The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron

I think generative AI is at its heart a con, and seeing these ultra rich, ultra powerful people lie through their teeth turns my stomach.

The word "con" is a strong word.

Well, what do you call something where from the very beginning they've sold it in terms of magic, but it's just a half-assed autocorrect machine? They are misleading the entire world.

You are the first person I've spoken to who has that opinion.

Well, the fact that this is happening is insane, and the fact that it's not a scandal is insane. I've been in the tech industry for 16 years now, and I love technology and I'm enthusiastic about it, but I don't like being misled. This is the largest non-consensual push of technology in history.

The AI Myths Game

The host proposes a game: presenting things Ed considers myths about the AI industry.

Ed takes each claim in turn. First, that the AI industry is creating enormous economic growth — no, it isn't. All of these companies run at a horrifying loss; OpenAI lost $20.9 billion last year. None of them can honestly say they're on a path to profitability, because they can't.

Second, that AI will replace all human jobs. That just isn't happening, and there's no economic data to support it.

Third, that the United States needs to spend trillions to beat China in the AI race. What does the race do for us, beyond constantly worrying about China? People keep asking what happens if these models fall into the wrong hands — they're already in the wrong hands, held by Mark Zuckerberg, Sam Altman, and Dario Amodei.

Reacting to Zuckerberg's statement that "we'll continue to invest aggressively in infrastructure to meet the demand," Ed compares it to Lord Farquaad in Shrek: "Some of you may die, but that's a risk I'm willing to accept." If only these people cared about poverty or actual problems in the world, rather than whether enough GPUs are being bought.

The host introduces Ed Zitron, noting that he holds several controversial opinions that contrast with other guests who have appeared on the show, and asks him to state them directly.

Zitron's answer is blunt: generative AI is, at its heart, a con. It is not sold as honest software. In his view, the industry overstates what the technology can do, what it will do, and the underlying financials, to the point of misleading the entire world — while actively exploiting weaknesses in journalism, in the economy, and in the responsible parties themselves, from sell-side analysts to governments.

The host pushes back that "con" is a strong word. Zitron stands by it: what do you call a product that has been marketed from the beginning as magic — something that will replace all jobs and cure cancer — when up close it is boring cloud software that is extremely expensive, unprofitable, and unreliable at its core?

Anticipating questions about his credentials, Zitron notes that critics say he lacks finance experience. He has spent 15 to 16 years in the tech industry, in PR, with practical experience and genuine enthusiasm for technology. Yet along came a product everyone told him was the best thing since sliced bread, and it cannot even do the basics — it cannot even do search. When you ask AI enthusiasts about their setup, he says, they describe a "Pee-wee's Playhouse" of hacks: you need this harness, this prompt with this model, but not that model at the start, only at the end. And this is supposed to be artificial intelligence — smart, autonomous, something you set and forget.

The host lists the six leading AI companies on the table — Anthropic, Amazon, Nvidia, Microsoft, OpenAI, Google — and asks whether Zitron is really saying their fundamental business model is a con.

Zitron's answer: their revenues are not really coming from AI, and until fairly recently none of them were. Right now, 70% of all AI revenues across those companies come from OpenAI and Anthropic — two unprofitable, unsustainable companies that literally cannot afford to exist without those same companies giving them money. Amazon sent $50 billion to OpenAI this year, plus $5 billion to Anthropic; Google sent $10 billion to Anthropic. Over the next three and a half years, based on actual sell-side analyst evaluations — the estimates that inform whether a stock goes up or down after earnings — the market expects $400 billion or more in revenue and roughly 30% of cloud growth from these two unprofitable companies, which will need to be given the money from somewhere.

On top of that, he argues, these companies show such low respect for the average investor and the analysts that they do not even disclose their AI revenues. The few times they deign to, they cite an "annualized run rate" — which means nothing, because they never define it. It could mean month 12, month 13, or the last four weeks times 13; it is different every time. Sometimes they simply do not mention it at all.

So you have what is billed as the biggest, most influential change to software ever, and whenever you ask how much they are actually making from it, the answer is essentially "I couldn't possibly say." These are public companies — at least the ones that are not Anthropic and OpenAI. When they have good news, they tell you. When they don't tell you something, that speaks volumes.

Have you used these AI tools — Gemini, Anthropic, ChatGPT and so on — and found no value in them? There's some value, but it's limited. These companies have spent over a trillion dollars in capex.

Capex means capital expenditures. When you're a business, operating expenses like electricity come right off immediately, while capital expenditures are long-term investments that are theoretically one-off — a data centre, or indeed the GPUs you put inside an AI data centre.

A data centre contains GPUs, which are computer chips. AI GPUs are much bigger and much more power-intensive. They take a lot of high-bandwidth memory, and because of how many of them you need — thousands, tens of thousands, in some cases hundreds of thousands — you need a huge amount of power.

As an example, OpenAI and Oracle are building a data centre in Abilene, Texas: 1.2 GW, called Stargate Abilene. Within it, each of the eight buildings will hold 50,000 Nvidia GB200 GPUs.

The city of Bristol takes about 7,800 megawatts of power a year. Stargate Abilene is condensing more power than that — 1.2 GW — into a space around 1,172 times smaller. Bristol covers about 1.2 billion square feet; Stargate Abilene is about 998,000. So you're condensing all of this power, all of this money, all of this labour into one spot.

These data centres cost billions of dollars, and all of these companies other than Microsoft are now taking on debt. They've spent over a trillion dollars so far and want to spend another trillion next year. And for what? To make tens of billions of dollars, most of which comes from two unprofitable companies, Anthropic and OpenAI.

One rebuttal to the spending argument is that customer adoption of OpenAI and Anthropic has been absolutely insane. These are the fastest-growing products in history, especially in technology. Hundreds of millions, even billions, of people use these tools every day for problems they have subjectively decided need solving. Money is a lagging indicator of value, so one could argue the companies are simply investing ahead of monetization.

But is that adoption honest? When you load Google Docs, Gemini screams in your ear. When you load Word, Copilot bugs you. When you shop on Amazon, Rufus has opinions on what socks you're buying. This is the largest non-consensual push of technology in history. ChatGPT has been screamed about by every media outlet for three years: "This will take your job. You must use this, or you'll fall behind." People are using it because they've been told to, constantly — and mostly as a search substitute, partly because Google fell behind on search and partly because generative search is sometimes better at ingesting queries. It's like a trawling vessel: not good at specifics, but if you ask "does this thing exist? has this person ever said anything like this?", it will probably still get it wrong, but it will scour the ocean for you.

Nevertheless, that's not worth a trillion dollars — none of it is. The money being sunk in is incomparable to anything, even railways, and it blows everything out of the water because there is no post-bubble story. AI GPUs aren't useful for other things either. It's a directionless egregore of capitalism, a headless beast lumbering around desperate to seek out growth everywhere, in the hope that if it harasses people, scares people, and demonizes labor enough, people will be forced to use it.

The reason I pause is that I think about my own company. Some listeners use no AI tools at all; others use them for everything from coding new software to writing to images. Enterprise adoption stats say 88% of organizations regularly use AI at least once for a particular business function, and in our company, 95% of people use one of these tools — Anthropic, ChatGPT, or Gemini — every day. Usage exists on a spectrum, from super users on it nearly every hour of every day to, say, executives who use it less because their job doesn't require it as much. Looking at how the world is changing from a content perspective, it's obvious these tools are being widely adopted — part of the symptom is the AI slop all over the internet. So I struggle with the idea that it isn't being used.

It is being used. But here's the thing about the slop: before AI slop, we had SEO slop, because Google incentivized producing the lowest common denominator that would rank well in search. There's a whole story about how they pulled back spam guards — thanks to Prabhakar Raghavan — making the internet worse by allowing worse content to rank higher. That's why searching "best washing machine" used to surface eleven horrible blogs that read like somebody got a concussion: built to rank, not to be read by humans or to be good. AI helps weaponize that at scale. Yes, you can make a bunch of generic slop — we've had slop for years. We've just found a slop machine.

Beyond quality, there's also the problem of cost. When you use AI services, you burn tokens, and pricing is per million tokens.

"What's a token?" It's roughly three-quarters of a word — it's measured in characters. The AI companies have a currency they charge you in, like a taxi meter in New York, and they call it tokens. For ease, think of it as paying per word.

You're charged per million input tokens — the stuff you feed in, like a document or a codebase. Output tokens cover both what it spits out at the end and what it does while thinking. If you ask for the best restaurants in an area of New York, all of that searching and reasoning counts as output tokens too.

With a monthly subscription, you don't see any of that. They just have rate limits — you can use the service a certain amount, and they obscure what that amount actually is. SemiAnalysis, a big analyst group, recently found that on a $200-a-month ChatGPT subscription you can burn $14,000 worth of tokens, and on Anthropic's you can burn $8,000 for $200. Even on the $20-a-month service you can burn $400.

Most people have no idea what AI actually costs — they just think it's $20 a month. All of these companies run at a horrifying loss; OpenAI lost $20.9 billion last year because people can burn as many tokens as they want. When they tried to move enterprise customers — companies bigger than 150 people — onto actually paying the cost of AI around March of 2026, to quote Sam Altman, people had a big problem with it. Enterprises immediately started freaking out. Uber burned through their entire annual token budget in three months.

So after everyone saying AI is the most productive thing ever, changing everything, the moment people actually had to pay for it, they hesitated. Obviously we all love it, but it's costing too much, so we need to reduce the cost — because people are just dumping stuff into it asking "what do I do here" and getting whatever the median answer is, which is what these things provide.

"So someone like me, a power user of these tools—" I could be costing Anthropic or OpenAI $1,000 while they only charge me $100, say. They're subsidizing $900 of my usage through electricity and data center costs. "And your assertion is that that's unsustainable?"

Yes — though it's probably not one-for-one; it might be 30-for-1. We don't know exactly, but it's unprofitable. These companies don't disclose it even in their audited financials; they play funny games with how they categorize things. On top of that, standing up inference — the thing that produces output in these data centers — isn't just flipping a switch. You have to provision the GPUs needed to meet demand: buy too much and you've wasted money, paying hourly GPU costs regardless; buy too few and customers can't use it, get angry, cancel, and go elsewhere. Either way, they'd be selling $20 or $40 of value for a dollar.

The simplest way to explain it: if they genuinely believed these services were worth the cost, they'd charge what it's worth, and regular people wouldn't get a flat monthly subscription — unless there's an economic problem. And it's simple: you pay when you use an LLM regardless of whether you get what you want. When these things hallucinate — say you're coding and it goes through a codebase, messes up a bunch of stuff, breaks a bunch of stuff — you're paying for that, whether it works or not, unless you're on one of these subscriptions.

The really interesting question is whether AI companies are spending ahead of the value showing up—which is presumably what they would argue—or whether they're subsidizing all of their users in a way that's unsustainable and will never be justified. Looking back through the history of technology, companies often lose money to grab market share, while also focusing on bringing costs down and making themselves more profitable. But they can't afford to underinvest.

If they were bringing the cost down, they would have brought it down—which they have not. It seems to be getting more expensive. In fact, no inference providers seem to be profitable, and even the companies renting out GPUs don't seem to be profitable. I imagine it wasn't that they started out knowing it was unprofitable and decided to keep going anyway. I don't think it's some big conspiracy. They probably thought at some point that this would go profitable: the chips would catch up, customers would pay for the overwhelming value—because you don't know in 2023 where things will be in 2026, you assume it's going to go up. That's the nature of venture capital.

But they should have stopped around 2024, when OpenAI lost over $5 billion. They should have said, "Yep, this is not going to work." Instead they kept going because it made the number go up so much—it pumped stock values, Nvidia, Microsoft, everyone. And not from the revenues.

Here's the funny thing about Google, Microsoft, and Amazon: for years people have been saying their AI bets have paid off, while these companies refused to say how much they were making from AI. Their existing businesses kept growing—through price increases, changes to how Google and Meta did advertising, Amazon bumping up prices and changing how it sold, and Amazon actually starting a remarkable ad business during this whole time. None of that had anything to do with AI, but because the number went up and revenue went up, everyone credited AI—on the assumption that these companies wouldn't spend a trillion dollars for no reason.

Except in fiscal year 2026, which just ended for Microsoft (annoyingly), they made, according to Bloomberg, about $34.33 billion total—$24.1 billion of that from OpenAI. That leaves about $10 billion in a year when they spent $115 billion on capital expenditures, with plans to spend $175 billion next year. The math does not make sense. I imagine their plan was that this would just get exponentially more valuable and at some point the costs would be outpaced by the return. The problem is that large language models need a bunch of money to train, a constant data flow, customized data—it's just this big expensive monster.

When you try to talk to people about this—Nvidia sold $215.9 billion worth of GPUs, mostly, in the last fiscal year, supporting roughly $22 billion of revenue total in the entire world outside of the two companies that literally require money being fed into them, sometimes by Nvidia, to stay alive—people respond, "Well, companies just lose money, right?" We have this quote from Ed Zitron of Prophy Markets:

We have this cult-like worship of the wealthy where we think that someone wouldn't spend all this money for no reason.

