As I predicted here in January 2024, with everybody building the same thing, there can be no moat:

That’s been catching to OpenAI ever since. The numbers don’t lie.

§

And it’s not just the dwindling market share. Microsoft was for a long time OpenAI’s biggest backer, but they have increasingly backing away.

Today they apparently just distanced themselves even further:

When your biggest investor starts openly contemplating outsourcing your main product to China, it is what they call in business a strong negative signal.

§

Sometimes of course bad things come in threes.

The third bad thing is that Ed Zitron just reported that OpenAI is burning money even faster than people thought:

You can’t multiply your losses every year by 8x (or even close) for very long.

§

To be sure, Washington might succeed in knee-capping Anthropic, rescuing Josh Kushner’s investment in OpenAI:

But if Washington doesn’t succeed in killing of Anthropic, and instead actually inadvertently helps them, as I joked here…

.. then I’ll stand by this set of predictions from two weeks ago…

… except to say that another dark horse bidder is Elon. Wouldn’t that be an interesting twist?

Subscribe now

Even some people on the right see this, like Dean W Ball, who helped the current Trump administration establish its AI policy

What the White House did seems arbitrary. And maybe even corrupt

It doesn’t help that Friday’s decision helped OpenAI, whose President, Greg Brockman, is a huge Trump donor.

It doesn’t help that Friday’s decision helped Jared Kushner’s brother Josh, who is a big investor in OpenAI.

It doesn’t help that Friday’s decision helped Amazon (and was sparked by a report from Amazon), a huge investor in OpenAI. It doesn’t help that it indirectly helped Jeff Bezos, who has ties to the administration.

And it doesn’t help that Pete Hegseth seems to have a personal grudge against Anthropic:

Which makes the whole thing feel petty.

To be sure, Anthropic contributed to the situation. The Twitterverse is full of memes mocking Anthropic like these

Anthropic certainly played a role both by overhyping Mythos and by alienating the White House.

But it is the job of the US government to be the adult in the room, and to make decisions that are sound for the nation.

Effectively — and ironically — they have actually vindicated Amodei’s claim that some AI might need to be regulated and might require export controls.

The ridiculous but fashionable idea of zero-regulation around AI, made popular by folks like Marc Andreessen, is (rightly) dead in the water. After Friday, the White House can no credibly say that no model ever should be regulated. The White House has acknowledged that some models might be risky; now we are just arguing about the details.

What Washington must do now is make those calls in a clear and transparent way, in a way that doesn’t reek of corruption.

§

The White House’s first effort at tamping down the firestorm seems to have been to have been a sort of half-apology. David Sacks put out a long statement on X explaining (from his perspective) what happened. Unfortunately. his statement is (a) desperately vague about how serious the problem was, or what exactly the problem was, whether anything was affected, or whether the problem was unique to Fable/Mythos and so on. There was no real justification for the speed at which a decision with such lasting consequences was made.

What really comes through above all else is impatience and ego. The White House wanted its way, didn’t get it instantaneously, and attempted to impose its will immediately — giving Anthropic almost no time to address the issues. (Why less than 24 hours rather than say 72?)

It looks like they were looking for an excuse to screw Anthropic, and blind to the larger consequences.

Thue rushed and seemingly unprincipled decision will leave the rest of the world rushing towards “sovereign AI” (made for example by European or Canadian companies) and even Chinese AI, rather than deal with the uncertainty in the US, in which any model can be shut down at any time without warning and without clear explanation.

Moreover, because the construction of the order targeted foreigners — even eminent ones who have long lived and worked in the US such as Andrej Karpathy – it may lead to a massive brain drain.

And it will slow down one of the leading US AI labs.

It is a feast for Xi Jinping.

§

As Anthropic put it on Friday : “the government should have the ability to block unsafe deployments as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. This action does not adhere to those principles.”

Kevin Frazier of the libertarian Cato Institute echoed this saying in a blog yesterday, saying

“Transparency, fairness, clarity, and evidence-driven enforcement should be core components of AI governance. So long as AI policy decisions are driven by a few actors taking action behind closed doors in response to non-public reports, we will be far short of that standard.”

That’s what Ball is also trying to say, and what I am trying to say.

At this point, there is no turning back. Governments going forward will have to consider which deployments are unsafe, and block those that are.

But they must do this transparently, and with fairness, and clarity, based on technical facts, and without the appearance of corruption.

§

The best thing I read today came from the California Congressmen Ro Khanna, who both distills the problem and points to what I think is the best solution is (in line with what I have long recommended):

An independent agency. Not last minute, opaque decisions made by a small cabal that hint at favoritism and corruption.

Without that, the US AI industry is lost.

Subscribe now

Where is the US with respect to AI policy? It’s a mess.

Let’s start with Trump’s Executive Order from June 2, and then turn to last night’s Commerce order; neither gets things right. (I will end with some suggestions about what might better).

As I noted when the EO came out (seems like a lifetime ago, even though it is actually less than two weeks!) I thought it was fantastic that Trump has issued an executive order that encourages AI companies to do some preflight testing.

But at the same time, the part of the EO on preflight testing is too weak. It is voluntary; and too narrow, focused primarily on cybersecurity. Companies can choose to ignore it, defeating the purpose of the order, and lots of things aren’t covered.

Florida’s recent lawsuit – the first state lawsuit against OpenAI– address many things that would not be covered, and highlights some of the ways in which the EO falls short. Here are some excerpts of the facts what the Attorney General for Florida—a Republican state—alleges:

(I omit the rest of that bit but encourage readers to read the full complaint)

and also alleges

they also argue for a kind of false advertising

And, of course, all of these risks were foreseeable. (Indeed I pointed to most of them in my 2023 Senate testimony and in many other writings).

Trump’s Executive Order is genuinely a step in the right direction. But it addresses basically none of these harms, and does not require companies to consider them, even voluntarily. Instead, it allows AI companies to release software that could cause all of these harms, and demands little or no effort to mitigate them. We need a stronger hand from the government in forcing companies to address those harms.

