The problem with his work is that a lot of the conclusions he comes to are a step too far, but the problems he identifies with how money is moving or how much datacenters cost or how no one has ever built a 1GW datacenter and we're expected to believe they can build a bunch of them, or even a 10GW data center - those are real.
I don't fully agree with Zitron on all his points but it's very hard to disprove the larger points will somehow resolve themselves. One point that I think will determine the future, moreso with data centers, is politics. Laws are not unchangeable and making decisions that determine the next 5-10 years of action based on the current law is a recipe for disaster. "Buying" a politician still has to account for the fact they can be voted out.
In a newly evolving situation, you have to assume that it's possible that the laws will change. Typically the way you account for this is hedging or keeping reserves to adapt.
Let us assume a person cannot change society, they simply do not have the leverage, even if they are a billionaire or most politicians. This is a systems problem, requiring enormous investment of time and resources to change trajectories of. Therefore, the choices available are to continue to live in a low trust society, or move to a high trust society and minimize your exposure to low trust society impact on your life. It's unfortunate. I would very much like to be wrong, please correct me if I am mistaken.
The propagandised idea that living in China means living under an oppressive communist regime, runs counter to the countries wild success at pretty much everything it does.
I cannot speak to China, but my recent personal experience is that Japan is high trust throughout Tokyo, Kyoto, and Osaka. Would love to hear about the trust situation in China from anywhere who has recent ground truth.
Even if we don't consider economics, the reason "AI" has not increased white collar productivity is because it's not necessary for knowledge work beyond aggregating information for research. Knowledge work has always been a lot more arguing than "doing".
Google was already the default option for this, and Google Gemini is a genuine improvement for Google Search. Google (and Apple) have sidestepped all the drama by just improving their existing products. Sometimes that's the right call in the shorter term, and this is one of those times.
Yes, and the scary thing is, when Google starts selling ads in Gemini results, you wont be able to see them, it will be a bigger business that Google Search ever was.
Update: Perhaps this is what Zitron doesn't want to say out loud, (and why investors are confused), the return on investment is going to come from having the largest mind control machine the world has even known.
There's just no easy way to see what makes LLMs able to duplicate the profits of the entire existing tech sector of the economy. This is the return they need to make based on what they're currently expensing and projecting. Couple that with the circular financing agreements and it's just preposterous. Zitron may have been first and most outspoken to this argument, but like all bubbles it's an open secret.
> the return on investment is going to come from having the largest mind control machine the world has even known
This is what Google already was and continues to be. They didn't need OpenAI and Anthropic as foils to maintain that position or expand Google Cloud. If anything, they did all this work with Gemini despite what's happening on the sidelines.
OpenAI secured their multinational defense contract with Project Stargate. I can easily see the extremely high demand to get rid of so many contractors with security clearances. No government believes that the risk of another Snowden-like whistleblower is worth the mediocre intelligence summaries.
The loser will be Anthropic unless Microsoft buys them out for a lowball price at the last minute in an attempt to make Copilot less terrible. For everyone else, it's business as usual.
The chimp brain cant handle unpredictability for too long. Just keep pushing people into unpredictable situations or specifically try getting someone to say I dont know constantly. Beyond a point story generation starts, to feel sane within an "I dont know" reality.
Truth is things are unpredictable. Predictability of anything hasnt increased.
Stories/Lies exist for cope. Cause of chimp brain dynamics responding to unpredictability.
So right towards the end Zitron says it isn't clear what Google, MS and Amazons plan is. But he just finished explaining that Anthropic and Open AI are paying for them to build data centers, and implied that on face value, this was to rent the data centers back to Anthropic and Open AI.
Seems to me they are just bleeding them dry so that when the money runs out for Anthropic and Open AI, they will be left with data centers that the can run their own (or open) models.
If you read more of Zitron's work, he says that there's little demand for the AI compute power outside of OpenAI and Anthropic. If they are bled dry, then the hyperscalers will not have enough customers for the data centers which they have blown so much capex spend on and will have to soon go into debt to ensure they can be built. The GPUs and the high bandwidth memory they use have little use other than to run and train these LLMs.
It's GPU compute, which is a lot more niche. There are only so many things you can render or simulate and only so many people/companies who want to render or simulate something. Not even remotely enough to create enough demand to saturate a trillion dollars worth of GPU data centres.
