Transcription
I don't know how they avoid a WeWork-like situation because when WeWork tried to go public, people saw the economics and said they look like a dog's bum on Thanksgiving. Sorry, Christmas. I guess it would be in England. And I think it's going to be worse for OpenAI. I think that they're going to have lost so much money. It's going to be so dramatic. It's going to be hard to get investors excited. And just to be clear, if OpenAI or Anthropic fails to go public, it's bedtime for this whole thing.
Hello and welcome to the Tech Report. I'm Louis Sykes. We're joined today by Ed Zitron, the author of Where's Your Ed? and of course the host of the Better Offline podcast. Ed, welcome back.
>> What's up? How you doing?
>> Good. Thank you, Ed. It's great to be speaking with you. Well, Uber, T-Mobile, Walmart, and and other large companies may be starting to look at scaling back their usage of AI and and LLMs over concerns it could start to cost them more than just hiring a person. What part of this is that there are these big concerns about, you know, tokenization, just the costs of running these things. It seems like a pretty big product flaw, doesn't it? If these things can't uh work any more efficiently than than a than a human being would.
So the thing is with AI, the big problem with large language models is it's actually very different, difficult even across harnesses like Claw Code Copilot CLI and all the different ways you can use them and also all the different models to measure the cost of one unit of work of what one thing will take. Everyone's estimating, which immediately creates a problem of how do you measure what something might cost? Okay, well, how do you measure that? How do you measure the actual return on investment? And this conversation that has started, and it all started last week with Andrew McDonald, COO of Uber, saying that they were having trouble justifying using using LLMs because, well, they couldn't really track them to any actual use. So, you've got a problem where you can't measure what something costs. You can't measure its return on investment. You call that something without ROI.
So, all of these current token caps, like the T-Mobile one at 2K, same with Uber, sorry, Uber's, 1500 a month, I think they're capping engineers at. They're doing these limits as preliminary things. These are going to drop further. People think that this is the limit of it. They're doing this right now because they don't know what to do because they can't just say don't use it at all. People will freak out. That's actually what happened at T-Mobile. I reported this earlier in the week. They tried limiting to people to as little as 30 bucks a month and people started flipping them, flipping a little bit. What I think is going to start happening across the board is you're going to see 2K, 1.5K, 1K, and then question mark from there because these companies don't actually know how much AI costs. Don't know how much it costs. Don't know how to measure the return on investment. Don't really know what to do. But it's kind of cargo cult stuff. They're doing it because the market kind of obtusely rewards them using AI. They're investors who don't do any real work. Business idiots, I call them. Go like, "Oh, yeah. You should use more AI, mate." No one really knows why they're doing it. So, everyone's doing things in a very symbolic way. They're like, "Okay, we should limit it to 1,500 bucks." Is that good or bad? And the problem is is no one knows. No one knows the correct number.
But the really funny thing is that I reported earlier in the week on BRICS. BRICS is a credit card company that actually got acquired by Capital One. I believe they have limited people to, I think, around 2K a month on OpenAI's Codex. But the funny thing is is that OpenAI has been giving businesses a few months of free Codex. So you're now getting businesses that have not been paying for Codex at all who are now facing token austerity. And as this spreads through the industry, I think there's going to be a lot of anxiety. And that anxiety is going to be most profoundly felt by OpenAI and Anthropic who must grow. They cannot slow down.
And how far could this go, Ed? I mean, you know, if if developers are continuing to struggle uh with with with the tokenization of these systems, what is stopping them from just dropping these tools and going back to to just doing, you know, what they did before AI and doing the work themselves?