For now, and until the US government takes a more comprehensive approach to addressing AI risks, states will have little choice but to file lawsuits like these; unfortunately, those can only be filed after harms have been caused.

And indeed, in a major scoop yesterday from The Wall Street Journal that got eclipsed, New York, working with several other states, has just subpoenaed OpenAI in a massive way, demanding information on ““a range of its activities and impact on users, including advertising, user engagement and retention, handling of consumer data and health data, activities related to minors and seniors, deep learning models, model sycophancy and company policies”.

States have no choice but to pick up the slack if they are to protect their citizens.

Subscribe now

Now let’s consider last night’s Commerce Department cut-off-your-nose-to-spite-your-face export control order, which has basically led to the (temporary?) shutdown of Fable and Mythos, perhaps sparking a crisis in the US AI industry.

You can tell almost nobody outside some very small circle even considered the order before it was issued, because almost everyone I have read — right and left, pro-tech and not — has panned it. It was both bad policy and bad politics.

§

To be sure, Anthropic brought this on themselves to a considerable degree, by (a) overselling doom and their own wares and (b) by literally and repeatedly calling for export controls themselves. If you say “Ohhhh it’s so dangerous something something beat China something something 10% chance of destroying all civilization something” often enough, sooner or later people might indeed want to put limits around your product.

That said, no industry can thrive if the government appears arbitrary and capricious, as they did last night, (effectively) abruptly shutting down an entire thriving industry, and appearing very selective and reactive in the choices it made.

What we need now is something I have not seen of the Trump administration: carefully considered nuance. Shoot first and ask questions later is not what we need right now.

Instead, the government appears to have panicked when Amazon showed them that Anthropic’s Fable could be jail-broken. Which is naive, since essentially every LLM-based system can be jailbroken. This is not a new problem. Whatever you might think about Mythos or Fable or AI in general, what was just identified was not new and likely not specific to Mythos or Fable and in any case should not have been addressed in such a hasty, arbitrary way, with such little attention to downstream consequences.

I still believe, absolutely, that we cannot count on the AI industry to self-regulate, and that government must look—in a comprehensive way considering the full range of known risks—at new advances that are likely to be massively deployed models before they are released.

More broadly, what we should all demand from our leaders is strong but measured AI oversight that is (a) bipartisan, (b) technically well-informed, and (c) calm and considered, proactive rather than reactive.

Share

Anthropic’s reply as posted on X (fuller link here )

Dean W. Ball wrote

My own take, concurring and elaborating:

I have my reservations about generative AI, but this is not the way.

Subscribe now

Chaser, from 404 Media:

Bonus, from :

Share

Here’s the scoop itself:

To longtime readers here, this should come as no surprise, because it fits 100% with bullet point three in the premortem I wrote of OpenAI in January 2024:

[Rare note for nonpaying customers: I am inserting a paywall here; it’s only the fourth time I have ever done so, and I don’t plan to do so a lot. But I believe that prescient analyses like the one above I made in January 2024 have considerable value, and I hope you will consider supporting my work, if you aren’t already. If you already are, I am grateful! Your supports means a lot to me.]

Read more

Section 230 of the Communications Decency Act is, to my mind, one of the worst laws the US Congress has passed in a long time; it has shielded social media companies from liability, and caused all kinds of chaos.

A lot of Senators from both parties agree. It came up a lot when Altman and I testified at the US Senate in May 2023. For example, here’s a fun interchange that I remember well, back when Sam Altman was pretending to be a crowd pleaser:

Sen. Dick Durbin (D-IL):

Thank you. I think what’s happening today in this hearing room is historic. I can’t recall when we’ve had people representing large corporations or private sector entities come before us and plead with us to regulate them. In fact, many people in the Senate have based their careers on the opposite that the economy will thrive if government gets the hell out of the way. And what I’m hearing instead today is that ‘stop me before I innovate again’ message. And I’m just curious as to how we’re going to achieve this. As I mentioned Section 230 in my opening remarks, we learned something there. We decided that in Section 230 that we were basically going to absolve the industry from liability for a period of time as it came into being. Well, Mr. Altman, on the podcast earlier this year, you agreed with host Kara Swisher, that Section 230 doesn’t apply to Generative AI and that developers like OpenAI should not be entitled to full immunity for harms caused by their products. So what have we learned from 230 that applies to your situation with AI?

Sam Altman:

Thank you for the question, Senator. I, I don’t know yet exactly what the right answer here is. I’d love to collaborate with you to figure it out. I do think for a very new technology, we need a new framework. Certainly companies like ours bear a lot of responsibility for the tools that we put out in the world, but tool users do as well. And how we want and, and also people that will build on top of it between them and the end consumer. And how we want to come up with a liability framework, there is a super important question. And we’d love to work together.

Of course we all know by now (but didn’t know then) that Altman is a scoundrel. And as far I know he didn’t do diddly squat with Senator Durbin to find an alternative to Section 230. His company did however recently support state law in Illinois that “would shield AI labs from liability in cases where AI models are used to cause serious societal harms, such as death or serious injury of 100 or more people or at least $1 billion in property damage”– pretty much the opposite of what he said to Senator Durbin in his deceitful, lie-filled testimony.1

To my great pleasure, though, a bipartisan subset of the Senators who were in the room that day – Lindsey Graham, Dick Durbin, Josh Hawley, Amy Klobuchar, Richard Blumenthal, and Marsha Blackburn – have in fact introduced a bill to “Sunset 230”. I wish them luck — lots of it. (As far I know, though, it has never been voted on; govtrack.us gives it a 1% chance of being enacted, 4% of making it out of committee, but given the extremely-well-funded opposition from the tech titans I wonder if those numbers might be a bit optimistic.)

In the meantime a reader, who wishes to be anonymous, just said something really interesting.