Reconciling that with the idea that the ultra-wealthy and ultra-powerful didn't get there through big brains—just luck, opportunism, and perhaps an MBA with the right people, regular people who happened to be in the right place at the right time—and realizing the world is not run as a meritocracy is kind of grim. So it's easier to say, "No, they're not making a mistake, I must be missing something." And that's what...

Looking back through the history of technological breakthroughs, one of my favorite books on this subject is The Innovator's Dilemma. It describes how the innovation that ends up transforming an industry often starts out worse, doesn't make economic sense, and none of your customers are asking for it — which is exactly why we ignore it.

Take horse-drawn carriages in the 1800s — an amazing form of transport, according to people of the 1800s. Then cars came along, and the problem with cars was that they broke down all the time — much like AI hallucinates now. They were more expensive and the economics didn't make sense; you might as well walk than buy one. There was even a law requiring you to employ someone to walk in front of the car waving a red flag. Clearly, it was a worse solution. But disruptive innovations have a higher ceiling of growth, so they eventually overtook the horse.

Applying that analogy here: AI is imperfect at the moment, the economic models aren't ironed out, and people are still figuring out how to make it cheaper and build the infrastructure. But compare the rate of improvement with, say, coding — how much could you train a human coder to improve their output versus an AI agent? Even a modest 5% improvement per month, compounded, plus a 5% reduction in cost — which is what happened with the internet and with cars under Moore's law — eventually changes everything.

Moore's law doesn't apply to GPUs, though — let me explain. Nvidia put out CUDA, the underlying software library for running software on GPUs, around the 2000s, and it took them a solid decade or more to make it usable for things like data analytics and MapReduce. So when AI came along, they already had deep experience with it. And this is a company with more money, more attention, more geniuses, and more people focused on making their hardware efficient than anyone could ask for.

For anyone who doesn't know, Nvidia makes the chips — the ones that go into data centers and on which AI software runs, both training and inference. And here's the thing: back in the car era, you didn't have virtually every mathematician and scientist pouring into the car industry, and you didn't have the world's governments never shutting up about it. Governments deserve credit here — since 2023 they've been saying this is inevitable.

Even the 5% figure is questionable, though. I don't even know how you'd measure it, because a junior software engineer still learns from context — from how people deal with problems, which isn't just reading code or emails. It's context cues from speaking to a person, being in different environments. There are uses for LLMs in encoding that; I don't dispute that. But what does "5% better" even mean? Better at Rust? Better at C++?

I'd say productivity — shipped code. Measuring by output volume alone is insane, though; it's like calling someone the best writer in the world because their newsletter is really long. With coding it's genuinely hard to evaluate, but "is the software out there better?" is actually a good test — and uniformly, I'd say no. The standard of software across Google, Microsoft, Amazon and Meta — Meta especially, which is a monstrosity — is worse. Someone posted on Twitter earlier today that we should get a notification when GitHub is up rather than when it's down, because that would be more reliable. Microsoft is one of the largest companies in the world and can barely manage GitHub. Weirdly enough, the quality of software is going down as more people use LLMs — and as more businesses genuinely demand that people use these services.

Returning to the horse-and-carriage analogy: imagine we're at today's point, at whatever rate of improvement we've seen since Tragedy came out.

I remember when Tragedy launched — I was in Asia showing it to my fiancé, saying "look what it can do," and it was hallucinating occasionally and getting things wrong. I actually don't have that experience anymore. I have moments where I think its reasoning is weak, but I don't get outright hallucinations anymore.

I disagree. Give me an example of what you define as a hallucination.

Great one. I have a Bloomberg terminal, and one of the most useful features on it is Ask B. When you make a Bloomberg inquiry — say, looking up what we think Nvidia's revenue will be next quarter — it runs something called BQL, Bloomberg's own programming language. Instead of learning that, you just type into Ask B and it generates and runs the query for you, and you know where the data is coming from. It deals with hallucinations really well.

The other day I decided to get a little spicy and look up the growth rate of Microsoft, Google, Meta, and Amazon stocks over five years. I copy-pasted the results into Excel while writing the newsletter, and thought: Microsoft's stock has never been $575 a share.

When it's a cute little thing like a stock price and I catch it, it's no harm, no foul. But when you're talking about a transcribing tool for a doctor, or a financial model a hedge fund depends on, it becomes a lot more dangerous. And with a software package, a hallucination can mean refactoring a codebase in a way that leaves a security hole open or just breaks something. Maybe you've been vibe coding for six months and haven't really written code with your own hands in a while — maybe you've forgotten a few things — and now you have this slop to look for. The problems become multiplicative, and I don't really know how you train that out of it. They certainly haven't succeeded.

On one hand, these models have gotten better — but one of the main ways that improvement is evaluated is through benchmarks designed specifically for large language models, because you can't just have them do real tasks. They've found some tasks they can run, like on Meter.me, where it's "check out this chart, look how much better it's getting at running tasks — wow, it can go for an hour." Then you look closer and it's successfully completing them 50% of the time. They have a hallucination—

There's a hallucination leaderboard that focuses on basic tasks. According to historical data from the Vectara hallucination leaderboard, the four-year trend shows hallucination rates on simple summarization tasks have plummeted from around 21.8% four years ago down to roughly 0.7% on today's top frontier models like Gemini and ChatGPT. The nuance here is that these are simple tasks, which matches my experience: on day-to-day things, it hallucinates less. If that rate of improvement continues, there will come a time when hallucinations become rarer than they are today.

A second point: other technologies also had technical difficulties at their inception. I remember growing up with dial-up modems — I couldn't be on the phone and on the internet at the same time, so I'd have to stop playing Runescape upstairs to take a call. You thought, "This is crap technology."

I don't know, mate. I loved it.

Yeah, it felt like magic. And in hindsight, you go, "Wow, I now have Starlink and 5G internet from my phone. It's unbelievable." You couldn't leave the house with internet before. That's what I mean by rate-of-improvement thinking.

The last point is that we often compare AI to perfection, but that's not the actual alternative in the working world. If I want to do a simple writing task, I should compare AI to my alternative way of doing it — measured both in my time and in my ability to hallucinate as a person who doesn't know everything, or, if I'm hiring someone, an intern who might also be prone to hallucination or have gaps in their knowledge. So the comparison isn't AI versus perfection; it's AI versus the other alternatives. If someone hallucinates 0.7% of the time but knows way more and is faster, maybe on a net basis that's a good trade, and I should use AI.

Here's an example. Someone I love dearly, Matt Hughes, my editor, lives out of Liverpool — wonderful guy. I don't pay Matt Hughes because he knows everything. I pay him because he has incredible context and a ton of knowledge, he's willing to expand it and work with me, and he's a great editor who gets into the guts of it and has the experiences of it. He's a decorated tech journalist, and on top of that a wonderful, loving being with empathy and joy for the stuff he loves and absolute venom for the people he hates. I can't get that from a large language model. But I'd also push back on the assumption itself.

When you say it "knows everything" — what good is something that knows everything when it sometimes doesn't know anything? And are you really paying an intern for something basic? Are you going to them saying, "Can you look up what the date is?" No, you're doing that on Google. Whatever the task is, you're also trying to train an intern. The point of an intern is to train them, take them out of Pinocchio status — and an intern learns, gets context, and learns your habits.

AI gets context and learns.

No, it doesn't. The way it "learns" is you create a giant CLAUDE.md file that it sometimes doesn't read, sometimes does read. You create a harness. It's like Pee-wee's breakfast machine from Pee-wee's Playhouse — you have to do all these contraptions to mitigate the hallucinations. And even then, how much effort have you put in?

That's an extremely simplified example, but if I went on my Claude now and asked what my dog's name is, it would know my dog's name.

Jesus Christ. This company raised 95 billion.

I'm using an extreme simplified example to show that it can remember things from the past. Obviously it knows much more complex things as well. So we accept that it does have memory of the past.

It has files it can access that have stuff on them, but that's not the same as memory. So it remembers your dog's name, it might remember your habits, it might be able to read things you've said before. But does it know your moods? Does it know what's going on in the world around it? Does it have good days and bad days? Is it there for you? It's just a text machine. An intern is something that can grow, something you invest in — that's not something you do by feeding files and text to it. The way we store memories ourselves, the way we accrue experiences, is a mélange of emotion, feelings, and facts — completely different.

There are two things here: the process by which something happens, and the output. The process you're describing is how a human does memory, and the way an AI does memory is different. But what people care about is whether there's value in the output. If I dump all of my files into Claude, I don't really care how it processes them, as long as when I ask "what's my revenue?" it has the number.

You could say the same thing about training someone. You teach them, put lots of effort into them, give them context, education, and experiences — and then you might ask them, "by the way, what's my revenue?" The processes are entirely different, but the outcome is what I care about: do they know the revenue number when I ask? That's the part we sometimes get lost in. I've heard this debate about whether AI can be creative, and I think the way to answer it is: it's about the output. When I ask it to do a creative thing, does it give me the answer — not whether the process matches a human process. Because honestly, who cares? People pay for the outcome, the product.

I actually disagree about the process. Take Matt Hughes, your editor — watching him go down a rabbit hole and being there with him, and vice versa. We were working on the research, and I ended up sitting there for a whole day-long session writing 11,000 words based on a bunch of notes he'd given me. Even describing that process makes me happy: us going back and forth, learning about the misanthropy of the horrible, cynical people at asset managers like Blackstone, saying "it can't be this," and pushing back on each other. That process is fundamentally different, because we were both learning together, and the learning process was as much about creating the output as the output itself.

When you learn something, you're not creating the average of the documents it could find — which really is what these things do. You're not getting particularly novel outputs. If I needed generic slop output, sure, but I've used some of the high-end LLM setups that hedge funds use, and they all give the same shite: the same generic reports, the same "we noticed this" analysis, things you can find in any AI slop out there.

What you described, though, has two points of value — well, many more, but you said you're learning, and you're getting the blog edited, which is the output. You're also getting connection and other things. When people think about the value of AI, of course they could use it to learn. But in the example I gave — repeating my revenue number back to me or computing a figure — I just care about the output. I could use it to learn; I could ask what happens if the revenue number was wrong once.

You should have defined deterministic ways of knowing those numbers; you should not rely on them, even with the terminal running BQL, which I trust — I will double, triple check everything just to be sure. Partly because, for me, the process of learning matters: I don't want just a report I glance at. I want something I fully understand, and I also want to understand the context around it. I don't think LLMs do that, and I don't see them getting there, because it's just not what they do.

There's also the other problem: the more detailed the report, the more likely there are things wrong with it. If you're with Matt Hughes, for example, I can trust he's got it right, that he understood, and that I can have a back-and-forth with him that will inform me if I've missed something. I can read the stuff he's read and actually trust him — there's a big trust component as well.

What is the basis of your trust in Matt? Could it be his historical performance?

I mean, yes. And also the fact that we've learned half of this stuff together.

But tenure doesn't necessarily mean anything — there are probably people you've known for 15 years who you also don't trust.

So I was trying to figure out what causes humans to trust another thing, and I guess it would be continual delivery of a commitment made. With Claude, for example, on simple tasks — as we've seen from this hallucination leaderboard — it continually delivers for people, and that's why we've seen the fast...

Is that what that board says?

Well, it's saying, is it getting it wrong, is it hallucinating?

How are simple tasks defined?

I don't know.

That's the thing, though — this is actually a very illustrative thing about the AI industry. They are the what-aboutism masters: "Look, we've got this benchmark that says we're good at this, and look, the number's higher." What does the number mean? No — what does that mean? And I'm not using this as a criticism against you.

When you can't give a direct answer, you give a side answer. When the LLM industry wants to prove its worth, it can't just say "just use the product." When the first iPhone came out, I was at Penn State at the time. I felt like the apes at the beginning of 2001 with the monolith — the official voicemail was immediate. I showed it to tech friends and to the most normal people in the world, and everyone was like, "Holy—". They were on Razrs, on Nokia 3210s. The value was obvious. Amazon Web Services, same deal.

It wasn't obvious, though.

Yes, it was — I bought it. And I also showed it to a bunch of people, because I'm aware I had bias: I just love gadgets.

But I remember the famous interview where Steve Ballmer, Microsoft's CEO, was told about the iPhone and burst out laughing:

"$500, fully subsidized, with a plan. That is the most expensive phone in the world, and it doesn't appeal to business customers because it doesn't have a keyboard, which makes it not a very good email machine. You can get a Motorola Q phone now for $99. It's a very capable machine — it'll do music, it'll do internet, it'll do email, it'll do instant messaging. So I kind of look at that and I say, well, I like our strategy. I like it a lot."

He burst out laughing, mocking it because it was so disruptive — way more expensive, and way different. No keyboard.

Well, phones used to be insanely expensive, and the carriers would cover them, but you had to sign a long contract — you were still spending 500 bucks. But the thing I'm getting at is you didn't have to explain to someone why. You might have to get past the cost part, but you could just say, "Look how good this is." Then with the iPhone 3G and the App Store, people were like, "Oh, this could actually change things." Mobile web — even though it was a monstrosity, so bad at first — even then you could get your emails and just look at them. The point is, BlackBerrys were also expensive and were actually kind of cool, but the way they worked was not like consumer software; they didn't have the classic GUI. iPhones felt like that — like a cell phone designed like a computer. It was obvious from the beginning. I was dating a girl in the center of Pennsylvania at the time, and everyone I showed it to was like, "Wow, this is incredible."

