That's exactly what the first (titled) section does?
Which of the hyperlinks provided at the beginning sounded like what you wanted, and after you clicked it how did it disappoint you?
The information you are describing is stuff I would not expect anybody to repeatedly duplicate across periodic blog-posts.
I scrolled down a bit to read. A popup took up my screen, asking me to subscribe, having read essentially nothing at this point.
I just left. Life is too short.
(And he's not Gen Z anyway is he; he's among the older millennials. He's appropriating it for muckraking purposes.)
If you are a consumer and you have a Mac or an iPhone, what do you need from AI that Apple's new offering won't provide? Why would you pay for ChatGPT, or even tolerate its inevitably increasingly desperate ad placements?
Assume Google will have similar tools in their phones, and Google search will continue to have the offering it does.
In short, where is the evidence that once Apple's tech exists, consumer AI is worth, to Anthropic or OpenAI, anything noticeably more than that $1B a year?
Maybe OpenAI strikes a deal to put something in Samsung phones. Let's say Samsung is ten times as desperate as Apple (which is how it looks, often). Still only $10B a year?
2026 consumer revenue projections from OpenAI are pitched at $14-15 billion, apparently. If they get that, it's the only year they will get that, because by late this year, everyone with an iPhone will have something useful built in.
Ed Zitron is a mouthy British rabble-rouser, but I think he is probably mostly on the money.
That it is so absurdly ambitious and so likely to run up against reality strikes me as really indicative of the quality of the envelopes these calculations are being sketched on.
> AI Cannot Afford To Slow Down — It Needs $3 Trillion Or More In Revenue By End Of 2030 To Sustain Its Existence
Is this true? With the total 2024 wages being 11.7 trillion USD [0], and nonfarm payrolls totaling 158,000 in the same year [1], it's an order of magnitude higher than my back of the napkin guesses I've made that AI needs to take or create 1/20 jobs minimum to break even.
[0] https://fred.stlouisfed.org/series/BA06RC1A027NBEA [1] https://fred.stlouisfed.org/series/PAYEMS
We sure do need a reset.
At this point I'm trying to believe there's a middle ground where the level of individual capability this unlocks, leads to major discoveries.
Take any stock index, remove AI stocks, what do you see? That's right! Nothing...
So where is all the productivity going? Where is the value? Where are the massive unemployment stats or the millions of new startups making big $$$?
That being said, AI seems kind of miraculous sometimes.
Similar to cars. So enticing that we make everything else in the world worse in order to maximize the profit, make it indispensable, subsidize it, and make the dependency on it irreversible.
And it's not even something to blame individual people for.
Driving away from all the other cars to spend a weekend feels like freedom.
Using AI to answer a question feels like a "bicycle for the mind".
But in fact it's more like a car. It requires massive resources and creates perverse incentives, and the result is ineffective and corrupt.
Both cars and AI are amazing technology and extremely useful, but using them is not an individual responsibility. It requires societal subsidy.
But with AI what is the exact price? My understanding is that R&D is extremely expensive, but running non-SOTA models is not that bad. We are getting pretty close to models which can be useful locally in many applications.
Or do you mean that at scale running them locally is not possible and hence the infrastructure price is in data centers, which will be expensive to maintain and scale for demand?
Because it was not so clear, and maybe I just wanted to highlight some observations without delivering a real argument for or against things.
The utility/leverage aspect for AI seems more esoteric than the one for cars because, apart from Chatbots, it's more hidden.
And also, similar to cars (or many other phenomena of industrialization), yes, my first vague point was the subsidization of infrastructure. But also, the power gap: that's something not only associated with AI or cars, but with a lot of technologies we all hold dear: sewage, powerline, logistics, etc etc.
What reminds me of cars in the current AI frenzy is the fixation on cementing infrastructure. And also, I think, a lot more people agree on, for example, some kind of universal right to, for example, clean water.
But all of industrialization confronts people with questions of efficiency, inequality, and collective support.
Most people would, for example, support a right to get a minimum amount of clean water when you are living and working in a tradionally inhabited space (if you're on the social-darwinist side) or at least not harming society (if you're more of a social democrat).
And, similar to the buildup of car infrastructure, and the procurement of resources, space etc for maximum building, giant data centers can obstruct people in buying drinking water. Or walking outside (AI obstructs traditional methods of online collaboration).
Neither did I want to say that a car is always more wasteful than some alternative.
But defaulting to the behemoth is inefficient, unless everyone is driven to do it: then it's in some way reasonable.
By adding "corrupt" and "dependent", as well as the economic terms, I wanted to offer a broader critique and create an analogy, not just talk about energy usage on its own.
What I had in mind was: it's easier to go many places that are a mile or less from me, by car. Because everything is obstructed by cars. And I'm atrophied by lack of movement. Best would be to drive somewhere to move/walk.
People already do that in masses.
And doing shopping by car, because everything else seems unbearable, also takes away your time, apart from wasting energy compared to more, smaller shops that would be reachable by foot, bycicle etc.
I guess you know the argument.
Today, people's thinking atrophies because their LLM is probably right in their summarization of some Wikipedia article, plus 2-3 other random sources.
Or so.
Using the Wikipedia search function is not expensive.
But, I mostly had a bigger picture in mind than what is the cost of inference.
I am concerned about the environmental impacts that AI poses, but they don't seem to me to be so catastrophic. Solar and battery tech has made enormous leaps in the past couple decades, and we will need to pivot to clean energy future irrespective of AI.
*This said, I have become gradually more alarmed over the past decade at the lack of epistemological rigor in the general public, as made apparent through the rise of social media. I don't know that AI becoming a truth-seeking crutch for people wouldn't be more good than bad.
Hand wringing about AI datacenter's environmental impact is well and good. We should keep the data centers accountable for their consumption and waste.
I just wish the same people had been upset the last 20 years with poor water resource management in a lot of areas (the west US especially) with urban, ranching and farming development.
> That's true, and I am not anti-AI.
Me neither!
80% of generative AI queries wouldn't even exist as google searches.
One claim of the parent comment was that AI is ineffective. For the purpose of finding answers to questions, it is more resource-efficient than the alternatives, and, to your point, capable of answering questions that were impossible to answer via other means before. In what way is that ineffective?
I awe at the capabilites of generative AI.
I also enjoy sitting in or driving a car.
