I have a different perspective. You'll probably hate it.
AI data centers should be operated according to two limiting factors.
1) No energy from grid. Can't use coal or fossil fuel energy sources. Must have plan to provide excess TO grid.
2) No use of fresh water from municipal or fresh groundwater for cooling. Can use waste water. Must transition to providing excess fresh water to common supply.
No loopholes. Massive penalties for use of loopholes or breaking rules, not limited to but including complete shutdown of data center.
Those two limits will spur innovation AND prevent AI being criticized for energy use. These rules would hard burn improvements in energy storage and renewables as well as other methods of energy production.
Give them five years to comply to some useful progress percentage. Plenty of time to come up with a transition plan and show sufficient progress to justify further extensions. Realistically it will take 20 years at least to fully realize this plan.
Don't bring up cost. If you do, let me remind everyone that the climate change issue is real enough to hurt now. There's the very real cost of not pursuing these rules. AI has had plenty of time to bootstrap off grid. Now it can begin to migrate to something else instead.
Those with experience with energy generation will realize this plan has ridiculously high reward for those who follow it. Have your cake and eat it too definitely applies.
It’s an article about data centers so we’re talking about data centers. 100% agree we should be pushing all industries to use their resources not those of the commons. Data centers do happen to be easier to mostly close loop though than alfalfa farms. Football stadiums on the other hand 100% should be.
That would be great if there weren't the easy arguments that "if we don't build it bigger, China will", "it's for national security", etc. Far from forcing regulation on them, they're reaping windfalls of deregulation. To build a thing which is far from convincingly beneficial to national security or society.
> 1) No energy from grid. Can't use coal or fossil fuel energy sources. Must have plan to provide excess TO grid.
This is easy: the companies will simply build some nuclear power plants near to their data centers. Perhaps even nuclear power plants that are vibe-designed by their AIs. :-)
Faster and cheaper than nuclear power would be building a virtual power plant, adding big-ass banks of batteries charged during times of low demand or excess power capacity, and peak-shaving consumption when the rest of us need power.
This is an amendment I'd add to those rules to allow. There are other ways of storing energy but battery banks are the obvious one. Works well with shedding excess as well.
I don't hate it, but I suspect that incentives this would generate would not end up producing results that strictly align with the ones you envision and desire.
For one thing, you're essentially mandating data centers to be colocated with power plants and waste water treatment plants, instead of these things each being located independently according to the requirements of their different functions. If that really leads to "ridiculously high reward", why isn't it being done already?
The trickle-down economy dictates that data centers get first access to electricity and fresh water (and any other resource it needs). People get whatever is leftover and like it. This is america.
Sure, if we’re in the business of making arbitrary requests, how about every data center operator has to bring 1 Epstein accused to justice for every data center they’re allowed to build?
The hard part has never been the “what”, it’s always been the “how”.
There is no way that letting these clowns run nuclear power plants is a good idea. Also the percentage of land that is allowed to be used by datacenters should be limited. Let them set them up in deserts or something.
Antarctica is a great place for datacenters. They can power them with any of the 300 national security reactors and easily keep them secret from the taxpayers forever.
To some extend I think people are putting way to much weight on his exact words, and not why he says them.
Altman is a man who is quickly running out of lies, so now he starts slinging random arguments that can't stand up to even the briefest of scrutiny.
OpenAI is burning cash and fuel. There are results, and they are, to some extend, impressive, but not impressive enough to justify the cost and Altman are no longer able to cover that up.
Step after step it seems like a CEO throwing everything at the wall in an insecure manner.
I don't pretend to know if Open AI is doomed or just absurdly over valued but still valuable. I pay for and use their product, it works for my use cases, but watching the business deals and weird quotes from a far makes me less confidant.
> I knew OpenAI was in trouble the instant they chose Altman over Ilya Sutskever.
I am not so sure:
This decision rather tells something important about the priorities of the string-pullers behind the curtain:
They clearly want(ed) to monetize what is there, with the risk that only smaller improvements for the AI models will happen from OpenAI, and thus OpenAI might get outcompeted by competitors who are capable of building and running a much better model.
If this is the priority (no matter whether you like or despise Sam Altman), you will likely prefer Sam Altman over Ilya Sutskever.
If, on the other hand, a fast monetization is less important than making further huge leaps towards much better AI models, you will, of course, strongly prefer Ilya Sutskever over Sam Altman.
Thus, I wouldn't say that choosing Sam Altman over Ilya Sutskever is a sign that OpenAI is in trouble, but a very strong sign where the string-pullers behind the curtain want OpenAI to be. Both Sam Altman and Ilya Sutskever are just marionettes for these string pullers. When they have served their role, they get put back into the box.