>> So, the first problem that they have with pulling this thing out entirely is kind of like a tick. You can't just wrench it out or leave bits in there. People have got used to using these tools for better or for worse. Doesn't mean the code's good. In fact, uh talking with people at Zillow last week, it's made the codebase worse across the board. It's just because when AI AI code is written because these things don't have thoughts, they don't have feelings. They don't have knowledge. They just pull from a corpus of information that they've been trained on. So code is very verbose, meaning there's lots of it. And it's written with no intention. There's nobody who's because when you're in theory, when you're writing good code, there's a certain logic behind why you built things a certain way. If you don't do this, this will break. If you don't break, if this breaks, these three things will break. So, you think of it like that. So, you've had years now of people screaming, use AI to write as much code as possible. So, you can't just wrench it out immediately because, well, AI wrote most of this code, so no one really knows how it works. So, there might have to be some kind of exploratory level of, okay, how does this work? Putting aside the LLMs, but fundamentally I think it's inevitable that costs get pulled back because you have, I think someone at Nvidia said, you've got people spending as much in token uh and LLM tokens as they are in employee headcount. That makes no sense. It's just a nonsensical thing to do. But even if it's 10% of headcount, if that's not providing any ROI, if it's not making things ship faster, everyone talks about lines of code being written. That's a terrible measure of software. It just means it's like, "Wow, you wrote a 2 million word book, but I can't really understand it." I realize my newsletters are long, so I'm somewhat of a hypocrite on that one. But think of it like this. If you're just writing a lot of code, that doesn't mean the code is good or stable. And in fact, the more code there is, the less control you have over it, the more cumbersome it will become.
So, this is actually going to create a great deal of uproar within these companies because you've kind of you've kind of got this mess of code that's been built with LLMs and you've got all of these engineers who are kind of used to using it just because they've been yelled at to do so. And I think that there are many secret imbeciles within the software engineering community. People that could barely cope before that really likely conned their way into the job or were very low level or they're just good diplomats who have used LLMs to appear more useful than they really are because they're able to fudge their way through certain operations. What this means is these companies are currently built on sand. They're built, they have rotted away their codebases. They've rotted away their institutional knowledge and indeed anyone who has gone from writing code regularly to using LLM has weakened themselves as a software engineer. They have removed their ability to do their job on some level. Moitar, a fantastic YouTuber, has made this point many times that the more you give up to these LLMs, the less proficient you are as a human. And so this doesn't mean that these things are replacing jobs at all. These things are not doing that. There's no actual data to prove that. It's just weakening people, giving them a kind of easy button that actually creates more problems in the future.
>> So, if the output is is maybe not going to be as big as was hoped, is there at least a way to bring down these costs?
>> That's the thing. I don't actually know if there is other than just cutting them entirely because think of it like this. You don't know how to measure the return on investment. You don't know how to measure the cost of a task. How do you bring down costs? This is actually the biggest problem that OpenAI and Anthropic are facing. Just think of it like this. You have said, "Okay, we've making all this money. We're growing faster than any software company ever." If you believe that their revenue projections and all that, you need people to keep spending as much as possible. The problem is people are spending an absolute butt ton right now, but they're doing so without any reason other than they are like the pays of Anthropic. They've just been brutalized by the media and investors into using as many tokens as possible. Problem is is that's not a stable business. It's hard to say, okay, if I'm spending $12 million a year on LLMs, why, what would I get if I spent 24 million or indeed what would I lose if I spent 3 million or 1 million or 500 grand? There is no compelling answer here other than you'd be able to use it less because you pay for an LLM's outputs whether they're good or not. If an AI makes a huge mistake, if an AI just as regularly happens with software engineering, spins its wheels and goes in a loop and kind of sits there and burns a bunch of tokens for no reason, you pay for that regardless. So companies are right now just basically experimenting with this. And so it's not clear how they would reduce, what costs would they reduce? They could potentially use cheaper models. DeepSeek. China clearly doing some geopolitical moves there, permanently discounting DeepSeek V4. And uh another one of the models, I can't recall, by 75%. Clearly just to weigh down these companies. Uh someone close to me described it as like financial Afghanistan for the American AI labs. It's really remarkable actually how weird this has become because you have Anthropic theoretically growing faster than anyone. And I think their annualized revenues are weird, but they're doing so off the back of just a bunch of people just with a fire hose going tokens everywhere. We use as many tokens as possible. Use even more tokens. And I I have this theory, the era of the business idiot theory where it's the people running these companies don't do any work. They have no connection to productivity or production themselves. So all they know how to do is hire people, fire people, spend money, uh and also go to lunch and read emails. That's really all they do. So they think, "Oh, more AI is good. I don't do any real work." When I pissed around with it, well, I made an app kind of come out of it. So if a software engineer uses this more, even more big good. These people do not have nuance thought or planning. So we've got this supposed AI boom that's mostly coming out of executive incompetence.