I did not want to make a moral argument, unless you consider each and every form of utilitarianism as moralism.
We got addicted to the convenience and overuse, and have started a mass extinction event because of it.
The perverse incentives will come for us all.
Where did all the stock gains go before AI?
FAANG / MAG-7.
Was everything from 2012-2020 fake, too?
The question is, is AI leading to massive productivity gains in companies that implement it? AI productivity gains take time to diffuse, but so far companies in the S&P 500 are seeing very high growth. YOY earnings growth rate for the S&P 500 is 21.7% https://advantage.factset.com/hubfs/Website/Resources%20Sect...
Now remove the companies selling the AI shovels: https://pbs.twimg.com/media/HIAjbZxacAARHwD.png
> Not sure what your point is.
My point is that they're selling us Skynet and the end of employment as we now it, things that we shouldn't even have to measure to perceive the results of, yet no one is able to measure any of it
Pointing a finger at nvidia, google, and the other few companies stuck in circular investment schemes that shouldn't even be legal and saying "OOGA BOOGA line go UP, UP GOOD!" doesn't count in my book
I do value having some naysayers in the mix generally, because we do need balanced critique in what is otherwise a very frothy hype cycle. I just don't think he's making sound arguments, and that's even assuming you even agree with his premises in the first place.
My biggest gripe with his napkin math is that he treats inference gross margins as something novel that you can't compare to normal SaaS margins. He's right in part: the constant carousel of R&D costs from model training, related infrastructure buildout, and other adjacent costs required to stay competitive do change the analysis a bit.
But he takes this way too far when he says this is structurally different from normal SaaS margins. The business model definitely doesn't look like Dropbox, but it absolutely looks a lot like AWS, especially early AWS, CDNs, telecom, etc. I can speak to the telecom bit personally, since it's been over half of my professional career as an engineer and, in this specific case, also as a founder. You can have a brutally capital-intensive infra business where profitability depends on utilization, oversubscription, peak-capacity planning, segmentation, and recovering capex over time.
The math he presents gets even more questionable as we see explicit segmentation happening for cost-saving reasons. Many forward-thinking orgs are waking up to the fact that they don't need to use the best, most expensive model for every task. They can route easier tasks to cheaper models, use caching, batch non-urgent workloads, and reserve frontier models for the subset of work that actually needs frontier intelligence. That directly undermines his claim that providers always need to chase frontier intelligence in order to maintain current demand, utilization, and pricing curves.
But that is not the full argument he is making. If the claim is that the labs will not be able to pay their creditors because inference is structurally incapable of becoming profitable, then he absolutely needs to be right about the technical economics of inference.
One part of that is the balance-sheet argument (which already shows insanely good margins). But it also depends on how inference-time compute actually works: routing, batching, kv cache reuse, model segmentation, different latency tiers, etc. Much of those details he's just been straight up wrong about in his writing, so as a result I have to call into question the rest of his reasoning as well (in part to avoid Gell-Mann amnesia).
Could you share what tells about it? I.e. where he was wrong about it?
I'll cherry pick a couple:
“When these new models ‘reason,’ they break a user’s input and break into component parts, then run inference on each one of those parts.” [1]
This is not at all how test-time compute works. At best, this is a very loose metaphor that he may have used out of convenience. This might sound a bit pedantic to point out, but this is a very basic thing that he's getting wrong (presumably at least, again it could be that he just used a poor metaphor).
A less pedantic example would be his claims related to gpt-5/chatgpt auto-routing. He argued that having a router means OpenAI can no longer cache static prompts, because the user prompt has to come before the hidden instructions [2]. This is just not at all how this works at inference-time. There is no evidence that the standard approach of system>developer>user instruction hierarchy has changed, the public API and caching docs maintain this.
But even more broadly, it suggests he is reasoning about kv/prefix caching at the wrong level of abstraction. It's true that conventional prefix caching does require a stable prefix, so yes, if you literally put variable user content before the static prompt, you would destroy the cacheability of that static prompt.
But that is exactly why inference systems are designed to preserve reusable prefixes where possible (via checkpointing or similar), and why serving systems care so much about prefix caching. This is also a big part of how disaggregated prefill/decode infra works where cache-aware routing is critical. His argument treats a bad prompt layout as if it were a necessary consequence of routing, rather than an avoidable implementation choice.
A router can read the user request, decide which model path to use, and then construct a normal downstream model call with stable static instructions first and user content later. Treating that as impossible implies a fundamental architectural misunderstanding.
[1] https://www.wheresyoured.at/how-to-argue-with-an-ai-booster/
The thing people I think have a hard time seeing is that "I go faster" does not mean "more features get finished".
It's a scale issue, and one scale is better than the other. People only pay for finished features, they do not pay for how much code you emit.
However, most of the engineers I respect have gone from being skeptics a year ago to convinced today. I don’t personally know any true holdouts any more. If there are studies that disprove productivity gains more than six months ago, I’m happy to believe that it was true of the AIs that were available at the time. But I’m going to need something much more recent before I disbelieve my lyin’ eyes where it pertains to the AIs available today.
Here is the report:
https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways
And my commentary:
I currently don't have work access to Claude Code, but most of my teammates do. Watching from the outside, the cycle seems to look like this:
1. Experience some success, which hooks you into relying on AI.
2. The AI keeps failing at some task, but you don't want to stop. Keep trying over and over again.
3. Run out of tokens and take a break.
Now, sometimes 1 doesn't happen. Sometimes 2 doesn't happen. 3 is a certainty though.
Now, if you told me that the productivity gain from 1 is enough to offset the loss from 2 and 3, I could believe you. But I also wouldn't be surprised if it didn't.
EDIT: In fact, parent comment has a link to some numbers.
[EDIT: Most] people don't want to go through the numbers. Ok. But there's a history here. When people don't want to see the numbers, certain kinds of things tend to happen.
Code acceleration is great, but.... something precedes that. Vision and strategy re. expansion of offerings and businesses. Once a firm reaches maturity in what it offers and is only touching the edges - this code acceleration is literally useless when you factor in all of the trade-offs.
This is a good thing - it means fat and slow incumbents are sitting ducks to be out-witted by creative and imaginative founders, which is healthy for a well-functioning economy.
Now the economics of existing frontier models are not sustainable - its looking like a mix of the airline (supersonic vs subsonic) and EV industry with China in the background providing decent offerings at much lower prices.
