Transcription
The man, the myth, the legend. Ed Zitron makes his triumphant return to help us burst the AI bubble.
This country was stone cold dead.
Good day again from Newsweek headquarters at the top of the World Trade Center in New York City. I am Carlo Versano and this is the 1600. So yes, as you no doubt are aware, we have Ed Zitron back with us in the cut today. Ed, good to see you, my man. Uh, so Ed is the guest of our most viewed episode on this podcast. So, we knew we had to have you back. Um, we want to talk AI. There's been a lot of developments in the AI space, or perhaps not, but there's certainly been a lot of stuff has happened in the AI space. Um, and Ed, you are one of our great you are one of America's great AI skeptics. Yes, I think that we I think that that's fair.
That's fair, right? Um, so let's there's there's a lot to get into. There's a lot to talk about. I think that
I kind of want to start with and and I should preface this by saying I'm all over the place with this stuff, right? Some days I'm like, "This technology is going to change the world and I'm going to be, you know, I I'm going to be shoveling coal soon cuz I'm going to be of no use to the to my profession." Some days I'm like, "This is totally vaporware, you know, crap that they're just hyping. Uh, nobody actually wants this." So, I don't know where I am. It's I I'm generally a skeptic like you. Uh, probably not as as hardcore of a skeptic, but that's where you come in. Um, and part of the reason why I I keep changing is because I'm influenced by there's so many of these sort of like viral subtag subt substack essays and exposts, right? So, we had one the one that I actually wanted to bring you on originally to talk about was this post that went viral a couple weeks ago. I think it was on X and it was titled something big is coming.
Yeah. From the scam artist from the guy who lied about his open source model reflection 70B.
Yeah.
Okay. Well, well, so so I asked you I said I was like can you come on and talk about this? uh you referred to it as abominable nonsensical hog slop.
Yep.
Uh and you actually helpfully annotated it uh for your subscribers. So
I wanted to talk about that one. But since then there's actually been a new another one of these sort of like viral essays or I don't even know what you would call them. It's not really like an op-ed. It's not really like an essay necessarily. It's almost like science fiction
ghost story.
So this is the one I'm referring to by Catrini Research. Uh it was published on Substack this week. Uh it was titled the quote 2028 global intelligence crisis. So th this this article by the way moved the stock market on Monday. It fills me full of worry that it did so because having read the 7,000 plus words and I know I can't judge anyone for writing that much several times and annotating it myself. I think the best part of it is this whole thing is about ostensibly AI replacing a large chunk of white collar workers. So you'd expect two things from that. one what the AI does to do that and of course two what it replaces. It is not obvious in the piece. Most of the in fact all of the technological framework is yeah AI gets cheaper and then it gets better. How much better? What does it do? That isn't really clear. At one point he suggests that Mastercard and I think AMX their shares crash because Agentic Commerce just becomes agents. Just to be clear, none of this is real. Complete fanciful science fiction bull. Uh but then he says that the reason that they'll eliminate them is that because they will do transactions on the Salana and Ethereum blockchain. What year is this? What year is it when we are like, "Yeah, AI agents are going to use crypto. You godamn do do a real job." But he doesn't actually explain much at all. And that which he does, he lies about. He says at the end of 2025 that Claude code can basically clone any software as a service company. This is just factually incorrect. He's not the only one to say this. Tons CNBC said this numerous times. It's complete nonsense.
They're saying that cloud code can essentially get it can now selfiterate. That's that's the
Also by the way AI cannot selfiterate. It does not learn from itself. Recursive self-learning is a a concept. It is not a real thing. They could like it doesn't have intelligence. It does not learn,
right?
It doesn't have thoughts. It's not conscious. But the big thing in there they were saying is it just AI gets cheaper and better and it can do anything. Interesting point though about that piece I want all listeners to really think about is you'll notice that two jobs never disappear in this piece. One analysts. Two, stock traders. The stock market. The one thing that they talk about in there that is actually algorithmically traded that actually has AI in it. N Capital back in 2012 I think lost a [ __ ] ton of money off of that. The only thing that has algorithms like that has like a strong AI thing and he just n just stocks are traded the same way. Insane. Also he starts it just before the midterms or just after the midterms. Nothing about politics. What's crazy though takes 3/4 the way to mention open AI and anthropic. It's not clear what anyone's doing. It's just ghost stories. There's one bit that I keep seeing quoted as well. God, this boils my blood. It's like we really we should have known that 10,000 GPUs in a data center in North Dakota being able sorry a GPU cluster in North Dakota would be able to do 10,000 PE uh white collar workers jobs. We should have seen that as an economic pandemic, not a panacea. What's great about that is two things. Number one, it's meant to make you think that that's what's happening. that a GPU cluster in North Dakota is doing that. It's replacing white collar workers. It isn't really. None of these productivity studies that have come out have actually proven that AI can do that. Number two, there is actually a data center in North Dakota, Ellenale, just to be specific, Applied Digital's data center that they rent to Corewave. I've heard tell they're losing a crap ton of money and based on my own data center model I worked out, they're losing tens hundreds of millions of dollars a year because of the ownerous debt they've gone into. Now, you would think that Citrini Research $999 a year would have I don't know looked at any of the data center clusters or the cluster that is in North Dakota and I don't know maybe extrapolated from that. No, he doesn't extrapolate anything other than from his own hole.
Okay. All right. Back up for a second here. So, if you if you haven't I know you have a lot of thoughts on this so many thoughts and that's great. That's why we brought you on. But so if you haven't if you're not familiar with this this piece that we're talking about by Catrini research the thesis is essentially this right their argument is that we are heading towards this sort of like economic calamity where you know you're going to get these sharp increases in productivity thanks to the AI
that is going to lead to sort of this big de sharp demand for human labor that of course is going to lead to elevated unemployment and then it begins this sort of like cascade of defaults in the credit market and then in the in the mortgage market And then you have this sort of like negative feedback loop that once you start it, you can't stop it. Right? So there's going to be layoffs. Layoffs are going to lead to weaker consumption. Weaker consumption is going to lead to more automation which is going to lead to more layoffs. But the the the point that they're making is basically that we're going to get this positive productivity shock.
Yeah.