> little demand for the AI compute power outside of OpenAI and Anthropic.
I think every other business is going to slowly wake up to how powerful machine learning can be, and or, how important it will be to run your own models you can trust.
I don't doubt that there is a bubble, but once the hardware is built, there will be plenty of customers.
I guess if you overpay for hardware, you can be undercut by data centers built after the bubble pops, but that might just mean your return on investment is slower.
I don't think so. AI is a general purpose tech and you have many companies still doing discoveries on how to apply the technology in a useful way. It's an expensive solution in search of a problem. Not to mention no one wants to pay the actual cost of what it takes to deliver the technology. The whole business model is unsustainable.
and he wrong on all of those accounts. but he’s got the best gig, he has been writing literally one same article for years, gets LLM to shuffle the words and publishes it under a different title :)
His follow-up point was, "that seems like a ridiculously nice double-dipping situation" until you realize that money depends on only two "unsustainable" customers, spending other people's money, that now have to become absurdly massive to ever make this work.
I don't think Zitron has been right about anything in the last 5 years. But he talks a good talk.
I cannot understand why people listen to him? I feel like I must be living in an alternate universe to a substantial amount of people? It's bananas...
The problem with his work is that a lot of the conclusions he comes to are a step too far, but the problems he identifies with how money is moving or how much datacenters cost or how no one has ever built a 1GW datacenter and we're expected to believe they can build a bunch of them, or even a 10GW data center - those are real.
Well he's saying what a lot of people desperately want to believe.
> It's bananas...
Actually, it's lemons (https://en.wiktionary.org/wiki/Zitron).
Sorry, I couldn't resist:)
I don't fully agree with Zitron on all his points but it's very hard to disprove the larger points will somehow resolve themselves. One point that I think will determine the future, moreso with data centers, is politics. Laws are not unchangeable and making decisions that determine the next 5-10 years of action based on the current law is a recipe for disaster. "Buying" a politician still has to account for the fact they can be voted out.
> making decisions that determine the next 5-10 years of action based on the current law is a recipe for disaster.
How else can one make business decisions? Or any long term decision, financial or otherwise.
If you are correct, which I fear you may be, then we, in the West generally, are now living in a low trust society.
Which is not a good thing at all, and should be of utmost concern.
I feel like it’s been understood that we live in a low trust society for a while now. If not for all of us, at least for a LOT of us.
In a newly evolving situation, you have to assume that it's possible that the laws will change. Typically the way you account for this is hedging or keeping reserves to adapt.
All day every day you make decisions based on things you think you can safely assume, and things you know you can't assume.
The most likely behavior of a political office is just one of the things you know you can not safely assume, that's all.
Let us assume a person cannot change society, they simply do not have the leverage, even if they are a billionaire or most politicians. This is a systems problem, requiring enormous investment of time and resources to change trajectories of. Therefore, the choices available are to continue to live in a low trust society, or move to a high trust society and minimize your exposure to low trust society impact on your life. It's unfortunate. I would very much like to be wrong, please correct me if I am mistaken.
https://www.visualcapitalist.com/mapped-where-people-trust-e...
https://ourworldindata.org/grapher/self-reported-trust-attit...
This data conflicts with my image of Japan and China, and I wonder where the discrepancy lies.
The propagandised idea that living in China means living under an oppressive communist regime, runs counter to the countries wild success at pretty much everything it does.
I don't think success directly translates into high trust.
I think of China as a hypercompetitive place, which to me should mean lower trust.
Explain why hyper competitiveness should mean lower trust?
Doesn’t competition imply rule and order?
The other direction is merely organised crime, which tends to have very few big players.
Hey, remember Naomi Wu? She used to post a lot here. I wonder whatever happened to her?
Good thing Western governments never Canadian truck drivers, Melbourne lockdowns, Covid lies, Edward Snowden, Julian Assange. Just to name a few.
I cannot speak to China, but my recent personal experience is that Japan is high trust throughout Tokyo, Kyoto, and Osaka. Would love to hear about the trust situation in China from anywhere who has recent ground truth.
Even if we don't consider economics, the reason "AI" has not increased white collar productivity is because it's not necessary for knowledge work beyond aggregating information for research. Knowledge work has always been a lot more arguing than "doing".
Google was already the default option for this, and Google Gemini is a genuine improvement for Google Search. Google (and Apple) have sidestepped all the drama by just improving their existing products. Sometimes that's the right call in the shorter term, and this is one of those times.