The new German decision that holds companies liable for their chatbots’ errors, might arguably be true as well under US liability law. Because Section 230 doesn’t exempt companies from what their own software does.

Section 230 is, and always has been about 3rd party speech. The German courts remind us that chatbot-product speech is not that.

If Internet Provider X carries BS Artist Y’s lies, Provider X is not responsible for what Customer Y said.

But what Germany just said, essentially, is the Google’s chatbot lies are Google’s problems. It’s not third party speech. It’s Google’s speech.

If American courts were to rule similarly, or if Congress made that the law of the land, all the LLM providers would be in deep, deep trouble. Because essentially all their software has a tendency to fabricate fairly regularly and even defame people and to sometimes give bad medical information, and so on. In Taming Silicon Valley, I wrote about one such case, a prominent law professor who was accused of sexual harassment by an LLM when in fact the whole thing was just an LLM fabrication, not based at all in fact.

Imagine if Google, and OpenAI, and Anthropic, and Microsoft, xAI, and so on, were held liable for the garbage their systems sometimes produce.

It might actually force the tech industry to get its act together, to try to find a sounder foundation for AI, in which hallucinations weren’t baked in.

In the long run, we would all be vastly better off.

Update 1: I was about to post this, AI law expert Ryan Calo hinted at the same thing in a reply to me on X.

Update 2: Neil Turkewitz made the same point last year, and drawing on an earlier essay in by .

Subscribe now

Coupled with my newsletter from earlier today on SoftBank and OpenAI (which is seeming more relevant by the minute) and another day on Wall Street like this, it looking like it could be a tough week for generative AI.

Subscribe now

P.s. My thought just now on Masa and the industry as a whole

:,

Part I: Flashback

A couple weeks ago, Steve Eisman, who predicted the subprime mess (Steve Carrell played him in The Big Short), interviewed me, and asked me how the AI boom might all fall apart.

For reasons I am about to explain, that interview, recorded May 22, now seems supremely relevant.

Here are two snippets. (Some parts of Snippet 1 are currently paywalled in a separate video but may come out soon for free; I will update with a link if they do.)

After that, I will explain why all this seems super relevant today.

Snippet 1: Stumbles from OpenAI could precipitate a tidal wave

Eisman: Assuming it’s a bubble. I mean if it’s not if if we’re wrong and it’s then and it’s not a bubble and and … they’ll they’ll be able to charge for the tokens and it’ll all become profitable and it’ll be and that’ll be clear within a year or so, then …

Marcus: Yeah. I mean I like to think of [the giant GenAI investments[ as it’s a bet. I think it’s a low probability of the bet coming through, but it’s not zero. It’s not

Eisman: Zero. But if it’s going to break, like in, you know, what what broke Subprime? What broke Subprime was the credit quality got so bad that the end user, the investor, stopped buying the paper. And that was the end. Because if the end user, the investor stopped buying the paper, the whole machine stopped dead in its tracks. The question I ask for you is: I mean, I wonder if maybe what will break it if it breaks is that the industry moves to this token pricing and people say, Screw you. I don’t wanna I don’t I don’t wanna pay.

Marcus: So that could certainly happen.

Eisman: And if or and if maybe that what what else do you think could happen?

Marcus: Well, so in a version of yours, and then I’ll give one of my own, what think might happen is that at some point one of these companies can’t really afford to keep losing money. Right? They’ll run out of VC money, they tap the public markets as much as they can.

My best candidate for this is open AI. Open AI, I think, has a problem right now. They are burning money the fastest. Okay.

Eisman: Which is what?

Marcus: They do not have the deep pockets that Google does. Right. They had a lead but they squandered it.

Eisman: Google got caught up and passed them.

Marcus: And then Anthropic did. I mean, maybe now they’re all kind of in a tie, but think about it compared to a few years ago. You know, everybody thought (not me, but almost everybody) thought that OpenAI was amazing. Sam Altman could do no wrong. Right. Now nobody trusts Sam Altman, there was the New Yorker story, there was the testimony in the Elon Musk trial. It’s pretty clear he should not be trusted. So I’m a company thinking you know, hypothetically, thinking about buying these services.And my choices are OpenAI, Anthropic, Google, maybe Amazon. I might not go to OpenAI anymore. And you know, th that’s reflected in the choices people are actually making, right? Anthropic has taken a lot of the business market and OpenAI has lost a lot of it. Google is fighting hard for that. So there’s a scenario. I tend to think that OpenAI is the one that goes down first. They’re also spending most money with least assets. they’ve made the most commitments. At some point they might not be able to pay their bills. Now, what they have done is they’ve kept raising the valuation. They’re starting, as your listeners might actually understand, to do weird things on preferred shares and contingencies on that funding, like some of the Amazon funding is contingent either on achieving AGI or IPOing, et cetera, et cetera. So they’re they’re making more and more concessions to the investors. They’re I saw one report that they’re offering private equity sa guaranteed seventeen point five percent returns. Which reminds me of Madoff and, you know. somebody else said maybe that’s like a first in kind of money thing, but there’s something that Reuters reported around this. You know, none of the paper is public, so we don’t know exactly. But you know, it’s pretty clear that they’re making more and more concessions to investors. They’re more and more urgent. I mean, the the biggest news just in the last few hours is that they want to rush their IPO in front of Anthropic…. they’re clearly afraid of anthropic. I think you will [see] when the numbers are really fully out, then anthropic is more efficient. we don’t know exactly why that is, but they’re more efficient. they’re less leveraged out to the hilt. They’ve made fewer promises. And so if I had the prospectuses of the two, you know, independent of what I think about the companies anyway, I’m pretty sure if I was had to invest in one, I would do Anthropic and not OpenAI. And so at some point, I think OpenAI is gonna have a problem where they can’t really meet their obligations.