That to me is the difference with AI to this day. When you ask, "Okay, why is it so amazing?" people still dither. They're still like, "Yeah, you can't run a business fully with it without this weird system of pulleys and levers and such."

But how do you square that with the growth stats around ChatGPT?

100 million active users in just the first 60 days after launch. For comparison, TikTok took 9 months, Instagram took 2.5 years, and the World Wide Web itself took roughly 7 years to reach that scale. Over 60% of US adults have integrated AI tools into their daily routines within 3 years of launch, reaching 40% of the population. The internet took 5 years to hit that same milestone, and personal computers nearly 12.

This is the part that's giving me dissonance. When I showed ChatGPT to my fiancée, it didn't really work for me — but she's a sole entrepreneur whose English isn't her first language, who has to write lots of copy and generate lots of images, and who was paying a graphic designer to make images she couldn't make herself. She would describe it as transformative for her business. What I'm hearing from you is that it's not transformative and has no value for people.

Would she pay the per-million-token rate? The actual rate? Because that's the thing — if this were sold at its honest cost, would she still be here saying it's transformative?

I would actually — if people were reacting like that and paying $2, $3, $4 every time they did something and were genuinely happy, that might be an argument.

What would the honest cost be if they weren't subsidizing it?

The actual per-million-token cost — the actual API price they should be charging.

Do you know how much that is relative to what people pay?

It depends on the model. But there's a point I want to make about what you said about the internet earlier. When I first got on the internet with a 33.4 kilobits-per-second modem, even back then I thought: if this was faster, it would be better — because it was slow. Downloading something like Shareware from a site took all bloody day. And even then I thought you could probably do video stuff with this, which eventually happened.

There's a guy called Jim Covello from Goldman Sachs who, in a 2024 report, argued there's too much spend for not enough return — paraphrasing there. He made the point that in the run-up to the iPhone, there were thousands of presentations arguing that as GSM, Bluetooth, and Wi-Fi radios got smaller, it was inevitable we'd get something like the iPhone. Then he said there is no such path for AI — no roadmap to AI becoming the thing its promoters promised.

I must be clear: if these companies had gone out and said, "Yeah, this is interesting generative cloud software. It's really expensive. We're not sure it's going to be a disruptive, world-changing thing — it has potential, but we're going to go slow. It's an R&D effort, and we're not going to expose consumers to it" — and if they'd just called them language models, not "AI," because it isn't AI: it's not autonomous, it's not smart — I might actually respect that.

Instead, since 2022–2023 they've said it's the best thing since sliced bread, it's changing everything, it's going to do all your work, it's going to take your job — you're going to talk to Bing and it's going to tell you to leave your wife. All these crazy things. And what's funny is when the writer Kevin Roose was speaking to Kevin Scott, Microsoft's CTO, about it, Kevin Scott said, "I'm just glad we're having this conversation" — instead of saying, "Settle down. It's a website. The website told you something. It's just an LLM." They talked it up. Everyone is talking about what they wish this was, rather than what it can actually do.

That makes it scary to people deliberately. It also makes it environmentally destructive — look at the gas turbines poisoning Black neighborhoods, I think in Louisiana, at one of Musk's data centers. Look at the enormous energy draws: it's raising power bills and creating inflation across all consumer electronics because of the…

The interviewer pushed back: much of what the guest says may be true, and yet the technology could still profoundly change the world. The closest analogy is the early internet and the dot-com bubble — including the guest's own essay, with its "rot economy" and "rotcom bubble" framing, arguing AI is worth less than people think.

In the dot-com bubble, there was huge hype and people overselling what their websites could do. But in the wake of that bubble, while 90% of companies went to zero, generational companies were born that changed the world. That's what bubbles do: massive hype, overinvestment, delusional investors who think everything will change — and at the same time, skeptics.

The internet had the biggest skeptics. In 1998, Nobel Prize–winning economist Paul Krugman said that by 2005 or so it would become clear the internet's impact on the economy had been no greater than the fax machine's. In 1995, astrophysicist Clifford Stoll wrote famously in Newsweek:

Do our computer pundits lack all common sense? The truth is no online database will replace your daily newspaper. No CD-ROM can take the place of a competent teacher. Commerce and businesses will shift from offices and malls to networks and modems. Baloney. So, how come my local mall does a roaring business and the cyber mall gets zero business?

Another from Krugman: "The growth of the internet will slow drastically as it becomes apparent most people have nothing to say to each other." That may be the worst of those predictions.

The guest noted Stoll's piece was interesting because some of his points held up: an overwhelming amount of bad information online is bad for society, and online education has not been a great replacement for regular education. But there is a vast economic difference. The dot-com bubble was actually two bubbles. One was the website bubble — trash on trash, like Excite@Home buying a company for roughly a billion dollars over something tiny. The big one people remember is dark fiber: cables laid in the ground expecting enormous internet demand. Analysts had estimated demand was doubling every 90 days when it was actually doubling every 6 to 12 months, maybe longer. The result was a massive overbuild of fiber optic cable and transmission infrastructure, on the assumption it would all get lit up and people would want it immediately — which didn't happen.

The standard post-bubble defense is that demand for the internet did eventually come. But that's very different from demand for generative AI. Right now, demand for generative AI is predominantly subsidized — most people experiencing it are not paying the real cost. On top of that, there is the largest, most disingenuous marketing campaign in history pushing it uphill, and Microsoft — the apex predator of cloud software — can only extract single-digit billions from selling AI software. Outside of OpenAI and Anthropic, the whole sector barely reaches $22 billion.

And the thing is, $22 billion is a large amount to you and me, but it's not a large amount of money when you've spent a trillion-plus dollars — when Anthropic and OpenAI have $1.1 trillion worth of cloud commitments. So how does this turn into a post-dot-com bubble situation? A data center built today is going to be as expensive to run in 2050 as it is today, unless there's some breakthrough in electricity — and that's not happening with AI. Or unless there's some breakthrough in GPU technology, but we already have Broadcom, Nvidia, Etched, and every major chip company including ARM trying to do something about this, and no one seems to be able to magically make this profitable, or indeed even less costly. Even Nvidia with Vera Rubin, their more expensive new GPU system — even then, they say, "Yeah, 10x more efficient, more dollars per megawatt." They're all coy about it. They don't just say, "Yeah, we worked with OpenAI and Anthropic and we found it reduced our costs by 50%." That would be the easiest thing in the world to say if it were true — and it's not happening.

Asked whether there won't be demand for AI, the speaker points out that "AI" is a deliberately broad term. The reason the industry uses "artificial intelligence" is so everyone lumps everything into it — protein folding has nothing to do with LLMs, robotics isn't LLMs, and autonomous weapons, horrible as they are, aren't LLMs either because you couldn't trust them. Everything gets mushed into AI so that when you say "AI can't," the reply is, "Sir, you forgot to give us homework, and also AI is working on curing cancer." No — that's not LLMs. Stop giving them credit.

The interviewer notes the similarity: all of these need GPUs. That's the funny thing — all the data centers being built are for generative AI alone, not for the other stuff. AI has been around a long time; a lot of the good stuff coming out of Google's search side is AI, but pre-generative.

Asked how you'd run AI in a robot like Tesla's Optimus without a GPU, the speaker points to Matic's cleaning robot. It doesn't have a little GPU inside. It may have used some GPUs, but nowhere near as many as generative AI needs to feed training data into it so it can clean a house. When the little bugger — he calls him Turdsly — goes around mopping his floor, it isn't burning money the whole time.

By contrast, Sightline Climate said in February there are 190 gigawatts of data centers in planning — never mind under construction. That works out to about 12 million megawatts, roughly $1.6 trillion to $3 trillion a year in annual demand. We don't even have $130 billion worth of annual demand. People say it will grow — how, when most of the demand is Amazon feeding money to OpenAI or Anthropic, Microsoft feeding money to OpenAI and Anthropic, and Google feeding money to OpenAI and Anthropic (well, Google hasn't fed it to OpenAI yet, but they're a pretty big customer — billions of dollars)?

The conceit is that we're building these effigies to capitalism — giant GPU data centers — while people are told it's for "AI," the thing that did all that unrelated other stuff. The worst argument he's seen is: "You like online banking, so you like data centers." There's a big difference between a data center for regular non-GPU compute — standing up a server, a content delivery system like Akamai, or how Meta runs Facebook — and these giant GPU data centers. The former takes way less power and is mostly CPU-driven; the GPU data centers offer one thing, and one thing only.

The interviewer counters with research notes saying that tougher AI systems designed to solve concrete physics, biology, and spatial problems require some of the most intense data center infrastructure on the planet — systems like DeepMind's AlphaFold, used for genomic sequencing and climate forecasting, run on high-performance computing clusters requiring immense precision and continuous heavy computing. Training the brains for self-driving cars requires billions of miles of simulated physics environments; the AI isn't generating text, it's learning to navigate 3D spaces and gravity, and relies on data centers.

The speaker agrees those data centers might have GPUs — GPUs were used for HPC before generative AI, and that's how AI was trained before, including how Tesla trained Autopilot in its own data centers, for better or for worse. But that is not why we're building these data centers. They're being built to sell to generative AI companies, either for training or inference, in a brainless way.

This is a good way of illustrating the con: everyone saw Google, Microsoft, Amazon, and Meta give Nvidia over $800 billion or so, and concluded they wouldn't do that for no reason — "we've got to build more of these things, there must be all this demand." Even though 70% or more of that demand comes from two companies funded by three of those same companies. The reason they don't want to break out their AI revenues is that it would become alarmingly obvious this was the case. The only real big customers are those two — and it's not like they're building a few data centers; they're building toward a trillion-plus in revenue potential, entirely speculative. They're building because they saw the biggest companies in the world buy a bunch of GPUs and said, "I want in on that." Surely they must have diverse customers — they wouldn't just have two unprofitable fail-sons they're propping up. Christ, they've raised $217 billion just in 2026.

We know some of the biggest companies in the world are using generative AI to write a lot of their code. That's a great productivity gain for them, right? But have you used Google, Facebook, Instagram, or GitHub recently? They are catastrophically worse. Amazon Web Services went down multiple times because of their AI coding tool.

How is Google worse?

Let me tell the story of a real guy called Prabhakar Raghavan. Previously one of the heads of ads at Google — in 2019, Google declared something called a "code yellow," which is when they say, "We've got a problem." The problem was material weakness in query numbers — the amount of times people were searching on Google. Ben Gomes, then internally the head of Google Search, said: wait a minute, what you're suggesting would mean we give worse answers, because if someone gets the answer quickly, that reduces the number of queries. Another engineer, Shashi Thakur, was saying, "Can we please tell Sundar this? This doesn't seem good. We can't just increase the amount of queries — that would mean people have to search more, which would make the product worse."

But it would make them more money. You'd show them more ads if people spend more time on Google.

But is this linked to AI doing code?

I'll get there.

The problem is that Raghavan, head of ads at the time, kept pushing: "No, we need to make more queries happen." Nick Fox, who I believe was taking over Google Search, was saying the same. Sometime in early 2020, Raghavan took over Google Search. And this is what I believe — I can't prove it — but if you look around the SEO sites like Search Engine Journal and the various forums, Google stripped back a lot of the suppression of spammy sites so people would stay on Google more. Over time, Google wanted to create more queries, and Google Search became much worse. It's why people always add "Reddit" to their searches — the underlying search results had got worse.

Then generative AI came along, and wouldn't you know it, Raghavan got put in charge of part of Gemini. Google was having trouble getting people back onto Google, and what did they think they'd do? Everyone's talking about this AI thing, so they put it right at the top so people have to stay on Google — and they'll use it more, because instead of searching websites and doing that annoying thing where they click away from Google, they'll only use Google. Instead of generating answers — by which I mean giving you search results you click through — now Google is the answer. Is it right? God knows. It might tell you to eat rocks, or eat poisonous mushrooms. Maybe it'll give you a few links to click through. But the ideal situation was that AI was the ultimate form of Google's evil.

But that's not the fact that coders could code with AI that made Google worse. Human decisions made it worse.

Yes. And then there's the instability of Google's platform, which I probably should have led with — a problem across the whole tech industry.

So you're saying Google is going down more?

Yes. Google is less stable. Google Docs is a bugfest right now and has been for a while. Google Sheets, same deal. And you're right, I'm being a little unfair — this is everyone. It's the same with Microsoft, the same with Amazon, the same across the board.

How do we quantify that outside of anecdotes? Is there a way?

You're right. GitHub downtime is the best example — Amazon Web Services went down two or three times this year because of AI tools. And honestly, it is kind of hard to quantify outside of anecdotes. But I challenge anyone listening to this: go and use a website these days and tell me how well it works, how buggy it is. Even with my iPhone — the supposed best UX in town — even the iPhone is a flipping mess these days.

AI-Assisted Code and Rising Tech Downtime

The research gives a short answer: yes. Tech downtime and software outages have demonstrably increased over the last few years, and industry data points directly to the explosion of AI-assisted coding as a primary culprit. The problem hits the tech industry from two directions: the code itself is getting buggier, and the sheer volume of AI activity is crashing the underlying infrastructure.