Yes, and the scary thing is, when Google starts selling ads in Gemini results, you wont be able to see them, it will be a bigger business that Google Search ever was.
Update: Perhaps this is what Zitron doesn't want to say out loud, (and why investors are confused), the return on investment is going to come from having the largest mind control machine the world has even known.
There's just no easy way to see what makes LLMs able to duplicate the profits of the entire existing tech sector of the economy. This is the return they need to make based on what they're currently expensing and projecting. Couple that with the circular financing agreements and it's just preposterous. Zitron may have been first and most outspoken to this argument, but like all bubbles it's an open secret.
LLM advertising consuming all current advertising dollars plus more is realistic. That's a solid chunk of the entire tech center it eats right there.
> the return on investment is going to come from having the largest mind control machine the world has even known
This is what Google already was and continues to be. They didn't need OpenAI and Anthropic as foils to maintain that position or expand Google Cloud. If anything, they did all this work with Gemini despite what's happening on the sidelines.
OpenAI secured their multinational defense contract with Project Stargate. I can easily see the extremely high demand to get rid of so many contractors with security clearances. No government believes that the risk of another Snowden-like whistleblower is worth the mediocre intelligence summaries.
The loser will be Anthropic unless Microsoft buys them out for a lowball price at the last minute in an attempt to make Copilot less terrible. For everyone else, it's business as usual.
Only a person who does no actual work could hold this opinion.
The chimp brain cant handle unpredictability for too long. Just keep pushing people into unpredictable situations or specifically try getting someone to say I dont know constantly. Beyond a point story generation starts, to feel sane within an "I dont know" reality.
Truth is things are unpredictable. Predictability of anything hasnt increased.
Stories/Lies exist for cope. Cause of chimp brain dynamics responding to unpredictability.
[dead]
So right towards the end Zitron says it isn't clear what Google, MS and Amazons plan is. But he just finished explaining that Anthropic and Open AI are paying for them to build data centers, and implied that on face value, this was to rent the data centers back to Anthropic and Open AI.
Seems to me they are just bleeding them dry so that when the money runs out for Anthropic and Open AI, they will be left with data centers that the can run their own (or open) models.
If you read more of Zitron's work, he says that there's little demand for the AI compute power outside of OpenAI and Anthropic. If they are bled dry, then the hyperscalers will not have enough customers for the data centers which they have blown so much capex spend on and will have to soon go into debt to ensure they can be built. The GPUs and the high bandwidth memory they use have little use other than to run and train these LLMs.
So his big insight is that if you take away the two largest customers the demand is much lower? Wow, such informative.
>The GPUs and the high bandwidth memory they use have little use other than to run and train these LLMs.
It's compute, and can be used to compute anything.
It's GPU compute, which is a lot more niche. There are only so many things you can render or simulate and only so many people/companies who want to render or simulate something. Not even remotely enough to create enough demand to saturate a trillion dollars worth of GPU data centres.
> little demand for the AI compute power outside of OpenAI and Anthropic.
I think every other business is going to slowly wake up to how powerful machine learning can be, and or, how important it will be to run your own models you can trust.
I don't doubt that there is a bubble, but once the hardware is built, there will be plenty of customers.
I guess if you overpay for hardware, you can be undercut by data centers built after the bubble pops, but that might just mean your return on investment is slower.
I don't think so. AI is a general purpose tech and you have many companies still doing discoveries on how to apply the technology in a useful way. It's an expensive solution in search of a problem. Not to mention no one wants to pay the actual cost of what it takes to deliver the technology. The whole business model is unsustainable.
The business model has not been discovered yet.
Go have a look at that Seedance thread. Movies and TV will be completely AI in the future.
and he wrong on all of those accounts. but he’s got the best gig, he has been writing literally one same article for years, gets LLM to shuffle the words and publishes it under a different title :)
Now, now, let's be fair. I don't think LLMs are used for writing or editing the articles in any way.
(Each reader can decide themselves whether what you wrote is unfair to Zitron or to LLMs.)
His follow-up point was, "that seems like a ridiculously nice double-dipping situation" until you realize that money depends on only two "unsustainable" customers, spending other people's money, that now have to become absurdly massive to ever make this work.
Another view of that is, hedging their bets. Which businesses do all the time.