"AI — not future, speculative AI, but AI as it exists today — has nonspeculative economic value. I pay $100 per month for Claude because Claude provides me with comfortably more than $100 per month of economic value."
Golly, how the goal posts have moved! I thought AI was going to cure cancer. I thought AI was going to result in 20% a year GDP growth, at a minimum. I thought AI was going to end death. Where's all that stuff? This is the point that AI hypemen can't respond to: the outsized claims about what AI is going to do have not been met by demonstrations of what AI can do. It's still all speculative. And we're nowhere close to justifying the economic investment involved. Who cares if you're get $100 of value out of Claude, if as a society we're paying far more than we get back?
Freddie you're doing the thing again where you lump a bunch of people with different opinions together and call it goalpost moving, when in fact it's... a bunch of people with different opinions. Kelsey never claimed that AI would cure cancer, in fact she has noted in one of her last articles that AI has not yet really begun to speed up scienctific research.
I don't recall being asked "as a society" to pay anything. Investors are investing as they usually do. If it doesn't pay off, they made a bad investment. They didn't ask society if they should invest their money, but then I don't expect them to.
I don't think this is a great counterargument. When investors as a whole get too excited about a single type of asset, the accurate revaluation of that asset can take down so many investors at once that financial infrastructure takes significant damage like with the 2008 housing bubble. Right now it's all just investment firms playing games that don't matter to any of us, but it might not stay that way.
It's weird that, to defend the statement "AI delivers significant economic value" one is expected to defend the specific marketing pitches of every AI company that has ever put out a goofy interview, and every random zealot who has made wild speculations on Twitter
There were two sets of goalkeepers, one saying “AI will someday create 20% GDP growth” and the other saying “AI will never create any economic value at all”. AI has already grown to earn tens of billions. This is obviously compatible with one of those goals and not the other!
Freddie, what kind of evidence would you accept as probative of claims that AI will one day be capable of doing [x]?
The linear passage of time ensures that 'claims about what AI is going to do [will not be met] by demonstrations of what AI can do', but you probably need to be open to evidence based on progress to date or something, rather than just insisting 'that's still not [x]!'.
By way of example: while they still can't cure cancer, any moderately recent AI could have told you your argument was dumb and suggested not making it. That feels like progress.
LM models are widely used and developed in biology and medicine and they are absolutely delivering value, and this is not my niche (I worked on genome models for genetic diseases) but I'm sure they're being widely use in cancer research in this exact moment.
So the fact that you don't know this doesn't mean it isn't happening.
The fact that Amodei (iirc) said the 10-20% GDP thing is not an argument against the usefulness of AI, but I mean, it's obviously a bad faith argument.
Same with the end of death straw man.
The usage data tells use that indeed a ton of people is getting value out of these models otherwise you wouldn't see such numbers, and nobody is paying "far more than what it gets back", because the vast majority of users are not paying anything and in general the societal costs of running these things are pretty minimal compared to basically any other kind of production.
Unfortunately, there’s a very large audience of people looking to confirm their priors that AI is useless and/or fraudulent and/or unethical and/or ruining the youth and/or hallucinates and/or, and/or, and/or ….
I think this is actually the biggest obstacle to Zitron doing better argumentation. Sure, it's possible to imagine an AI critic making the financial criticisms Piper would like for Zitron to be making, but that wouldn't serve Zitron's actual audience.
Zitron's audience is the people you're talking about. They want to read about how AI is fake and how it's going to go away any minute now once the bubble pops, and that's why they read Zitron. A bunch of sober financial analysis based on the premise that it's real and won't go away doesn't meet that market's needs, and they'd unsubscribe. And it's not like Zitron could just pivot to writing for an audience of practical investors looking for real financial analysis, as that's super not his wheelhouse.
So he's pretty much locked in to writing about how AI is fake and on the verge of disappearing; I suspect he rides that into the ground until even those skeptics have given up on the idea, and then finds a new beat.
AI is useless, fraudulent, unethical, and ruining the youth. And it hallucinates in very seriously problematic ways.
There is very easily discoverable evidence of all of these points via scientific studies.
That's not what is at stake in the article. Is the LLM industry financially viable? Probably not because all of it is circular, but I have no idea if it is a bubble that will burst.
It just doesn't feel like a healthy economy/marketplace. And if the last 20 years have taught me anything about Silicon Valley, its that none of these guys are trustworthy or good people and we should stop listening to them.
My company is working through an enterprise agreement with Anthropic and I (head of tax) am one of a group of people testing out Claude enterprise. It’s almost magical. Claude Code built out the full tax return workpapers for one of our jurisdictions. Claude Cowork is solving major issues for us on the tax provision. Their excel plugin (sans Cowork/Code) has been rolled out across finance and is a game changer in and of itself. I’ve been using Claude Pro for personal projects for a while now, and it’s amazing to see how much better these models are getting in real time.
I am not an AI skeptic and use it for work and pleasure every day. That said, the bullet point list of what you pay for AI to do makes little sense to me. Finding out about a soccer club or missing books could be done much more easily by “asking someone” or “looking at the books.” I can’t figure out what kind of analysis you were even doing there.
I can't speak for Kelsey, but unless I have a specific website in mind whose name temporarily escapes me I don't bother with Google anymore. Claude is just better and faster than manual search. Upgraded Google is not worth $100/month, of course.
But if you’re looking at your homeschool library you can see what books are missing. If you’re watching a soccer team play you can ask someone “hey what team is this?”
I can see how some of these could be slightly faster. For the example of book cataloging, it could be a significant time saver if you're trying to individually check each book on your list and then check where to buy it if you need a copy and you want to do this for a list of 100+ books. Still, it's notable that all of these are low-skill tasks that could be done by any literate human considering that much of the hype in favor and against AI is over the prospect of replacing/enhancing expensive white-collar professionals.
The thing is that most households don’t have spare low skill literate labor on hand to do things while the adults focus on tasks like cooking and playing with the children. Spending a few cents having Claude look at a photo and tell you what’s missing is better than spending 20 minutes writing down a list of what you have and googling what’s missing.