And well, one one person that seems to agree with you, funnily enough, uh, Ed, on the issue of cost is Sam Altman, who himself has admitted the cost of running these these models is a huge issue. That's a direct quote at an enterprise event on Tuesday. Has that set the alarm bells ringing yet?
>> Little bit. Little bit. Mr. Altman's allegedly filing to take his company public soon. And Mr. Samuel formaldehyde Altman, you can't be saying, "Yeah, our our our customers have a huge problem with the costs of our our software." You, your OpenAI has projected to make $284 billion by the end of 2030. Like this is less than four years away. And he has predicted that to do that, you cannot slow down. None of this can slow down. Anthropic is in the boat. They're meant to make $174 billion in revenue in 2029. You, you can't slow down, mate. You can't be saying, "Oh, oh, the cost. People are really freaking out about our costs." Because that suggests that there is a price ceiling on all of this. The amount of data centers being built are being built under the impression that this will never slow down. Because here's another fun fact about all the data centers being built. For there to be enough demand to to satiate the amount of compute being built, we need three or four other OpenAI and Anthropic sized compute spenders. We don't have those. No one. Google isn't spending. The only reason Google, Microsoft, and Amazon are spending that much on compute is to serve it to OpenAI and Anthropic. And so this whole boom, this cannot slow down yet. It it has to keep growing. They have to be at 100 billion of revenue in the next year or two. There is no, you can't avoid this. So, if Mr. Altman's saying this, well, that's probably a bad sign. But the other thing is is OpenAI only moved enterprise customers along with Anthropic to token-based billing, which is paying the actual cost of tokens versus subsidized subscriptions a few months ago. So, this is all happening very quickly and it's mysteriously happening exactly when people have to pay the real costs. And this is because the AI industry conned everyone by offering subscriptions that allowed you to burn way more in tokens. Like $5,000 worth of the 200 buck a month plan on a subscription. Every single journalist that has written about AI has done so using a product that will not exist. They have done so using a product that they are not experiencing in reality. Imagine this. Imagine you pay for a cab service. You don't get charged by the mile. Just point to point and you go, "Oh, I'll go from LA to San Francisco. I'll go to San Francisco to New York and it's the same cost regardless." And then one day some schmuck says to you, "Well, no, no, no, mate. We're moving to a new mile-based system, so you're going to have to pay by the mile." You might suddenly not value the car service quite as much. That is what's happening to the AI industry. It's nothing like Uber. When Uber raised their prices, they raised them from ridiculously low, like 10 bucks for a journey that should be 50 bucks. But it's still the cost of a cab. This is the equivalent of if Uber went from charging 10 bucks for a ride to charging five grand for a ride, even 500 bucks for a ride. Just a significant jump that is beyond the realm of what users expect and indeed have been trained to experience. Think about like this. If you've never thought of how much your tasks are costing, you've never acted as if tasks cost anything. You've run, you've basically done the all you can eat model. You don't know how much your work costs and when you have to actually pay those costs as GitHub Copilot users are finding, you can't do anything. You can't do any of the stuff that got you excited about this nonsense to begin with.
>> And the the other end of the of the question from from costs then Ed that we have we have touched on I think today already um is is what is all this cost going towards? What's coming out the other end. Um, invisible output is a term we're starting to hear from some industry watchers. Is there some trick that we're missing to to to quantifying what the actual output is that AI is is creating? Because, as we've discussed here, there is, you know, a lot of work going in. There must be something coming out the other end.