I admit that if a small team or an individual uses an LLM, it's likely they can create value faster.
I think as soon as you don't own the responsibility for the defects you generate with an LLM, their use starts to destroy value. Regardless of product maturity.
This is what I think the data says.
How can something so undeniable have zero scientific evidence? Are there any large peer reviewed or meta studies confirming your claim?
I think the surest sign of productivity gains is the sheer volume of adoption. If you look beyond headlines, adoption is just incredible. Of course adoption does not necessarily point to productivity gains, but if this was some sort of FOMO or smoke and mirrors you would not see this much retention and this feverish a pace of adoption. You would not see a large segment of the profession using coding agents exclusively. All of these companies track productivity, again with imperfect proxies, yet everything points to a pretty consistent picture. Same with benchmarks, again a lot of crappy benchmarks but a lot of high quality ones too and a very diverse collection of tasks and capabilities they probe.
Adoption meaning productivity supposes there are no other dominant factors for the AI push nor AI retention. It is possible for practices to be picked up or continued in spite of causing productivity DROPS. What studies have suggested are factors that make for productive work environments and what is actually enforced in the workplace are different things.
Adoption implying at least some significant productivity gains doesn’t contradict there being other factors. You’re seeing entire companies reshaped. The argument is this is all for show or CEOs are in some sort of idiot class?
“It is possible for practices to be picked up or continued in spite of causing productivity drops” well of course. I just find that incredibly far away from Occam’s razor.
My point is: we have lots of evidence that’s highly consistent with real productivity gains, and I don’t see many pieces of evidence to the contrary.
LoC: people argue it’s not what’s important
PRs/day: same as LoC
Getting projects done faster: oh but what about the quality.
Solve the technical problems and actually be more productive, the social systems build around the old way of doing things will hole you back.
Finish a PR in 10 minutes doesn’t matter if you’re waiting days for a human review.
The fact he’s never reflected on the glaring failures in his analysis tells what we need to know about his intellectual integrity. There’s truth in some of his words about financial risk, but if you can’t acknowledge that there’s upside too, you can’t evaluate risk properly either.
I find it difficult to take him seriously.
Do you think it's not slowing? Do I miss anything really important?
My understanding is that we have now is incremental improvement on thinking models which appeared more than a year ago. Of course, a breakthrough might happen, but I don't see one yet.
I think it's dangerous to rely on claims made by people who financially profit from you believing them without checking.
[0]: https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-v...
It found hundreds of vulnerabilities in Firefox, according to Mozilla: how does Mozilla benefit? It found a 27 year old vulnerability in OpenBSD. How do they benefit from that? Is that made up? Are the maintainers of those codebases lying for the benefit of Anthropic’s IPO? Is copy fail a fabrication by big AI? The 12 OpenSSL vulnerabilities found in January?
https://venturebeat.com/security/mythos-detection-ceiling-se... https://www.wired.com/story/mozilla-used-anthropics-mythos-t... https://cyberscoop.com/copy-fail-linux-vulnerability-artific... https://www.schneier.com/blog/archives/2026/02/ai-found-twel...
Im not sure whose claims you think I’m relying on. I trust Firefox that they’re not overstating the number of CVES they’ve found. Same for OpenSSL. The OpenBSD folks definitely don’t seem like the types. I’ve not known Linux to fabricate CVEs either. I think my sources are fine.
It's not that the utility of it put in question. What is however a giant question mark is how the heck any of the big AI companies are ever gonna get that ROI? Given how many of us are becoming more and more fine with local models that run just fine especially on a good enough computer which most developers have anyway...
Why should someone pick Opus 4.8 when Qwen3.7 Plus produces similar results for about 1/20th the cost.
That sort of pricing disparity is across the board. But further it's becoming more and more apparent that they are doing more with less parameters. That's what's giving the local models their super powers.
I'd say that yes, ignorance plays a role here because a decent number of engineers are looking strictly at the benchmarks and choosing Opus just for that reason.
But I'd also say that a major factor for Opus use is because Opus is being purchased for the engineers by their employers. They don't get to pick which models they are using.
The jury is still out on that.
The way you make a viable service that eats 300bn annually is to have enough demand to service that. Anthropic underbought compute. That tells you something.
The question is: what does "underdeliver" mean here? the pro-AI arguments I am seeing in this thread are equating mass adoption to agentic coding. Er, I dont know of any trillion dollar cap companies that sell dev tools. The point is Zitron doesn't have to be 100% right for his central prediction to come true.
* robotics (need to close data gap and release first viable product to get a data flywheel)
* conversational ai (no one is ready for this and we’re getting closer and closer to natural speech. The quality still isn’t good enough but it’ll be soon).
* other agentic use cases, openclaw adoption was crazy and that had a ton of barriers to entry
* ai products, like the one OpenAI is working on with Johnny Ive
Anyone thinking it’s unreasonable to hit whatever revenue requirements is just not that aware of what’s happening. Not to mention were capacity constrained already!! This is barely speculation at this point.
- RL is extraordinarily sample-inefficient.
- distribution shift/catastrophic forgetting aren't solved. only off-policy learning with giant decorrelated batches works.
- the breakout success of transformers as an architecture doesn't neatly translate to robot motion policy models.
the field is missing fundamental breakthroughs.
I also find it very interesting that conversational AI has taken this long. where are the models with good turn-taking? passive listening? the ability not to respond in paragraphs? has Anthropic simply not gotten around to it?
For conversational AI these labs do have lots of things to do lol but you’re right; it likely also requires some architectural improvements but you see the infancy: look at the llama4 speech duplex model. Very unimpressive yet all of the components are there. Just a matter of pushing on them, licensing and commissioning better data, etc. takes time and compute is stretched thin.
And where are those? They seem particularly hard to actually observe and only appear in anecdotes.
> I'm trying to believe
For every exponential increase in compute capacity you see linear gains in output accuracy. This is a death spiral. Anyways, you see "massive productivity gains" so why is "belief" a function of your viewpoint?
Just because you keep repeating something doesn't make it an undeniable truth.
This, combined with his extreme ignorance, makes him unreadable. The only reason people read his stuff is because it validates and confirms their own anti-AI beliefs. It's why every time he publishes an article, it reaches the front page in an hour or less.