Yeah, digital personal secretaries for everyone is a real value add. Still, it's notable that this isn't a big point in any faction's argument over what LLMs are or aren't worth. Everyone's speculating about if gen AI is as good as the average artist and if that's a good thing, if Claude or whatever is the equivalent of an omni-PhD in your pocket or if it ever could be. Meanwhile, the average person is using it to do meal planning faster.
Kind of reminiscent of how hype about the internet was grounded as its reach expanded. What feats will the typical person accomplish with zero-latency access to the complete breadth of human knowledge? Turns out that most of the time, most of us just want to watch TV and gossip about each other and increases in our abilities doesn't change our basic drives that much.
A lot of the little everyday uses for LLMs seem trivial and unimportant, because they are. But like, look at all the Google searches you do, and 95% of those are also trivial and unimportant (and could often be done by using other tools).
I think Zitron's most compelling argument, by far, is that OpenAI and Anthropic are currently selling their products at such ridiculous, artificially low prices--spending $5 billion to sell $2 billion worth of services, that sort of thing--that it's impossible to infer anything positive about their business model from the mere fact of their revenue, or revenue growth. This article's only response seems to be a hand-wavey "everyone's got unprofitable users." Sure, but that ducks the question, which is: are either of these companies likely to assembe enough profitable users that they will one day be able to survive without constant infusions of billions in VC cash? Zitron may not be right about everything, or even most things, but that seems like a very important question no other journalists are asking.
I agree with most of your article. I've read Zitron sporadically over the past couple of years, and agree that his dismissal of AI as useful technology was always off-base, and is basically delusional at this point.
Where I think he's on firmer ground is still the economic arguments. While I grant that a lot of usecases can be sold at sustainable rates for the value they provide, when it comes to coding it's clear that the most powerful (ie, agentic) workflows have been heavily underpriced. It's absolutely insane what I can do with my $20/mo Claude Pro plan, especially when I compare to what it'd cost me if I was paying API rates (which I can preview by enabling "extra usage"). For a while power users using the plans like this may have been in the minority, but it's clear that as agentic coding has achieved a step function in capability and awareness of this has grown throughout the technical and non-technical ranks, so has the demand for tokens. Yes, tokens get cheaper over time, but the biggest gains in capability have come from massive increases in the number of tokens used to complete tasks.
We're already seeing Anthropic toy with restricting access to Claude Code, and Microsoft Copilot just killed off subscription-based coding altogether and raised the cost of all the good models substantially. Part of this is simply a capacity issue, but it's a capacity issue because they've been basically giving the product away in these subscription plans. Having heavy demand for something you're offering at well below cost is not some great achievement.
The bubble argument that I find convincing isn't that AI isn't capable or won't continue to be a compelling product or won't continue to disrupt many knowledge industries, but that building and providing models remains a very tough business with awful margins. The fact that coders will flip from provider to provider (Claude was the hotness a couple months ago, now everyone's running to Codex because Anthropic is throttling harder, etc.) shows that any competetitive edge is fleeting and they're competing mostly on who can give the most tokens and context away for a given price. They have to expend insane amounts of capital to stay competitive, and then it's a race to the bottom on pricing.
The capabilities are incredible. I regularly max out my Pro plan writing personal software and as a professional software engineer, it's an incredible experience. I also am far less productive with this in my day job because the impediments to moving fast in a professional software environment are not primarily the cost of writing code. It is absolutely still worthwhile for some things, even at "more realistic" API rates, but a lot of these heavily AI-dependent workflows feel wasteful and like they're just optimizing one small part of the overall process with dubious ROI. With time I expect it will be revolutionary, but right now the net gains still feel incremental in an enterprise development environment.
I think the persistent skepticism of AI’s value is downstream of skepticism of the value of the “knowledge work” economy overall.
If you believe that most white collar employment is a kind of theater (the David Graeber “Bullshit Jobs” thesis), then automating white collar work will feel valuable to white collar workers and yet also not create much in terms of economic benefits to people at large.
I’m not as skeptical as Zitron, but it does seem like in seeing how this progresses we may learn if all those emails and PowerPoints were ever creating economic value after all. I expect some were and some weren’t.
Graeber really should have gone for the modern Marxist take - just as a 19th century factory worker could be alienated from the end product of their work and therefore feel that their job was bullshit, now we’ve got big information management enterprises that have commoditized white collar labor in such a way that the workers are alienated from the end product and therefore feel that their job is bullshit.
I honestly don't get this complaint even from Marx's time. I know I'm alienated from the end product of my labor, that's how I can contribute to projects far out of the scope of my understanding and provide value to people I don't know. What else am I supposed to do, only provide goods and services directly to people I know? Do communists think I have neighbors who can provide all basic material goods and would be eager to do so in exchange for custom software applications?
I think there are some anarcho-syndicalists who do think one should only exchange goods and services with people you know. I think more orthodox marxists just think that this sort of alienation exists on a continuum and that inhuman capitalist systems tend to produce worse versions but that with better structuring you could have better connection to the fruits of your labor. That’s probably right to some degree, but I don’t know what degree that is, and I suspect the gains from trade make up for at least some of the soul crushing of alienation.
As someone who generally likes Graeber's other work, he has a weird hate for delivering pizza in a way that makes me assume he just didn't like pizza very much. This, amazingly, calls him out as committing your Thesis #2 over the course of hundreds of pages.
There's also the closely related thesis #3: many jobs do things that other people value but I think they shouldn't value, therefore the work is bullshit.
Fwiw the Internet bubble of the late 90s was driven substantially by the phenomenon of weak/poorly conceived and/or poorly run businesses attaching ". com" to their name and receiving a windfall of investment. I was CFO of a small company in that time we were seeking capital to expand and both our investment banker and the investors we talked to literally said "we'll give you a 5 times bigger multiple if you can make the case that this is actually an online business" [which it wasn't].
This is the same logic of "no doc loans" as shown in The Big Short.
In both cases, the wasted money is going to the "wrong" growing business, not to a whole business category that doesn't make sense.
There's other ways for bubbles to form, but that's not what's happening with AI, so I think the.com analogy is poor. People aren't throwing investment at catchy startups because they think they'll dominate a new category by "AI-ing" it.
Interestingly, I think part of the reason that the investments aren't being driven by "AI-for-X" startups -- though those exist -- is the belief by investors that the labs could just sherlock whatever product those startups might build before even getting to a place where an acquisition would make sense.