But that's somehow going to lead to economic collapse. I guess that's the where all these people are out of jobs. That's the thing that I don't understand.
So here's the thing. Also, another great part of it is when he talks about how AI agents will be buying and selling houses now. Just there's so much in there when you look at you're like, man, you don't know how any jobs work. You just you don't know any you only know other people on Substack, mate. You got got to go out and talk to someone else.
Well, the problem with all of these things, the these these articles that go viral, it they're just it's basically just vibes.
Well, it's just vibes, but the problem is is that they're written in this serious tone from a respectable quote unquote person. And thus you get the media writing about it. You get the Wall Street Journal, CNBC, Bloomberg. And I saw someone, a CNBC person, I respect them, so I won't name them. They said like, well, it it's um thoughtprovoking. What the does that mean? It's just lies. It's just making stuff up. Now, to be clear, if AI existed that could actually do all the stuff they said it would, it can't. Maybe this would be a worthwhile conversation to have. And it's a thought exercise. as I imagine people have had before. What if automation replaced white collar labor? And their big thing is, oh that white collar labor, the credit markets stop really functioning because credit is based on how likely someone is to pay, which is based on salary. Salaries no longer these the most interesting parts of it are the flimsiest. He also kind of talks about private equity and overvalued valued assets. I think that's valuable, but that's not what this was about. That was just the chum to get you to the scary parts,
right?
There's a lot of scaremongering. No question.
It is only scaremongering. He begins it by saying, "This isn't AR. This isn't dumerous. This is Yes, it is. You attention-seeking little goblin."
[laughter]
So, okay. Let's let's pretend for a moment that this argument that, you know, we're going to basically be in this we're going to get this productivity shock that's going to crater the economy is accurate. Just bear with me. What all these people seem to to mis price in these, you know, in these arguments is an underestimation of the societal upheavals that you would see in a in a place where you're going to have entire sectors of the white collar workforce disappearing if not overnight than than in some people aren't just going to be okay with that, right? upper middle class people who have spent their lives in in, you know, jobs like we have, you know, writing, using your brain in the so-called knowledge economy. The these these are people that have political power for one thing. They they have economic power. They're not just going to sit back and like take a UBI check from Elon Musk or whatever and let their entire careers go down the toilet without fight.
But here's the other thing. There's I mentioned earlier, it doesn't acknowledge the fact that the midterms exist. That's gonna be a significant thing no matter what happens like probably gonna be a blah blah blah. It's gonna be pretty bad.
Would look that way for the Republicans. Yeah. Glammy Sammy. Uh but it's the other thing it does is it does not talk about data center financing. And I know that this sounds like a very me thing like hey you missed out on it.
Yeah.
178.5 billion of data center credit deals done in 2025. Meaningful part of private credit. Meaningful part. Why? And also I assume the major part of the AI bubble and boom sorry
right
doesn't mention them at all doesn't mention the economics doesn't talk about that because I'm serious I have calculated and I've talked to a ton of experts infrastructural people financeers no one can actually square the math on the data centers the data centers do not make sense in Oracle's estimate in their best case scenario they're getting like 35 40% gross margins which doesn't include marketing costs or many starving things point I'm making is this econ economic analysis, what have you, scenario, what have you, doesn't include the major economic factors that will matter. Your own thing about, yeah, if white collar labor is replaced, that would cause political upheaval in a way that would outweigh the stock market. I don't know. I think white collar labor invests in the stock market,
right?
I think that there's a thing called a retail investor. Just leave that off. Nope. Nope. Too much work. I'm just going to say a bunch of vague. And why would you bother? Why would you actually try to write something that's thoughtful when you know your only point of this is to go viral and take advantage of a media industry that does not know what they are talking about and does not do the bloody reading to just understand the basic facts and it's just it's exploiting a weakness.
Okay, so let's talk about data centers for a minute because this has become sort of a, you know, an interesting political story. Um, you've got two parties. Uh, basically neither of them are up to the task of legislating around anything, let alone, yeah, you know, cutting edge technology. Um, but when it comes to data centers, you're starting to actually now see a backlash among Democrats, even some Republicans. Santis is has been out there. This guy Spencer Cox in Utah has been sort of vocal about it. Uh Bernie Sanders yesterday, I think, just said for the first time, he's demanding, he actually may have said this before, but demanding a moratorium on on data center buildouts nationwide. You know, you had you had Democratic governors like who were potential 2028 candidates like Wesmore and Pritsker who were giving tax breaks to these data centers just a couple years ago and now they're saying no. So they they see where the the the public perception is is turning here. And I guess the point I'm trying to make is, you know, is are data centers the the are they the battle that the the right battle?
I think they're the next great financial crisis. Okay. I think that the political side is doing the right thing. I think we need to stop these things because putting aside my own views, they yet to demonstrate the economic value that they claim to. They're getting billions in tax breaks. Where is the economic value? Whereas the
Well, there's no labor. There's no they're not they're not job creating machines. They're just warehouses.
It's the West looting the South.
It really is. They're building in Louisiana, Mississippi, Texas. And what they're doing is they're flying in specialized talent. So these data centers aren't creating local jobs. They might create an economy of the new people who now have to live there to build a horrible loud data center, but when they leave, gone it. It's why Oracle's having to tweet out that they love the residents of Abene and Port Washington, Wisconsin so much, cuz they know that they're creating no economic value. But there is a much bigger problem, which is that the math on data centers, which is actually pretty simple, doesn't make sense. So to build an AI data center, it's about [sighs] it's about 30 million in critical it so servers, GPUs, mostly GPUs, networking and all that gump about $30 million per per megawatt and about 14 plus million for construction labor and all that and actually getting more increasing because it's becoming more expensive to get the stuff. It's becoming harder to find the labor. So you have to pay more for it. Electricity is actually quite expensive to lead to a place because it's not about the power existing. It's about the power physically getting there and working. So you've got all of these massive costs. So about $44 million a megawatt
and how many megawatts per data center?
A couple hundred, sometimes a thousand.
We're talking about billions of dollars.
