens, the difference nets out to 20x cheaper. I then point out that GLM-53 flash is more pleasant to use and is 29 times cheaper than GPT-5 in bench runs. You're getting better intelligence for a similar cost, not cheaper intelligence overall generally
19:07The old rates go away cuz there's no infra to run those old models or they don't want to use it because it doesn't cover any costs. And those 10x better intelligence numbers are just benchmarks, which we all know aren't reliable. I will agree that benchmarks aren't the most reliable thing with modern models, but it's not for the reason he's implying here. The reason benchmarks aren't reliable anymore is cuz they're all either saturated or stupid. Do you know why we like benchmarks so much though? We like them because last year models weren't smart enough to saturate them, so they were
19:35very useful. The fact that cheap models today can saturate the benches that models struggled on last year does not mean the bench is bad. It means as my chat is pointing out that what we need is better and newer benchmarks. Then he says a 1% gain might be a quote 10x improvement.
19:53Thank you, S John. I was trying to be charitable to him before, but arguing more intelligence for the same cost doesn't imply the same intelligence is cheaper is genuinely crazy. Yep. He's contradicting himself. Yes, he is. But this is the thing and what you have to realize with the anti-AI dev doomer, contradictions don't matter to them.
20:09Facts don't matter to them. Numbers don't matter to them. Usage doesn't matter to them. Progress doesn't matter to them. Words that align with their beliefs is the only thing that matters to them. So if somebody writes an essay, even if they used AI to write a slop essay, if the headline is something that makes them feel better, they will gladly share it with the world.
20:28I'm not saying it's not getting cheaper, but folks who quote those numbers do so without considering the practical rate of improvement and cost of applied AI.
20:37If I had a problem that a model could have solved a year ago for 10 cents, it's probably still costing me close enough to that. That's what I mean by quote, it's not actually getting cheaper in a lot of scenarios. It certainly hasn't gotten like 30x cheaper for that kind of space. And then there's problems where models still struggle to solve, and those haven't gotten 30 times more reliable. They've gotten better at it, but the numbers don't actually scale to outcomes. This might be the worst part.
20:59I tried something fun recently. I plugged GPT-4o into Open Code to see if it could update a page. It couldn't figure out how to call tools. It spit out broken HTML in the chat log and then stopped responding cuz it didn't know what to do with the history. I think every model released this year is minimum 30 times more likely to solve a task in a harness than a model like 408.
21:25I honestly think people like this need to be forced to use a model from a year ago to remember how bad things were.
21:31GPT-6 Luna is 100 times cheaper than GPT-5 and it's legitimately smarter and better in every way I can think of. 61 Soul is 10 times smarter than GPT-5 and it's still cheaper than five was. If I had a problem that could be solved a year ago for 10 cents it's still the same, I just outright said this is false. The only way it would be the same cost is if you let the modern equivalent model do more work than it did before, probably work you did by hand before, like running the dev server and checking it in browser. I'll self-own a bit here.
21:58I remember making fun of WinSurf when they added the ability for the model to run commands. The way I referred to it is like, "Oh, this is baby beginner mode for people who don't know how to run a dev command. Why would I ever want this?
22:09I can tell the agent which file to edit and I would do it with command K in the file." I just didn't get why you would want to ask a model to run a command.
22:16You can just run the command. I was super wrong on that. I did not see the future where models would use tools and bash to do their work. It seems like he's forgotten about that time even though that time was not long ago. He challenged me to go find a task from three years ago that a model would nail in accuracy. I'm going to hop back to chat. Can y'all name a task that models from three years ago could nail in accuracy? Ugly to-do apps? Three years ago? No. An AI model's not making a to-do app from scratch.
22:44We are so used to what AI could do today, we don't know how to think back to what AI did three years ago. Three years ago it couldn't make a to-do app.