I’m far from a right wing troll, but I can think of no better category for Ed Zitron than “shitlib”.
He occasionally rolls through some sports podcasts I listen to, and there’s no other way to really describe it. He’s not a prog — he’s too bought in on conventional 90’s-00’s liberal narratives — but he’s fundamentally just as toxic as the worst progs. So he’s a liberal, through and through; even if he thinks Kamala Harris is a sellout, on a secret survey he’d probably have more answers in common with her than a genuine tankie.
He’s also a shithead. He “knows” a whole bunch of things to be true, so everything gets filed into those buckets, and he’s the noble one exposing it all for our benefit, if only we’d all rise up and do the hard work of revolting against his enemies for him. His aggressive posture is a chickenshit substitute for having a meaningful critique of those enemies.
Crucially, he’s not as smart nor profound as he thinks he is. He doesn’t understand the systems he criticizes, but also doesn’t correct for that epistemic shortcoming by learning or asking questions; whatever “research” he does is little better nor more rigorous than an antivaxer who’s read all the papers and can cite chapter and verse of their results, but couldn’t actually teach you a fucking lick of microbiology, organic chemistry, or advanced stats.
It's ad hominem, but I think the vast majority of the hard skeptics on this are just ideologically against AI and just want to see it fail. This is of course even more true of the radically-hard skeptics.
Sometimes I wonder how can someone still say that AI is useless or broadly useless, given the usage data that we know of.
Just to make an example, I live in a 50k city in Europe, so it's not like the epicentre of the Silicon Valley, and basically every person I know whose job involves processing information uses it. And this is not from December 2025, but it really got traction with a ton of people since mid-late 2024. Everybody uses it, every programmer (of course), I think in this case it's a 100% adoption rate, every guy who needs to process excel sheets, pdfs, ready through bureaucracy and legal stuff. Just everyone.
So again, I don't know in what kind of world they live in.
Personally, I started paying the subscription in January 2023 and never stopped, of course it's immensely more useful and powerful today than it was when I started the subscription, but it always delivered value. It's not like it was "dubiously useful" back then and now it's clear. Imho it was always useful, just now it's much more.
Of course I'm not touching the AGI/ASI thing that for me is mostly bs, at least as presented now. The point is that it doesn't need to be AGI to be economically valuable.
* This sector is a bubble, these days companies are bubbles.
* The technology hype is wildly overblown.
1) The economic/Wall Street argument. Bubble. Will these companies ever be profitable? How many of them? Are these pets.com or amazon.com?
You yourself expressed skepticism about investing in them, so you seem to agree that there's a lot of froth in this sector. You agree with him.
2) The hype? The promises? Are they being met? Well, it depends on which promises you look at, right? Is he wrong that there *no* there there? Yeah, he's wrong. They are good for coding. They are useful for some things. Generative AI is not useless. (e.g., I use chatGPT weekly to develop cartoons to illustrate my blog entries.) I just asked three LLMs to give me the inverse of a matrix, and one of them even got it right! AI's biggest critic's criticisms go too far? Is that your point? Or do you think that the AI hype and promises are actually being met? How important is coding other the economy?
AI's biggest critic goes too far in hs criticisms. That's your point, right?
"But we desperately need better skepticism." OK. Do we have criticism, or is your effort to argue through a single anecdote just lazy painting with a broad brush? Why don't your ask your favorite LLM to tell you about the other critics and other criticism. Maybe better skepticism is being well expressed out there, but your guy isn't *the* guy.
***************
Can you get off this argument-by-anecdote approach? Please.
Ms Piper is not claiming that skepticism is unwarranted. This is an article about Ed Zitron's brand of motivated skepticism specifically. Mr Zitron worth writing about because he's high profile and thus widely read, and his bad faith skepticism can have the effect of crowding out well thought out skepticism.
Ed Zitron gives the likes of Sam Altman an opportunity to portray *all* their critics as being out of touch with reality. Which is correct in the narrow case of Zitron, but wrong for more thoughtful criticisms, several of which are explicitly expressed by Piper in this article.
This is sort of like people concerned with data center environmental impact focusing their attention on water use instead of energy use. Data center water use is large in absolute terms, but negligible in relative terms. Data center energy use is enormous by all measures and will have a much larger environmental impact than water consumption.
Companies will use the focus on water consumption to dismiss all environmental concerns, even the energy concerns are legitimate.
1) There will always be crazy critics. Muzzling one crazy critic doesn't prevent that.
There's a term for this, "nut picking." There are always nutty critics that one can point to to paint the whole lot as crazy.
So, no, I don't believe that Zitron is is some linchpin that uniquely prevents AI critics from being taken seriously.
2) Ms. Piper wrote, ""But we desperately need better skepticism." Again, let me ask, does better criticism and better skepticism exist? If so, why is she so focused on on particular nut and why is she implying that he is the best skeptic out there?
"Nutpicking" is when you take a random post on Twitter by someone with 89 followers and pretend like that person represents your opposition. Ed has more than 100k followers on X, and (as Kelsey notes) supposedly more than 80k subscribers to his newsletter. He is at the absolute forefront of AI skepticism - he and Gary Marcus are the two names I hear referenced *constantly* in these conversations. And the arguments he makes are absolutely being made by mainstream AI skeptics (as well as the nutty anti-AI fringe).
Like, come on. If the New York Times does shitty coverage, and people complain about it, is it reasonable to respond "well they're just one paper. Lots of other papers exist. Why wouldn't you talk about other, better papers instead of focusing on this supposedly bad one?" Because it's the New York Times! If there *is* a NYT of AI criticism, Ed is it! It is worth responding to bad arguments made on vary large, popular platforms!
1) You want to respond to one guy, respond to one guy. Don't pretend you are doing more than responding to one guy.
2) "if there *is* an NYT of AI criticism..."? Let's just stop there. No, there is not an NYTimes of AI criticism. But if there is, this is not the guy. You are claiming he is the most popular and perhaps most widely cited. But you are not claiming that he is the most established and respected. You are saying...pardon me for the dated reference...that he is the USA Today of AI criticism. Popularity, not quality.