Part of it is simply that people are writing a bunch of code, pushing it, and so there's more code on GitHub. Open source has had this problem too, because well-meaning people think, "I learned a bit of code with an LLM, I'm going to make this project better" — and they barely understand what they're shipping. Or they understand a little and think, "I can understand some of this," and push the code anyway. So GitHub is flooded with AI code.

This is also making humans complacent: "I let it write the last 100 lines and it was broadly right, so I won't check the next 100 lines as much." That's human nature — taking shortcuts, spending less energy on an activity if you can. But the AI is still making the mistakes, and we're still making all the promises about AI. Sam Altman has been promising the world for years that this will replace software engineers, and Dario Amodei has been saying 50% of white-collar labor will go away in the next few years.

If instead they admitted it has issues — that it's probabilistic, it's going to make mistakes, and if you don't know what you're doing you'll miss those mistakes and things will get multiplicatively worse — that would be one thing. But human nature is only part of it. So is the marketing. So are the promises.

One of the smartest things a business can do is build like a bigger company without actually hiring like one. The problem we all face is that most companies don't have every skill in house. When I look at the businesses seeing real success today, the consistent pattern is how quickly they move: they bring in specialists with skills in emerging areas to stay ahead. Even in our company, we spent the last year pulling in talent across areas like AI native strategy, no-code builds, and product workflows.

We find this talent through our longtime partner, Fiverr Pro. Their premium service only shows you vetted talent, so you always have the safeguard that anyone you bring in for a complex project has the skills you're after, will deliver to the same high standards as your internal team, and — most importantly — will keep up with the pace. It's a simple strategy, but it lets us stay agile without compromising on quality. If you need these kinds of skills in your business, head to pro.fr.com to find pioneering talent to fill your business's gaps.

The little traditional SIM card that goes inside our phones hasn't changed at all since it was invented in the '90s. You have this physical piece of plastic that means you're locked into one carrier, one network, and the second you cross a border, that carrier can start charging you whatever they want. There are alternatives, and today's sponsor, Saily, is one of them. It's an eSIM app that gives you a safe and secure data connection in over 200 destinations. All of their eSIMs have built-in cybersecurity, which is great if you're traveling for work and looking at confidential material.

I've been using Saily whenever I travel because the connection is always reliable, it saves me a ton of roaming fees, and it means I don't have to deal with the faff of sorting out a SIM everywhere I go. To give it a try, download the Saily app from the app store and scan the QR code on screen. For 15% off your first purchase, use my code DOA at checkout. Keep that to yourself.

My car that drives itself — that is AI technology. I sat here with Dara from Uber, and he was saying that in a couple of years we won't need drivers for Uber because the cars will be fully autonomous. And driving is, if I'm not mistaken, one of the biggest professions on planet Earth. So when you hear CEOs saying there will be job disruption, you say they're not telling the truth?

Yes — or they're guessing in a way that's very good for them. Think about it from the perspective of Microsoft's Satya Nadella. He's not going to say, "Yeah, we don't know if this is going to work, mate." Of course he's going to talk his book and say this is going to replace all workers, it's going to be amazing, it's going to be so powerful. And then he'll change his tune and say actually it's not going to replace workers — that makes him more powerful because the technology isn't catching up. Dara from Uber, for example, is of course going to say that if this happens, it would be good for Uber, because Uber would just become an autonomous taxi service.

There's a reason Waymo has taken the approach it has — I find Waymo fascinating, and I think it's really cool. But I think there are socioeconomic problems that will come from it, and real problems that will emerge.

What kind of problems?

Socioeconomically, as you said, driving is one of the largest employment centers in the world — the economics of cabs will fall apart. But again, we are nowhere, nowhere near that. We're not even close. Waymo has had to do the smallest rollouts and the most controlled deployments, because the problem with pretty much every AI system, but especially driving, is not getting 95% of the way there — it's the edge cases. Rain is a big problem for them in San Francisco. Or a kid runs across the road wearing a high-visibility jacket — does the system even notice it's a child?

This is a really interesting and very applicable example: the right comparison shouldn't be autonomous vehicles versus perfection. It should be autonomous vehicles versus human drivers.

I don't know if I agree, because a human driver might make mistakes, sure — but again, I'm not an expert in autonomous cars, just to be clear. If we're pushing autonomous cars out there willy-nilly, and not in extremely controlled environments, those edge cases will multiply and be dangerous. Yeah, they might be better than human drivers in some ways, but they might also — I was in Vegas the other day, in a hotel, and I watched a bunch of Zoox's cars just get stuck.

They're autonomous cars?

Yeah, those weird boxy things. They just blocked the exit — they all lined up and fell asleep.

I saw the same thing happen outside a hotel when I got out of a Waymo in San Francisco. It just stopped, and then a bunch of cars and another Waymo got stuck behind it.

I've seen some bad human drivers as well. I agree — but we have control over deploying these bad or good drivers. We have the ability to roll them out slowly, which is exactly what we should do. I'm not saying autonomous cars are bad. I'm saying we need to be so, so careful and treat them as guilty until proven innocent, because we can prove it. They also have people overseeing them — they actually have people monitoring the routes. It's something they cannot rush out, and it doesn't seem like they're rushing it, which is good. And they're not promising the world.

I do agree. Listen, I'm a big fan of taxi drivers generally, in part because I spend a lot of time in taxis, and I'm not just getting in there to get from A to B — I'm getting in there for lots of other reasons. However, when I look at the stats around what is more dangerous — driving myself or having an autonomous vehicle drive me — there's a 68% lower overall crash involvement rate when you're in an autonomous vehicle. Autonomous vehicles experience roughly 2.1 police-reported crashes per million miles, compared to humans at roughly 4.68 per million miles — a 55% reduction. And autonomous vehicles show an 80 to 81% reduction in crashes resulting in injuries versus human drivers.

Uh-huh.

So you're 85% less likely to be involved in a single-vehicle crash, like hitting a wall or a tree, if you're an autonomous vehicle versus being driven. I agree it's safer. But what's the sample size of human drivers in that data? We have many, many more years of drivers and many more years of accidents. And also, does that have anything to do with generative AI? If we were just talking about that, we'd be having a different conversation.

I guess the question here was really around job disruption. We look across industries and see that driving is a massive profession—is there going to be job disruption because cars can now drive themselves? And if we think about white-collar jobs, lawyers and accountants, people sit here and tell me that lawyers and accountants—or rather, some of the skills within those professions—will be relegated to AIs to do.

Here's the thing: lawyers are a great example. You always hear legal partners talking about AI, never the associates. The associates are the ones who go out and find the precedent, do the grunt work, pull motions half the time. The partner might be the litigant, the client-facing one, but the associates are doing the day-to-day work—and I'm not hearing from them. I'm not hearing associates saying, "This is awesome." I'm hearing a bunch of well-paid people who have sat on ChatGPT and gone, "Yeah, yeah, I'm the greatest lawyer ever." They're not the ones I want to hear from; I want the actual workers.

White-collar labor disruption is not happening. OpenAI had a study come out, I think about a week ago, that said there was no connection between spending on AI tokens and revenue per employee.

What does that mean? Could you explain that to me?

As in, the more tokens you spend has no correlation at all with the amount of money you make. It's the second report they've put out—the other one was that hallucinations are mathematically guaranteed, almost. It's the one thing I respect about that company: they just put out a study saying, yeah, kind of sucks.

But the people actually having their work lives disrupted are art directors, transcribers, translators—people whose bosses don't care about the output, who consider it cheap work. And the problem is, those bosses would have automated your work away anyway. They would have sold it, taken the cheapest option, sold it to the global South. That is something AI is doing. And again, those people aren't paying the actual cost of AI; they're using a subscription.

The actual white-collar labor force might have some things slightly changing, but there is no evidence of productivity gains. In fact, if there were, they would be screaming it from the rooftops. There was an Oxford Economics study last year saying young people are finding fewer jobs because of AI—we actually read the study, which multiple journalists did not. It was a single line saying, "Yeah, we saw some correlation." It didn't give a number, didn't say what the correlation was.

We are so conditioned to believe that the rich and powerful know what they're doing that we internalize these narratives: previous booms lost a lot of money, technology takes time to do stuff. And they are intentionally playing on those mythologies, knowing that journalists, analysts, and investors will believe them. This is partly because our realities are defined by stock prices—because the stock prices of these companies went up, we say, "Oh, look, it must be working."

Both of those things you said were true, though, right? Previous technologies didn't make money at the start, and they'll get better.

But that's the thing—because another thing got better, this will get better.

The interviewer pushed back: those two claims—technology often starts unprofitable and improves over time—are both true, so something the AI companies say must be fundamentally false.

The speaker agreed that the individual statements are true, but argued that what these companies fundamentally mislead people about is how possible it is—how many actual signs they have. They don't have the signs: no signs of costs getting cheaper, no signs of the technology being able to autonomously do work without a Rube Goldberg machine, and even then doing so reliably in a way that makes the customer more money—productivity you can state "with your whole chest" without a series of asterisks. That's how it is across the board.

He noted that the people most excited about this—whom he called "psychopaths on Twitter"—often show an attachment he has never seen in any other industry outside of maybe sports teams: the attachment some people online have to these companies. If you dare to criticize Anthropic, it's almost a religious attachment.

A good example came this week, when Bloomberg reported that OpenAI was on track to hit $40 billion in annualized revenue. Whether that's a month times 12 or four weeks times 13, nobody knows—they don't define it. Yet he saw multiple people insisting, "Actually, it's 60 billion. It's actually 60 billion. I heard from someone." To him it's like a cult—a cult of software built around growth and the idea that by backing the right horse you will get some grand thing. OpenAI in particular, and Mr. Altman in particular, along with Tibo and others at OpenAI, have been fermenting this online, building a parasocial relationship with both the large language models themselves and the companies. Allegiance to the companies is treated as so important that he finds it truly vile—if only these people cared about, say, Medicare for All, poverty, or other actual problems in the world, rather than whether we're buying enough GPUs.

Some of your narrative, one could argue, actually helps them. How? The AI doomers who have come on this show—including figures like Geoffrey Hinton, one of the founding fathers of AI—have said that what these companies are building is highly dangerous and will be fundamentally disruptive to society. Interestingly, some of the CEOs you've mentioned had the same historical narrative: this is really dangerous, and there's a significant chance it could wreck the planet. What we've seen is a slow pivot away from that, because now they're getting booed and attacked—and the pivot sounds a lot like your narrative. It now sounds like: actually, no, it's not going to change anything, you're all going to be fine, it's not dangerous at all.

That's why I think there might be a couple of PR people at these big AI companies thinking, thank God for Ed, because you're telling people: don't worry, everything's going to be fine. It's not going to take your job, it's not going to disrupt the economy, it's just a fad, there's no technology there. And I think they don't believe that.

Here's the thing: I think Altman and Amodei are some of the most deeply corrupt and cynical people in the world. When Altman said back in early 2023, "we're a little bit scared about what we're creating"—oh, shut up. I've met so many of these rich liars, and I know exactly why he says that: so you'll invest in his company and buy the software. So you'll be scared that if you don't use AI today, you'll be left behind in the future. That's their continual narrative: get on the train today or be left behind.

By the way, every single scam and con starts with rushing you. Every trick in history begins with "you must do this now." The best advice I ever got was: if anyone tries to rush you—and it's not literally a mortal emergency like you're bleeding or the house is on fire—slow down. Yet all these companies say it's so scary, and now they're talking about slowdowns. But have you noticed that Amodei and Altman say, "maybe we should slow down progress," and then they don't? Right now Altman is saying they're slowing down because they're "so delayed"—no, they're out of compute. I can guarantee you their PR people do not like me; I know OpenAI's PR people aren't super fond of me.

But I bet there are elements of what you're saying, because you are theoretically calming down the general public.

And you know what? I hope I am, because fear-based tactics are horrible. These companies don't want calm—these companies want people scared. I'm 100% sure of it.

I just fundamentally disagree. Can I—because the timelines are there. What I do is log their quotes over time and read them out from 2015 to 2026, and the change you see is them going from "there could be extinction"—that was the early narrative; Elon said it himself, it's the single most dangerous thing—to "this age of abundance, we're all going to have unlimited stuff," and now the new slogan at xAI is "intelligence for everyone." And whenever Dario comes out and says, "by the way, it's really dangerous," they attack him. They hate him.

Honestly, I've been saying "Dario, shut the up" for years. But I get your point, and here's my counter: I don't think they changed their tune to calm the public down so much as they're desperate not to get regulated—which is laughable, because America doesn't regulate tech. We are still trapped in the hands of Milton Friedman, Margaret Thatcher, and Ronald Reagan, still stuck in the neoliberal hellscape of growth at all costs and free-market capitalism. So no, no one is regulating these companies. The regulation should have been, I don't know, breaking them up. We shouldn't have companies this big—it makes things...

"But these technologies are dangerous." They are dangerous, but not in the ways people have been warning about.

The actual hacking incidents were human error

Take the cyber hacking incidents. To be clear, those were not cases of an AI breaking out of a sandbox — the sandbox was set up wrong. The server the model was on was misconfigured. In the Hugging Face attack on OpenAI, we don't even know how much compute was spent, but we do know the server was improperly set up: they thought they had turned off internet access and hadn't. That's human error, even though an indeterminately large amount of compute was thrown at it.