Yeah. Yeah. And then you get the other. So that's your first problem. So now you have this massive upfront cost. So this means that no one building these data centers is doing so without debt. And because these AI data centers are relatively new, they're mostly getting low rated debt junk grade in some cases. You've got cases where they get around that Crusoe building open AI's server um sorry OpenAI's data centers out in Stargate Abene. They were able to get better terms because they were able to use Oracle's credit rating.
Okay.
So currently coreweave's raising money. They can use Meta's credit rating. Point is all these data centers require all this massive upfront debt, tons of it. And then they have to buy all of these servers full of GPUs that depreciate at best in six years. And also they're obsolete in like one. And then the data centers take a year or two to build. So you have stuff that when it turns on the tech insight's already obsolete and then you've got, I don't know, hundreds of billions of dollars of these things. So you've just got this massive commodity of these very expensive data centers full of expensive stuff that you haven't really started paying off cuz you haven't got the money in,
right?
So you've got this massive upfront cost requiring all this debt. Everyone has to get debt. Debt's quite expensive because it's brand new. Then you have the problem of who are the customers. Where's the demand? I've mentioned it on the show before. It was like 6065 billion of total AI revenue last year. Maybe if you remove all the hyperscalers, less than a billion dollars of AI compute revenue. Let's just say the revenue even exists though. All of your customers are unprofitable AI startups. So you the debt holder who has lots of debt that you need to pay off your revenue source
is another company that cannot pay its bills and must take on equity.
So the great news is that you have gross margins of maybe 20 to 30% if you have constant customers.
You lose millions of dollars. Not put it put aside depreciation spreading the cost out. I just mean the cost of debt and the cost of leaving the servers idle. you can lose millions of dollars a day waiting for that customer to spin up. The customer could be ready, but perhaps they need to set up the servers. And so you're in this situation where that's every AI data center. Every single one is built on these very shaky economics and we're building tons of the and on top of that, a lot of them just don't have the power to them. And even when they do, great. You want have you ever heard of a construction project that finishes on time and on budget?
Of course not. So all of these companies have raised all this money to build these data centers that require their customers which are unprofitable to pay them on debt raised for construction timelines that they'll likely miss and need money for. All it's going to take for this to fall apart is one or two of those to collapse. One or two of them. It will be one that never gets started cuz the debt falls through Lancaster, Pennsylvania with Core. It will be one that's in progress that's not fully built yet. Don't know what that one will be. And it will be one that's finished that goes out of business because the customers aren't there anymore.
You're saying that'll be sort of a
those will be your three those will be your three pale horsemen of the data center apocalypse.
And the thing is, there is no real argument against me. People will say well the economics are better than that. And I'll say prove it. And they can't. Or they'll say well there's going to be insatiable demand for compute. In my example the one where they collapse that's when there is demand
right?
If there isn't demand they'll collapse even faster. Right now, the largest user of compute is open AI. They just I think that they just estimated they're going to spend $600 billion on compute by 2030. Where's the money coming from, Sam? Who's got the money? Where's the money going? They owe Oracle $300 billion. How are they going to start paying that? They have to start paying that next year. Don't worry, though. Oracle hasn't finished the data centers.
Well, Open AAI says they're going to do ads now. It's going to come from ads and chat. GPT.
Yeah. FD estimates low billions of revenue from that. OpenAI Sam Alman nove it was October 31st Halloween last year he said on the B2 Gen podcast I believe he said we're making way more than $13 billion and that was in October turns out CNBC reported down he made $13.1 billion of revenue that's not way more Sam it's almost like you're making up your numbers on the spot also my own reporting says numbers are different point is now this makes sense now that the money doesn't exist it's all coming out of someone's pocket right
it's all relying on debt and equity And who's issuing that? Private equity who is having trouble exiting venture capital which and I think since 2018 uh the return from venture capital has been on average.8 to 1.2 uh per dollar invested like real dog numbers. Everything is relying on a system that doesn't exist anymore.
So Derek Thompson, who I know you have thoughts about uh the the I guess he's not at the Atlantic anymore. He's independent, but he's he's he writes a lot on AI. He's he's more of a he's less of a doomer, I guess you could say, right? But but he has this thesis that basically to your point, one of two things is happening, right? Either we are in the world's biggest private sector spending bubble in history and the bottom is going to fall out because, as you say, there's where's the money, where's the revenue, where are the customers, right? So that's that's option one. Option two is the that the revenue does come along, right? Which at the scale that it would have to to justify these investments would imply a huge investment from from corporations and governments which would imply a huge amount of labor displacement.
And this is actually a fascinating point I was thinking about on the way here. So Claude code, which we'll get to in a bit, I imagine. Claude code, the most no one will shut up about cla code. Claude code everywhere. You got the Atlantic writing these furious pieces, but you got the Verge, the NBC, everyone's saying Claude code this biggest thing ever. Everyone is supposedly replacing Codus. Don't believe it. But the point is it's doing this 2.5 billion in annualized revenue. So about $28 million in that's
in a month. Yeah. $28 million. That is a lot of money to you and me.
But for the product that's allegedly the most important thing that everyone uses that's changing everything, that's it. And that's the thing that's with them subsidizing it.
Claude code. So there's a researcher called Chillac who did a great they mathematically worked out claude code allows you depending on your subscription. So $20, $100 or $200 a month. They allow you to spend $8 to 13.5 per dollar of your subscription. So on a I think on the $20 subscription it's over $140. uh that $200 a month, you can spend over $2,700. Everyone using Claude Code right now is in an illusory environment. They are using a product that will not last long term, that they cannot afford to serve. So that this is the point you were making last time last time you were here, which is that the thing with with these these LLM companies is the the their power users cost them more money, right? as opposed to like a Facebook situation where like your power you your power users have these you know you have these network effects where you're you're making money from Instagram that's feeding into to Facebook.