1) Well, point #1 stands. She should just admit that she is just trying to discredit one guy. She shouldn't use that to suggest that there are not better critics.
2) Yeah, I think so. That shifts him from being the most credible guy to being the most popular guy. That removes the suggestion that others are even worse. And then the question becomes, "Why is this idiot the most popular guy? He's clearly a fool." And *then* we need to think about why this fool gets so much attention, despite his idiocy. It shifts the focus from the low quality of his arguments (i.e., which really matters if he's the best out there) to their attractiveness (i.e., which really matters if his popularity is the question).
3) No, I clearly don't agree with that you the nuts have to unpopular for it to count as nut picking. When we have Glen Beck or RFK Jr available, we can nut pick with prominent people. The nut picking strategy allows us to look for obscure nuts, we can can pick popular ones if they are nutty enough. Nut picking is about picking some real crazy, without being limited to popular or prominent nuts. But they ain't off limits.
Is she implying that he is the best skeptic out there? If not, why write, "But we desperately need better skepticism."
If she wants to take down best known nut, fine. She should own that. But she should not suggest that his criticism tells us about the nature of other criticism.
It comes down to this: What is her purpose of this piece? Is it the narrow purpose of taking down this particular nut? Or the broader point of demonstrating that we need better skepticism and criticism?
I think that Ms. Piper is at least two steps down the road of a methodological claim that one can discredit a whole field by picking on one example. She is nut picking, herself. Rather than owning that she is making a narrow argument, she tries to claim that her takedown of one nut proves the whole field is lacking. And doing so with a single anecdote—ironically after bemoaning sample sizes issues in unspecified studies.
Look, I think that Ms. Piper is factually challenges and argumentatively weak. I think she is at least two step down the path or proving that definitively. But I don't think that that proves that substack needs better authors or that The Argument's contributors are all softheaded. I am making a narrow argument, and I know that I am making a narrow argument. She clearly is claiming broader arguments, but only arguing narrower ones.
I mean, I guess that's one way to read it. My read on her argument was something that blends the two things - basically that we need better skepticism with a broad audience. I don't think anywhere in the piece did she suggest that the entire field of AI skepticism is lacking, exactly the opposite, really, just that the person with the biggest microphone is one of the worst examples.
I count this sentence as really important. "But we desperately need better skepticism."
In my view, that strongly suggests that this is as good as it gets. That suggests that this skepticism prevents the existence of better skepticism—an attack she also made on the much broader field of educational research, just last week.
That sentence does *not* say, "We need more attention paid to better skepticism." It's not about attention; it's about existence.
I listened to an episode of his podcast, Better Offline, last week because Cal Newport was a guest. Listening to this guy talk for a few minutes will tell you all you need to know. Every other sentence is some flippant line about it all just being fancy autocomplete. Its a tired set of ideas about the technology, and it really does call into question whether he has used any of these tools in the last two years.
Kinda surprised all this newfound attention on Ed hasn’t causes more folks to stumble upon his, umm, dubious business practices as a “fake it til you make it” PR person….
There are plenty of reasons to be skeptical that we are on the verge of AGI, which is conventional wisdom in Silicon Valley at this point. So it’s deeply frustrating that many of the most prominent skeptics are so extreme and “head in the sand” about current AI capabilities. The discourse doesn’t elevate nuanced positions.
I appreciate Kelsey Piper’s focus in this piece. I think there are a range of questions about AI that get conflated into one big up or down question. Is there an investment bubble in AI? Is AI a “normal” (general purpose and potentially transformative) technology that is subject to the same implementation frictions at bottlenecks as past technologies? Will AI inevitably be more transformative than past transformative technologies (e.g., the printing press, electrification, the green revolution, the internet)? Does AI’s capacity for recursive self-improvement make exponential improvement in AI models inevitable? These are different questions that deserve different answers, even if the answers to some questions have bearing on the answers to others.
I know you're being sarcastic, but when it comes to AI, it may as well be decades ago. If you're using a study on GPT-3.5 to claim that AIs in 2026 are dumb and useless, you're not a serious person and have nothing to say about the topic that's worth considering.
ChatGPT 3.5 was superseded by GPT-4 in March 2023, so I’m not sure why that’s relevant to the question of studies from 2024 or the distant past of (gasp) five months ago.
Because Opus 4.5 came out in December and is a game-changer for coding. Lots and lots of people who were skeptical about coding tools in 2025 changed their mind with Opus 4.5, because the facts changed.
(I was not skeptical in 2025, and while I liked the METR study that everyone cites, I also think that AI skeptics have massively over-interpreted it by focusing on a couple of headline numbers, rather than engaging with what it says and how it got there.)
It’s just absurd to say “he’s using studies from 2024 and 2025” as if that’s somehow invalidating in April 2026. You may as well say “there’s nothing you can offer that would be convincing” and leave it there.
I was explaining why an idea that seems absurd -- that everything is legitimately different now than it was six months ago -- is true. If you were in the software field, you would know this as a commonplace, widely-accepted fact of life, and you would be exhausted by having read a zillion road to Damascus essays in January. I know not everyone is in this field, so was giving you that context.
But I'm not interested in having an argument on this. If, having been led to water, you aren't interested in drinking it, that's up to you. Reality is what it is, and eventually you'll see that yes, the software field is not living in 2025 (never mind 2024, when Claude Code and Cursor agentic mode didn't even exist!), and things are different.
"AI — not future, speculative AI, but AI as it exists today — has nonspeculative economic value. I pay $100 per month for Claude because Claude provides me with comfortably more than $100 per month of economic value."
Golly, how the goal posts have moved! I thought AI was going to cure cancer. I thought AI was going to result in 20% a year GDP growth, at a minimum. I thought AI was going to end death. Where's all that stuff? This is the point that AI hypemen can't respond to: the outsized claims about what AI is going to do have not been met by demonstrations of what AI can do. It's still all speculative. And we're nowhere close to justifying the economic investment involved. Who cares if you're get $100 of value out of Claude, if as a society we're paying far more than we get back?
Freddie you're doing the thing again where you lump a bunch of people with different opinions together and call it goalpost moving, when in fact it's... a bunch of people with different opinions. Kelsey never claimed that AI would cure cancer, in fact she has noted in one of her last articles that AI has not yet really begun to speed up scienctific research.