22:52You could set up an environment where it could build, give it the main file and tell it to make changes, but you couldn't tell it to go build the ugly to-do app. You had to set everything up for it and then let it edit a file. What else do we have in here? There was no GPT-2 to-do app now. I like that someone figured out a good example here. Skate Bench. Skate Bench, funny enough, isn't like the best example in the world here. I'm not going to pretend otherwise. Like, Skate Bench is letting models name skate tricks and its ups and
23:21downs, but I ended up running numbers for this because of who I am as a person. I ran Skate Bench on O1. If you don't remember, O1 was the first ever reasoning model.
23:33It was when OpenAI figured out that letting the model spin in circles before making its response could increase the quality of its answers. This model came out in December of 2024.
23:45O1 cost $15 per million tokens in and 60 per million out. That makes it more expensive than Fable. Even better is the cash read price was only half off. It was 750 per million reads. Now, the norm is 90 95% off, not 50% off.
24:02So, this would have been $1.50 modern times worst case. And don't get me started on O1 Pro, which is $150 per mil and 600 per mil out. But, I I could have destroyed him by doing that, but I chose not to. I used normal O1. And here is what it cost to do Skate Bench.
24:19It got a score of 67.7% and it cost 12 cents per task. 1 year later, almost to the day. Actually, it is today. It was December 17th, 2024 where O1 came out and December 17th, 2025 when Gemini 3 Flash came out.
24:33Higher score, not a tenth the price, almost a hundredth the price. 58x lower costs exactly 1 year later. This is where I realized I was wrong. I thought the number was a pretty consistent 10x.
24:46It's not. It's closer to a hundredx year-over-year decrease at a given level of intelligence. And I told my agents to keep going. What is the cheapest that gets close? GLM 5 3 Flash, crazy to have that come back, got a slightly lower score, 70.8% versus Actually, no, it's higher. 01 was 67.7. So, 53-Flash got a higher score at 70.8 at $0.016.
25:13That is 0.00016 per question. 680 times cheaper in under 2 years. How the [ __ ] is intelligence not getting cheaper? It's not just getting cheaper, it's getting cheaper faster. Anyone who thinks otherwise is stupid and wrong. Like, I'm sorry. The numbers speak for themselves, and I was wrong about this, too. I thought we would hit a ceiling for how smart the models got and a floor for how cheap they could be. I couldn't have been more wrong. We're nearing the point where the computer that you run your agents on,
25:42not the one where the inference happens with the models, but the one where it's calling bash, the VM that runs on is going to be more expensive than the inference at this rate. Back to what Zig said here. So, yeah, intelligence is getting cheaper, but in practice, that doesn't mean tasks are getting cheaper at the same scale because the measurement in those benches is not related to actual outcomes. Yeah, because the difference is if the task was fix this bug, previously, you would try out the app, find the bug, go to the
26:10file in the code base, open it, select where the bug was, press command K, and tell Sonnet 3.5 to edit the code, save it yourself, command S yourself, rerun the dev server, open it up in the browser, and see if it worked. The same task now is me handing a screenshot of the bug to the model and saying fix. And yeah, in that case, the task didn't get cheaper, but I'm not doing the [ __ ] work anymore. Sure, closing that ticket on Jira costs the
26:39same amount, but I didn't do the work this time. If I was to keep using Cursor the old-fashioned way, where you select text and then tell it what to do to it, I bet it would have been way cheaper.
26:50But since we live in the real world, where applications are trying to free us up of our engineering effort as humans, no one builds that way anymore. Here is where I started to crash out enough that I realized I had to make a video. He gave the number of 3 years ago, which I thought was insane. Like nobody should compare to things 3 years ago. So I did 2 years instead. 3 years ago, there were no reasoning models. Tool calls didn't exist yet. It made no sense to do this comparison at that time scale, because the tasks models could do were crappy auto complete. O1 came out in December
27:18of 2024, gave the prices. I genuinely struggled to find real-world tasks that O1 could complete. I chose to make it name skateboard tricks. The cost was around 11 cents per question. 3 flash came out in December of last year. It got a higher score and it cost 0.2 cents per question. IDK why you're fighting this, man. It's so obviously true. Try using O1 for any real code work, then switch to GLM 53 flash, and tell me honestly which you prefer to use. It's going to be the one that is more than 10x cheaper. And this is the part that's
27:46hardest is David was very careful to frame things in ways where he is not angry or upset in his verbage, while also triggering me really badly to be very angry in my verbage.