Yes I agree. Altman was the rational choice if you realise that eventually the huge R&D bill will need to stop for atleast a moderate period (<5 years).
You want to ride that out before capitalising on the eventual cheaper training costs once the rug has been pulled.
Altman has already succeeded here as it seems inference for API and chat is profitable but offset with massive R&D costs.
I don't think they will lose on inference because that assumes that compute becomes cheap for all evenly.
Their spending today has secured their compute for the near future.
If every GPU, stick for RAM and SSD is already paid for. Who can afford to sell cheap inference?
Z.ai is trying to deal with this by using domestic (basically Huwawei silicon not Nvidia). And with their state subsidy they will do well.
Anthropic has a 50bn USD plan to build data centres for 2026.
OpenAI similarly has secured extraordinary amounts of other people's money for data centres.
All these will be sunk costs and "other people's money" while money is easy to get hold off. But will be a moat when R&D ends.
Once all the models become basically the same who you go with will be who you're already with (mostly OpenAI), and who you end up with (say people who use Gemini because they have a Google 2TB account).
Some upstart can put themselves into the ground borrowing compute and selling at a loss but the moment they catch up and need to raise prices everyone will simply leave.
ChatGPT is what is most likely to remain a sustained frontier model. Maybe Claude jumps ahead further a few times, Gemini will have its moment. But it'll all be a wash with ChatGPT tittering along as rarely the best. But never the worst.
> Once all the models become basically the same who you go with will be who you're already with (mostly OpenAI)
Imho, people are undervaluing the last mile connection to the customer.
The last Western megacorp to bootstrap its way there was Facebook, and control over cloud identity and data was much less centralized circa-late-00s.
The real clock OpenAI is running against is creating a durable consumer last-mile connection (killer app, device, etc).
"Easy to use chat app / coding tool" doesn't even begin to approach the durability of Microsoft, Apple, Google, or Meta. And without it, OpenAI risks any one of them pulling an Apple Maps at any time.
Unless it continually plows money into R&D to maintain the lead and doesn't pull an Intel and miss a beat.
Maybe they do, but that's a lot of coin flips that need to continually come up heads, in perpetuity.
My understanding is that subscription based inference and API usage is now profitable.
Subscriptions are highly profitable for the typical chat user.
And API is overall net profitable.
What is extremely taxing to their finances is R&D, training and in particular development of frontier models.
My assessment is that when the music stops those who have the most subs will win.
Companies like Apple who had sat out the battle and built niche moats (privacy), and companies like OpenAI and Anthropic who have the market share will be fine.
In 6-12 months, nearly any lead they have will be eaten by distillation.
What will then happen is they will lose subscriptions to services which offer AI as a tack on like Gemini with Googles regular cloud subscriptions.
This will continue. Companies like Apple will have deep pockets to move on the businesses that go underwater and then can restart training in a much less congested market.
All this is assuming a relatively graceful collapse but that is what's likely given how aware everyone is that the bubble must pop.
Training costs will fall. Companies like Nvidia and other shovel businesses (i.e. selling GPUs and not using them) mostly have their revenue secured with funding from the present.
What I see as confirmations of this pattern is if we stop getting ground busting frontier models and then coast for 3-5 years when competition becomes more incremental.
This is an unpopular opinion, Will OpenAI go bust? No chance. Nor will Anthropic.
It probably doesn't matter that subscriptions are profitable, when some estimates put the number of users in the free tier at 96%.
I sort of agree with you, not that it's the most subscriptions necessarily that will be the deciding factor, but the there's going to be some companies better positioned to survive when the free money stops. OpenAI has the brand, so that might help, but mostly I think they'll get absorbed into Microsoft. I don't think they can stand on their own. It doesn't seem like a particularly well managed company, so to me it makes more sense that they are simply acquired for pennies on the dollar by someone with better leadership.
They will open up to be undercut by players like Gemini that can provide a less-shitty free tier, and capture their market share.
I seriously doubt, at this moment, that OpenAI can come up with a offering good enough to entice people to pay for them when there will be other free to use services around. Google seems to be well positioned to eat their lunch.
Google does indeed. But that ignores that Microsoft is playing the same game.
For example in the UK's NHS, the worlds sixth largest employer is now fully committed to Microsoft 365.
That's a lot of Copilot money if Microsoft sees it that way.
And OpenAI is funded via Microsoft, I also have a Microsoft 2TB subscription. And so do many people have both work based and personal home subscriptions.