It's also just not that expensive to serve an Instagram or a Facebook user unless you attach an LLM to him which they did. But that's the the other thing is just what people consider claud code what they have been impressed by which we'll get into is running at an egregious subsidy at a subsidy that if they had to pay the real API rates I think every single person flipping their proverbial about this would walk away. I I think that
it's funny money is what you're saying
it is but it's also manipulative. So think about it like this. If you did not pay the cost of gas and you drove a Humvey, you'd just be like, "Yeah, I should I'm just going to driveing Hummer wherever I want. Drive 300, 400 miles. Who gives a crap? I pay him a hundred bucks a month for my gas." You detonating 500 to $1,000 a week in gas. Doesn't really matter to you. You wouldn't think about, you wouldn't get anxious about that. Similarly, you're around with clawed code, you're going to forgive it for a lot of its problems. You're going to when it takes longer, when it messes something up, when it gives you something useless, when it doesn't understand something you've done, you'll be like, "What if it was costing you 50 cents each time? What if it blew 10 bucks, which it does, just going in circles?" Because these are large language models, and the more they reason, the more they have a chance to hallucinate, right? When you're coding and it hallucinates, it might set up a service wrong. It might ignore a very crucial security document. That means that it will just leave a door open on something. it just won't set up the security settings and it may tell you it did. These things you might forgive if you know what you're doing. You go and fix it. Fine. But what if it was actually costing you money every time? What if that dinky little website you made that you think cost you part of your $20 a month subscription actually cost you 15 bucks or 10 bucks or even five bucks? What if every amendment you made to it was 50 cents to a dollar? These are the real economics and that's based on the API costs of what anthropic charges. We have no proof that they're profitable on inference. I don't think they are. But nevertheless, even if they are, they are allowing their subscribers to burn 8 to 13 and a half times every dollar they're making on Claude Code. That is not a real business.
Well, so this reminds me of Uber, right? So Uber,
it's so much different.
Okay, so tell me. So Uber, so you know, Uber famously makes a ton, you [snorts] know, it does all of this customer acquisition by offering rides that are highly subsidized, right? So you get used to taking in New York City, you get used to taking an Uber versus taking a yellow cab because it's much cheaper. Then once you're in the system, they jack the prices up. Now it's now it's way more expensive. Isn't that what's going to happen with these with like cloud code, which is $20 a month? Yes. But with Uber, the cost difference was still within the realm of what you would pay for a cab ride, right? It would be much more. You're no longer getting those $5 drives anymore. Passively 10, 15 bucks.
That cost multiple is still within the realm of the psychology of what a cab ride should cost.
I'm not talking about Claude code costing 200 bucks anymore.
I'm a month at least. I'm talking 1,500 to three half to$5,000. I'm talking about the people using the $20 a month one who are probably spending more like 150 bucks. And these aren't overestimates. These are real numbers. Even if Anthropic is making a 15 50% gross margin on on the cost of running these tokens, I don't think they are. There's no proof. Dario Amade said in a toy example, gross margins would be 50%. He's just doing it so people take that and run with them. Point is, even if you cut the cost in half, they're still blowing several hundred% each. This isn't something where, for example, with Uber, it would be like instead of you paying a flat rate for a for a ride that you know on the spot, it would be that you're charged a metered fair and that metered fair was four or five times what you were paying and jumped arbitrarily. If the cab driver, I don't know, he was on his phone, maybe he didn't really understand. He he misread his GPS. He was thinking of something else. He went down a weird road. Yeah, that little trip just cost you another tenner. Why? Because that's what it costs, baby. I don't think people realize how different Claude code would feel. And this is even before we get to efficacy if they had to pay for every mistake. Right now, they do not. People will argue, well, anthropic will just they'll just jack up the prices, right? Five 10. You think any of these people are going to act do you think that people are going to go from paying $200 a month to $2,000 a month and be like, I love this. This is good. I'm going to do exactly the same thing. And you've got the situation where people are getting used to their workflows using the subscription. They're not ready to pay API costs. Now, the tempting thing to say there is, ah, it's the drug dealer model. It's the drug dealer model.
Yeah, but you're not going to have coders selling the stripping the copper from the walls and like stealing from their grandmother to pay for Claude code. They're just not going to use it. Or they'll use whatever the next subsidized model is. The moment the subsidy stops with one, they move over to another. Then they move over to another. Then they have all of them have like multiple accounts and multiple things so they can run their limits hard. Open AAI codeex is doing the exact same thing. I just don't know exactly how bad Shell's working on it. It's just it's an illusion.
But is it the idea that if you're going to use something like Cloud Code in to automate all of these different parts of your life,
then you're going to be in this you're in their ecosystem. So you will pay.
Am am I though? Because seriously, what actually differentiates anthropics models from any of the others? People really can't tell you. Now, people like Claude models for coding, but they also kind of like codecs for coding. And there's also Moonshock. There's Kimmy K2, I think is for not remembering the Chinese models. People like all sorts of crap. Yeah.
Claude code as an interface can be run with different models if you use the API. The API, by the way, is just how you pay for the tokens directly.
Okay. So like if you're a company, sorry I really should have leed this. If you're a software company and you add an AI feature, you would pay anthropic or open AI per million tokens. And I think Claude Sonnet 4.6 is $3 $3 per million input. So stuff you feed into it. $15 per million output. So stuff it spits out. So with code,
you're burning a lot of tokens. And those little little mistakes, very expensive little mistakes. Also, anthropic doesn't charge, they don't count. There's also cache reads, which is when you know how RAM works.
No, not really.
So, it's actually pretty simple. So, if there is a file your system needs to use regularly, instead of reading it off the hard drive ever every time, it has random access memory. So, it just puts it in there so it can quickly reference it.
So, the cache is kind of like that. It's kind of a dilutive thing. Someone in the comments is going to say, "Ah, it's not perfect." you do an interview. But it's just what is the Claude ecosystem? What is it? What like Claude Co-work? It does. Claude Co-work. They said it was written 100% in Claude Code. Yeah, it shows it deleted a guy a guy's wife's photos, all of them, because he tried to clean up her desktop. Like, it's like, yeah, it sucks and it doesn't really And that was one of the examples from the website. Yeah, I would believe that that was written 100% in include code and that it it appears to be crap and no one can really tell me what it does. I mean, so I and again, I I'm not a a hardcore user of these models. I I I find Claude to be the the better one, right? Um, you know, I I I fed a bunch of blood results into it the other day and it was, you know, it it saved me from having to go to the doctor to get those results, although the results are probably that, you know, completely wrong.
I got to say, mate, if I got a blood test, I'd probably want a doctor to read it just in case.