I don't recall being asked "as a society" to pay anything. Investors are investing as they usually do. If it doesn't pay off, they made a bad investment. They didn't ask society if they should invest their money, but then I don't expect them to.
I don't think this is a great counterargument. When investors as a whole get too excited about a single type of asset, the accurate revaluation of that asset can take down so many investors at once that financial infrastructure takes significant damage like with the 2008 housing bubble. Right now it's all just investment firms playing games that don't matter to any of us, but it might not stay that way.
It's weird that, to defend the statement "AI delivers significant economic value" one is expected to defend the specific marketing pitches of every AI company that has ever put out a goofy interview, and every random zealot who has made wild speculations on Twitter
There were two sets of goalkeepers, one saying “AI will someday create 20% GDP growth” and the other saying “AI will never create any economic value at all”. AI has already grown to earn tens of billions. This is obviously compatible with one of those goals and not the other!
Freddie, what kind of evidence would you accept as probative of claims that AI will one day be capable of doing [x]?
The linear passage of time ensures that 'claims about what AI is going to do [will not be met] by demonstrations of what AI can do', but you probably need to be open to evidence based on progress to date or something, rather than just insisting 'that's still not [x]!'.
By way of example: while they still can't cure cancer, any moderately recent AI could have told you your argument was dumb and suggested not making it. That feels like progress.
> Who cares if you're get $100 of value out of Claude, if as a society we're paying far more than we get back?
Who is paying more than they get back? Is society paying something over and above what the customers and investors are paying?
LM models are widely used and developed in biology and medicine and they are absolutely delivering value, and this is not my niche (I worked on genome models for genetic diseases) but I'm sure they're being widely use in cancer research in this exact moment.
So the fact that you don't know this doesn't mean it isn't happening.
The fact that Amodei (iirc) said the 10-20% GDP thing is not an argument against the usefulness of AI, but I mean, it's obviously a bad faith argument.
Same with the end of death straw man.
The usage data tells use that indeed a ton of people is getting value out of these models otherwise you wouldn't see such numbers, and nobody is paying "far more than what it gets back", because the vast majority of users are not paying anything and in general the societal costs of running these things are pretty minimal compared to basically any other kind of production.
The frontier labs aren't finished yet.
It will probably do all those things eventually. I've read people from labs saying AI is doing stuff in hours that would take months or years without
But it will still take time
It's amazing but not magic
Unfortunately, there’s a very large audience of people looking to confirm their priors that AI is useless and/or fraudulent and/or unethical and/or ruining the youth and/or hallucinates and/or, and/or, and/or ….
I think this is actually the biggest obstacle to Zitron doing better argumentation. Sure, it's possible to imagine an AI critic making the financial criticisms Piper would like for Zitron to be making, but that wouldn't serve Zitron's actual audience.
Zitron's audience is the people you're talking about. They want to read about how AI is fake and how it's going to go away any minute now once the bubble pops, and that's why they read Zitron. A bunch of sober financial analysis based on the premise that it's real and won't go away doesn't meet that market's needs, and they'd unsubscribe. And it's not like Zitron could just pivot to writing for an audience of practical investors looking for real financial analysis, as that's super not his wheelhouse.
So he's pretty much locked in to writing about how AI is fake and on the verge of disappearing; I suspect he rides that into the ground until even those skeptics have given up on the idea, and then finds a new beat.
Yeah, the most charitable explanation for what Kelsey describes in the 2024 -> 2026 evolution of him is straightforward audience capture.
AI is useless, fraudulent, unethical, and ruining the youth. And it hallucinates in very seriously problematic ways.
There is very easily discoverable evidence of all of these points via scientific studies.
That's not what is at stake in the article. Is the LLM industry financially viable? Probably not because all of it is circular, but I have no idea if it is a bubble that will burst.
It just doesn't feel like a healthy economy/marketplace. And if the last 20 years have taught me anything about Silicon Valley, its that none of these guys are trustworthy or good people and we should stop listening to them.
My company is working through an enterprise agreement with Anthropic and I (head of tax) am one of a group of people testing out Claude enterprise. It’s almost magical. Claude Code built out the full tax return workpapers for one of our jurisdictions. Claude Cowork is solving major issues for us on the tax provision. Their excel plugin (sans Cowork/Code) has been rolled out across finance and is a game changer in and of itself. I’ve been using Claude Pro for personal projects for a while now, and it’s amazing to see how much better these models are getting in real time.
I am not an AI skeptic and use it for work and pleasure every day. That said, the bullet point list of what you pay for AI to do makes little sense to me. Finding out about a soccer club or missing books could be done much more easily by “asking someone” or “looking at the books.” I can’t figure out what kind of analysis you were even doing there.
I can't speak for Kelsey, but unless I have a specific website in mind whose name temporarily escapes me I don't bother with Google anymore. Claude is just better and faster than manual search. Upgraded Google is not worth $100/month, of course.
But if you’re looking at your homeschool library you can see what books are missing. If you’re watching a soccer team play you can ask someone “hey what team is this?”
I can see how some of these could be slightly faster. For the example of book cataloging, it could be a significant time saver if you're trying to individually check each book on your list and then check where to buy it if you need a copy and you want to do this for a list of 100+ books. Still, it's notable that all of these are low-skill tasks that could be done by any literate human considering that much of the hype in favor and against AI is over the prospect of replacing/enhancing expensive white-collar professionals.
The thing is that most households don’t have spare low skill literate labor on hand to do things while the adults focus on tasks like cooking and playing with the children. Spending a few cents having Claude look at a photo and tell you what’s missing is better than spending 20 minutes writing down a list of what you have and googling what’s missing.
Yeah, digital personal secretaries for everyone is a real value add. Still, it's notable that this isn't a big point in any faction's argument over what LLMs are or aren't worth. Everyone's speculating about if gen AI is as good as the average artist and if that's a good thing, if Claude or whatever is the equivalent of an omni-PhD in your pocket or if it ever could be. Meanwhile, the average person is using it to do meal planning faster.