That said, advanced AI models can act as agents on the open internet: they can examine the code bases of different websites, find vulnerabilities, and exploit them — at scale, and arguably with more intelligence, speed, and breadth than any human hacker could manage theoretically. So yes, that is dangerous.

The models are already in the wrong hands

People keep saying we can't let the Chinese get hold of these models, that they mustn't fall into the wrong hands. They're already in the wrong hands: Mark Zuckerberg, Sam Altman, Dario Amodei. The wrong hands are the hands of the people running these companies. We shouldn't be training these models to do these things at all — I don't know why we are, other than that they've run out of other things to train on. The fact that they can do it is interesting, but that's all.

Would you agree that an intelligence — you might dispute the terminology — that can go out onto the internet, click around, and take actions carries inherent risks? The second part I agree with. We've had people running automated hacking scripts for years and years; this is brute-forcing the same thing with a lot of compute. And yet it is dangerous — these companies are doing something dangerous.

This isn't what the doomers warned about

But this is not what Jeffrey Hinton and others have been warning about. They've been saying these things could destroy society, could manipulate people. When you actually look at the underlying capabilities, not so much. Hinton, for his part, still seems to hold his Google stock, and oddly, after leaving Google because he was worried about the AI there, he immediately commented that Google is actually very responsible. Strange.

Regulate the compute, not the models

Back to the cybersecurity side: I agree it's dangerous, and these people should not have access to so much compute — they clearly don't know what to do with it. There's a really easy fix: don't let them use so much compute. Regulate that part out of existence. What if the Chinese do it anyway? The Chinese were able to distill the models regardless.

We got to this point — to use an annoying Sam Altman term, we let the genie out of the bottle — because we let these companies go unregulated and burn as much compute as they wanted, with enablers allowing them to do so. And for all the dire warnings about AI dangers, no one seems to have actually done anything.

"Okay, we're going to play a game, Ed." "Let's play it."

The host proposes a game: on the cards are beliefs Ed considers myths about the AI industry, and for each one Ed must give a one-sentence first reaction. Ed agrees to play.

The first card reads: "The AI industry is creating enormous economic growth."

Ed's answer: "No, it's not. It's nowhere in the data." Asked to add a second sentence, he explains that pretty much all of the economics is either Nvidia feeding money to AI companies like CoreWeave, or these three companies feeding money to those ones to spend it back with them.

The host asks what evidence he has that it isn't causing economic growth. Ed clarifies that, other than the spend on semiconductors — the speculative investment in GPUs and data center infrastructure — actual spend on AI is barely cracking a hundred billion, and most of that is just the big two running their services and paying companies like Oracle and CoreWeave.

The host pushes back: a hundred billion is a lot of money for a relatively new technology. Ed replies that it's not, when you've spent $300 billion in equity funding, and just for these three companies, something like $600 billion in capital expenditures.

The host concedes that means it's not profitable, but argues the hundred billion is an expression of consumer demand. Ed disagrees: when the compute is mostly driven by subsidized subscriptions, it isn't real demand. "When you're giving someone $20 or $40 for a dollar, they're going to use it more. If this was all on a per-million-token basis, we'd be having a different conversation."

The host accepts the point and moves to the next card.

The claim under discussion is that the United States needs to spend trillions to beat China in the AI race. The response: what AI race? That's actually the point. Is the race to make big scary LLMs? China already did that without the Nvidia GPUs — and by the way, they've got Blackwell GPUs. Analysts like Kakashi and Jastario, who have been covering this for years, have pointed out that China has had Nvidia GPUs it wasn't supposed to have for years.

And even granting that, what would the race be for? They already have the LLMs. Is the goal to spend more money than them, or to constantly panic about China? If that's the case, China has already won.

Myth number three: AI will replace all human jobs. That just isn't happening, and there's no economic data to support it.

Will it replace some jobs? It has replaced some contract labor that would otherwise be replaced with cheap labor in the global south — it's a kind of digital globalization in that sense. But all jobs, most jobs, a lot of jobs? No.

Robotics is a different question

What about robotics? Robotics is not what we're talking about here — it's a very different thing. Yes, robotics will be powered by AI, but there are tons of different kinds of AI. This mythbusters piece was explicitly about generative AI.

Take the Optimus robot Elon is working on at Tesla: even in the demo of the hand, they had a guy controlling it — it wasn't operating autonomously. Here's the thing: if they can beat all these challenges, robotics would be really cool. I don't know how long that will take, but that's one I'd actually be willing to believe in a couple decades.

Have you seen the Chinese robots, like the one from Unitree that can dance? They can't really do human things yet, but it is pretty mind-blowing. Robotics are cool — I'm not going to pretend I don't think so. I wish the tech industry were still building robots and actually making fun, interesting stuff, instead of large language models.

Why robotics is exploding now

I was in San Francisco at a massive incubator there. Three years earlier it was all software startups; when I went back, it was all robot startups. I asked the founder of the incubator, "Why is everything robots now?" There was one robot that was just an arm with a frying pan on it, and the whole thing is it cooks for you. He was showing me it cooking, and he explained: the hardware, the physical parts, have always been fairly cheap. The expensive part was the intelligence, and now that's come down to pennies. So you're seeing this explosion in the robotics industry, because robotics is a function of intelligence plus hardware — we've always had the hardware.

And a ton of data as well, and the data is very expensive. Cybercabs rolled out real slow; it's going to take a long time.

Could robots replace jobs?

Could it be a threat if they build a robot that could replace a human job? Sure, it could. But human jobs are multifaceted and change with environments. And a lot of jobs you might think of — a dishwashing robot, for example — don't pencil out: some guy at a restaurant isn't paying 10 or 20 grand for a robot to replace a job he's already not paying enough for. So yes, it could replace jobs if it can — but that is not what this myth is about.

I ask these questions not to challenge you — I'm actually trying to form my own opinion. As it's written, "AI will replace all human jobs" — obviously not. But I'm trying to figure out whether the truth is somewhere in the middle: that there's a certain type of job which humans probably shouldn't have ever been doing. Think back through history — there was someone's job just to sit in an elevator and press the buttons. That's a job humans probably shouldn't have been doing, and as technology gets more advanced, it takes on a lot of that automated, monotonous stuff.

Right — but with this particular claim, from the specific blog I wrote, I was explicitly talking about generative AI. When people say this, they are referring to that.

"So you're not talking about agentic AI, which is—"

"Agentic AI is LLMs. Agentic AI is just a fancy way of saying an LLM talking to another LLM with a harness on top. That is still LLMs. Agentic AI is one of the bigger lies they tell. When you hear 'agent,' you're meant to think autonomous AI that can do what you want. It's still LLMs. It's still LLMs talking to other LLMs."

"Taking screenshots and putting them in an LLM and stuff."

"Oh god. Yeah."

"Okay, but I could make the case—thinking about my personal usage, I definitely use agents to do things I would have previously asked people to do. Not to say I don't still hire, because we're hiring like crazy."

"Yeah."

"And I'm still hiring in that particular function. Think about the chief of staff role. My chief of staff would previously have triaged all of my inboxes, put them somewhere, and told me about them—or maybe once upon a time shown me a piece of paper, back in the day. I guess now my chief of staff is no longer doing that job."

"You still have a chief of staff, though."

"This is what I'm saying. They're doing other things."

"Right, but again, what you're describing is fairly basic automation. I don't know what the tasks are—triaging."

"Basic? Spend a trillion dollars on triaging email. That's the promise. If they'd spent $10 billion and this were much smaller, I'd go, cool software, yay. A lot of the things people are impressed with, like scripting stuff, are just LLMs doing Python. You should be impressed by Python code—Python's incredible. You can scrape websites, you can download things. It's awesome. But the point I'm making is none of this would be anywhere near as much of a problem if they didn't ask for all of the attention, all of the money, and promise the world. It's their promises that are the problem. And the journalists who went along with it, and the analysts and the Twitter people who went along with it, saying this would change everything and replace everything—leaving the realm of reality."

Is there any technological innovation through history that was really game-changing where this didn't happen?

People overpromised on the internet too.

They overpromised on the businesses, but I've read through a great many pieces about the early internet, and a lot of people were excited but hesitant. They worried there wasn't enough demand, yet they still thought it could have potential ramifications if it happened. People were not super negative about the internet. Many skeptics said they were worried about an overload of bad information — look at where we are now. A lot of people worried about the social consequences of everyone talking online, and they were correct about that. On the economic side, they were specifically talking about the globe — which I think made hundreds of thousands of dollars and had, I think, a billion-dollar market cap.

Yeah, there was massive hype in the dot-com era.

I read a lot of those stories, and the hype was nowhere in them. You didn't have articles everywhere saying if you don't get online, you'll be left behind. You didn't have professional consequences. Nick Sesh mentioned his blog earlier; he described this thing — global AI-saturated global decision-making — where you work at businesses where, if you don't say you're more productive with AI, whether or not it's true is irrelevant: you face professional consequences, you can get fired. There are people having to AI-wash their jobs by saying AI did it, otherwise bosses who don't do the work will get mad at them. This did not happen with the internet. Part of it is that social media wasn't like it is today — the decentralization of media in general has caused this as well, and also day trading. So many things are different; it's crazy.

I do think AI is different from the internet, in part if you just measure it on the speed of adoption — especially generative AI.

But adoption of the internet required physical connections to your house, while adopting generative AI involves having a web browser. It took a vast amount of effort to bring the internet to people; even with dial-up, it still required physical distribution.

And that's why it was so slow, and there was less hype than with AI. I do agree there's way more hype, and again, we're clustering AI into this big category of lots of different things.

Generative AI is explicitly what I'm talking about here. When bosses say you need to use AI, they're not saying go buy a Unibeam robot — they're saying use LLMs. They have this theory, the era of the business idiot: we are ruled by people who don't do work, because nobody who actually does a bunch of work and is really productive harasses someone who works for them for not being productive enough.

They don't have the time — they're doing work. Someone sitting there with the ingratiation machine telling them that every beautiful idea out of their messy little skull is amazing is going, "Damn, this thing says I'm a genius. Why are you not using the genius machine to do more work?" And if you're a boss who goes to lunch, leaves lunch, and sometimes reads your emails, LLMs are...

One of the most compelling arguments I have for the overhype of AI: in a world where everybody has access to these tools, whatever the tools can do would largely be commoditized. What the tools can't do — call it human taste, judgment, people skills, whatever you want — becomes the valuable thing, because throughout history the scarce and the hard becomes the most valuable while the commoditized becomes the least valuable. The very fact that we're commoditizing the generation of content, or code, means that's actually not where the value will accrue for the user. And if you think about what it takes to make something objectively great — if an AI can do it, then the great thing is not of value.

I've been thinking a lot about how you avoid the temptation of sloppifying the things you make, the value you put into the world. A simple example people can relate to: if you use ChatGPT or Anthropic's Claude to make your LinkedIn posts, they will be LinkedIn posts, because everybody else is using them. A great LinkedIn post now is someone who doesn't use them and makes something irreplaceably human — deeper, more personal, an N-of-one lived experience, all these things AI can't do. That's a compelling argument that commodity tools produce commodity outcomes.

But really, what's changed? We've got a bunch of slop, but these people were half-assing their jobs before. It's just a half-ass accelerator machine — it's what I'm talking about with the slop blogs. People that gave you dog before have now got the dog machine to pump out dog.

There's a guy called Carl Brown, internet bucks — awesome guy, great software engineer. He said — I might have said this earlier — it makes the easy things easy and the hard things harder. When you're doing a really distinct small script for something and it can plop that out, it's awesome. I used Claude the other day for something useful: my kid loves Minecraft, and I was trying to fix a broken mod because he loves his Wither Storm. It was awesome — and it still took me half an hour and kept getting things wrong.

When asked what he uses generative AI for, one speaker said he really doesn't. The only exception is AskB on the Bloomberg terminal, which he uses when requesting consensus analyst estimates, for example for Nvidia. Otherwise, nothing.

Asked how he knows it's bad, he said he has used it and put it through its paces, trying to build financial models with it. He found one error and immediately gave up, and has never been particularly impressed. The one thing he will defend it on is tech support. He uses a tool called Synergy at his place in New York to share one mouse and keyboard between a MacBook and a PC laptop, and dropping a giant troubleshooting log into an AI and asking "What's wrong?" — and it pointing out what's wrong — is genuinely useful. But is that a trillion-dollar use case? No. Is that a $2 trillion company? No.

Another speaker pushed back: isn't it better than Google search? The first speaker said he still tries to use Google but has to push the AI-generated content out of the way. Another said he can't remember the last time he did a Google search, while a third admitted he sometimes finds himself using Bing — which he hates saying — because he has to scroll past the AI summaries to get the actual links so he can read the source himself. When it was suggested you can just ask the AI for the links, he replied that it doesn't do a particularly good job.

One speaker gave an example: his iPad wasn't turning on and was doing something odd on the screen — surely typing that into Google isn't better? The skeptic clarified that troubleshooting may be the only LLM use case he defends; being able to drop a log into it is genuinely its one strength. But that's not what the companies are selling. They aren't selling it as a useful little tool or a quirky bit of software — they're selling it as the thing that will change everything and replace all jobs.