Extreme ignorance?
No, he's not, he's making tons of money every month from his Substack subscriptions. In fact, the AI bubble popping would be the worse thing ever for him, he would be out of a job.
Just like the who have predicated the US dollar will collapse any-moment-now and which pushed gold for decades.
Funny how people always say "oh, you are an AI lab, of course you are going to hype AI", but never "oh, you make sooo much money from predicting the collapse of the AI bubble..."
How are they undeniable? They're very deniable. One example is the (seemingly) increasing maintenance costs for AI-generated code[1]. Another is the cost incurred by everybody reading AI slop instead of actual communication.
I don't have hard data as to whether these cancel out the benefits, but it's not as rosy as some seem to think.
[1] After years of people understanding that LOC is not only a poor productivity metric but also a negative indicator of code quality (shorter code for the same thing is better), we now have people touting how many LOC their LLM agent is generating. It's like everyone forgot what LOC actually represents and what it means for long term maintenance costs.
Predicting the timing of such a thing is notoriously difficult. I don't think being wrong about timing 2 years ago means there won't be a correction.
They were right about all of that but it took 15-20 years and the companies involved grew 100x in that timefold, eventually reaching trillion-dollar valuations that would've seemed insane in 2007.
There is a tremendous amount of money to be made in destroying society.
I'm not open-minded to arguments about utility, given that I personally witnessed LLMs evolve from interesting but useless toys to insanely helpful tools I use every day.
But from the article I linked back in March 2024:
"Generative AI models are expensive and compute-intensive without providing obvious, tangible mass-market use cases. Murati and Altman's futures depend heavily on keeping the world believing that development and improvement of their models' capabilities will continue a rapacious pace of progress that has unquestionably slowed, with OpenAI admitting that GPT-4 may be worse on some tasks.
As I've written before, hallucinations are a feature not a bug. These models do not "know" anything. They are mathematical behemoths generating a best guess based on training data and labeling, and thus do not "know" what you are asking it to do. You simply cannot fix them. Hallucinations are not going away."
Since then:
- hallucinations are dramatically less of a problem
- several mass market use cases have emerged, most notably coding
- rate of progress has increased
> - hallucinations are dramatically less of a problem
Sure, but it remains a big enough problem that human intervention and review is still necessary for any serious work across all use cases and industries.
> - several mass market use cases have emerged, most notably coding
Coding seems to be the only one, but there are still a lot of open questions about how the market can sustain the costs, and that's without considering the market dynamics that could emerge once costs are lowered enough that open source models start to become an attractive option.
> - rate of progress has increased
Debatable.
From my perspective, the model gains are mostly incremental now and a lot of the gains are just from things like improving the agent harnesses. I could be wrong though.
Every facet of the field is being pushed on and advanced at the same time.
No they aren't. The models still hallucinate just like they always did. You cannot trust them, ever, to get something right.
> several mass market use cases have emerged, most notably coding
They aren't really useful for coding based upon the above. Since you can't trust them, you have to carefully review everything they make, which in turn destroys any productivity they could've given you.
> rate of progress has increased
I have yet to see any progress. Opus 4.8 that you get today is no more effective than GPT-3.5 was. Much less would I agree that the rate of progress has increased. Only hype has increased, but there has yet to be a drop of substance.
Most notably? This is not a mass market use case in the way the author is describing. They are asserting that the amount of spend they need to get this off the ground necessitates the entire world coming in on it, and I would say that opinion has aged pretty well. There are a lot of coders, but there are more people scratching their heads as AI is shoved into every part of their lives.
> I believe that artificial intelligence has three quarters to prove itself before the apocalypse comes, and when it does, it will be that much worse, savaging the revenues of the biggest companies in tech. Once usage drops, so will the remarkable amounts of revenue that have flowed into big tech, and so will acres of data centers sit unused, the cloud equivalent of the massive overhiring we saw in post-lockdown Silicon Valley.
We have seen 8 quarters since. Has any of that come to pass?
Instantly close the tab as soon as the popup to subscribe to his newsletter pops up.
One other thing that’s working against the model makers is the hardware is getting better and the models are getting smaller and more capable. I don’t think we’re going back to the mainframe days. Local will be the endgame.
Is Ed right? Probably because in the end it’s unsustainable the companies left will be the companies that have income coming from somewhere else and there’s one large tech company that isn’t even participating in the boondoggle unless you count $1 billion dollars a year as participating ultimately there is no moat in AI model making.
Nvidia and Microsoft trying to introduce another Arm processor in a laptop of all things won’t change the tide either.
So, judge the book by it's cover?
> arguing that AI is failing, is a waste of money, is bad, will never work, etc.
Then the opposite should be easy to prove. AI is succeeding, is efficient, is universally good, and is working everywhere it's tried. Are those true?
It is literally judging the book by it's author, which is an extremely rationale judgement to make.
How is that better?
> which is an extremely rationale judgement to make.
So it's "rational" to take bias into reading? Why even read? If you know what you think and refuse to accept new information then what purpose is there in consuming anything?
You should just read the comments and get a warm fuzzy that the crowd, for the time being, agrees with your intentionally static ideology.
Comments like these obviously hope they can sway the crowd before they can take an unbiased reading of the article. If the author is that wrong then the crowd here should be able to discover that on their own. If the author convinces the crowd then I'd think you'd want to present a better argument than "well, he was wrong _before_." Post hoc, ergo propter hoc, in action.
We are only five or six years into the leap LLMs represent. For reference, radio waves were discovered in 1886, Marconi used them for communications in 1895, and while telephone and radio coexisted for many decades, it wasn't until the 1995 that mobile phones and wireless technologies started picking up. It took so long not because of the physics of radio waves required time to mature and improve, but because everything else needed to profit from it did require time.
To me, LLMs are not so much AI as it is a building block. Radiowaves maybe, or the equivalent of transistors. We are already seeing that it's possible to chain LLMs into agents. Currently, price is a strict limiting factor for coding and agents.It's probably fine-ish if all you want is Claude Code or Codex, but there are many other possible compositions of LLMs that most people don't dare to experiment with. For example, LLMs to drive NPC dialog and world mechanics in games is not a thing due to cost. Were prices of inference hardware go down and inference algorithms keep improving, I'm convinced (and afraid) we would see things very difficult to imagine today.
Hah, I'm actually working on just this problem.