And if they do, that’s gonna have ripples throughout the market. And it’s gonna be interesting sets of ripples. So

Eisman: Ripples … it’s gonna be [a] frickin’ tidal wave. I mean, as an example, Oracle, and full disclosure, I own a little bit of Oracle. I sometimes I lose sleep over the fact that I own Oracle. When Oracle reported its third quarter numbers, the stock was like 230 before they reported. And they reported these big numbers and then they came out with I can’t remember what the word was, but it is is like a version of backlog.And the backlog was like five hundred and fifty billion and it was up like eighty five percent and everybody went absolutely insane. And the stock went from like two hundred and thirty to three hundred and thirty. Like like overnight. Well

Marcus: I wrote about that on September eleventh [2025], I know exactly when it was.I wrote a piece that I called Peak Bubble. Right. And my claim was not that Oracle would never go higher than that, but that we would look back on that moment which you’re now talking about as that kind of height of absurdity of all of that. Well the prob…

Eisman: [the] problem was it turned out, you know, two days later everybody figured out that of the five hundred and fifty billion [projected revenue], like three hundred and fifty billion was just [from] openAI. That’s right. Everybody said, wait a second, I’m now completely dependent if I own Oracle, I’m now completely dependent on V C raising money for open AI to fulfill its obligations to Oracle. And the stock went from three thirty to one hundred eighty.

Marcus: Right And [Oracle] has kind of hovered around there. Maybe it’s two hundred now or something. But it basically dropped within a few days or a few weeks of I think it took a few weeks. Right. But within a few weeks it had had dropped down. And so it was at like 310 when I wrote my article Peak Bubble.

Eisman: So if you’re telling me that openAI is a problem, the ripple effects are big. I mean, this is not like, you know, some little subprime mortgage company. I mean, this is a massive company with massive obligations to massive tech companies whose names we all know. Right.

And the ripple effects would be … are going to be huge.

Marcus: That’s right.

So that is my, you know, most likely scenario. There are others, but that that is the most likely scenario to me is that openAI at some point can’t really make ends meet. Anthropic eats their lunch, Google eats their lunch, people have less confidence in them, et cetera. Right. And then you have all this stuff booked for Oracle, for NVIDIA.

….

Eisman: Yeah.

Marcus: You know, and I frequently use the metaphor of Wily Coyote over the edge of the cliff

Eisman: where he’s kinda like flailing his legs and doesn’t fall.

Marcus: Yeah. And then he looks down and he falls. That could be the moment where everybody’s just like, Okay, we’ve got this wrong.

Clip 2: The death of tokenmaxxing

We also talked about token economics:

Marcus: So agents are basically systems that will do things on your behalf. So they might make travel plans for you, or they might run code for you, they might write code for you. So we had this paradigm initially with ChatGPT, and it actually goes back to before ChatGPT, but where you would type in a question and you get an answer. And it might be like a yes or no query or, you know, tell me where you know where William Shakespeare was born.

The agent paradigm is different. You basically say, do stuff for me.

And the reason that those tend to be more expensive is that they tend (with the help of some good old fashioned AI that nobody wants to talk about) to run multiple large language models in the background many times until they get the answers that by some criteria seem to correct. And so sometimes they end up taking, you know, a thousand times more computational or compute or or a million times or billion times or whatever. And so the prices of the agents have been quite high.

Eisman: So let’s talk about the economics of this, because what I’ve read is there’s a there’s a cost for the tokens, and any service that anybody has signed up for, which is the a subscrip most b basically a subscription model, is way lower than the actual cost to create the answers that that we know why did Julius Caesar cross the Rubicon and it’s certainly a hell of a lot less than any agent. So now the industry is starting to shift to token pricing. I think Microsoft is about to launch it any day now. And my question for you is this is going to be a lot more expensive. And how are people going to react to this? I can’t I can’t imagine positively.

Marcus: It’s a question of who assumes the cost. So maybe a metaphor here is an all-you-can-eat buffet. You know, these these companies were serving all-you-can-eat buffets and people were eating a lot, especially when they started using agents. And the margins here are either very slim or negative in many cases, meaning that it’s costing the companies more to serve the tokens than than than you’re paying them to use it. And so they’re in a bind and it’s especially kind of pointed or poignant bind because these companies want IPO. And so they’re trying to, you know, make their numbers lookgood in some way. And it’s this trade-off because basically they’ve been running this stuff at a loss. And with the agents come along, it’s an even bigger loss. And so they’re like, you know, on the one hand, they want to get customers to use their stuff, and they’ve beenpretty effective at that. But on the other hand, they want to make a profit. And those are actually kind of in tension, right?

They’re always in tension to some extent, right? You know, the the smaller your margins, the more your customers like it. but it is turning out that it is a little bit like an all-you-can-eat buffet. And some customers were not eating that much. But when people started using things like OpenClaw, which is a kind of system for using LLMs as agents, some people started eating an awful lot at the buffet

The question is, how do you make this economically viable?

Part II: What just happened

First of all, domino one: tokenmaxxing has indeed clearly died, just as Eisman and I had anticipated a few weeks ago. The LLM providers no longer want to offer all-you-can-eat and customers no longer want to pay for a la carte. The fundamental problem is that to make a reliable version of the product you either have to charge more than customers want to pay, or the providers have to take a loss. Then again, if you read this newsletter regularly you already knew that. (Others perhaps learned it today; confirming what I had been saying for weeks, ZeroHedge just reported that Citrini Research “ has written a follow up on the status quo of the AI ecosystem, noting that in just weeks we’ve gone from tokenmaxxing to tokenpanic.”)

The death of tokenmaxxing means less revenue for the LLM companies, which can only hurt the IPOs. And more and more people are noticing.

But, domino 2, something else just broke. And it may be deeply important.

It needs some context.

The context is that around a month ago SoftBank tried to take a $10 billion margin loan based on its (very large) stake in OpenAI. This didn’t go over well with the banks, as Bloomberg noted (and I as mentioned in earlier essay):

The new, breaking news is even worse. Bloomberg is now reporting that the banks don’t even want to do the smaller loan:

SoftBank took an immediate hit; dropping the ask reeks of desperation, and its a terrible vote of nonconfidence that the banks still said no:

That’s domino 2: we can infer that the banks don’t really think that OpenAI is worth its current valuation. They don’t want OpenAI stock as collateral.