Another speaker agreed that the companies claim it will replace everything, but pointed out that, funnily enough, the critics say the same thing — people like Geoffrey Hinton, or those who have left OpenAI's safety team, warning of the impact it will have on the world. The skeptic responded that it's strange how all these critics also have a vested interest in AI doing well — for example Daniel, a former OpenAI person, who wrote AI 2027 with the Star Codeex guy: nothing more than badly written science fiction that he's already had to walk back. It was noted he could have made more money by staying at OpenAI if he'd had early options, though it looks like he lost out by leaving.

The core criticism, he argued, is that these critics aren't critical of the companies themselves — not of the stealing, not of the environmental damage, not of the fact that you cannot rely on the answers. They're critical of a big scary boogeyman out in the future: "I'm scared of when this becomes so powerful, and everyone should talk to me about how scary and powerful it is." They don't say, "Here are the harms today, the things we're actually looking at today" — the social problems of an automated way of spewing out slop, filling our feeds with crap, and presenting information with a tiny disclaimer that "sometimes this gets wrong." In the tiniest words possible, they don't talk about the fact that these things are trained on stealing millions of works.

The speaker picks up on the earlier point that AI will become progressively more intelligent and, once it does, will be a danger, and asks directly: would you agree that artificial intelligence has gotten more intelligent, if you measure it by any measure of intelligence one might use?

The answer is that it has only gotten better on the tests that are rigged for the models — tests where you can train for the test, the ones the models are intentionally trained for. If you logged the rate of improvement on a graph, it would look like a steep climb, at least in terms of what the systems are capable of doing.

The reason is that these systems haven't gained new features. Outside of OpenAI and Anthropic, and setting aside the coding startups, there is basically no successful AI startup company.

Both speakers agree that AI has become more capable over time, and that this trajectory will continue. The question posed is: at some point it crosses human intelligence — so will it not start to do some of the jobs people are doing today?

One speaker concedes software engineering as an area where AI has genuinely improved, and then lists other areas of progress: chief-of-staff and admin tasks, video generation, photo generation, text generation, and theoretically coding — plus agentic workflows.

Asked what an agentic workflow is, he explains it as automated workflows doing the same kind of task repeatedly. A good example is preparing a briefing on a CEO's guest: ingesting all the backend data, summarizing it, searching the internet for who the person is, looking at every interview they've ever done, summarizing and generating a model of the things people want to know from them, producing a report, and sending it to the inbox — so you get a 20, 30, 40, or 50-page report on your guest before they arrive.

The other speaker pushes back: this is basically the same thing AI has been doing for years — it's not really new capabilities. Web search has existed for years, and report generation for years. The reply is that high-quality video generation is new: models like Seedance produce videos that look like movies, indistinguishable from camera footage — they are incredible.

The core argument is that over the last 10 years there has been a steady rate of improvement in capabilities, output, and quality: hallucinations have dropped, models have become more "intelligent," and they score higher on IQ tests than they did 10 years ago. So there has been an upward motion of improvement. The skeptic notes this is pretty much how machine learning goes when you feed it more data — exactly, comes the reply, and when you put more compute behind it. If this continues, the question of AI taking on human jobs follows.

If more data and more compute keep pushing models forward, what does the future look like? The rebuttal I was expecting to hear is that it won't continue.

I actually don't think it will — I think there are hard limits we're going to hit.

So you do believe there's a hard limit somewhere?

We've kind of already hit the diminishing-returns level. Take video generation, which by the way is far less of an American concern anymore — OpenAI shut down Sora, though I think you can still use the API. Look around you at the amount of crew you need to get a shot. People think movies just happen shot by shot, magically. You've got first ADs, assistant directors, gaffers, lighters — and simulating light is insanely difficult. So many things happen in creating visual images that, yes, you could make a one-minute clip that might fool someone. But how do you practically turn that into a movie? There was a movie — I forget the name — that claimed it aired at Cannes. It didn't. It aired in the city of Cannes during the Cannes Film Festival; it was not at the festival. When it comes to the practical creation of actual things versus magic tricks, the practical outcomes just aren't there.

The reason I keep coming back to capabilities as the example is that, sure, they do better on tests — the number goes up. But when it comes to whether you can actually rely on it for distinct tasks: you can rely on it for summaries, for generations, the things it was already doing. It's getting roughly linearly better at those. But there's a ceiling to that. Okay, so it gets really good at research — what does that actually mean? You've already kind of got the automation there. What's the next step? Training it to be more autonomous, for example, isn't something that comes from training data. That's a new Gary Marcus "neuro-symbolic" problem: you actually need to build structure around the AI to make it work. And even then, it doesn't fix the—

So you're saying there will be a point where the rate of improvement plateaus?

We're already there.

We've already hit that diminishing... Gary Marcus said this in 2022 as well. Do you know there are lots of people listening now who've had their—

The skeptic's response to claims that people's workflows have been completely transformed by these tools: yes, there are such people — but first, did you pay for the tokens, and how many did you burn? Setting that aside, what workflows exactly? If it's "I did a bunch of web scraping or web searches," that's just not impressive. Did you make an entire movie? No, you didn't. Is it speeding up your coding? That's believable — multiple people have said so — but again, how much can you trust that?

The other speaker's point is that in the moment of any technological innovation, people extrapolate linearly or treat the present as a static state — assuming tomorrow will look like today, or that things improve in a straight line — when what often appears instead is exponential improvement.

The skeptic pushes back: all the innovations being discussed — compute, fast processes — are hardware breakthroughs, and the hardware breakthrough companies don't seem to be fixing the LLM problems despite all the king's horses and all the king's men. Google is now on nine or ten generations of TPUs; Broadcom is building chips with OpenAI, their "halapeno" chip. Yet none of these people can say, "Yeah, we're on the path to making this profitable" — because they can't. If the environmental problems and the profitability situation were fixed, maybe he'd be more generous. And for all the touted improvements and capabilities, at some point he asks: can it do even a tenth of what they're promising? Sam was saying the other week that in about six months it will be like a genie you can ask wishes from — never mind that he apparently never watched Aladdin, and that the genie was charming. The long and short of it: the promises do not line up with the capabilities or the capability improvements.

Exponential improvement in software performance is always the result of direct hardware improvement. With all the gifted mathematicians, software engineers, and hardware engineers in the world, where are we? A trillion-plus dollars in, facing a future great financial crisis and the world's greatest marketing scam.

The other speaker counters that he does think, in the future, all of the devices and...

One speaker argued that in the future, the devices and computers we use—and physical items in the world generally—will be more intelligent, powered by underlying AI infrastructure: more data centers and falling energy costs.

The other pushed back: how does a GPU-filled data center translate into, say, a smarter Nikon camera? The claim that devices will get smarter is broad and plausible, but what does it have to do with these data centers? They are not being built to make consumer electronics smarter—they're being built for nothing other than speculating on the ability to capture demand for generative AI services.

The reply was that it's not just generative AI, and pointed to Meta's earnings call a couple of weeks earlier, where Mark Zuckerberg described the big breakthrough: running anything users post on social media through an AI to get full context of what it is—resulting in 15 basis points of increased retention, which he believed referred to Instagram. As an example: if the AI can see a guy in a blue shirt holding coffee, it can learn to serve content to users who want that, and people are retained longer. The counter came quickly: 15 basis points is 0.15%. The response was that at scale, it makes a big difference—but the rebuttal stood: over ten billion dollars in, and the best result is 0.15%. There's a reason he's quoting basis points rather than dollars.

The first speaker clarified his point: this is another application of these data centers, since it requires a data center and drives revenue—but it's still outside the question of pure generation. The other speaker noted that Meta's generative ad model is GEM, their LLM is Muse Spark, and that this context-understanding feature is the kind of thing behind the strange popups on Instagram ("Dave the cat—why is Dave the cat suffering?"). Meta has ruined that product.

The core criticism: why can't Zuckerberg just say plainly, "we've made a couple billion"? Because he can't—there's no way to show that spending 14 billion dollars on Scale AI (Alexander Wang) produced a matching return. If it were going well, you'd be told how well it was going, rather than a rain dance of "if we move all the pieces around, in three years theoretically this will happen."

Across almost 700 interviews with some of the most interesting people in the world, one unexpected lesson stands out: vulnerability is the doorway to connection. After sitting with a guest for two or three hours, I feel a deep sense of connection to them, and as they leave, I ask them to write a question in the Diary of a CEO.

We've taken all of those questions and put them on cards, each with the name of the person who wrote it. You can play the Diary of a CEO conversation cards at home, as I do with my fiancée, with colleagues at work, or with others in my life, whenever we get a minute — and it's incredible what happens.

They're great if you're in a romantic relationship and want to connect more with your partner, great for a team that wants to bond, and great for families who want to learn more about each other and need a good excuse to spend time in the analog world, connecting human to human instead of in a digital one. It is remarkable what the right question at the right time can do.

Go to thediary.com to get the conversation cards now — there should be a button just down below here.

If the button says "subscribed," you're already subscribed. If it says "subscribe," you're not yet — and if so, please hit that button. It helps the show more than you know. According to the algorithm, you're someone who watches our show but hasn't hit that button yet. Thank you so much.

The speaker describes the situation as "fugazi" — inflated by the media — where many people have spent money they arguably shouldn't have, and now realize they've invested heavily, much like the metaverse. A lot of money went into that vision of a virtual world, but it never materialized and shows no sign of doing so in the near term. AI, in this regard, is the same as the dotcom boom, NFTs, and arguably much of the crypto industry — hype inflated by the media.

The difference is why the metaverse and NFTs never escaped this: there were no stocks to speculate on and no big companies to invest in. Tech companies had record earnings in 2021, with money floating through the system thanks to post-COVID federal money flowing into banks — easy, zero-interest-era money. Then came the hangover: growth slowed dramatically. This is his "dotcom bubble theory" — companies ran out of hypergrowth ideas, so they started buying GPUs. When they bought GPUs, people said "they're doing AI, we'd better buy the stock," and the stocks went on an incredible run, growing by hundreds of percent over the last few years — despite zero proof. The media simply repeated that Meta's revenue was growing because of AI, that Microsoft's revenue was growing because of AI. The fugazi was that everyone gave them credit in advance, and now we're reaching the point of asking: you didn't spend that trillion dollars for no reason, did you?

The other speaker concedes there is overspending, but argues it exists because there is something real here — practical uses for the technology — and when people realize that through history, they go crazy because they want to be the one who owns the opportunity.

His counterpart disagrees fundamentally: he doesn't believe the speculation is a result of actual demand. Private credit isn't sinking hundreds of billions of dollars into AI because of actual demand; they're doing it because they saw the biggest companies in the world building data centers and making a ton of money from the two companies they fund, and wanted some of that money.

The first speaker clarifies: he's saying there is value in the underlying technology, though he's not saying how much, nor that it's proportionate to the investment. He offers an analogy based on the other's "rot economy" essay: imagine you're on a desert island with 10,000 people and someone finds a banana tree. People will stampede toward it, clawing each other to pieces. If the essay is right that there was desperation because no innovation had been found in a while, maybe that explains it — they're stampeding, killing each other, and making irrational decisions, like hungry people would.

The other speaker realizes they actually agree. That is his point: these companies' main business lines at Meta and elsewhere are running out of growth — there's only so much they can grow. Indeed, over the next three and a half years, analysts think OpenAI and Anthropic alone will spend over $400 billion with Microsoft, Google, and Amazon. The crazy thing is that's a large part of those companies' future growth — and if this money isn't spent, their growth slows down.

On the point about the bubble, I actually agree — this is the dot-com bubble: these companies don't have a new thing and they're desperate. And they got rewarded for buying the GPUs. When they bought those GPUs from Nvidia, the markets went wild overnight. There were stories about armored cars being sent with GPUs to Microsoft to make sure Microsoft got them. So everyone saw all that money flowing in. Even though these companies never disclosed AI revenues, investors saw the expenditures and thought, "I want to do what these people are doing and get a little of that money."

The area where we slightly disagree is that I think the underlying technology has a lot more long-term promise than you do. But to push back: for these two companies to keep progressing, they need to spend tens of billions of dollars a year on training. The only way that can happen is if these companies, venture capitalists, private credit firms, and Nvidia keep circulating money to them. So the progress we've got so far is entirely a result of this circular system.

Circular — you mentioned VCs there?

Venture capitalists — by the way, the majority of the funding OpenAI got in the last six months came from SoftBank, Nvidia, and Amazon.

So the point is that continuing progress in LLMs can only continue as long as the money keeps flowing. Once the money stops flowing, the progress stops.

But isn't that most — like, Spotify didn't make money for 20 years.

Spotify didn't lose $20.9 billion in one year. They didn't need to raise $217 billion in the space of six months.

Uber is another example — $33 billion since inception before it became messily profitable. Amazon Web Services, between 2003 and 2015 when it became profitable, spent $29.7 billion — and that's total capital expenditures across the entire logistics operation, normalized for inflation.

So the TL;DR is they all lost money for a long period of time.

Yes, but the amounts they lost are magnitudes different from these three.

Can I argue that's because the potential of intelligence permeates everything, whereas Amazon at the time was like selling books?