27:58It sucks because this is dumbass versus [ __ ] and they look equal, because one person's being a dick and the other person is being nice, but is entirely wrong. And I just I I need you guys to understand this isn't one of those like right to your opinion things. This isn't two different takes on the same thing.
28:14This is somebody who spends way too much time and energy. Like make fun of me all you want for being too into AI. I know my [ __ ] with this now. I do a lot of work with these models. I have built tons of benchmarks. I am evaluating this stuff obsessively. I have a terabyte of my own history with models in a clickhouse database that I set up on my network, so I can audit all of this. And I am thankful you guys are reflecting on this thing. I'm being too nice if anything.
28:39I appreciate you guys, cuz I feel I felt so insane when I was going through this this afternoon, which is It's I closed on IDK why you're fighting me on this.
28:47It's obviously true. If any real code work was done with these, you would understand." To which she responds, "Okay, then. We can all agree that suggesting intelligence is getting 10x cheaper a year is a silly statement to make. You can run the same exercise aligned to the benchmarks, by the way, and you won't find anywhere near the same pricing delta in applied use. Sorry to push back so much on this. I know you're smart and I respect the hell out of you, which is why it's so genuinely confusing that you have this take. A year ago, the flow was highlight text in editor, command K, read the code by hand. Now, the flow is "Hey Claude, I
29:17want to do X. Figure it out." I asked him to give me a set of real tasks that he used AI for 12 months ago, and I'll turn it into a benchmark to show it is actually 10 times cheaper. He responded with more essays. Yeah, I'm I'm annoyed.
29:28I My issue isn't that like some people are dumb about this. My issue is that these dumb takes have been going viral so often because people want to believe this. They want to ignore AI and say this is just the labs wasting tokens because they can, and they're trying to trick us all into wasting a bunch of money. No. These things are getting cheaper to the point where the labs are losing their margins fast. If you hate OpenAI and Anthropic, you should be listening to me here. You should be joining me here. You should be shouting louder than me about how much cheaper inference is getting and how Anthropic
29:57and OpenAI will have no margin as everything gets driven to the floor.
30:01Here's where I get to my final points on this.
30:03With my version of the numbers, I priced her usage at around 100K a month. With just a few months of further improvements, if she wasn't to reach for smarter models and instead used equivalent models that level of intelligence but cheaper, she could get her costs down to as low as $3,000 in just a few months. And if the trends continue, as low as $100 a month by this time next year. But we're not going to do that because we're chasing the dragon. We want to use the smartest. We want to improve and change our workflows. We want to solve tasks differently in ways that involve us less
30:33or perhaps solve harder tasks that models couldn't have solved before. But here is where my final point is. No one said this is normal. I've seen a lot of people claim that folks like Lauren are trying to normalize the idea of spending millions of dollars a month in tokens.
30:49No, they're not. Lauren is not trying to seem like a normal person. She is very strange. She is a weird one-of-one individual. It's why I've always found her so exciting and fascinating. She's just She's a phenomenal engineer. And I'm thankful nobody's trying to go after her engineering capability right now cuz I would crash out so hard I'd be canceled for life. She's one of the best engineers I know. Way before this AI stuff, too.
31:13Turns out she's also really good at the agentic engineering stuff. I said she's in the .0001 top percent of agentic engineers. Measuring her work as dollars in software out is reductive and wrong.
31:23Her spend should be looked at as a combination of product and eng, research, FTE work with P stack, marketing, and like product research and improvement. Is this worth $350,000 a day? Probably not.
31:36Is it worth a mill or two a month?