And OpenAI wants to IPO at even higher valuation. That’s going to be hard given this news.

So maybe domino 3 is this: OpenAI may well wind up as the WeWork of AI. (Trivia fact: Masa famously invested big in WeWork just before they fell).

Here’s a version of the argument I made two years ago:

Two years later, competitors have caught up. OpenAI has had to cut prices. More key employees left, Opensource models (though perhaps not Meta’s) are only a few months behind. Sora briefly shipped but ultimately was pulled from the market altogether. GPT was delayed (until August 2025, as predicted). The latest models (though not 5 specifically) are costly to run, and cost is a growing issue. Core problems of reliability and factuality haven’t been solved,. And OpenAI still has yet to turn a profit, and it’s still not clear they ever will.

Every bad sign is worse now. The big difference is that people are finally noticing.

For perspective, my tweet above got a measly 34,000 views; now the entire world is worried about whether AI is a bubble. And rightly so.

That Wile E Coyote moment could turn out to be soon.

Which takes me directly to Part III.

Part III: An Updated Scenario

What happened, of course with WeWork is that they were planning an immense IPO, and everything fall apart; in October 2019, not long after raising an enormous round from Softbank, they withdrew their IPO.

If the market gets squirrelly on OpenAI’s IPO, which after today seems like a real possibility, OpenAI might be forced to withdraw its IPO (Domino 3), or to cut their aspirations sharply. If so, I don’t see how they could meet their immense ($600 B) obligations. (Domino 4)

If they can’t meet their obligations, I expect Nvidia and Oracle and many others would take a big hit as collateral damage (Domino 5). And the damage might affect the banking system, retirement plans and more (too many dominoes to count). There would definitely be discussions of bailouts.

(Disclaimer: This is not investment advice, it just what I am seeing. Use these thoughts at your own risk.)

and how Tom Friedman freaked out about it at the New York Times?

And how Anthopic said Mythos was too dangerous to release, and held hushed meetings in Washington?

And how Anthropic’s valuation leapt up after that?

Well here, we are, two months later. And guess what….

(A) the world has not ended (remember how I told you not to panic?) and (B) suddenly, two months later— just as tokenmaxxing has come to end and companies have begun to limit their AI token budgets — Mythos is no longer too dangerous to release.

Nope.

Instead, Anthropic tacked some guardrails on, and pushed it out the door.

Anyone can now use it, for the right price. A press release I just got in my email gives up the game:

Yep, Anthropic — that once said they wouldn’t work on frontier models because they were dangerous, and that called for a pause “option” that would probably never in a million years invoke unilaterally or so long as China exists — just introduced their new frontier model, hellbent on racing everyone on all the benchmarks.

Anthropic once again played the media like a fiddle. And so many folks ate it all up.

§ 

And worse, this routine is not even new. Anyone remember 2019?

If you scroll down far enough in the original 2019 GPT-2 is too dangerous to release blog post, from OpenAI, you will recognize some of its authors who moved on other companies—including Anthropic’s Dario Amodei, Daniele Amodei, and Jack Clark.

They have been running literally the same play for seven years: scare, hype (evoking media interest), and (eventually) release.

Scare, hype, release. And repeat.

Is it that hard to see?

Subscribe now

As someone on X just pointed out (I am rewriting slightly for clarity, and trusting their numbers, which sound about right to me), Amazon returned 2,538x the original investment. If SpaceX were to do the same, its Market Cap would be $4,442 trillion, or 36 times the current total world GDP.

That is absolutely not going to happen.

It’s insane.

§

Another way to think about it.

I have to think that anyone who buys the SpaceX IPO just isn’t that good at math. Or history.

§

At least a few people have figured that out:

§

A new study , from Jessica Wachter (Wharton) and Jonathan Wachter (Point 72), aims to calculate how much productivity needs to grow for the whole thing not to fall apart. (In technical terms they sought to derive “the cumulative increase in productivity required to justify this level of investment, assuming no change in discount rates, and zero economic profits … Were then boom to be smaller than this, the investments would be NPV-negative.”)

I haven’t looked carefully at the math, and the paper’s lead author, Jessica Wachter, added some nuance on the tweet above’s characterization when I asked but they don’t change the overall situation all that much:

The tweet is correct in that 2.7 is the required multiple of productivity. The missing context, however, is that the 2.7 number applies to the AI sector, broadly defined. Also, large companies with relatively low debt have ways of staying out of bankruptcy, so there would be some leeway in how long that increase in productivity would take to play out. That said, we do assume it would occur by 2028 (which is reasonably quickly).

In any case, I share the paper’s observation that what we have now is “a productivity boom that is in investment data, but not yet in productivity data. If the boom fails to materialize, the current buildout will be the largest misallocation of capital in history” (a phrase and conclusion almost identical to one I used here in April).

§

If there is a boost in productivity, we will perhaps see it first in coding, but 2.7x relative to the entire economy is an enormous ask.

§

And that doesn’t even of course count the many other risks to society, many just summarized in this massive review:

§

Yet donations to Jensen’s retirement fund keep pouring in. Data below are from https://isaiprofitable.com, which I strongly encourage to you to visit.:

The best part of https://isaiprofitable.com is the dollars “spent since page load” counter, which tells you how much money is wasted while you linger on the page. If you stay for around a minute, you can be witness to burning of million dollars of capital.

AI is not profitable yet, and may never be. And, breaking news, costs are rapidly rising too:

GenAI is a risky bet that could pay off, or it could well tank the economy. (A more reliable form of AI might be a safer bet.)