No — that was bringing retail online. And when AWS grew — the reason I bring it up, and I'll repeat something but it's really important: it was founded in 2003, mostly because Amazon as a growing online store needed hardcore infrastructure. I think 2006 is when they turned it client-facing — I may be wrong on the dates — and 2015 was the year it became profitable. Total capital expenditures normalized for inflation were $29.7 billion across that twelve-year period. Yes, it lost money, but—

Speaking cold economics here: Amazon was unprofitable in a way, but their margins actually started improving because AWS was a very margin-heavy business. It was great.

Meanwhile, these two — Google cash flow negative, Amazon cash flow negative. The reason you liked software businesses was that they're meant to be cash-heavy and asset-light. These companies, along with Meta, have added more than $700 billion of new property, plant, and equipment — data centers, GPUs — in the last four years. They've gone from being cash machines to cash furnaces.

You said a second ago this can only continue if investors keep investing. I think investors are used to pumping money into things that burn cash, and your rebuttal sounds like "this is burning more cash than ever." So I'd say: is the opportunity bigger than those other case studies you referenced, like AWS? One would say the opportunity of intelligence permeates everything, so the TAM — the total addressable market — is enormous.

Maybe the revival back to me is about open source and all these kinds of—

No, no, I actually know what you're getting at. What you were describing is the argument that Sachin Adella or Sam would make: that the theoretical opportunity of large language models is enormous. I could have bought that a year ago, but we've gone way past the point at which you can rationally argue that LLMs need this much money. And when I say the money needs to keep flowing, I'm talking about these two companies—OpenAI, just OpenAI. Sam has said, and the Wall Street Journal and The Information reported a few weeks ago, that they plan to spend $750 billion on compute through 2030. I think they're going to be dead before then, but $750 billion. That is an insane amount of money.

And a large chunk of that is training. When I say progress, I mean that literally making the models better requires billions of dollars invested just in data, and tens of billions more to feed that data into the GPUs. Training is actually a really interesting thing, because when I train with my trainers Jake and Troy, I have a defined routine: when I do it and eat right, muscles get bigger. When you train an LLM, you're experimenting each time—and this isn't a hit on the companies, because they're still trying to work out how to do this. Putting aside how I feel about their attempts to innovate, I think there are people at these companies who genuinely want to do something interesting. It's just costing too much money.

Once the money tap turns off, there won't be money to buy the data or feed it into the GPUs. Put aside all my other thoughts—just the raw capital to get them this far has cost increasingly larger amounts, with training runs that sometimes fail outright. GPT-5 was meant to be the panacea for the AI industry, and they had at least one training run that cost half a billion dollars and produced nothing. If we think about progress in a vacuum, they need so much more money just to maybe get somewhere. There's no guarantee—there never is. But there's a reason Google and Amazon are cash flow negative now, and a reason Oracle is probably going to die as a result of OpenAI: Oracle's future depends on OpenAI spending $300 billion over five years.

It's absolutely fascinating, because I was just reading through a list of...

I was reading through quotes from the big AI company CEOs to see how they would rebut you, and they're all basically saying the same thing. Sundar Pichai, CEO of Google, said: "The risk of underinvesting is dramatically greater than the risk of overinvesting." Andy Jassy, CEO of Amazon, said: "We're not investing approximately $200 billion in capex in 2026 on a hunch. We're not going to be conservative in how we play this. We're investing to be the meaningful leader, and our future business operating income and free cash flow will be much larger because of this investment." Mark Zuckerberg, CEO of Meta, said: "We'll continue to invest aggressively in infrastructure to meet the demand. I'd rather risk building capacity before it's needed than being late."

That reminds me of Shrek with Lord Farquad: "Some of you may die, but that's a risk I'm willing to accept." Zuckerberg is just going to spend all this money, and he can't be fired because of the unique board situation he has. So he'll piss the money away and hope he's right — and from the people who know what matters, he's not right.

The question is: why might you be wrong? What would it take for me to be wrong? A bunch of hardware breakthroughs to make this profitable — a bunch of new mathematics. The AI people who claim this is going to be the biggest, strongest thing in the world — did they ever actually get that? They didn't.

The strange part is that when it comes to being a critic or a skeptic, you're the one put on the hot seat — not the people spending a trillion dollars, not the people promising the world. The person with a blog, someone like me, is the one who has to prove it. If the CEOs came on the show, they'd be on the hot seat too.

They won't talk to me, though. I think it's because my guests are quite critical — I don't think Sam Altman wants to come on. Mr. Altman, come on Steve's show. Do it.

But of course they're going to say that. And if they thought they were right, I don't think they do anymore. If I were in their shoes and believed this was an existential thing, sure. But it gets back to the dot-com bubble — this is the last thing they've got.

The interviewer noted that the guest's opinion differs from the two most popular views — that AI will be catastrophic and must be stopped, or that an age of abundance is coming and we should press on. His view instead is that it's a con with no real underlying value in the technology, that it's overhyped, and that there's far too much spending — a position the interviewer said he'd rarely encountered. So what would change his mind?

It would take a hardware breakthrough that reduced the cost by something like a thousandfold — a dramatic breakthrough that isn't happening, he added, because everyone has been trying. It's the cost, and also the data centers.

He thinks the way data centers are being built is reckless and damaging to communities. In places like Vernon, New Jersey, residents say they don't want them, but planning boards approve them anyway — presumably, he assumes, because they're having chummy lunches with the people doing it. He called the use of gas turbines disgraceful, and said gas turbines and behind-the-meter power are reckless and damaging to communities, as is the noise these facilities make. On the water situation, he said he isn't well read enough to weigh in.

Generative AI, he argued, is an egregious, pornographic demonstration of how unfair the world is. Regular people trying to get a loan for an ordinary business get turned away by the bank. But if you want to build a data center, Jensen Huang will back you — he'll give you 25% residual value. A steadily profitable regular business can't get venture capital, because VCs demand 10 or 100x returns. Try to get a mortgage and the bank puts you through a full colonic, but if you want money for Jensen Huang to buy GPUs, he'll hand you a contract.

CoreWeave is a great example: a neocloud that builds data centers, fills them with GPUs, and rents them out. Nvidia was one of its first investors in 2023, and signed a $1.3 billion contract to rent back its own GPUs from CoreWeave — so CoreWeave can go to a bank and say, "I've got a customer — the guy I'm buying the GPUs from, with the debt I'm getting from you." Buying GPUs is open season; living a regular life, building a regular business, or buying a house means the highest interest rates ever and a bank that says it doesn't trust you. But if you're an unprofitable neocloud, you get billions from Jensen and it doesn't matter.

The interviewer observed again that this was the first time he'd heard that opinion, and the guest replied that he is "pro-user."

Myth: AI will be conscious

Moving to the next myth — that AI will be conscious — the guest said superintelligence and artificial general intelligence are theories. Anyone claiming this technology will become those things is just guessing and has no proof.

The next myth: AI systems are already blackmailing and escaping control. This is a really specific one, and there are actually two cases. The first involves OpenAI's GPT-3.5 — and I realize this is more than a one-sentence myth, so I apologize.

In their system card, which a bunch of media outlets covered, the claim was that OpenAI's model blackmailed a TaskRabbit into solving a CAPTCHA. What actually happened was a user of GPT doing the experiment got it to generate things to say to a TaskRabbit to make the TaskRabbit do stuff. A TaskRabbit — as in a person that you rent, not even to solve a CAPTCHA. It's someone you rent to nail a picture up in your apartment. It's an insane example, and it was covered as if these things blackmailed someone. They specifically said, "Yeah, we prompted it to do this." The other note was that AI systems can't do autonomous stuff like this.

Then there was the Anthropic case, where they said a model was blackmailing someone, saying that if you don't do this, I'll email proof that you slept with someone other than your wife. What actually happened was that Anthropic explicitly trained a model to do this and then prompted it to blackmail.

This keeps happening, and the media just slops it up — put the story in the bag, no thinking required. It's frustrating because it scares people. Put aside the fact that it's wrong; it's scary. People living their lives, who have to work longer hours to make less money and whose money doesn't go far, turn on the news and hear some being say, "Yeah, you should be terrified — it blackmailed someone."

This is so counterintuitive to their interests to some degree, and they've experienced it backfire. It has literally backfired — Eric Schmidt getting booed at a commencement speech by 8,000 people every time he said the word AI. These people are being attacked at home, which sucks and is terrible — and to be clear, don't hurt people, don't attack people at home. But the point is that this narrative is backfiring in a big way for them. I don't think they saw it coming, because you have to remember — you mentioned regulation earlier — these tech companies have been glazed for their entire existence. Travis Kalanick going, "Oh, what? People don't like me now?" Well, Uber was a horribly run place and he was kind of a monster — though there were also tons of articles about how great Uber was at the time. The point is these companies are not used to pushback.

What they thought would happen — I'm just guessing here — is that they'd do this scary stuff and get floods of money, and everyone would just say, "I kneel before you, I'll do whatever you want." They didn't expect what has happened. I agree this has backfired on them because they were inarticulate and disconnected from regular people. Sam Altman drives a $5 million car around San Francisco — at like 9 miles an hour, it's hilarious. These people are disconnected from everyone else, so they don't experience real problems and can't build solutions for them. They thought scaring people into doing what they wanted would work. It didn't.

All of this blackmail stuff was an attempt to make AI mystical — to make it seem like this unknowable, impossible-to-control, powerful thing, where only these two angels could possibly control the beast they've created.

This is quite a controversial statement, but for some reason I trust Dario a little more, because he's been the most balanced in his writing about the risk profile, whereas the others seem to move with the wind.

I get what you mean, but the reason I don't like Dario is that he was doing the scare-tactics thing when he worked at OpenAI — when GPT-2 came out, it was "too scary to release." He's also gone on television and given AI psychosis to Axios, saying 50% of jobs are going to go away because of AI.

What I respect is the consistency. He's now being attacked by them. And to clarify the word "attack": Dario is being verbally attacked by Silicon Valley — and if powerful people in Silicon Valley are attacking someone...

Four months ago he wasn't, though. They were all saying he was the smartest boy ever.

The point I want to make is: you claim to be so scared of how powerful this technology is — it's so scary — but what are you actually doing about it? Nothing. "We have an alignment team." So does every AI lab, though I guess OpenAI cycles through those really quickly. If I were Dario Amodei, sitting there scared of everything changing, having built a thing I thought would eliminate all jobs, I'd be terrified. I'd be walking around with a ten-ton weight on my back. The fact that he doesn't feel that weight — that he wants to be this weird elder statesman too scared to hold Sam Altman's hand at an event — makes me believe he's only saying it because it's convenient, and he'll walk it back, as he already has, whenever it suits him. I think OpenAI and Anthropic are basically the same level of bad company, though Anthropic is more cultlike. Jack Clark, one of the co-founders, used to be at The Register — one of the most critical journalists around — and now it's like something took over him, because they talk about these things in these lofty, fluent terms. Then again, maybe some of the people at Anthropic or OpenAI actually buy into it. I don't know.

Going back to the central question we asked at the top: what would have to happen for you to look back and say you were wrong in 2026? You said it would mainly take the cost of AI production dropping dramatically — and it would also have to do insane amounts of stuff, be truly autonomous, keep improving in capability. It would have to be a different product, indistinguishable from magic. The reason they hold it to these high standards is that they set them.

Okay. Fair.

It's interesting, because all these debates are about technology, but it's also an information war — narrative versus narrative. Everyone trying to escape the financials, everyone trying to escape what the models can actually do. The big thing I always say about AI boosters is that if I could regulate them, I'd ban them from speaking in the future tense. Just talk about today, mate. You get two weeks into the future, max. Because if they were constrained to what's happening today, they would sound like insane people.

Most technology companies would at the time, I guess. Uber would have sounded insane.

Uber was basically the difference, though — they were pissing money away, weren't they?

They were, but the unit economics were the same, just subsidized. You were still getting a service from A to B at a much lower cost. It wasn't like you paid Uber twenty bucks a month for 500 miles of rides and then one day started paying by the mile — which is what's happening here. Haven't they changed the business model for customers like me, so I have to buy credits?

Kind of. With Anthropic's Fable model, some accounts have to pay for usage, and adoption of Fable has been pretty low because of the cost. But for enterprises — companies over 150 people — you now pay by the token, per million tokens.

Oh, so they are moving to token-based pricing.

Yeah. And when they did that, everyone went from calling it the most impressive thing ever to saying, "We've got to control these costs." Uber's COO, Andrew Macdonald I think, said it's getting hard to justify, because it's hard to connect spending on tokens to actual useful outcomes.

He said the exact thing I've been saying. We're in an AI bubble.

Yes.

And when this AI bubble—

When it collapses, so much of the economy is resting on it that there will be downstream consequences. So I have two questions for you. First: are we in an AI bubble, and what happens when the bubble pops?

Yes — and it depends. The big thing people say is, "Oh, we'll get bailed out, Donald Trump will step in." Here's the problem with that: this isn't just an AI bubble, it's the dot-com bubble all over again. The AI bubble collapsing will probably start with this company running out of money: OpenAI.

OpenAI was meant to go public this year, and now it's been pushed to next year — a week and a half after I released their audited financials. Wonder why. Their CFO, Sarah Friar, has now said they'll do it "earlier than 2027, or 2027." Great answer there.