Yeah, I think you're probably right. and but but it was notable that I was able to put it in it. It it it spit out to me in English what the the results were were showing me. And again, maybe it's wrong and I probably should have a doctor look at it, but here's [snorts] another example. I I have a friend who told me that he and again, I wouldn't do this personally, but he fed all of his like the last year of his bank statements in there and he asked it to create a you know, a program that tells him where he's doing recurring money that he's not using, right? subscriptions that he's paying for that he, you know, he forgot about and he said that it was able to flag a few things, some streaming subscription he forgot about. I'm not saying this is worldchanging stuff,
right?
But it's it at least strikes me as as having some sort of value prop that is more than just it being a, you know, the plagiarism and slot machine that it has been.
What is the value prop though? He was able to make your life a little easier.
Rocket money. I pretty like there's I get like I get like 18 ads for the subscription thing a day on Instagram. I I don't know. I'm just kind of like and also Anthropic just raised $30 billion. You're going to like it. That's the thing. It's like
what is the thing I'm looking at here? What is the thing you are doing? What is the thing you are changing? Cuz Spotify for example. Spotify said uh in December I think it was that their best coders so very vague bucket basically don't write code anymore.
Right. So I spoke to someone on Spotify actually went and found out found fun fact uh they pay from what I understand they pay for claude code subscriptions so they're already their cost of doing business is completely subsidized. The way I worked it out is they've got 2700 engineers if they spend half of the limit of the $200 a month subscription it would be about $40 million [snorts] a year in API calls. So that's like 10% more than that. I think they spend like 400 something million dollars on salaries a year from those engineers. It's like you're just increasing your employment cost by 10%. But then there's the other thing which is okay I don't know if this is true. I actually don't believe it. I think there are some people who are doing it maybe but I don't know how many of them. Let's say 100% of code is written. They're writing it in AI. So what are you shipping more software? Is the software better? Are you rapidly iterating? How is this changing your team? what is it doing? And I don't mean this in a dismissive way. I mean there is if the situation is is that instead of writing all the code again don't know if this is happening just saying instead of writing all the code a machine's writing it and the output is basically the same if the productivity is basically the same.
Haven't we just changed the modality a bit? Haven't we just moved one thing here and there? and they'll say, "Well, it solved 11 bugs." Were all of those like huge bugs? Was one of them just like a best standards thing?
Yeah.
It's just if 100% of code was written by Claude, for example, in this scenario, what is the actual reason we should be impressed? That it can do it, which is I guess cool, but like like okay, that it can do it. How good is that code? How do they know it's good? Because if you're writing all your code with with Claude, aren't you eventually going to stop learning how to write code and solve problems with code? Because
But isn't that sort of isn't that sort of the point? Isn't the idea here that it there are, you know, and who knows how true this is, but the idea is there are all these people out there who have, you know, great business ideas or app ideas or whatever and they weren't they haven't been able to, you know, they don't know code, so they haven't been able to actually build these products. If the coding can be done and you can just say into the model build me this app with these parameters and then you know run tests and and and iterate it that isn't that going to open up some new level of innovation amongst the you know, the plebes out there
so I think well one it isn't happening
right
like we have where is the vibecoded company that's so amazing where is the vibe coded feature vibe coding is a scam
right so
v coding is just this just what we're talking
just the idea that you can just build software without there's two things. So with the vibe coding thing so I think every single article I've read about a non-coder which I know I am uh building software with code and being amazed is just it's key jingling. I'm sorry it's not that impressive when they build a website. You'll notice that a lot of these websites are purple. It's really weird. It's the constant use of purple. That's because a lot these models are trained on openly available open source software and on website templates like Tailwind CSS. Tailwind CSS, I think in the mid2010s, they had a lot of purple. You had a lot of startups at the time that used that purple. Stripe, for example.
When you look, it's always the same kind of boxy thing. It would be like
animation behind big text, little text, little text, box, box, box,
box, box, box.
This is the coding version of what what these LLM are doing with text, which is they're spitting out these the these cliches and you know, it's not this, it's this. You can tell something.
Exactly. And it's you can kind of see it with vibe coding software. All of the web apps look the same. All of the mobile apps look the same. And also software engineering is much more than just code. To get something functional, stable that works in every region that is operable when you change things that can be changed in the future to grow because you're going to be any software company is constantly maintaining and changing bits. What happens in 5 years? Do you remember what you told the model? because you didn't build this with any intention. And then there's the other thing of this whole software being like that companies will be able to vibe code their own Salesforce or SAP. SAP is a great choice. So I said SAP down like 10% I think last month. I forgive you.
Remind me what SAP does. I know there's one.
No, no, I'm going to because basically no one knows because they're so monolithic. Nevertheless, SAP down about 10% I think in the last month. So SAP is enterprise resource planning. It is software for everything. It is HR, procurement, it is accountancy, it is compliance, which is a big one if you're a giant company or if you're in anything like
trans anything related to transit or shipping or anything like that.
No, I've used these and it's I mean
they're horrible. They they suck. They suck absolute donkeys. It's so bad. However, those systems are not just code. They are interlinking pieces of information that are and they are massive databases. There is actual server hardware they need to run on. And that server hardware needs to be specific to the database. Different coding languages have different requirements of memory and require different things when you set them up. Depending on when you build a software, you may have to update it if a currency changes.
Mhm.
Does your vibecoded software know that the French stopped using the Frank? Do they know that that's different to the Swiss Frank that does exist today? These may
and those are the things that AI gets wrong all the time.
Or maybe it gets one of them wrong. Maybe it gets something wrong, but even then you have to when you build something that's sock two compliance on account, you have to call in an accountancy team to verify it. You can't do that on your own.
And then when you've done all of this, great. I hope that you have a law firm
because you're going to need a law firm for this ERP system, especially if you have your customers touch it. And woe beside you if you have your customers money touch it, cuz then you've got a whole another legal thing. And man, if you have one there, one little up there, um, who do you sue? Yourself. So, I assume you're going to buy just an entire new engineering team to run your internal SAP to your internal Salesforce, your internal, I don't know, Netswuite, uh, your internal Microsoft 365, your Slack, your Canva. Canva, I guess, wouldn't be a great example of a big SAS company, pardon me, but
Adobe.