>> Well, that's the problem. So, capitalism only does this when it's very sick. SemiAnalysis put out a report saying that there was dark output and it mostly came down to, yeah, AI is so special and it's a AI stuff is so good that you can't measure it in traditional ways like return on investment. You have to come up with alternative ways of measuring it. This is nonsensical. This is just it's it is cop it is full copium. It is ridiculous that we're even humoring it. But if you look, we've looked for years, the only way that they've ever been able to measure how good these models are is using benchmarks that are rigged for them because these things can't use computers and use tools in the way that we do. So, you have to coddle them like a rich kid. You have to be like, "Oh, oh, wow. Look, it can beat this benchmark that it has to use because it can't do regular work. Oh, look. Oh, we've done Anthropic claim this week." Well, not claimed, they reported that they said, "Wow, they're shipping eight times more code than they were in the last four years." That's actually a bad thing that more code means you have to review it more. And if it's too much for a human to review, you have to use an LLM. And an LLM can hallucinate, which means that there might be serious bugs in there. By the way, Claude Code's source code leaked a few months ago. Wonder how that happened. Point I'm making is the reason that they're having to come up with weird mystical ways of measuring this is because they don't have a way of just pointing to it and going, "That's good." And actually, I think the easiest way to describe this problem is why are we still describing the value of AI? Why do we have people in the news saying, "AI is here and it's real"? You don't have to say that when something's real. You don't have to be like, "Hey, this thing's real over here." You don't have to convince people that something's real if it's real. They just look at it and they go, "That's real."
But the iPhone, when the iPhone came out, I I bought the first iPhone, bought it in State College, Pennsylvania when I lived there. Bought that iPhone and when I showed it to people, they went, "Wow, voicemail that you just hit a button and the voicemail plays versus going through the number and hitting three to skip the ones you have to go through. Wow, I could just look at my text and my text look like an an app." Well, didn't use the term back then. The application looks like an email. Looks you can send an email on your phone and it's just a touchscreen. Wow. People immediately saw the value. They saw the internet in their phone. They went, "Wow, this is really cool." Then the iPhone 3G came out and people were even more excited. There were very tangible reasons to be excited about mobile cloud computing. Cloud computing previously, Amazon Web Services, for example, before that you'd have to run your own Sun Microsystems server. Now you can do an AWS storage bucket. You can run your services and project it to the entire world. Obvious value. LLMs. Well, you can you can write all your code using another way and maybe it'll be good, maybe it won't. You'll be able to ship more software in theory, but even then I don't see anyone shipping good software with this. In fact, if you look across the board with software today, software is mar like it's obviously worse. Pretty much every app is worse. Everything. BlueSky. I like BlueSky a lot. But that website like for a while you just couldn't crop images and they are big time Claude Code people. I only in fact I saw someone mention this in social media. I think AI is actually reducing productivity because I don't see we don't see the benefits but we can point to a great many of the harms. We can point to software that's less stable, software that keeps going down. Amazon Web Services went down twice because of an Amazon AI coding tool. An AI coding tool deleted a guy's database. Like there are tons of examples of this thing, these things causing trouble. No one really can be like, "Oh yeah, and we were able to completely ship this product. It's great. And we did so in a week and it's stable and good." Anthropic, they took a week to ship Claude Co-work. Yeah. And then the very same week it deleted all of a guy's photos. That's what happens when you ship software fast and you do so using a tool that doesn't know anything. It breaks. It's bad. It sucks. And I think that they want us to come up with special little special little measurements that make them feel so special because these are rich kids. Dary, Sam Altman, and their investors. They're just rich kids. They're just rich kids that have never had to live efficiently, never had to prove themselves cuz the press coddled them, cuz they captured the business and tech media. And now that they're actually having to prove themselves, they can't because they haven't got anything.
>> So Ed, the the the stories and and and and the idea around AI is that this increased efficiency is going to shift the world economy in part because, you know, AI is supposed to be able to do the work that that other people would do, things like writing emails, word processing. Um, we are starting to see some of those shifts in the world economy. There are things like pretty significant issues with youth unemployment across the West. What is causing that shift if it isn't AI?