Kind of reminiscent of how hype about the internet was grounded as its reach expanded. What feats will the typical person accomplish with zero-latency access to the complete breadth of human knowledge? Turns out that most of the time, most of us just want to watch TV and gossip about each other and increases in our abilities doesn't change our basic drives that much.
A lot of the little everyday uses for LLMs seem trivial and unimportant, because they are. But like, look at all the Google searches you do, and 95% of those are also trivial and unimportant (and could often be done by using other tools).
I think Zitron's most compelling argument, by far, is that OpenAI and Anthropic are currently selling their products at such ridiculous, artificially low prices--spending $5 billion to sell $2 billion worth of services, that sort of thing--that it's impossible to infer anything positive about their business model from the mere fact of their revenue, or revenue growth. This article's only response seems to be a hand-wavey "everyone's got unprofitable users." Sure, but that ducks the question, which is: are either of these companies likely to assembe enough profitable users that they will one day be able to survive without constant infusions of billions in VC cash? Zitron may not be right about everything, or even most things, but that seems like a very important question no other journalists are asking.
I agree with most of your article. I've read Zitron sporadically over the past couple of years, and agree that his dismissal of AI as useful technology was always off-base, and is basically delusional at this point.
Where I think he's on firmer ground is still the economic arguments. While I grant that a lot of usecases can be sold at sustainable rates for the value they provide, when it comes to coding it's clear that the most powerful (ie, agentic) workflows have been heavily underpriced. It's absolutely insane what I can do with my $20/mo Claude Pro plan, especially when I compare to what it'd cost me if I was paying API rates (which I can preview by enabling "extra usage"). For a while power users using the plans like this may have been in the minority, but it's clear that as agentic coding has achieved a step function in capability and awareness of this has grown throughout the technical and non-technical ranks, so has the demand for tokens. Yes, tokens get cheaper over time, but the biggest gains in capability have come from massive increases in the number of tokens used to complete tasks.
We're already seeing Anthropic toy with restricting access to Claude Code, and Microsoft Copilot just killed off subscription-based coding altogether and raised the cost of all the good models substantially. Part of this is simply a capacity issue, but it's a capacity issue because they've been basically giving the product away in these subscription plans. Having heavy demand for something you're offering at well below cost is not some great achievement.
The bubble argument that I find convincing isn't that AI isn't capable or won't continue to be a compelling product or won't continue to disrupt many knowledge industries, but that building and providing models remains a very tough business with awful margins. The fact that coders will flip from provider to provider (Claude was the hotness a couple months ago, now everyone's running to Codex because Anthropic is throttling harder, etc.) shows that any competetitive edge is fleeting and they're competing mostly on who can give the most tokens and context away for a given price. They have to expend insane amounts of capital to stay competitive, and then it's a race to the bottom on pricing.
The capabilities are incredible. I regularly max out my Pro plan writing personal software and as a professional software engineer, it's an incredible experience. I also am far less productive with this in my day job because the impediments to moving fast in a professional software environment are not primarily the cost of writing code. It is absolutely still worthwhile for some things, even at "more realistic" API rates, but a lot of these heavily AI-dependent workflows feel wasteful and like they're just optimizing one small part of the overall process with dubious ROI. With time I expect it will be revolutionary, but right now the net gains still feel incremental in an enterprise development environment.
I think as the models continue to get better and people become dependent on them eventually the prices will rise to cover the costs.
But that doesn't mean AI companies will make a ton of money. I don't see what the moat will be
I think the persistent skepticism of AI’s value is downstream of skepticism of the value of the “knowledge work” economy overall.
If you believe that most white collar employment is a kind of theater (the David Graeber “Bullshit Jobs” thesis), then automating white collar work will feel valuable to white collar workers and yet also not create much in terms of economic benefits to people at large.
I’m not as skeptical as Zitron, but it does seem like in seeing how this progresses we may learn if all those emails and PowerPoints were ever creating economic value after all. I expect some were and some weren’t.
There are two parts to the bullshit jobs thesis:
1) "an increasing number of people feel like their jobs are bullshit." That seems plausible, but is very complicated.
2) "a large share of jobs are actually bullshit." This is just juvenile ODD, a form of Gell-Mann amnesia that flatters academics and creative types.
Graeber really should have gone for the modern Marxist take - just as a 19th century factory worker could be alienated from the end product of their work and therefore feel that their job was bullshit, now we’ve got big information management enterprises that have commoditized white collar labor in such a way that the workers are alienated from the end product and therefore feel that their job is bullshit.
I honestly don't get this complaint even from Marx's time. I know I'm alienated from the end product of my labor, that's how I can contribute to projects far out of the scope of my understanding and provide value to people I don't know. What else am I supposed to do, only provide goods and services directly to people I know? Do communists think I have neighbors who can provide all basic material goods and would be eager to do so in exchange for custom software applications?
I think there are some anarcho-syndicalists who do think one should only exchange goods and services with people you know. I think more orthodox marxists just think that this sort of alienation exists on a continuum and that inhuman capitalist systems tend to produce worse versions but that with better structuring you could have better connection to the fruits of your labor. That’s probably right to some degree, but I don’t know what degree that is, and I suspect the gains from trade make up for at least some of the soul crushing of alienation.
As someone who generally likes Graeber's other work, he has a weird hate for delivering pizza in a way that makes me assume he just didn't like pizza very much. This, amazingly, calls him out as committing your Thesis #2 over the course of hundreds of pages.
There's also the closely related thesis #3: many jobs do things that other people value but I think they shouldn't value, therefore the work is bullshit.
Fwiw the Internet bubble of the late 90s was driven substantially by the phenomenon of weak/poorly conceived and/or poorly run businesses attaching ". com" to their name and receiving a windfall of investment. I was CFO of a small company in that time we were seeking capital to expand and both our investment banker and the investors we talked to literally said "we'll give you a 5 times bigger multiple if you can make the case that this is actually an online business" [which it wasn't].
This is the same logic of "no doc loans" as shown in The Big Short.
In both cases, the wasted money is going to the "wrong" growing business, not to a whole business category that doesn't make sense.
There's other ways for bubbles to form, but that's not what's happening with AI, so I think the.com analogy is poor. People aren't throwing investment at catchy startups because they think they'll dominate a new category by "AI-ing" it.