Adobe. Well, even then, yeah, no one's no one's building their own Photoshop. Well, I mean
there are there's the open source one, [ __ ] That's a real one. I'm serious. But this is the thing. People seem to think you just go, "Okay, cl code build me Salesforce." You hit enter and Salesforce just erupts onto something because that thing's going to have to run on on like a cloud service and it's going to have to connect to all your other systems.
You're saying there all of this plumbing that takes years and years and and skills to build. And that's before you get to the fact that not everything's written in the same code. And even those things that are
and now there are software engineers who barely understand some code. I'm not saying all of them. They may love this. But the truth is that there are practical concerns that LLMs cannot fix. LM can write some code. They they are able to write functional code even sometimes. The problem is it doesn't always work. It isn't always perfect. So your vibe coding thing is just a mess. But even then, let's say you're you have a team of software engineers who really know their code. They're sharp. They're using LLMs. That team's skill is going to atrophy. And this isn't an insult to them. But think of it, if you are used to doing a a manual action and you have a machine start doing it, the muscle's going to break down. You're not going to learn it. And the thing is LLMs also they need to be constantly updated on the model providers end so they don't experience model drift which is when it's not even anything wrong with the model necessary. It happens with all machine learning. It's the world changes so they need to change so coding standards might change or it may just be the things you feed into it are new problems it's never seen. the everything the coding models and SAS itself is this massive infrastructural plumbing effort that is not just code it's and it's very frustrating because you read all over the business press saying this thing about there's anxiety is over AI replacing things cuz anthropic launched a plugin
Charles Schwab got dumped because of an accountancy tool [snorts]
it's just IBM had its worst day in 20 years because of that to treat me
cobalt so cobalt runs on mainframes. You familiar with mainframes? There are specialist computers that basically never
go down. They run on cobalt. IBM sells them. Cobalt also is not just a programming language because it does so many different things. It has different compilers and plugins that work with different things. I sure hope Anthropic No, wait. They didn't train on anything. It was a single blog post cuz Anthropic's biggest business model is lying.
They did a This is a somewhat subject change, but it's relevant. They put out a thing sometime last year where they said it coded for 30 hours without losing focus and it built a Microsoft Teams Microsoft Teams or Slack competitor. Now you may think, can I see it? No, you can't. Can I see the code? No, you can't. What'd it do for 30 hours? We're not going to tell you. If only we had a tech or business media to ask these questions. One or two of them limply put, yeah, and they didn't show us it.
These people are just saying also 30 hours, man. How many tokens do you burn? No, we don't need to tell you that.
Truth is that if the AI companies actually cared about telling people the truth, they would go out of their way to correct these things. They would push to make Open AI and Anthropic would both push to make it less frothy because they would want to sell the product they actually have. But they sell on deceit. They sell on hype. And Anthropic for all of this importance only made $4.5 billion last year and they lost 5.2 billion. Dog company. Terrible economics.
Okay, so let's talk about let's let's keep talking about anthropics. So there's a big story in the news right now um related to uh anthropics um contract with the Pentagon, our our boy Pete Hegath, our Secretary of War. >> Uh so they have a contract to use Claude with the military to to basically embed Claude into the military. It's the only LLM that they say can actually do this, can actually work with classified. >> Do what? Do what though? They never say that's the thing. They never say >> well. So here's the thing, right? So, so essentially the story here is that is that HEG has given Anthropic, which is a customer of the Pentagon, a deadline of Friday. So, today if you're watching this, >> so the Pentagon is a customer of Anthropic, >> correct? Right. Right. Uh so they uh uh they've given them a Friday deadline to open their model to open Claude for unrestricted military use or else uh the Pentagon says they're going to kill the contract or worse or they're >> going to make them a supply chain risk. >> A supply chain risk which that would be a big deal. I mean that that's effectively nationalizing one of our frontier AI labs. >> I think it basically says that they that no military adjacent contractor would be able to use them. I'd need to look into that. I've heard from military experts who say it's not a big deal and some who say it's the worst thing in the world. Swag I >> so anthropic is basically saying we don't want we're not going to let the the US government the defense department get you know get into the source code and use it for specific things like the automated killing of people or mass surveillance of Americans.
So I got to be honest first of all it's not even the source code they're just changing the system prompt. It's just the thing that it loads in with. And also what do you mean anthropic? your LLMs don't do autonomous robotics. They can't. That's not how autonomous robotics works, >> right? >> Also, no one seems to have really cracked that within Silicon Valley. Andreal loses money. It doesn't and they had to discontinue use in in Ukraine because it wasn't particularly useful. Putting all that aside, I think the funniest part about this is anthropic. Anthropic is doing this for marketing. I don't think >> that so that was going to be my question because they Anthropic. No, no, no, of they they market themselves as the you know the moral AI company, right? So, it's in their it's in their interest to get these headlines. I think that like, you know, they're standing up to to the Trump administration saying that they're you're not going to use our technology to surveil Americans. But you're saying that's BS.
>> I mean, what do you mean surveil? What do you mean autonomous robots? And the answer is anthropics LLM do basically the same thing as everyone else. They are the same. They use a lot of the same training data now that they have post-training which is when you feed you you kind of correct the pro you correct the outputs and you feed more specialist data into it. They don't have a special thing like Grock is going to work with them and I I assume say generate something illegal but it's just like what is the difference between anthropics and Grock and GPT? Well, presumably he H heath wouldn't be giving them a deadline if he didn't if he didn't think that they needed that technology. They're >> no offense to the secretary of war or whatever he is. I don't think Pete Heesgth is an LLM expert or indeed user, >> right? >> I would be shocked. It would be really funny if he was just like a hot he like subscribes to semi analysis. He's he's he's like reading a Dan nice like reading about Taiwanese ODMs. >> Yeah, I don't think so. >> Oh my god. I can't believe Williams. Anyway, that's that's a bit for later. No, I just think it's they're just trying to make all of them do it. They're just like, we want to be able to use you for whatever and Wario Amade is just like, yeah, well, you couldn't possibly use it for war. Mass surveillance. Okay. What does that mean, though? Because if the answer is feed a bunch of data in and looking at it, you can do that already. I don't >> That's Palanteer's whole business model, isn't it? >> That's I mean, Palunteer is just evil. They're evil. They're evil. Um, Accenture, >> right? there they have forent >> they have forward deployment engineers that go in and customize the service. It's like the it's a natural outgrowth of the 1980s the insaw promise software and you know who ran that [panting] Robert Maxwell. No, we're not getting into Epstein stuff but um >> it all comes back to Epste No, this is weirdly separated from this AI bubble stuff. But other than like Ellie Yudkowski who was in the emails, we got to stop.