>> Well, in 2022 there was massive overhiring. So 2021, 2022, there was the zero interest free era. Cash was readily available. Massive overhiring there. But also we we broke mentorship in society decades and decades ago. We don't have fellowships anymore. We don't have mentorships. We don't even really have managers. Managers mostly exist to be cops. AAB includes managers. Middle managers are a pox. They're a poison in the veins of society. We have built a very manager-heavy society, but managers are meant to be kind of signal callers. They're meant to move things out of your way. They're meant to help you organize your your work, ideally your organization, and they're meant to actually move things along. What managers actually do is managers sit there and go, "Oh, that's good. That's mine. I did that. Oh, good. Yeah, my team did the work. Yeah, I did that too. That's what a manager does. A manager does not mentor. Organizations do not mentor. So, finding employment as a young person is absolute hell. And it has been for honestly over a decade. 2008, most millennials grew up into that. I graduated into that myself. It was hell. It's been hell for a while. Finding a job is also hell because most systems use, I think it's ATS or something. It's like a No, that might be planes. There's an automated HR system that most companies use. Indeed, is a cancer on society. Monster, all of them, all of the job boards, all they do is obfuscate actual employment. They don't help people find jobs. They help sell advertising. And they help there are tons of fake jobs as well. And the HR systems automated. So, you're a young person trying to find a job. I saw a job listing the other day that said 5 years experience with AI agents. That's a term that was invented two years ago. What you, what do you mean 5 years? What would which you completely insane? But that's because the means of production are disconnected from the companies themselves. HR departments don't know how the job works. The person that you're interviewed by is a middle manager who doesn't do anything or maybe the CEO who also doesn't do anything. Young people can't find work because the people that run companies don't do anything. So you've got this weird cludgy thing where no one can quite get in and those who do do so in the most desperate abused terms. So you're in this situation where yeah, they're going to say, "Oh, AI did this. AI is the reason this is happening." Absolutely zero proof that's the case. They want that to be the case. The companies want you to believe that. Any company that or any investor who mentions this is just doing so for their own gains. And any journalist who says it has not done the research. And it's frustrating because it's yet another way that workers are being suppressed. And AI itself, AI's actual outcomes and efficiency uh efficacies even is not the reason that people can't find work or are getting laid off. People are being laid off because bosses don't understand anything and want to save money. I have some people at Meta I've talked to. There are people being fired at these companies who are really good. These aren't people who were on performance improvement plans. These aren't people who were kind of just diddling around. These are people who were very important cogs in the machine. And it's breaking these companies. And it all comes back to one thing which is that the people that run companies, managers and executives and vice presidents are not doing work. They are lazy. They are disconnected. They don't care about the customers and they don't care about their employees. And what happens when you run a company like that and that's most companies is the company rots from the inside. AI is this exciting thing for them where they go, "Oh, finally I all that's standing between me and greatness is work and this thing will do the work." And then much like my employees, I can just steal it from them. So, it's a real scumbag economy. And I think that anyone covering this needs to go into it and realize that pretty much every person saying that AI is taking jobs is doing so so that you'll buy or invest in something that they have money in.
>> Well, yeah, indeed. The the investment side of that is certainly going to be uh in in the spotlight in the in the coming days and and this week certainly. Um, AI giants are charging ahead with their initial public offerings. Anthropic, I believe, becoming the first to file theirs this week.
>> Yes.
>> There's going to be a lot of money flowing into this, smart or dumb money either which way, isn't there? Despite some of these concerns about costs and the economy more widely.
>> Yeah, I think SpaceX is horrible. SpaceX is a big lossy mess of a company that kind of several different companies stapled together. If we had a functioning SEC, that company wouldn't be going public at all. They're going to try and merge it with Tesla. Again, functioning SEC would stop that. Anthropic and OpenAI cannot wait to see Anthropic's S1. OpenAI, well, maybe I might have seen some stuff. I don't know. I couldn't possibly say, but let's just say these companies not ready for the public markets. They have they never will be. They are unsustainable, unprofitable. Their financial stories that they've shared with the press are deliberately manipulated to make things look good. They are non-GAAP, so generally accepted accounting principles. These companies are dogs. To quote Gordon Gekko from Wall Street, they're a dog, pal. These companies are terrible. They're awful money burners. Worse than SpaceX. SpaceX is a terrible company that loses billions of dollars because they stapled an AI company to it. If you ripped XAI out, it's not not a brilliant company, but could be much worse. OpenAI and Anthropic are absolute dogs. OpenAI especially, I think these companies have probably lost tens of billions of dollars in the last year. And I think that it's only getting worse and the only way that these companies can become any kind of profitable is financial engineering. So in reality, I mean, OpenAI and Anthropic shouldn't be allowed to go public. There should be laws and regulation against companies that are this unprofitable, this unsustainable and have no path to profitability ever seeing the markets because that is dangerous for retail investors and it's dangerous for regular people that are in indices that might or might not include them.