Certainly, SpaceX will never return the way Amazon did. OpenAI is in trouble, and may not make it. Anthropic has a chance of being profitable, but reaching 10x the IPO price would be amazing. To get to 100x the expected IPO price of $1 trillion would mean a market cap of nearly the entire world GDP (about $125 trillion), another manifest absurdity; 1000x is out of the question. (Also the whole thing Jensen Huang said is an example of a kind of post hoc fallacy, in which you take the most extreme outliers in history post hoc and project from there. Elon is presumably counting on retail investors to succumb to the same fallacy.)

I do have to wonder whether the guys talking this stuff up believe what they are saying, or whether they are saying it even though they know it can’t be true – or whether they are just bad at math.

Subscribe now

The particular graph from the FT happens to be about mobile apps (h/t Jen Zhu). But you could make similar graphs in many domains: loads of nominal productivity, but not much real world impact.

You could, for example, say something similar about books. GenAI has led to a lot more books, but it’s not clear that they are better books or (sorry I don’t have the data immediately available) more books sold. Here are numbers of books available over time, from The Washington Post:

Book sales, on the other hand, have declined slightly over the same period. And I don’t hear any argument that books are getting better.

The Post article that I drew the above from had similar graphs for the rise in music tracks uploaded, self-represented lawsuits filed, scientific papers submitted, and web content generated.

Again there is no reason to think that any of it has added to the GDP or quality of science or music. Most of it is slop (and in some cases work slop):

The graphs from FT and The Washington Post show the same thing over and over; we are being flooded with AI-generated content, from apps to music to scientific papers, but that doesn’t make any of it any good. Most of it (in fairness, not all) isn’t.

And then you have slop inundating Wikipedia, libraries and so on.

It is very likely that the same thing is happening in math. A group of mathematicians, who are fearful about many aspects of the impact of AI on math research, have written an open letter called the “Leiden Declaration”, reported by the NYT, expressing (among other concerns) their fear that

Current automated techniques can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs. This applies not only to informal arguments, but also to formalizations, where the difficulty lies in the translation between computer-encoded and human presentations of concepts. These fast-moving developments put our present system of review under increasing pressure, jeopardizing our ability to implement traditional standards for the correctness, transparency, and independent verifiability of proof.

In other words, math slop.

Subscribe now

Not long ago, when I was thinking about the official productivity measures and whether they would change with AI, an investor friend wrote to remind me how narrow the technical definition of productivity with respect to GDP is, noting that “In the technical world of national accounting, paying people to dig and refill holes adds to GDP.”

Using GenAI runs up tokens costs, and floods the world with content. But all too often that content — created by a giant but unreliable word prediction machine that is more regurgitative than innovative — is of little lasting value.

Maybe the biggest exception is coding, but even there, it’s not yet clear how many of the systems that have been created with the help of agentic coding will endure.

§

Moreover, especially in agentic coding, there are massive losses, up and down the food chain. The providers (Open AI, Anthropic, Cursor, etc) lose tons of money, and are scrambling to raise prices, Customers are balking at the new usage models.(In a new analysis, Gerben Wierda has argued that Anthropic and OpenAI “may actually pay $1000 for every $100 you pay them”. Data center intermediaries like CoreWeave are also losing money. It’s hard even to say how the chip companies are doing, once you factor out all the circular financing and shrinking cash flow. The one possible major productivity story (coding) also happens to be the one burning the most cash.

It could well be that once the providers (Open AI, Anthropic, Cursor, etc) attempt to charge enough money to cover their losses, the cost of AI could become more expensive than the humans it is replacing.

Talk about digging holes without creating real value.

Share

PS. See also and Paul Kedrosky for new pieces reaching somewhat similar conclusions, from different angles.

PPS Apropos tweet, tongue presumably partly in cheek, from yesterday:

I doubt Friday was as bad as it will get, so the title of this essay (“AI’s Black Friday”) is a bit tongue in cheek. But Friday was bad; across the tech industry something on the order of half a trillion dollars of market value disappeared overnight.

The Dow was down, but only 1.35%. Chip companies (NVidia, Broadcom, Micron) GPU leasers (CoreWeave and Nebius) and some other major AI related companies (Oracle, Microsoft, Meta, etc) all took larger hits:

The S. Korean market, heavy on semiconductors, was also down.

And that wasn’t even the worst of it. Maybe this revelation was:

Remember this tweet?

The game is on. “Investment” is just another word for bailout.

§

I had this to say about that:

In short, it stinks.

Bailing out OpenAI is crony socialism and corruption (remember those $25M donations from Greg Brockman and his wife?), preventing capitalism from taking its natural course.

§

But it’s not just OpenAI:

As an important aside that is more about geopolitics than economics per se, the ramifications of this may be huge. As I noted on X, nobody is going to trust an American AI company that is partly owned by the US Government. Just the way the US doesn’t trust Huawei, Europe and Asia are not going to trust companies like OpenAI and maybe even Google. The new cold war will be about AI. (And Google might perhaps be wise not to accept Trump’s proffered investment, given how much of their income is global; OpenAI may be so desperate they have no choice.)

§

But back to economics. Who can forget the hapless Meta, which may ultimately win the prize for most money burned vs least results?

§

Meanwhile, it also came out yesterday that Musk is leasing another 110,000 GPUs to Google, on top of the 220,000 it is leasing to Anthropic.

Last year, people had to steal GPU’s from armored trucks.This year, SpaceX’s AI division is leasing GPUs left, right, and center, because they can’t figure out what to do with them.

If scale was “all you need”, or if AGI was actually nigh, Elon would be hoarding LLMs, not leasing them.

Some people might ask whether Elon is making money on the deal (I don’t know), but the real question is why he is leasing the very chips that he hoarded last year in the first place.

Whether SpaceX is making money on the deals with Google and Anthropic or losing money, they are waving the towel on winning the frontier model race— by arming their competitors rather than themselves.