Um, but back to Hezg, here's the thing. All these LLMs basically do the same thing now. They have different features. They have different strengths. But I don't know, Google's worked with with CBP before. Do you hear anything about Gemini? I'm sure they'll just use Gemini. But also, what are they doing with it? What are they doing with it? What what what's the thing they're doing? Oh, control robots. >> Screw you. It's not real. You're not controlling robots with this thing. You generate and kill this. They do that already. >> Yeah. >> Yeah. They're going to Oh, will Anthropical be better at it? How? No, they just honest to god I think they're just with them. I think that they just >> Yeah, I think No, I think that >> Yeah, I >> I think he just doesn't like him. >> Yeah, >> I mean if you've heard Dario and for my I did the haters guide to anthropic last week. >> I can't stand any of these guys. >> No, but Dario is the most annoying of them. >> You think he's more sanctimonious than Alman? >> 100 [snorts] times. He sounds like Elizabeth Holmes's voice that she did in front of investors. He sounds like it's really weird when you hear it. You can't unhear it. No, I think that I think Anthropic is milking this opportunity to make themselves seem special and different. I think that they know that there's no real value and like if they did it, it wouldn't cause any more damage than is already being done. I don't know. And by the way, I also agree that you shouldn't use these things for these things. Obviously, I don't think that's a good idea.
>> I also don't think Dario Amaday cares. He said he wouldn't take money from the SS. What did he do? $30 billion round 37 different companies. Pretty strange. changed their whole all their AI pillar, their safety pillars. >> They don't give a none of them give a none of them give a AI safety is being used as a marketing thing. Notice that AI safety never talks about the economic problems, the bubble, the stealing, the ecological damage. It always talks about this theoretical future AI that's going to I don't know make you leave Skynet or I don't know. None of them have watched like Upgrade. That was more of an that was a device. That movie rocks. But these people aren't really thinking about anything at all. People are just like, "Well, that confirms my biases or it says the thing that my investors want to hear, so I'm going to go along with it." It's just very cynical. It's very, it me off when I see someone be like, "I want to keep you safe." And this involves you buying something from me and my friends. It's It involves me taking you taking me very seriously and buying my product. Not Oh, by the way, this product I'm talking about today does not do any of the things I'm scared of. None of them, right? None of them.
>> But it might in one day. >> But what if it does that? Even anthrop my favorite anthropic thing is the thing they did. I forget if I brought this up last time. They do this thing where they're like, "Yeah, it tried to blackmail us. Tried to blackmail us and then take advantage of the system." When you actually read the paper, they prompt it to do that, >> right? >> And it's like it used an email tool. No, it didn't. It generated an email. Well, you see these sorts of So, here's a headline from just today we were looking at a minute ago. This is a new scientist. AI can't stop recommending nuclear strikes in wargame simulations. Leading AIs from Open AI, Anthropic, and Google opted to use nuclear weapons in simulated war games in 95% of cases. Uh, okay. Well, that strikes me as bad. But you don't have any idea like it what's the is the prompt like >> do a nuclear strike? >> Well, you can probably find the prompt, but also it's trained on the internet. It's trained on social media posts. It's trained on whatever crap they can get. It's going to be the summation of human human consideration online. So yeah, I think most idiots online would be like >> blow them off the face >> hell yeah blow them just because the chump and also if let's take this to its logical end point. So this is a war game scenario. Where do you think they got them from the various war game forums? Do you think the people on there are like couldn't possibly could possibly no. I swear to God if they it's like there's this one game that I really should remember the name of where it's a very realistic tank simulator like very very realistic and they constantly I think it's is it war I forget it's war something and they constantly have problems where people will post classified information cuz they'll be in an internet argument someone will be like I think that tank truck they'll be like no based on this document I have and it'll be a classified document and it'll be something minor but that it's just like it happens all the time might be war fury but anyway way. That's the kind of they're training off of. It's like, yeah, the model the and I wouldn't be surprised if they deliberately let that slide cuz they've definitely tried because they were like, "Yeah, [ __ ] it. This will create a headline one day." And it's just >> so much of this is kayfabe. It's just everyone freaking out but not really believing anything. Not really having the proof for the stuff that they're freaked out about, but being freaked out about it allows them to claim they're ahead of the game.
Well, it just seems like nobody really knows anything. So that because that it it creates these the these opportunities for for doom and gloom or for you know these utopian you know Elon Musk talking about how the you know Tesla robots are are going to come and and save us all from having to go to our lowly jobs and whatever. It's what's going to save Elon Musk from the billions of dollars he's going to have to pay Vista Equity Partners for the giant uh Colossus 2 data center he's building. Good luck with that. I I genuinely think the Colossus project may actually destroy Elon Musk. >> That's a good topic for another time. I don't even >> so much money.
>> Um well, okay. So, before I let you go, Ed, this is theoretically a political show, although seems like a tech show today. That's all right. Um it seems to me like the the biggest problem that these AI companies have, well, one one of their many problems. >> This stuff has no natural con constituency, right? People don't want this. I I I think right I mean I think people are interested in hey oh yeah maybe using claude code to you know to to to get my blood results better or whatever there's some of that but I don't know anybody who's like I see the future and this is going other than the people who are actually in the to stand to make a lot of money regular people don't want this they don't want their jobs to be displaced >> so they're being they're obviously very skeptical um based on a lot of these people being the same people that sold them a line bull about social media being, you know, connect us all. So, I think everybody knows that they're being hoodwinkedked by these from some of these guys. And, you know, talking about Altman, Mickey, can we bring this up?