Cost isn't the issue. There are only so many coherent (in context) responses and scenarios, that you don't need an LLM to generate text in the game, in real time. Instead, you can have LLMs build a vast corpus of "atoms" (dialog messages, fragments, cues, etc.) that can be stringed together in a deterministic way in response to player input. These can also be pre-screened and subjected to various tests prior to implementation.
To a player interacting in the game, a system like this would seem functionally indistinguishable from generated text within the game's designed interaction envelope. And it has huge advantages: Although it can expose seams if the player breaks character and decides to probe it, it won't be exploitable the way an LLM would be.
Worthless statement. Wow, you suspect something can make things better, worse, or both? That's a keen insight there.
> For reference, radio waves were discovered in 1886, Marconi used them for communications in 1895, and while telephone and radio coexisted for many decades, it wasn't until the 1995 that mobile phones and wireless technologies started picking up.
We are still so early.
I mean, we have advertised them in multiple super bowls, have companies that basically own tech news (incredulous journalists will repeat any stupid insane shit a CEO wants to say), that say they're valued at over a trillion dollars and nobody with the power to argue those finances seems willing to do anything but agree. We have built hundreds and hundreds of acres of data centers (and made deals for data centers that are never going to happen) that demand *billions* per month. They are devouring all the silicon to where people are visibly seeing the price of hardware double, triple, more in price. Work places insist on employees using AI (then pulled back because it turns out this stuff costs money and it's not fun anymore when it's not subsidized).
But we just need more time, more eyes, more people looking at it.
Where in the radio wave timeline did this happen?
Coding seems to be one of the core use-cases for LLMs (as Simon Willison pointed out recently) and even if that's the only real use-case for LLMs, they're wildly useful. I do understand that useful != profitable and that's where I think Ed has a real point: until inference becomes much cheaper these companies cannot be profitable. Some mega-players will pay the API token price, but most will not.
If the AI companies need $X billion in revenue to stay afloat, it doesn't matter if 0.5% or 5% or 50% of that revenue is from transforming the State of the Art. It's 100% irrelevant: what matters is that, transformation or no, these companies won't have the income to pay their bills. And if they can't pay their bills, a whole lot of other companies can't either.
So again, transformation or no, it's still a house of cards waiting to collapse. The only thing that would change that is not more "transformation" ... it's a feature set that lets them multiply their current user base (or multiply how much they charge them) several times over.
I find it quite refreshing in some ways. Lots of people, when they start complaining about this or that aspect of this AI stuff, are wont to add in a little disclaimer that, despite all of the above, they actually really like AI and use it all the time. I assume this is to avoid the scenario of a bunch of pragmatic builders turning up and calmly shipping nuance in the comments (or whatever you call it these days when you get brigaded by a pile of angry keyboard warriors with chips on their shoulder) - and it sure is tiring having to wade through the equivocation.
That's a criticism that'd be hard to level at Zitron! Say what you like about the man, but he's unafraid to appear to take a side.
This is often repeated but comes from ignorance mostly. You have * zero * reason to believe inference is costly other than just vibes. If you go by data and intuitions - the margins are high.
This kind of thinking really reinforces my belief that people have no idea and are using this whole [AI is not profitable and too costly] thing as a cathartic way to deal with immense progress.
https://www.wheresyoured.at/oai_docs/
However, it needs to be said that he received those numbers. I personally have quite a few issues with him, but there's no reason to doubt his journalistic integrity. Because of that, I believe he reports truthfully on data he receives by informants.
Additionally, none of the frontier models actually publicly talks about inference costs in anything but broad, "let's just forget that"-like takes. Which does not exactly spark confidence.
I'm eagerly awaiting anthropic's public disclosure of their financial details. That should be rather interesting in any case and finally put the inference-discussion to rest.
BTW, one thing for sure he is right about are the economics, as of today there is no way these massive investments are gone be paid.
Also because we now have a massive demonstration that vastly more efficient hardware is desperately needed.
Similarly other effective efforts towards on-device AI like Nvidia RTX Spark PCs and 2bit quants of strong models like DS4.
So inevitably, significant investment will be going into vastly more efficient CIM efforts like Mythic AI and new FeFET devices etc. in order to make human-level and beyond AI at scale feasible. There is so much demand for this and the power requirements of current hardware are so excessive, it seems unlikely that the data center build-outs will be able to recoup their costs before the more efficient paradigms make it out of the lab and start scaling.
So when I see monthly budgets in the thousands for developers at some larger companies, I'm curious to learn how they are managing to spend that kind of figure: how much code/documentation are they feeding into their prompts, are they using agent orchestration systems to make the code factory run 24/7, and how much value is coming out the other end versus before?
And, if they are pouring thousands into LLMs per developer, have they considered looking at alternatives like having LLMs running locally on own hardware with their own agent harness?
Those are the kind of questions I'd love to ask - I just wonder how much stuff is truly cutting edge and how much might be wasteful?
As for how to spend that much -- not that hard, to be honest. Just give it a lot of context and some relatively open-ended problem and it will easily eat through tons of tokens.
I have $200 subscription for Codex and it is crazy what it can do in terms of debugging. I have a pretty complex Electron setup with some native code linked via Node addons, a few App Extensions and it can easily read the source code to see how the builder works internally (e.g. if your end Info.plist is not correct), debug the xcodebuild output to see at which step something is not linked correctly (like after XCode major version bump), etc.
It is not a silver bullet but if you are not the one paying for it, there is no downside to throw a problem at it and see if it can come up with a fix.
> And, if they are pouring thousands into LLMs per developer, have they considered looking at alternatives like having LLMs running locally on own hardware with their own agent harness?
I am curious about that myself. I have a good machine now (Macbook Pro M5 Pro with 48GB memory), so I'll give it a try; I don't have high expectations so if it is actually helpful would be very neat.
Anthropic is growing way faster than doubling yearly so don't think this is entirely implausible
They have ai glasses and integration into instagram and facebook as the other avenues. I don’t see ai glasses as compelling yet, and don’t know how much more ad revenue or user engagement they can squeeze out with llms baked into the IG of FB flows. They are spending a lot and not seeing any returns. Am I wrong in being pessimistic about meta with AI?
I interpret the exact same evidence in the opposite direction. A year ago the idea that a company would spend $1,500/month/employee on AI tooling felt absurd, what could people possible want to do with AI that would cost that much?