>> Which which which begs an even bigger question of why now? I mean, all major areas in the US and global economy are struggling with with pretty major problems in their supply chains. Um, around the war in Iran. Other global crises as well that have been having an impact for years now. AI is absolutely no exception to that. Things like helium shortages we're hearing about are causing problems. These it's dangerous for any country where any company that's that's going into an IPO like this. What at this moment it doesn't look like there is much resilience in these in in in global economies, let alone these very new companies. So this is exit liquidity for investors. They are being told by investors it's time to go public. Also, both OpenAI and Anthropic have raised at too high a valuation to raise again privately. They could, but right now both of them are bonking their heads against the trillion dollar mark. That's very high for a private company. It makes the odds of a public offering in which you'll see any kind of reasonable multiplier. Because think about it, if you invest at like around a trillion, you need that company to be three, four, five trillion, which would make them bigger than like TSMC and Meta. It's just nonsensical. I mean, 5 trillion would be Nvidia. I think Google's a little under five trillion. Like, do does anyone seriously think Anthropic is worth as much as Google? I mean, if so, you might have a gas leak or mold, black mold, because these companies are desperate. These companies are desperate for more money. Daniela Amadeay, uh, who I think is the chief of staff, Anthropic's CEO. I forget her role. She's Dary Amadeay's sister who's the CEO. She said the other day that the reason they're going to the public markets is they need more access to capital, which is insane considering Anthropic has raised $75 billion in the last 3 months. I mean, these companies need more money. They need to dump onto the public markets. They need to become some kind of respectable so that they can raise debt. The problem is they can't really raise much debt because when you lose, I don't know, 10, 20 billion dollars a year, just picking a number random, uh, yeah, the credit markets might be a little bit like, "Hey, you're going to exist in two years? You're going to run out of money anytime soon? Let me, you sure you can pay me back? Because I don't know." And in fact, we saw this this week because Broadcom, which is a chip manufacturer, is doing a deal where Google buys the chips from Broadcom with an SPV that then buys them from Google, which then leases them to Anthropic. But part of the deal involves Anthropic borrowing money for it. I know the thing is, Anthropic, when asked by some investors to share financials, declined. Yeah, just a not not for me, thanks. What's funny is some lenders said no. Some said yes. And I think any lender that is just like, "I don't need to see your papers, mate. I think that they deserve to lose everything." I think that or I don't know, maybe we should make it that that's not legal. Maybe we should regulate this. Maybe we should regulate lending. Maybe we should regulate how companies can borrow money because people will say, "Oh, it's going to cut research and development. Look at what we've got from these companies allowing them to have as much money as possible." Can you point to any of this? Can anyone point to any of this and tell me it's worth a trillion dollars? It isn't. Not even close. So the only reason they want to go public is to make themselves rich, to make their investors rich, and to spread the ri to socialize the losses, to get these losses out of their own private investors' bags. I personally do not see how OpenAI goes public. I think when they publish their S1, I think Anthropic might not be far behind, but I I have some feelings about OpenAI's finances personally. I get, I don't know how they avoid a WeWork-like situation because when WeWork tried to go public, people saw the economics and said they look like a dog's bum on Thanksgiving. Sorry, Christmas. I guess it would be in England. And I think it's going to be worse for OpenAI. I think that they're going to have lost so much money. It's going to be so dramatic. It's going to be hard to get investors excited. And just to be clear, if OpenAI or Anthropic fails to go public, it's bedtime for this whole thing. It's time to wrap this up. It's closing time. And that's if we make it because right now the market is souring. The I think the revenues are slowing. I don't see how this continues. And to be clear, putting aside my bare case, for this to do the thing that everyone's promising, these companies will have to be the size of Google in four years. Both of them. Both of them will. They both projected that. Is that I don't think that that's very likely. Oh, and also they will need to have founder business model. They they've also not done that. They have no path to profitability. They're dogs.
>> Well, uh, as we've mentioned, those uh those IPOs are almost certainly incoming now. Um, so I'm sure you will be watching those very closely in the in the coming in the coming weeks and months, Ed, and we can uh we can discuss them further. Ed Zitron, author of Where's Your Ed? and the host of the Better Offline podcast. Thanks very much for joining me again on the Tech Report.
Thank you.