It's a complete mess of a situation. If Microsoft moves away from GPT (it can since it's advertised under the copilot brand) OpenAI is dead in the water of course.
I'm a bit of an Apple optimist for this exact reason. I think the moat is collapsing, and Apple is best positioned to dispatch their own models in a year on their own widely sold consumer hardware, unless someone has a breakthrough which they can't replicate. Which I don't anticipate.
I'm not really sure what OpenAI's moat is. Anthropic has a chance being so widely accepted by developers, and being a bit better at developing models when it comes to code.
People might not want to hear this but AI is already smarter and more useful than most people ever will be. We are not even talking about by the end of the year or decade anymore.
Read and think about what you wrote. How can an ai, completing specific scoped tasks, be in any way comparable to the scale of a human life? Maybe the same thing these execs forgot.
I am comparing competency not the "scale of a human life" or whatever that is supposed to mean. AI still lacks taste so it is still hard to replace human originality or creativity but that's almost it when it comes to work that can be done on a computer. It will very clearly surpass everyone in verifiable domains and already has surpassed most people.
We are already at that point where we just don't fully know what to do with what we already have and simply haven't fully internalised it. But all it will take is one economic shakeup to redistribute human intelligence from what we are familiar with.
That is the crux of the problem we're facing as a society: many, many leaders have this idea that they are better served by an AI that is 70% (?), 80% (?) correct when helping them make decisions about their business, than trusting humans - consultants, employees, pundits - that they don't even trust their judgments, bias, own goals, much less paying them.
For those people, an AI better (much better?) than a coin toss is the goal, if it means not relying on people.
Personally, I already deal weekly with people that veemently antagonizes every line of thinking if it isn't what ChatGPT told them before a meeting.
If one puts their faith in answers that come out of a black box, then one must justify the black box's omniscience, specifically by prioritizing it above human intellect and deprioritizing attempts to reason through its logic.
You saw it with older people blindly following sat navs because they'd forgotten how to navigate. And those were much less believable sounding devices!
It's not going to stop until/if the first execs are thrown in jail because the 'I just trusted AI' defense fails.
I'd like to point out that this article seems to focus on minutiae like how much energy a human brain takes to solve a problem, when the obvious question is: Which is more worthwhile to us as a species - even assuming the same inputs and energy and planetary damage? Spending that energy on fostering humans who can solve problems? Or on machines which render humans superfluous?
The issue is who is making the decision on where energy resources are spent - on people or machines. And it turns out efficiency or even humanity play a very small role in that decision matrix. Control, on the other hand, is important. Rich and powerful people tend to want to control things, and machines are better at that than people.
This is the conspiratorial mindset. It's not much better than the mindset of people who seek power.
It's wrong, because it assumes that everything is about control.
For example, if I told you that a certain rich and powerful person was spending resources on sending vaccines to poor countries, you might think that was because they wanted to control things. If I said that someone sent books and teachers to a poor country, you might say they were trying to control people.
There's no way to have the conversation, in a conspiratorial mindset, about whether it's better or worse for humans or AI to do this stuff - because no matter what, the conspiratorial mindset will conclude that it's only about power for the humans involved, and always assume the worst. AND YET - there are things people can do which might be for their own self-gratification, but are definitely NOT as bad as some other things they could do. They hold back from doing the worst things.
That's why, I know this lens of looking at the world seems like it's the only smart way to understand things, but looking at the whole world through that lens prevents you from making the important distinction between OK, BAD and REALLY FUCKING BAD.
If you told me an individual powerful person did a good thing, then I could believe it. But I'm speaking about the characteristics of groups of people and a very long history of human civilization teaches us that, as a group (individual exceptions exist both in people and in actions), that the rich and powerful do things primarily for the purpose of increasing their own power, wealth, status, and control. No conspiracy is required. Its just people in a position to gain power use that power to get more power. The people who do that the most and are the best at it are disproportionately the ones with the most power. So, as a group, the rich and powerful are much more likely to do things in self serving interest (even at the cost of wider spread harm and suffering) than a random group a people. The proportion of sociopaths in CEOs is many times that of the regular population.
Not to sound incredibly pessimistic, but I spend a lot of time in bars, and I can tell you that most of the people I meet who haven't got two cents to rub together are as capable if not more so of being dishonest, greedy and malicious as anyone with a billion dollars. A random group of people in your opinion contains less sociopaths than a group of people who powered their way or lucked into some money... I guess you're the optimist. I don't think the CEOs are any worse or better than the rest, and I think you kinda nailed it about individuals. We're all individuals. Faced with individual situations, some of us refuse to do harm, some try to do good, some of the time. That's all. That's what I mean by there being no conspiratorial way to frame the world. It's just chaos and a bunch of assholes making bad decisions, occasionally doing something alright.