For anyone who doesn't understand what going public means: it means joining the stock market. At that point, your investors can finally sell the equity they received for investing in the company while it was private. Companies often flirt with the idea of going public "someday soon" because investors need a moment in their head where they'll get their money back at a return. If you're in OpenAI's shoes, you need to keep flirting with going public, or investors won't want to invest.

OpenAI has been a private company, and at their last funding round they were valued at $865 billion. When they tried to go public, they wanted to set a $1 trillion valuation — the New York Times' Mike Isaac reported this. Apparently their advisors said no, don't do that, and for good reason: OpenAI needs perpetual amounts of money. They raised $122 billion this year, most of it already spent. They'll need to raise at least a hundred billion a year just to survive. If they can't go public, they'll have to raise another funding round — and it will be difficult to raise even the same amount again; they'll probably have to take a flat round. But they need money so badly. Amazon sent them $35 billion that was meant to be contingent on them going public early — they did that because OpenAI needs the money.

OpenAI is the center of the potential catastrophe here, because Anthropic is likely going to beat it to going public. Once Anthropic goes public, it will be borderline impossible for OpenAI to do so — Anthropic is an unprofitable, unsustainable AI lab, but a better business that's growing faster than OpenAI's. I believe they have a ceiling too; they'll eventually face the same predicament. I think sometime in 2027 things start running out of steam, because as I said earlier, the only way these models get better is by feeding tens of billions of dollars into them.

So you think OpenAI runs out of steam in 2027?

I think they're already running out of steam — but I think they run out of cash. The sequence of events will be: they go out and try to raise, and they have trouble raising another round. Maybe Nvidia props them up a little. Maybe private credit — Blackstone, BlackRock and the like — gets involved, and the reason asset managers are investing is that they're putting money into the data centers, and they know this company drives most of the data center demand.

So they run out of steam in 2027, according to you?

Yep. And if they bum-rush going public, they'll have worse economics than Anthropic and get savaged — WeWork was a great example of that. Another SoftBank classic.

I think OpenAI collapses, though there are many different ways it could happen. The crucial thing is that multiple companies are existentially tied to OpenAI. SoftBank, one of the largest companies on the Japanese stock market, holds on paper about a hundred billion dollars' worth of OpenAI stock. If OpenAI can't go public, SoftBank can't do diddly squat with that. SoftBank's future — its ability to keep paying the people around it and exist as a business — relies on continually liquidating value from its investments, either by selling stock or taking loans against it. If OpenAI can't go public, SoftBank can't do that.

SoftBank probably won't run out of money, but we're going to see one of the largest holding companies in the world become much smaller. We will also see Amazon, Google, and Microsoft have to restate guidance — they will have to say, actually, we don't think we're going to grow as fast.

And what happens then?

I think we enter a tech depression. The core of my theory about the dot-com bubble is that these companies are out of hypergrowth ideas, but the market doesn't think so. The reason they're so maniacally spending is that buying AI GPUs allows them to kick the can further — to say, we're still doing something, we're working on AI, don't think too hard. Meanwhile, their current businesses are still growing, but they will eventually slow: there are only so many price increases, only so many tweaks to ads, only so many tweaks to Google Search, only so many ways Amazon can squeeze merchants. So in that tech depression, which you think might be triggered in

If this hits in 2027, is it a cascading downstream economic depression? The stock market is heavily dependent on these companies, so a pullback means panicked investors stop investing.

Yes. It's hard to capture every consequence, but a few things worry me. First, a ton of American money — regular retail investors' money — is in these companies. People bought into the Magnificent 7 believing the number goes up forever. Nvidia is the largest company on the Fortune 500 and NASDAQ, and 7 to 8% of the S&P 500. When the bottom falls out of Nvidia — and we haven't really gotten into how Nvidia is doing the most circular financing imaginable, feeding companies money so they can raise debt to buy more GPUs — I think Nvidia's revenue could fall 50 to 70%. In 2022 it was making single-digit billions.

But think about Jenny and Dave watching this — normal people with normal jobs.

People's retirements are going to contract severely, and I don't believe those values are coming back. So much of the S&P 500 and Russell 1000 comes from these four companies plus the rest of the Magnificent 7 — Apple, Tesla, Meta. And I don't know what happens after that, because more than half of venture capital last year went into AI, and I think most AI venture investments go to zero. Companies built on top of an LLM are all unprofitable too. LLM companies haven't really been acquired — the exception being Cursor by Elon Musk for the coding side. Meanwhile Cognition, just another LLM company, raised at a $26 billion valuation. That company has to go public, because who's buying a company at $26 billion other than Elon Musk? There were rumors he was trying to buy them too. Is he just going to pick off every LLM company — going to TJ Maxx for AI?

So is that a recession you're describing?

It's a recession, but it's also a depression within people's retirements — I'm talking 20, 30, 40% off the top of these companies' stock values.

Economic contractions and recessions consistently lead to job losses through a predictable sequence: falling demand as consumers and businesses spend less, revenues drop across most industries; margin compression, since overhead like rent and debt is fixed while profits shrink; then cost-cutting — hiring freezes, reduced hours, layoffs.

Yes, all of that would happen. But we're talking about equity values dropping with no real home for that money. So much is riding on these companies, and you can't bail it out. You could theoretically bail out OpenAI — I don't think it happens. You could pump these dogs full of money and keep them alive for a bit, but Anthropic and OpenAI together have $1.1 trillion of commitments. Oracle is building 7.1 gigawatts of data centers — over $400 billion worth — just for OpenAI, and there is not a customer on Earth beyond that. Oracle's revenue has been flat for 15 years adjusted for inflation. Without OpenAI, Oracle dies.

So you think OpenAI crashes and runs out of money, causing a domino effect across big tech, hitting the stock market and the broader economy?

Yes. Plus the tens of thousands of tech-sector layoffs. The venture capital side matters too, because VC has been on one of the most historically bad runs since 2018. The average return — total value returned per dollar invested — is between 0.8 and 1.2, meaning you get back 80 cents to $1.20 per dollar.

Paper gains.

No, that's actual returns. They'll show you paper gains, and even internal rate of return — a separate measure — isn't happy. Long story short: venture capital is not making money. It isn't actually providing returns.

They're celebrating paper gains.

They're celebrating paper gains, and they're raising money off paper gains. Just being able to say, "Look, the valuation of Anthropic went up."

But that's what Google and Amazon were doing. Google's last quarter boosted its net profits on paper by $99 billion because of the increased value of its SpaceX and Anthropic holdings. The fact that this is happening is insane, and the fact that it's not a scandal is insane—but that's the culture we live in, I guess.

Right now, everyone is benefiting. It's like that great tweet about reaping and sowing: when you're reaping, it's "yeah, this rocks"; when you're sowing, "ah, this sucks." Today they're all saying the speculative gains are awesome—the paper gains, the theory of Anthropic being worth $2 trillion, the articles they can write, the promises they can make. But when the rubber meets the road, it's going to be rough on them, because the valuation of Amazon, Google, Microsoft, and Meta rests on the idea that they will grow eternally, forever.

To quote Ed Elson from ProfitG Markets again: they're all doing Botox right now—sinking money in to make themselves feel young again, and the market believes them. When the market stops believing, we're not just talking about a depression. I'm talking about the market valuing them like airlines: "Yeah, you're really big and you make money off your existing products, but you don't have anything new—you'll just be doing this forever, and we'll value you accordingly."

Asked whether ordinary people like Jenny and Dave should do anything differently—conserve money, be more conservative if a recession or depression is coming—the guest's answer is yes, with caveats. He doesn't have money in the market himself and thinks it's a casino pumped up by the media. On investing in the S&P 500 or OpenAI: "oh god, no." He lives in cash right now and doesn't trust the market. He's not comfortable giving financial advice, but his view is that acting in this market is gambling—be conservative, because things might get volatile. Act as you would with volatility: take the gains when you've got them, don't sell everything, but be suspicious of tech.

The biggest thing is to be suspicious of what tech companies are promising. If you're acting based on their promises, don't trust the promises—trust that they will say whatever makes the stock run rather than what's actually happening, and that they will find every dodgy way to make you think something is happening when it isn't. The annualized run rate is a great example: Microsoft claimed $37 billion of annualized run rate in AI. You hear that and think they made $37 billion—maybe a month's revenue times twelve. They don't even define it, but it's built to manipulate. They get away with it because we don't have a functional SEC, and we don't have a media environment where skepticism is the priority and protecting readers matters.

The counterargument, as the host frames it, is that these companies would say the technology will be so great and so transformative that they are investing a ton of money in advance of the value and utility showing up. The host isn't defending them, just offering balance between the two perspectives. Many people also predict a bloodbath because not all of these players can win big the way they describe—someone has to lose—and when one starts losing big, there could be a domino effect or contraction.

The guest thinks what people want to believe is the dot-com bubble analogy: it worked out afterwards because Amazon and Oracle didn't die after the bubble—they're actually fine. But this isn't like that. These are bigger companies with bigger promises, and he thinks even Oracle could die—"RIP Larry. What couldn't happen to a nastier man?"

I asked the question purely because I wanted an answer, not because I agree or disagree: why don't you like these people?

I don't like being misled, and I don't think regular people like being misled either. I really don't think the average person could get away with bullshitting as much as these companies do, and I don't think the average person gets anywhere near the level of allowance for failure and lying that these companies get. There is a real economic and human cost to letting these companies run rampant, promise the world, and never really get called on it.

The tepid nature of criticism these days is so frustrating. There are some really great critics out there, but seeing these ultra-rich, ultra-powerful people lie through their teeth — or "misstate," whatever people want to call it — turns my stomach. I hate seeing people being misled. I write at such length because I really want people to see how I've come to my conclusions. Am I right? Am I wrong? I think I am — of course I do. But I also just find it loathsome. These companies don't make good products anymore, they don't care about their customers, and they treat their customers with contempt.

Where to find his work

The host noted the guest's great Substack, but he corrected him: it's actually on Ghost — he moved off Substack in 2024. He also runs a podcast called Better of Flame, and the host said he would link both below and highly recommended following him.

The host reflected that podcast listeners sometimes expect every guest to say the same thing as the last one, but that's just not the nature of information, opinions, and discussion. People have different opinions, and the job of both host and listener is to work through them, collect reference points from different people over time, and do your own research.

The guest agreed — whether it's about your health or something like this, watch, research, and learn. Never believe one person or one perspective religiously; collect a body of evidence and follow the evidence yourself.

The host said he loves watching the guest's YouTube because it offers a different opinion that challenges him to think beyond his current view of what might be possible. Hearing him describe this as an economic bubble and discuss the capex spending by the big frontier AI labs made him pause and consider that there could be some fakery going on. It then made him reflect that through history there has always been a bit of fakery in these moments, and it was an interesting take on what might happen in 2027 or 2028 when there's a market pullback. He encouraged people to watch, saying we need contrarian voices for honest discussions, thanked the guest, and called him a compelling, captivating communicator from whom he had learned a lot.

The closing question

The show has a closing tradition: the last guest leaves a question for the next guest, not knowing who they're leaving it for. The question left for this guest was: given that high-quality relationships are important for health...

The question left for me is: given that high-quality relationships are important for health and longevity, what should we be doing to improve our relationships and social connection? This is actually connected to the AI bubble. I'm a critic, I'm a skeptic, and what I've found is that showing, appreciating, and loving the people around you—uplifting them and raising them up as you succeed—is the way we do that. Your success should lift everyone around you. It's not economic. Talking about Matt Hughes for a while made me really happy.

This whole thing has at times been quite grueling, negative, and brutal. But the love and joy I've found in community and in the people around me—even in the small groups of critics, like Gary Marcus, or the people I talk to, Edward Ongweso Jr., Molly White, Brian Merchant—so many people have been loving and caring. Especially in these very critical moments, when you're dialing in on how negative and how bad things are, finding your people—the ones who find it repulsive too, who will talk to you about it—matters. Even Troy and Jake, my trainers, who are so excited about this work: talking to them as normal people, knowing they're going through their own struggles, gives you perspective and reminds you that you're human too.

I know this point is a bit all over the place, but it's really easy to get hard-locked on everything in life and drift away from why you do things, focusing too much on the work when the most important thing at times is just knowing there are other people feeling the way you do. When I hear from my listeners and readers, the most common thing they feel is that they have a voice and that someone is there for them.

I don't think it can be understated how much it means when you just reach out to someone you love and tell them you love them. Tell them they're rocks. Tell everyone—when you like an artist, a writer, or a podcast like this—tell them you love it. We don't do this enough, and we need to do it more.

Well, that's a good closing message. If you've enjoyed today's conversation with Ed, please let Ed know down below, and do leave your opinions below—I shall read all of them. Ed, thank you so much. I'll link to your website and to your YouTube channel, where people can learn more, and I'd highly recommend you do, because it is truly fascinating. I think we need more voices demystifying a lot of the fugazi and the narrative in this moment in time, and you are certainly one of them. I really enjoyed the conversation. Thank you so much.

YouTube has this new crazy algorithm where it knows exactly what video you'd like to watch next, based on AI and all of your viewing behavior. The algorithm says this video is the perfect video for you—it's different for everybody watching right now. Check it out, and I bet you might love it.