Interestingly, I think part of the reason that the investments aren't being driven by "AI-for-X" startups -- though those exist -- is the belief by investors that the labs could just sherlock whatever product those startups might build before even getting to a place where an acquisition would make sense.
I’m far from a right wing troll, but I can think of no better category for Ed Zitron than “shitlib”.
He occasionally rolls through some sports podcasts I listen to, and there’s no other way to really describe it. He’s not a prog — he’s too bought in on conventional 90’s-00’s liberal narratives — but he’s fundamentally just as toxic as the worst progs. So he’s a liberal, through and through; even if he thinks Kamala Harris is a sellout, on a secret survey he’d probably have more answers in common with her than a genuine tankie.
He’s also a shithead. He “knows” a whole bunch of things to be true, so everything gets filed into those buckets, and he’s the noble one exposing it all for our benefit, if only we’d all rise up and do the hard work of revolting against his enemies for him. His aggressive posture is a chickenshit substitute for having a meaningful critique of those enemies.
Crucially, he’s not as smart nor profound as he thinks he is. He doesn’t understand the systems he criticizes, but also doesn’t correct for that epistemic shortcoming by learning or asking questions; whatever “research” he does is little better nor more rigorous than an antivaxer who’s read all the papers and can cite chapter and verse of their results, but couldn’t actually teach you a fucking lick of microbiology, organic chemistry, or advanced stats.
It's ad hominem, but I think the vast majority of the hard skeptics on this are just ideologically against AI and just want to see it fail. This is of course even more true of the radically-hard skeptics.
Sometimes I wonder how can someone still say that AI is useless or broadly useless, given the usage data that we know of.
Just to make an example, I live in a 50k city in Europe, so it's not like the epicentre of the Silicon Valley, and basically every person I know whose job involves processing information uses it. And this is not from December 2025, but it really got traction with a ton of people since mid-late 2024. Everybody uses it, every programmer (of course), I think in this case it's a 100% adoption rate, every guy who needs to process excel sheets, pdfs, ready through bureaucracy and legal stuff. Just everyone.
So again, I don't know in what kind of world they live in.
Personally, I started paying the subscription in January 2023 and never stopped, of course it's immensely more useful and powerful today than it was when I started the subscription, but it always delivered value. It's not like it was "dubiously useful" back then and now it's clear. Imho it was always useful, just now it's much more.
Of course I'm not touching the AGI/ASI thing that for me is mostly bs, at least as presented now. The point is that it doesn't need to be AGI to be economically valuable.
So, you think he is making two arguments, right?
* This sector is a bubble, these days companies are bubbles.
* The technology hype is wildly overblown.
1) The economic/Wall Street argument. Bubble. Will these companies ever be profitable? How many of them? Are these pets.com or amazon.com?
You yourself expressed skepticism about investing in them, so you seem to agree that there's a lot of froth in this sector. You agree with him.
2) The hype? The promises? Are they being met? Well, it depends on which promises you look at, right? Is he wrong that there *no* there there? Yeah, he's wrong. They are good for coding. They are useful for some things. Generative AI is not useless. (e.g., I use chatGPT weekly to develop cartoons to illustrate my blog entries.) I just asked three LLMs to give me the inverse of a matrix, and one of them even got it right! AI's biggest critic's criticisms go too far? Is that your point? Or do you think that the AI hype and promises are actually being met? How important is coding other the economy?
AI's biggest critic goes too far in hs criticisms. That's your point, right?
"But we desperately need better skepticism." OK. Do we have criticism, or is your effort to argue through a single anecdote just lazy painting with a broad brush? Why don't your ask your favorite LLM to tell you about the other critics and other criticism. Maybe better skepticism is being well expressed out there, but your guy isn't *the* guy.
***************
Can you get off this argument-by-anecdote approach? Please.
Ms Piper is not claiming that skepticism is unwarranted. This is an article about Ed Zitron's brand of motivated skepticism specifically. Mr Zitron worth writing about because he's high profile and thus widely read, and his bad faith skepticism can have the effect of crowding out well thought out skepticism.
Ed Zitron gives the likes of Sam Altman an opportunity to portray *all* their critics as being out of touch with reality. Which is correct in the narrow case of Zitron, but wrong for more thoughtful criticisms, several of which are explicitly expressed by Piper in this article.
This is sort of like people concerned with data center environmental impact focusing their attention on water use instead of energy use. Data center water use is large in absolute terms, but negligible in relative terms. Data center energy use is enormous by all measures and will have a much larger environmental impact than water consumption.
Companies will use the focus on water consumption to dismiss all environmental concerns, even the energy concerns are legitimate.
1) There will always be crazy critics. Muzzling one crazy critic doesn't prevent that.
There's a term for this, "nut picking." There are always nutty critics that one can point to to paint the whole lot as crazy.
So, no, I don't believe that Zitron is is some linchpin that uniquely prevents AI critics from being taken seriously.
2) Ms. Piper wrote, ""But we desperately need better skepticism." Again, let me ask, does better criticism and better skepticism exist? If so, why is she so focused on on particular nut and why is she implying that he is the best skeptic out there?
"Nutpicking" is when you take a random post on Twitter by someone with 89 followers and pretend like that person represents your opposition. Ed has more than 100k followers on X, and (as Kelsey notes) supposedly more than 80k subscribers to his newsletter. He is at the absolute forefront of AI skepticism - he and Gary Marcus are the two names I hear referenced *constantly* in these conversations. And the arguments he makes are absolutely being made by mainstream AI skeptics (as well as the nutty anti-AI fringe).
Like, come on. If the New York Times does shitty coverage, and people complain about it, is it reasonable to respond "well they're just one paper. Lots of other papers exist. Why wouldn't you talk about other, better papers instead of focusing on this supposedly bad one?" Because it's the New York Times! If there *is* a NYT of AI criticism, Ed is it! It is worth responding to bad arguments made on vary large, popular platforms!
1) You want to respond to one guy, respond to one guy. Don't pretend you are doing more than responding to one guy.
2) "if there *is* an NYT of AI criticism..."? Let's just stop there. No, there is not an NYTimes of AI criticism. But if there is, this is not the guy. You are claiming he is the most popular and perhaps most widely cited. But you are not claiming that he is the most established and respected. You are saying...pardon me for the dated reference...that he is the USA Today of AI criticism. Popularity, not quality.