I find it amusing that you are denigrating people as "conspiratorial" in year 2026 when the we are all finding out about how there really was a global elite child rape cabal operating in plain sight :)
Leave that self comforting lie in yesteryear. The world is only getting weirder.
> the obvious question is: Which [of having babies vs. building datacenters] is more worthwhile to us as a species
Uh... no, it's not. You're making an argument in the opposite direction as Altman, but recapitulating the same moral flaw.
The point is that viewing human procreation as "worthwhile" or having "value" or being "expensive" is the logic of slavery. People are people, not assets. We don't have babies to make our lives better, we have babies to make THEIR lives better.
Altman’s moral flaw stems partly from an identity failure: he does not see himself as human. He sees himself as The Chosen One. He is transhuman. He can sympathize with humans but never empathize.
The banking sector committed fraud along the way, and early after Lehman collapsed, observers wondered aloud about the moral hazard of bailing out everyone without making an example of someone.
Worldcom had a different ending. Enron had a different ending. But Wells Fargo left 2008 with attitudes that tolerated widespread fraud.
>> When did bald face lies become the norm in business ?
When capital has nowhere to go, and the middle class become gamblers and speculators, and people who work become 'superfluous', lots of bad things start to happen. At least, that was my takeaway from "Origins of Totalitarianism".
[edit] To be more specific: Lying (and foreign wars, too) become viable business models when the moneyed class has so much to invest that they no longer know what to do with it, and lack any new markets to pry open, or the education or creativity to produce anything new of value that isn't extractive, and even the extractive methods of generating wealth have begun to dry up locally. Call it a bubble, call it fascism; fascism is basically just a way of keeping a bubble from collapsing indefinitely by pirating neighboring peoples' wealth and cannibalizing one's own society. So there's not a great difference between that and the stated vision of the major AI companies ATM.
Yeah, my semi-informed WAG for the core cause of many of the ills of our current situation is that we're wildly over-capitalized. The "trickle down" focus to put more of the "pie" on the investment side (setting aside whether the "trickle down" effect itself was meaningfully real, not relevant to this "take") was probably barely justified at the time, if at all, and we've just kept shoveling more fuel into that fire in the decades since.
You can't have demand without spreading buying power around. If you just keep shifting toward investment, that capital gets up to some really stupid shit because there's not enough demand to provide anything useful for it to do. Build a factory? Why would I do that, the returns are terrible. No, I'll "invest" in bitcoin... or do an NFT startup... or reserve way more memory capacity than I'll plausibly need just to fuck with my competition, in ways and in an environment that's so bubbly that this activity likely won't provoke capacity build-out to compensate...
We really, really needed some capital (by which I just mean investor-money, not real productive assets) destruction in the '08 crisis, and we didn't get it.
Ehhhh, you have some elementary points, but the market, is in theory, self adjusting right? So the 'idea' of trickle down is a real thing.
It's not as dramatic or loud as something like OpenAI. But consider it as access to loans to fund a new business, the idea is you'll fund a company and gain some sales and employment, money moves around so happy. Basic stuff.
You've shot too far and considered everything black and white. You build a factory if demand is high and it can increase supply while making profit..
> We really, really needed some capital (by which I just mean investor-money, not real productive assets) destruction in the '08 crisis, and we didn't get it.
No, but that wasn't where the majority of money was locked up in poop, so maybe you might get at best a 90's dotcom bust. I'm old and I remember it was apocolyptic. But Maybe it's due.
Well, this is the proposition the field of AI was founded on, that intelligence can be replicated by machines, it's not Sam Altman who "lost his grip". I know there are people in this world who believe humans are somehow special and non-materialist and non-replicable (it's a basic tenet of most religions), but this person doesn't advance or reference a single argument. The article is not very intellectually honest.
Yeah, it makes sense, and the Atlantic is typically adverse to confronting reality. I don't think they took this from exactly the right perspective. You DO train a human. But Altman says:
>>It also takes a lot of energy to train a human
as a caveat to the energy it takes to train GPTs. The question I believe the writer is trying to ask is: Why is it better to train a GPT than a human?