31:37Without question. It's not costing that much, but even if she legitimately wanted to spend two to three mill a month, if I was the CEO of Cursor, I wouldn't even think twice. I would ask if she wanted more. She's exploring what happens when you ignore the cost of intelligence. She's not doing it cuz Cursor expects every company to spend 10 mill a year on tokens. She's doing it because that 10 mill on tokens is going to cost $10,000 in a year or two. And of course, Lauren popped in. She strives to be a power user of Grokbot and Cursor so she can give good feedback and make changes and make the experience better
32:06for users. And she does this. I have watched this happen live in our shared slacks and DMs. I think Lauren is a rare gem in a space full of people who admittedly don't know what they're doing. When there are so many like agentic engineers that are just throwing tokens at things they don't understand, I understand the desire to go after these people.
32:27Lauren's not one of them. If you want to attack somebody for how they're wasting tokens, go after me. I'm a much easier target. I'm throwing away 500k in tokens on a rewrite of TypeScript just for fun.
32:39Lauren is actually pushing the frontier.
32:41Steer your frustrations towards people who deserve it more, like me, and let Lauren invent the future, because that's what she's here to do. And please, for the love of God, when you see people sharing numbers that are obviously wrong, or maybe just seem like they might not be right, maybe just screenshot the tweet and throw it at Claude or Codex and ask them to verify the numbers, because even the dumbest AI models nowadays could probably explain to you why none of the numbers being
33:10shared here make any sense at all. The lesson I really want y'all to take from this is how to notice people sharing these types of like fear-mongering things, where they care more about drumming up emotions than sharing information. And also, what one last thing to notice with this, watch for when people go after someone's motivations instead of trying to understand what they're saying. This is the thing I do my best to not do. I care about the information, and when I am mean to somebody, it's not because I
33:40think that they are lying to someone paid them to or something like that. I say that what they said is dumb.
33:45The folks I'm warning you about here are going to do the opposite. They will never engage with the points. They're going to steer around them and say something else, like, "Well, yeah, you're paid by the labs." Or, "Well, yeah, they're doing this for marketing."
33:58Or, "They're trying to normalize this thing." Or, "They are just using the benchmarks to make the numbers look better." They will say anything other than, "You're right." Those people are not good sources of information, and I hate to draw these lines, but it's now at the point where I have to call out a good friend with David Cramer, and say you should not listen to a thing he says about AI, even though I did a panel with him 2 weeks ago about AI, because I legitimately think the way he talks about these things are actively harmful to the space that we are in. If we
34:27cannot talk about the facts, then we cannot have a conversation. And sadly, this level of egregiousness around the facts has degraded the conversation to [ __ ] and people like Zeig/Kramer the founder of Century who has been a sponsor of mine for a while and Zeig's a good friend of mine and even Peter who I've had my back and forth with and like we did not get along initially but we have fully come around and are decently friendly now. I hate calling out my friends like this but I'm calling out my friends because
34:55it would feel a lot worse calling out the dumber people doing this and I know that both Zeig and Peter are smart enough to listen and stop doing this [ __ ] going forward. I hate drama videos like this I really do and know that I'm carefully picking my targets as people who I know and trust and am friends with because I think they might actually fix their [ __ ] but until they do until they publicly acknowledge how wrong they were similar to how I have been doing forever now. How many times have I made the meme that I was so wrong about the
35:23intelligence ceiling or that I had no idea how powerful Wind Surf's idea of letting the agent run commands would be or that I thought intelligence would stop getting cheaper and now it's getting more cheaper more faster or that I didn't think Anthropic would ever be able to do RL and now they have the best RL models in the world. I was wrong a lot and I will gladly publicly shout about this because what I care about is getting the right information to y'all and I am sure the comment section is going to reflect exactly that.
35:49Oh man I am scared to see this one when it comes out but uh I've said what I have to hopefully I've helped you guys better understand these types of things and that you now have some things to recognize when people are being really really wrong on the internet. Uh I'm going to go do literally anything else with my time and until next time peace nerds.