Does literally anything change about your reply if I sub in "The USA Today of AI criticism" for "The NYT of AI criticism"
1) Well, point #1 stands. She should just admit that she is just trying to discredit one guy. She shouldn't use that to suggest that there are not better critics.
2) Yeah, I think so. That shifts him from being the most credible guy to being the most popular guy. That removes the suggestion that others are even worse. And then the question becomes, "Why is this idiot the most popular guy? He's clearly a fool." And *then* we need to think about why this fool gets so much attention, despite his idiocy. It shifts the focus from the low quality of his arguments (i.e., which really matters if he's the best out there) to their attractiveness (i.e., which really matters if his popularity is the question).
3) No, I clearly don't agree with that you the nuts have to unpopular for it to count as nut picking. When we have Glen Beck or RFK Jr available, we can nut pick with prominent people. The nut picking strategy allows us to look for obscure nuts, we can can pick popular ones if they are nutty enough. Nut picking is about picking some real crazy, without being limited to popular or prominent nuts. But they ain't off limits.
Because this particular nut has a huge platform and is widely consumed among certain folks
Is she implying that he is the best skeptic out there? If not, why write, "But we desperately need better skepticism."
If she wants to take down best known nut, fine. She should own that. But she should not suggest that his criticism tells us about the nature of other criticism.
It comes down to this: What is her purpose of this piece? Is it the narrow purpose of taking down this particular nut? Or the broader point of demonstrating that we need better skepticism and criticism?
I think that Ms. Piper is at least two steps down the road of a methodological claim that one can discredit a whole field by picking on one example. She is nut picking, herself. Rather than owning that she is making a narrow argument, she tries to claim that her takedown of one nut proves the whole field is lacking. And doing so with a single anecdote—ironically after bemoaning sample sizes issues in unspecified studies.
Look, I think that Ms. Piper is factually challenges and argumentatively weak. I think she is at least two step down the path or proving that definitively. But I don't think that that proves that substack needs better authors or that The Argument's contributors are all softheaded. I am making a narrow argument, and I know that I am making a narrow argument. She clearly is claiming broader arguments, but only arguing narrower ones.
I mean, I guess that's one way to read it. My read on her argument was something that blends the two things - basically that we need better skepticism with a broad audience. I don't think anywhere in the piece did she suggest that the entire field of AI skepticism is lacking, exactly the opposite, really, just that the person with the biggest microphone is one of the worst examples.
I count this sentence as really important. "But we desperately need better skepticism."
In my view, that strongly suggests that this is as good as it gets. That suggests that this skepticism prevents the existence of better skepticism—an attack she also made on the much broader field of educational research, just last week.
That sentence does *not* say, "We need more attention paid to better skepticism." It's not about attention; it's about existence.
That's my read.
I listened to an episode of his podcast, Better Offline, last week because Cal Newport was a guest. Listening to this guy talk for a few minutes will tell you all you need to know. Every other sentence is some flippant line about it all just being fancy autocomplete. Its a tired set of ideas about the technology, and it really does call into question whether he has used any of these tools in the last two years.
Kinda surprised all this newfound attention on Ed hasn’t causes more folks to stumble upon his, umm, dubious business practices as a “fake it til you make it” PR person….
https://archive.nytimes.com/publiceditor.blogs.nytimes.com/2013/07/03/no-quid-pro-quo-by-youre-the-boss-writer-but-intrigue-anyway/
There are plenty of reasons to be skeptical that we are on the verge of AGI, which is conventional wisdom in Silicon Valley at this point. So it’s deeply frustrating that many of the most prominent skeptics are so extreme and “head in the sand” about current AI capabilities. The discourse doesn’t elevate nuanced positions.
I appreciate Kelsey Piper’s focus in this piece. I think there are a range of questions about AI that get conflated into one big up or down question. Is there an investment bubble in AI? Is AI a “normal” (general purpose and potentially transformative) technology that is subject to the same implementation frictions at bottlenecks as past technologies? Will AI inevitably be more transformative than past transformative technologies (e.g., the printing press, electrification, the green revolution, the internet)? Does AI’s capacity for recursive self-improvement make exponential improvement in AI models inevitable? These are different questions that deserve different answers, even if the answers to some questions have bearing on the answers to others.
Studies from 2024 and 2025?! That’s like literally decades ago!
That would never pass muster in any other field. Is he some kind of moron?
I know you're being sarcastic, but when it comes to AI, it may as well be decades ago. If you're using a study on GPT-3.5 to claim that AIs in 2026 are dumb and useless, you're not a serious person and have nothing to say about the topic that's worth considering.
ChatGPT 3.5 was superseded by GPT-4 in March 2023, so I’m not sure why that’s relevant to the question of studies from 2024 or the distant past of (gasp) five months ago.
Because Opus 4.5 came out in December and is a game-changer for coding. Lots and lots of people who were skeptical about coding tools in 2025 changed their mind with Opus 4.5, because the facts changed.
(I was not skeptical in 2025, and while I liked the METR study that everyone cites, I also think that AI skeptics have massively over-interpreted it by focusing on a couple of headline numbers, rather than engaging with what it says and how it got there.)
Lots and lots!
It’s just absurd to say “he’s using studies from 2024 and 2025” as if that’s somehow invalidating in April 2026. You may as well say “there’s nothing you can offer that would be convincing” and leave it there.
I was explaining why an idea that seems absurd -- that everything is legitimately different now than it was six months ago -- is true. If you were in the software field, you would know this as a commonplace, widely-accepted fact of life, and you would be exhausted by having read a zillion road to Damascus essays in January. I know not everyone is in this field, so was giving you that context.
But I'm not interested in having an argument on this. If, having been led to water, you aren't interested in drinking it, that's up to you. Reality is what it is, and eventually you'll see that yes, the software field is not living in 2025 (never mind 2024, when Claude Code and Cursor agentic mode didn't even exist!), and things are different.
I work in software.
There’s lots of things that could be meaningful, but studies from five models ago aren’t!
Then there’s nothing meaningful. By nature, research you can cite was done in the past.