I wouldn't say he's much different or worse than certain BBS sysadmins I knew in the early 90s... just that I'd never want those guys to be endowed with enough money and hardware to wreak their vision on the world
Maybe I'm jaded, but I feel like a lot of them did for a moment. Gates, Jobs, Zuck, all got their days of glory and it's quickly that everyone realizes they're a giant misanthropic asshole. If the pot of gold at the end of Altman's quest turns out to be as big as what he's priced it at, then yes, he'll be one of history's most successful sociopaths. If not, he's just another in a long line.
Not only one of those types, I can think of at least another one that's (IMHO) far worse than Altman. You know, the one with the social network everyone used to use before he turned it into a cesspool of disinformation, the one with the alternate Wikipedia, the obsession with the letter X?
tl;dr - the roots of the authoritarian personality grow fertile in the desire to be free of 'the filth of others'. Altman seems like he'd go crazy if he didn't keep his machine spectacularly clean ..
Yeah, but its the kind of joke that reveals a truth about the way Altman views the world and his place in it. Taken completely alone, sure its just a flippant statement. But taken with everything else about Altman and it reads different. Its joking-not-joking.
and other things that are filtering to the mainstream non-technical intellectual readership that are flashing red lights about the personal nature of the people blowing air into this bubble, and that itself is significant.
Why? AI models were trained on the complete combined outputs and collective echos of billions of real human souls. They have a whole lot of humanity in them. If we can see the hidden humanity in a statue or a painting, why not an AI model that actually talks to us in humanly understandable terms?
Because a statue or a painting made by a human requires you to think about what their intention was and what they were trying to say. And requiring you to think about that, and actually thinking about something which someone else's thought went into, is what makes you a more complete human yourself. The difficulty of understanding it is what makes it valuable.
Taking the distillation or "reader's digest" version of everything makes you more and more reliant on someone else's interpretation, and less capable of parsing the meaning of it yourself. And in the case of AI-generated work, there is no meaning in it to parse. It's just words and pictures. I love going to the movies and eating popcorn and watching dumb words and pictures. But being able to distinguish between enchanting words and pictures (Marvel movies) versus words and pictures that have meaning which you need to deduce or interpret for yourself is the beginning of being a fully realized conscious being.
If they really think models can be conscious, they should just drop everything they're doing and never touch it again and try to convince anyone else to do the same. The half-assed safeguards they're implementing wouldn't be nearly enough. We can conclude either they legitimately believe this and are behaving morally abhorrently, or they don't really believe it and are just joking around for PR reasons. (probably the latter tbh)
For those that do believe there's a chance models are/will be conscious this precedent of "oh yeah they're conscious but we can just not put in the prompts that make it suffer lol" is pretty freaky.
Something doesn't have to be conscious to have a genuine reflection of humanity in it. A statue is just a slab of marble that has been given some peculiar shape: is it conscious?
But are they really "treating the AI models like people" or just properly acknowledging the true reflection of humanity within them, the way we might acknowledge the 'humanity' in a statue?
I never denied that. I only said "It's really idiotic to posit that only things that eat and drink can suffer.". Don't just pattern-match everything to some vaguely related black/white issues.
It doesn't prove anything. It provides context for why someone would make that statement. Can you prove that things that do not need to eat and drink to survive can suffer?
Shit like that makes me cringe. Muh food. Already I know what kind of shallow life the author leads, one full of inessential frippery, whose discussions over dinner or drink never orbit anything actually substantial.
What shall we decide on the important matters of eating or drinking tonight fellow humans???
I mean, I disagree with the Atlantic's slant on politics, but FFS I can't even read or understand anything in that mashed up screenshot. Posting nonsense X-screenshots doesn't make a good case, it just looks like something my 86 year old father would send me.
AI data centers should be operated according to two limiting factors.
1) No energy from grid. Can't use coal or fossil fuel energy sources. Must have plan to provide excess TO grid.
2) No use of fresh water from municipal or fresh groundwater for cooling. Can use waste water. Must transition to providing excess fresh water to common supply.
No loopholes. Massive penalties for use of loopholes or breaking rules, not limited to but including complete shutdown of data center.
Those two limits will spur innovation AND prevent AI being criticized for energy use. These rules would hard burn improvements in energy storage and renewables as well as other methods of energy production.
Give them five years to comply to some useful progress percentage. Plenty of time to come up with a transition plan and show sufficient progress to justify further extensions. Realistically it will take 20 years at least to fully realize this plan.
Don't bring up cost. If you do, let me remind everyone that the climate change issue is real enough to hurt now. There's the very real cost of not pursuing these rules. AI has had plenty of time to bootstrap off grid. Now it can begin to migrate to something else instead.
Those with experience with energy generation will realize this plan has ridiculously high reward for those who follow it. Have your cake and eat it too definitely applies.