Then coding agents (and, increasingly, general purpose agents) happened and suddenly companies are having to set limits because otherwise the demand from their employees is too high.
The TAM of these AI companies just leapt up to $1,500/knowledge-worker/month, how is that "slowing down"?
Companies love to cut costs, and just like they axe employee numbers at will, they will just as well make that kind of budget quickly dissapear the moment they realize they can go a different path for same or better value... Or simply because share holder short-term value demands it...
I think it's a poor number to build an "AI is slowing down" narrative around.
I do however think that shouting "look, Uber capped pricing at $1500/engineer/month hence AI is slowing down" is a questionable position to take.
No they're not. In reality, actual 'explosive uncapped growth of unlimited agentic token spending' will result in valuations several times more than a 'mere' $1T.
And as you have written on your blog it's a soft cap that can be exceeded with justification.
Anecdotally, $dayJob consumes Anthropic models via Azure subscriptions which lend themselves pretty neatly to the spending dashboards Ed mentions are missing from Anthropic themselves, and finance seems ok with the current usage, but there's no real hard incentives internally for AI usage either.
I guess Q3-4 are going to be interesting to see where this all goes.
I have found agentic coding to be extremely useful for a bunch of small, middleware, very focused bits of software for small businesses:
* A company had a very specific scheduling need, they needed to move about 8-15 staff around with a bunch of different shifts, and have custom reports on who was working how many hours, and have the employees get a nice clean email summarizing their schedule
* A manager wanted a very simple "let me send a text to add a to-do to the group list" need
* A sales team of 3 wanted to be able to type pricing of raw goods into their phone, have it compared to other market sources, and have it text the other 2 salespeople and their manager when they were out in the field
All of these were coded with Codex in about 4 hours with further refinements over the next week of back-and-forth with the people using the tools.
I suppose yes we could have found some custom middleware solutions that did similar things, but it's nice to be able to make a web page or tiny mobile app that just does EXACTLY what the person wants.
It's hard to do that and then listen to someone who says it's all just garbage.
"Anthropic, OpenAI and every other AI company deliberately obfuscated these costs because they knew that the second a user actually had to pay for the fuckups of an AI model they’d scream like they were being stung to death by bees."
So some of the growth was purchased by underpricing, subsidizing the customers with venture capital. Uber did that, and eventually got out of it by raising prices and squeezing the drivers.
The "fuckup" problem is real. LLM-type AI exacts huge costs because it is terrible at reporting "I don't know". When it doesn't know, it generates noise and polishes it. If a "confidence too low for output" signal could be extracted, this whole technology would be a lot more useful. You could use small, inexpensive models on small problems, and only use big models when the small models failed. Most customer service bots fit that model. Needing ever-larger models to fix the noise problem is not cost-effective.
It isn’t thinking or knowing and then expressing the resulting understanding but just spitting out contextual words and hoping it reaches a conclusion or ending of some sort.
In addition, there's a lot of research on the hardware angle and actual prototypes are already being built such as AI-on-chip Cerebra and Taalas for one.
Don't want to ruin it but go read some old posts from the author about AI, the tone is the same and he is very much wrong.
I’m not attacking the piece. I’m not saying it’s right. I’m not saying it’s wrong.
What I’m saying is, the tone made it hard for me to judge the arguments fairly, despite finding some of them convincing. And as much as I dislike it, persuasion does partly depend on how an argument is made.
Of course that mentality is obsolete. Now we all have infinite access to perfectly correct information via the internet.
He's in the media business... its in his interest to amp things up.
Does the truth normally lie somewhere in the middle of it all?
Usually does when you decide what constitutes extreme.
How people take this seriously? Anthropic is at 45B ARR S-1 shows inference margin climbed to 70% (obviously could drop) So where that 200B number is coming from ?
Edit:
> If you’re wondering what the story is, [...] I expect it to be out in the next two weeks [...] I can guarantee you it’ll be worth it, and you’ll be stunned by what I report.
Ok, this takes clickbait to new lows. The headline is trying to sell the teaser here, with very limited meat in the middle of the sandwich.
A good analogy might be networking companies and infrastructure companies during the dot com bubble. It devalued a lot of companies but the internet stayed. A lot of dot com companies didn't make it. Much of the infrastructure investment did not go to waste, however. Nor did a the technology go away.
I think it will be the same with data centers, related infrastructure, GPU hardware, algorithms, OSS components, etc. for AI companies. More companies need that stuff than is currently available. The ones that don't make it will have a lot of assets that they can pass on to the one that still have a chance. I don't think a lot of that stuff will get decommissioned or will be underutilized. It might get a little hair cut in value though. And like during the dot com bubble, some companies actually survived and did quite well. Especially those in the business of selling shovels during a gold rush.
After the inevitable consolidation that follows the next logical stages in the hype cycle, I don't think AI will go away. It might be a bit of a bloodbath for some silicon valley investors that placed the wrong bets in the last few years. But that's the price of doing business over there. That doesn't mean it's all bad. And the smarter ones probably spread their risk enough that they still might come out looking alright.
And like with the dot com bubble, many financial types have no clue what is happening and are running around like headless chickens. Which is why they ended up sinking a lot of money in exactly the wrong things. You'd hope they would have learned something.
But articles like this suggest that that might be too much to hope. They still don't really get how technology tends to not stagnate and might continue to deliver potential for performance and cost optimization. The current level of investment is only unsustainable if that doesn't happen and nothing else changes. I don't think those kind of closed world assumptions are a safe bet at all.
That said, I think his voice is useful as a counter to the mainstream opinion.
Given the amount of investments, approaching AI from the angle of economics seems correct.
We all have some level of personal experience using AI/LLMs, both chatbots and coding tools, and I personally enjoy using them, but I am sure this experience is relevant in this discussion.
I also enjoy luxury hotels, gourmet food, jet skis and helicopters, but this is not something I indulge in often because of the cost-utility ratio.
The real cost of AI may or may not be lower than its utility. The bet is that utility is increasing while cost is falling.
Internet continued to thrive and grow even after the stock market came and went, it took 13 years to roughly nasdaq to recover but the explosion of GDP from internet has been largely decoupled from the previous bubble boom and bust.