§

Also, per a friend who read a draft of this, “If Space X is leasing capacity back to Google and Anthropic, this is now servers and chips those guys do not need to buy themselves. Elon clearly bought a lot of hardware he did not have demand for. He is probably not the only one that bought substantial capacity long before there was demand or a use case for it.”

None of that is particularly great for Nvidia, CoreWeave, Oracle, Nebius, Micron, and a variety of other hardware and semiconductor providers

And my friend adds, “Then of course the fact that even Google and Meta, who are [historically] very cash flow positive companies, need to sell equity to continue to fund their AI investments. Just shows you this whole thing is a black hole.”

§

Share

I’ll leave you with these predictions, written a few days ago, before the shit started hitting the fan:

I have no idea whether Monday’s market will continue the downward trends we saw Friday. But it should.

Subscribe now

Coming soon: more on why AI has had less impact on real-world productivity so far than many people had anticipated.

but did Anthropic actually call for a pause?

No. Not really. Read carefully:

Reading between the lines, they want it both ways.

They *don’t* actually want a pause - at least for now. Rather, they want to (and will) rush ahead, hinting at “least cautious actors” for justification.

Instead, they want people to talk about an “option” they don’t actually plan to take, and are unlikely to ever take. (More likely, they will like always hint at China, and continue rush ahead.)

It’s an incredible, cost-free piece of rhetoric — perfectly timed for the IPO.

Caveat emptor. Caveat lector.

Subscribe now

I believe that we’re only a few years away from that, maybe like 2030 plus or minus a year, which is astounding to think, really.” [@ 27:09]

Which is funny, because Sir Demis Hassabis [let’s call him Sir Demis Hassabis #2] , just a few months earlier in January 2026 at Davos, had something rather different to say, anticipating a somewhat slower timeline of 2031-2036:

Per the YouTube transcription at 6:26, Alex Kantrowitz asks Hassabis, “We’ve brought up AGI a couple times. Um so let me let me put this to you because I was speaking with Sam Alman towards the end of the year and I asked him l was like you know you seem to be saying two things. We’re not at AGI yet, but every time he talks about what GPT models can do, it seems like it fits his definition. And he said uh a that AGI is underdefined. And what he wishes everybody could agree to was that we’ve sort of whooshed by AGI and we move towards super intelligence. Do you agree with that?

Here is reply, boldfacing his definition of AGI which I share:

I’m sure he does wish that, but it’s um No, absolutely not. I don’t think AGI should be sort of turned into a marketing term or for commercial gain. I think there is always been a scientific definition of that. My definition of that is a system that an exhibit all the cognitive capabilities humans can and I mean all.

So that means you know the kind of highest levels of human creativity that we always celebrate the scientists and the artists that we admire. So it means you know not just solving a maths equation or a conjecture but coming up with a breakthrough conjecture that’s much harder you know not solving something in physics or some bit of chemistry problem even like alpha fold you know protein folding but actually coming up with a new theory of physics something like like Einstein did with general relativity right can a system come up with that because of course we can do that the smartest uh humans with their brain architectures have been able to do that in history and the same on the art side you know not just create a pastiche of what’s known but actually be Picasso or Mozart and create a completely new genre of art that we had never seen before right

Aand today’s systems in my opinion are nowhere near nowhere near that doesn’t matter how many you know Erdos problems you solve which for some reason. Uou know I mean you know that it’s good that we’re doing those things but | think it’s far far from what you know a true invention or someone like a ramen would have been able to do and you need to.

And you need to have a system that can potentially do that across all these domains. And then on top of that, I’d add in physical intelligence because of course, you know, we can play sports and control our bodies and to amazing levels, the elite sports people that are walking around, you know. here today in Davos. And we’re still way off of that on robotics as another example.

So l think an AGI system would have to be able to do all of those thingsto really fulfill the the the original sort of goal of of the Al field. And I think, you know, we’re 5 to 10 years away from that.

I am 100% with the more conservative Sir Demis Hassabis #2.

We won’t see AGI this decade, and we not even close, exactly for the reasons that he laid out.

Share

The twitterverse is all verklempt with Anthropic’s latest blog.

The blog, which shows considerable advances in the ability of their systems to write code, is interesting, and make an argument for slowing down AI research that should be considered.

But it should also carry an advisory. Here’s the advisory they forgot to include:

AGI is *harder* than “recursive self improvement” [RSI]

AGI: machine can do anything human can do, autonomously. That has not achieved]

RSI would be “an AI system capable of fully autonomously designing and developing its own successor” but what’s been shown us so far a useful coding tool that humans can leverag; it’s great for example at (some) code optimizations, but not even (as the blog makes clear) actual RSI, which remains speculative.

The results in the blog are about things that hint at the possibility of RSI, not AGI.

Getting to AGI will require new ideas, not just new code optimizations.

So we don’t need to panic yet.

To some degree, the new Anthropic blog is a bait and switch: Anthropic is trying to strike terror into everyone’s hearts (“full recursive self-improvement also might increase the risks of humans losing control over AI systems”) but all they have really shown is just faster coding — entirely under human control, along with progress on some benchmarks.

A faster coding tool will probably not end the world.

§

More good news. Terrific news, really.

I had written before in an essay called “This one weird trick might cost your retirement fund billions” that the S&P 500 was considering rules changes that would have allowed SpaceX to be included in the S&P 500 faster and with fewer restrictions than ever before, an idea that I thought could prove horrific for retail investors and retirees investing in index funds, given that many (thought not quite all) onlookers find the initial share price to be loopy.

To my great joy S&P did NOT change the rules!

You can click the above to read the full reasoning, but the bottom line is that Musk’s dubiously-priced SpaceX offer will NOT be immediately jammed down index funds without delay.

Instead, SpaceX will have to earn its keep; index funds may still wind up indexing it, but only after the market has had a year to evaluate its worth.

That is great news for fairness. Rejoice!