>> Oh, I know the one. >> One of the things that is always unfair in this comparison is people talk about how much energy it takes to train an AI model relative to how much it costs a human to do one inference query. But it also takes a lot of energy to train a human. And it takes like 20 years of life and all of the food you eat during that time before you get smart. So the fair comparison is if you ask ChachiT a question, how much energy does it take once its model is trained to answer that question versus a human? And probably AI has already caught up on an energy efficiency basis measured that way. >> God, he's a this guy. I can't stand this guy. >> What a first of all, anybody who talks like human beings that way should be nowhere nowhere near this technology at all. He looks like he was generated in an Xbox One game. He's very strange looking fellow. >> He is. Um, >> but what a what a moronic. >> But it's just like I've Yeah, I've I've consumed a lot of energy in my 41 years of of life and but [clears throat] and it hasn't all gone to economic productivity, right? It's also I don't know there's like what what's it's been expended in unquantifiable ways and how I love my family and help my friends. >> Just not how energy works. It's just the kind of thing that you say if everyone's told you you're smart your whole life. I think that also people would want this if it did what they said it did, which it doesn't.
>> I think people right now would like everything to stop being bought by private equity. They would like groceries to be cheaper. They would like health insurance that the government paid for versus them paying a company that then screws them. I think people would like their jobs to be easier and this doesn't even do that. And also they'd like to stop reading a headline every day that's like rich guy excited to make you bankrupt because that really is what it is. And also rich guy who has not worked in a long time who truly does not do work who has more money than you'll ever collect. If you got hundreds of people together they would never touch their wealth. They want those people well, they probably want some violent actions to happen to them but in reality they want those people for all of this talk to actually have something at the other end. >> Yeah. And I think that the regular people I talk to, they're just like, "Yeah, it's kind of like a search engine." I get that. Most people have that reaction. >> Yeah. >> Regular people I meet who are vibe coded are like, they get to a point when they're like, "Oh, right. There's all this messy infrastructure. Oh, right. My app looks like everything else and it doesn't the toy software that they build isn't practical." I think also the data centers are obnoxious. They are obnoxious. They are demanding and the output isn't interesting. the output isn't really inspiring. The only people that it's inspiring are deeply craven and cynical people who frame themselves as optimists >> who claim that they're excited about the future. Meta's running a commercial right now. I don't know if you've seen it or not, but it it's about it's I I forget. It's like it's a thesis of the commercial is this this this data center is coming in. It's going to like revive the economy. It's like what are you nobody works at these things? >> What are you talking about? It's just a big It's a big room with servers in it.
>> But regular people hear that and they go, "Wow, my groceries like 20 30% more expensive." Wow. >> Right. My energy my housing cost my housing my energy costs have gone up. Oh, well, how much is Meta putting into a Louisiana based Hyperion data center that they've deliberately used accounting to keep off of their balance sheet? $30 billion. How much do anthropic race? 30 billion. How much is OpenAI raising? $100 billion. We are told the money does not exist for basic social services. Yeah, these fucknuts can raise 30 to hundred billion dollars of just because they feel like it. And then they put out something to CNBC or the information or the Wall Street Journal. Oh, we're going to spend another $600 billion. Oh, and by the way, the reason we're doing this is so you won't have a job. No wonder people hate these people. Yeah. No one. And I think >> Well, they need a better spokesperson. That's for sure. >> They need a better product. They need to stop burning cash. They need to build a real business because what they are doing now is guaranteeing that nobody's going to be there for them. And David Saxs has made the point, not David Saxs, but he said there's not going to be any bailouts. I think the only company that could get bailed out is Oracle cuz they're they're >> well Ellison's Trump's boy. >> But even then, I genuinely think an Oracle bailout will not fix what Oracle stock price will be when it can't pay its debts. I think the collapse of Oracle could change the world for the better. But I'm deadly serious. That's the only bailout target because these distressed data centers, I don't think people understand how TAP worked. It was meant to be 700 something billion. They only did >> troubled asset relief program 2008 great financial crisis. We're not going to get those for data centers or if we do it will be quite small.
>> Well, there'll also be pitchforks in the streets if we have companies. >> That's the thing. If they think people hated bankers, >> people hate data center. banks at least provide some sort of service. >> Also, banks invested in the data centers, private equity. It's like, hey, we need to bail out the people you love the most. Tech people, private equity people, and banks. You love those, right? It's just like the worst poss and this will what uh I think this could pop this year. I think it's going to be a very I think there's going to be potentially months if not years of data center collapses. So yeah, going into the midterms, do you think Donald Trump's going to be like, "I got to save the I'm going to save a black d. They're the most beautiful company we've ever seen. We need to save Corweave. We love Michael and Trader. Don't we?" No, that's not going to happen. >> You do a great Trump. It's It's very It's really It's very strong. >> We're Dario. Um, sorry. >> My man my boy Dario. >> Have you seen him? Have you seen >> You see you seen Claude? I don't trust Claude. It's a French name. [laughter]
>> Um, so okay. Okay. So, bottom line, Ed, you remain, this is what I wanted to get out of this conversation. Yeah, you remain as skeptical, if not more skeptical than you were last time. You still believe the bubble will pop. Um, now I say all this, but have you seen this video of Trump how he would look as an Asian man? I will leave you with that. Just an ins just insane. were just doing racism with it and people like there look look we're doing something kind of racist like an a early '90s level of racism and this >> the first AI video that I actually stopped >> you had to feed the equivalent of 13 gram of Alsation meat into it to generate that like and and now we're doing '90s level racism what like it's Adam Corolla lm [laughter]
>> okay uh Ed tell us before we go um first of all where people can find your work and if you're uh if you're working on anything anything new coming up. >> So, uh, working on the haters guide to private equity. >> Oo, that's going to be good. >> It's a banger. It's already a banger. I'm going to start doing more regular pieces at where's your edge. And a better offline.com. B t o ff l i ne.com. That's how I spelled. That's the podcast. >> And yeah, check me out on one of those various vehicles. I'm also on blue sky. >> Not Twitter though. >> I am. It's just Twitter is just it's like a PvP video game. It's just you go on there if you want to fight. And I don't really have much fighting. And yeah, it's gotten so bad. Okay, let's leave it there. Ed Zitron, appreciate you, brother. >> Thanks for having me.