If you use the stock market as a yard stick to project new revolutionary technology we shouldn't have had trains, internet. In fact internet should've stopped with the bust of Nasdaq and everybody would've moved back to using paper but we didn't it gave rise to the next wave of economic output powered by this new tech.
I don't see AI to be any different.
Who writes like this? When you lead with "everyone who doesn't agree with me is a lying cheat coward imbecile" I think we should just turn the volume down on you to zero.
This is breakdown in dialog. If it leads like this then I I don't care how accurate the critical analysis to follow is. I didn't read the rest of the article and don't think anyone else should either out of sheer disdain for this argumentation style.
Ed is confused between whether AI is useful, and whether the current level of funding and valuations are sustainable. The following statements can both be true:
1. AI is already quite useful and will continue to be so. This is true even if AGI doesn’t happen.
2. The funding and valuations of many AI companies are too far ahead of their skis, and will probably roll back. Some may fail entirely.
About the “where’s the productivity in AI?” question: I think it’s entirely possible that the primary benefit of AI will not be top-line growth but reduced costs (through reduced human labor). Companies will need to reduce prices to prevent losing market share to existing or new competitors, meaning that GDP may not increase, but costs will.
Maybe AI is different. Certainly, the level scale of investment is on a different order of magnitude. But I'm wary of believing anything about the financial impossibility of AI being sustainable when I've seen such similarly confident arguments proved wrong in the past.
It's a pretty classic business strategy, and not directly comparable to any of the AI companies. There's a reason people compare the current situation to the dotcom era and not Uber. Also, don't take Uber as an example of a slam-dunk VC success story and leave it at that -- plenty of dumb ideas get pitched and funded and go bankrupt for every Uber.
It was only because Uber successfully bulldozed over all regulations that it was able to succeed ... and that was hard to predict before it happened.
Anthropic and Open AI could evaporate tomorrow and we'll still be using the models.
The market may collapse, but the people who think AI is going to disappear as a result don't understand what it is.
They are possibly in a winner take all death race against each other.
The stakes are so high that these cash rich companies cannot afford not to throw everything they have into this.
The sunk costs are irrelevant when it’s a question of survival.
Whether you hate or love AI computing is being completely reinvented - at the absolute core of this is computers programming computers.
Anthropic is winning this race by a country mile right now.
This is such an important future bet for these companies that the trillions must be spent because there’s no future or a greatly diminished future for some of them unless they have ownership of the technology.
Bloomberg is interested in what he has to say
But not HN commenters
The angry polemic that goes on and on and on with cuss words used liberally is just meant to evoke emotion and cathartic resolution to the type of people mentioned above. Not truth.
The thing is, there are a lot of people that find comfort in what he’s writing - primarily because it’s a coping mechanism against how quickly things are moving and a way to deal with being left behind. When you spend time, years, building institutional knowledge and making a whole identity out of it, you obviously will feel bad with the threat of it being commoditised.
I would write against the content of the article but I find it easier and more illuminating to write what he has said before instead. Then it shows how incorrect the guy has been and with what confidence he keeps speaking with.
-------
> While complex, generative AI is a technology that probabilistically generates answers, and has no "intelligence." It is inherently limited by its architecture, and in turn can only get "better" in a linear fashion. I see no signs that the transformer-based architecture can do significantly more than it currently does.
He wrote this in 2024 before reasoning models came out. Remember how ChatGPT was in 2024? Do you think this person is someone who gets predictions right?
> Furthermore, I hypothesize a race to the bottom in generative AI will significantly hamper OpenAI's ability to expand revenue, compounded by the fact that we're approaching the limits of transformer-based architecture.
He wrote this in 2024 and since then Anthropic's revenue increased by 160x to $40 B dollars a year and OpenAI's increased by 6x. Do you think this person gets predictions right still?
> I believe we're reaching the upper limits about what generative AI can do and how accurate its outputs can be,
He wrote this in 2024, do you really think we have reached upper limits? Huh?? What I'm using today is significantly more accurate and 2 tiers above what we had.
> And if there are true industry-changing possibilities waiting for us on the other side, I am yet to hear them outside of the fan fiction of Silicon Valley hucksters.
He says this about AI when we have with all honesty have had industry changing possibilities like agentic coding.
> There are indications that consumers have also lost interest. As pointed out by Alex Kantrowitz’ Big Technology newsletter, traffic to ChatGPT on both mobile and web has started to stagnate, if not decline. In January 2024, ChatGPT had 1.6 billion visits — 11% below the all-time peak of 1.8 billion. This makes it only modestly more popular than Bing, which had 1.3 billion unique visits during that period. On the mobile front, ChatGPT has an estimated 6.3 million US users — or 1.7 times less than the total of new Snapchat users added during Q4 2023.
He agrees with the claim that the consumer interest has declined. Since he said this, there was a 9x growth in active users.
-----
https://www.youtube.com/watch?v=_wStScmT748&t=1s
"AI Bubble Already Bursting?" (8 months back)
https://www.youtube.com/watch?v=T8ByoAt5gCA&t=1s
"A.I bubble is bursting with Ed Zitron" (1 year back)
He's been constantly crying bubble for years now.
-----
> AI video won’t get truly fixed just by waiting a year.
This is what he had said in 2024, and you just need to compare video from then and now to check whether the predictions came true. Why would anyone trust what this guy has to say?
The criticism goes both ways. The word "fixed", in Ed terms, can be translated to "become a viable business that justifies the spend".
In regards to AI video, I think the fact that Sora is no long around is an indicator. And there is seemingly no real appetite for AI video outside of memes, jokes, and misinformation, probably indicates that the prediction around AI video has come true.
Frankly this is anti-social and should not be tolerated here.
>In regards to AI video, I think the fact that Sora is no long around is an indicator. And there is seemingly no real appetite for AI video outside of memes, jokes, and misinformation, probably indicates that the prediction around AI video has come true.
His point was about the performance and accuracy and not about the community/market. He was wrong.
It's like someone arguing that cheese isn't real. Yes I can go to the grocery store and take a picture of cheese and show it, but what's the point? They can live in their own world. It doesn't change any of our lives. The world is what it is.
And for those who are all "but dur CCP get all ur data" you can use things like AWS Bedrock (at least for earlier versions of Deepseek and Qwen for now) and have more familiar people get all your data. Or buy (at obnoxiously inflated prices) your own HW and not send your data to anyone.