Subscribe now

P.S. Some commentators have interpreted the coding speed up results in the Anthropic blog as showing that AI or deep learning are not hitting a wall. It actually shows the former but not the latter. The progress is real, but this isn’t a victory for pure scaling per se, it’s victory for harnesses and symbolic tools, which are neurosymbolic systems. Pure deep learning has largely hit a wall; neurosymbolic AI is rescuing it. (Put differently and more personally, my own claim was never that AI would hit a wall, or that deep learning should be abandoned, but that we would need to supplement deep learning with neurosymbolic AI, in order to address the limitation of pure scaling. And that’s exactly what happened.)

Just over three years ago, at the US Senate, Senator John Kennedy (R-LA) asked me for my three most important suggestions for AI legislation.

Number one on my list was preflight checks, similar to the FDA review for drugs, for models with large commercial impact:

Today, most definitely not on my bingo card, President Trump signed an executive order that was very much along on these lines:

I can’t take credit for causality, but it’s thrilling to see a dream come true.

May any such oversight be bipartisan, and in the interest of American citizens, and in the interests of humanity as a whole.

Subscribe now

1. Everybody, even Google, seems to be treating AI as if it were some kind of winner take all competition like web search was, in which Google taking over 90%

2. But everybody is building essentially the same technical solution with essentially the same data, so there is no moat.

3. If there is no moat, nobody is going to take 90% of the market.

4. With no clear winners, nobody can charge monopoly prices; instead, you get price wars and commodity pricing.

5. Which means everybody will wind up overpaying compared to the modest profits they will be able to make in an intensely competitive regime.

Am I missing something?

Part II: The Psychology

None of that would matter in the near term if people weren’t noticing.

But they are. Consider:

1. When I posted the above last night on X, sparked by the jawdropping announcement that Google of all places was raising equity financing , over 750,000 (and counting) viewed the tweet overnight. And what’s more most people agreed, which in that crowd is very rare.

2. The financial world is taking a lot more notice of my skepticism than it used to. Just yesterday two significant (and excellent) podcasts with me just dropped, a return to the show hosted by the well-known Steve Eisman (played by Steve Carrell in The Big Short): and the other a four-hander with the well-known investor George Noble, the economist Julie Garran and the financial analyst/reporter who is known in the youtube world as @nobodyspecial. Everyone involved in both podcasts share my concerns, and had sharp arguments of their own to add. Garran’s distillation of why the LLM industry will never be commercially viable was crystal clear and compelling.

3. My critique of tokenmaxxing is getting picked up, too; everybody seems to think that trend is over. And companies like Anthropic are ending their all you can eat buffet, which will make many customers think twice.

4. Still more reports have coming out questioning the RoI for corporate customers. Bain’s conclusions yesterday almost perfectly the old definition of insanity that I keep quoting (doing the same thing over and over again and expecting different results):

§

Timing when all this might fall apart is hard, but the economics don’t make sense, and more and more people are starting to notice.

What is most upsetting though, is how retail investors and retirement funds are likely to be taken advantage of in all of this.

Same will happen with Anthropic and OpenAI. All of which led me to this prediction yesterday:

I seriously hope that I am wrong about this. But I doubt it.

To close, here’s one more economist, reaching basically the same conclusions; I leave the last words with him:

The technology is powerful but unreliable. It must be boxed into narrow, supervised uses. The economics of inference are unforgiving. The competitive environment is crowded, with open models and vertically integrated hyperscalers eroding any pricing power that an independent lab might hope to have.

What is the plausible, well‑specified path by which Anthropic or OpenAI grow into the kind of durable, high‑margin franchises that would justify the valuations their private rounds have implied?

There is none visible.

Subscribe now

The Noble Update

AI: The Biggest Capital Misallocation in History

Read more

15 days ago · 31 likes · 1 comment · George Noble

I could write a long essay about why this makes very little sense, similar to the one I recently wrote about Richard Dawkins, when he made a similar error, the crux of which was below. All I would needto change is the words in bold.

The fundamental problem here is that [insert name of well-known senior figure here] doesn’t reflect on how [AI] outputs have been generated. Claude’s outputs are the product of a form of mimicry, rather than as a report of genuine internal states.

Consciousness is about internal states; the mimicry, no matter how rich, proves very little. Dawkins seems to imagine that since LLMs say things people do, they must be like people, and that simply does not follow.

In his framing, [Senior Figure X] confuses himself, and does violence to the concept of consciousness [and what it is to be a being]. You can’t just look at the outputs, without investigating the underlying mechanisms, and conclude that two entities with similar outputs reach those similar outputs by similar means. And the differences are immense; one (the LLM) effectively memorizes the entire internet; the other (the human) builds a mental model through experience with world.

But why bother to rewrite that essay?

A couple days ago, before Hinton’s interview, Pope Leo XIV said it better, and more concisely, in, of all things, a tweet (following up his excellent Encyclical):

True comprehension comes from experience, not text approximation.

§

As Valerio Capraro notes, the Pope’s point is quite similar to a point that we made together with Walter Quattrociochi in February in Nature:

§

LLM researchers are NOT creating beings.

They are creating interactive fiction that is trained to predict the language of actual beings.

Those two are NOT the same.

And Hinton should know better.

Subscribe now

P.S. Want to know my thoughts on internal “emotion” states in LLMs, and why they are not compelling evidence? Stay tuned for a forthcoming essay. For bonus points, you can see my recent impromptu debate with the artist/musician Grimes on AI and consciousness, in two threads here and here.

and

Furthermore I saw bunch of other people saying similar things noting the apparent decline of toxenmaxxing, such as Fortune’s Jeremy Kahn in his newsletter:

So ok, if that’s true, what’s next?

Here are two takes. First my own:

And second, AI researcher and benchmark creator , who fired back in a reply to me with a totally different, far more optimistic set of predictions, spelled out in admirable detail:

Within minutes, the contrasting pair of predictions become a meme:

Share

We agreed to check back in 12 months to see how these predictions are going.

Stay tuned. And feel free to drop a comment below.

Leave a comment

Subscribe now