The funniest part of this is that people are often talking about how LLMs are now writing 100% of their code, then also saying that they don't want to expose their code to foreign government exfiltration by using foreign models.
But, uh, if an LLM is writing 100% of your code you have no actual secret sauce to hide from anyone, so why worry about it.
Meanwhile, like I think you suggest, I would assume everyone can generate similar outputs themselves. The idea that you can claim priority on your dream prompt and lock up the market on prompt responses sounds delusional to me. It's not novel invention when you're spit-balling at the same level of abstraction as every fantasy/scifi writer who ever was.
So I also have doubts about the sustainable business model. How long will it take for this fantasy to unravel, as people discover they cannot monetize their AI outputs as much as they dreamed, and in turn cannot afford to pay the AI services they use?
My absolute nightmare is that this becomes a "too big to fail" thing and oppressive/fascist governments decide to back full regulatory capture. That instead of letting it unwind, they grant and support enforcement of an increasingly absurd and arbitrary copyright/patent regime to support this monetization scheme.
> It's like someone arguing that cheese isn't real
I agree with your first statement (any being you) because of your second statement.
The current wave of AI unlocked language - the tools are now speaking and understanding. This, on its own, is astonishing progress. Language is the foundation of our culture and society; it is the very technology that got us, as a species, to where we are today. To have tools that can understand, manipulate, and produce it is a massive leap forward.
Once you see things that way, it is clear that we are not in a bubble; we are in a transition. Yes, there is tons of hype and over-investment, but the demand is real, and so is the impact. Unless you are deep in the tech and have that structural depth, it is easy to dismiss. This is like the invention of the personal computer, but with 100x the impact and speed.
The business model does appear to be viable for these labs. But that viability comes because they aren't wasting a bunch of R&D money developing worthless products like AI video production.
Regarding your comment about the business model—the people in Silicon Valley are not stupid. They know the playbook; we've seen it with social networks. The issue isn't the business model itself; it's that these companies need to dominate the market, and the big players are competing for that on a global scale. It's the exact same playbook that played out in financial systems and social networks, and now it's happening with AI. Once these technologies are deeply integrated into enterprises and the global economy, these players will dominate the market for decades to come.
I can assure you, the people running those companies are smarter than you, me, and the author of this article."
If I were to make a prediction, it's that ultimately these cheaper models are going end up eating their lunch. I don't think they'll make back the money they've invested and once that reality hits investors, those two companies are sunk.
That, however, is not the end of AI. Nor will it be the end of Nvidia/micron/etc. It will more just be a localized bubble pop that doesn't eliminate the product from the market.
These models are building deep integrations into companies and the entire economy. Once that stabilizes, it will be like the electricity grid—pumping tokens to fuel decision-making across the entire global society. Good luck unplugging from that.
Furthermore, there is a massive geopolitical aspect to it: those who are already on the Western financial and technical stack will get integrated even deeper now.
Much like the electric grid, what we are seeing is a convergence on standard APIs. For example, most of these cheaper models are hosted using APIs compatible with OpenAI. It's not a matter of rewiring your electric plug to work with a different socket standard, instead it's just the process of plugging it into a new socket.
> Furthermore, there is a massive geopolitical aspect to it: those who are already on the Western financial and technical stack will get integrated even deeper now.
Certainly the Chinese models appear to be some of the best when it comes to competition, but they aren't the only ones. There are European models and other US based models which all run for cheaper.
I remember one government project where we wanted to migrate a system from COBOL to a modern stack. The requirement was for the UI to stay exactly the same as the old green terminal; the evaluation criterion was pixel-perfect proximity to the original. We literally had to build terminals using web tech.
These models are not the same as each other. Once they are integrated and working, the incentive to change them is incredibly low. So really, the race is about who can integrate deeper, wider, and faster over the next couple of years—that is what will determine the long-term winners.
This is the exact same playbook we saw with social networks. There is a reason why we have only a handful of them dominating globally, and guess what? It's not because of the tech.
There is no incentive to rewrite working software in COBOL to something else. You don't really change the people cost of maintaining that code all that much and you incur a huge rewrite cost.
AI is different, it's an ongoing cost to the company. If that cost raises aggressively, you can bet companies will race to eliminate it, no matter how integrated it is. Companies can and do do this all the time.
And the models are close, not the same, but close. That's what matters in LLM stuff in general. If a model is capable of doing the same work for less, it will be chosen. Especially since the switch over cost is often on the level of "point the tool at this URL instead of that URL".
I get what you are saying if this were a more sticky concrete tech that is harder to move away from. But that's simply not the case for these LLMs. A big selling point they have is that they are super flexible.
I don't think the transition will be as simple as just flipping a URL. There is an entire legal and technical infrastructure being built around these models and their integration. I think you underestimate an organization's resistance to change once things actually work, as well as the sheer complexity of making that shift.
I also expect pressure will eventually drive the cost of running these models down. Power plants are being built, more capable chips are being produced, and a big chunk of the capital right now is being used to scale the physical infrastructure—the data centers and energy grid. Once that stabilizes, these companies will have positive cash flows. Again, it's highly similar to what we saw with the expansion of social networks, just with more aggressive and widespread adoption.
Ultimately, a handful of companies are going to provide these core capabilities, just like we have a handful of major cloud providers right now. Why do you think this would change? If anything, the trend toward deep vendor lock-in is even stronger now.
The biggest competitors aren't small models, they are just the traditional players that already have an "in" with enterprises. That I think will start to show its face once this initial round of buildout is complete, which may not be for another 5+ years.
I disagree. Mainly because those small models are exactly what erode away the moat of needing a giant data center. Those smaller models have been proving themselves to not be far of from the SOTA models.
As OpenAI and Anthropic look to raise their prices, businesses will be much more compelled to looking at cheaper models. And if the narrative is "do the same as you did with OpenAI at 1/20th the cost" that's going to sell to a lot of businesses.
It certainly cuts into what exactly these companies can sell in general. For example, if I wanted to integrate AI into a product I'd almost certainly not chose OpenAI or Anthropic. That's because they are simply way too expensive and what they'd give me is a lot less. We've actually ran into just this. We needed a classifier for a lot of records, we picked a free model because, as you can imagine, we didn't need something as good as what OpenAI and Anthopic offered and free works.
This is fire erasure
/s
