Million GPU clusters, gigawatts of power – the scale of AI defies logic Next year will see some truly monstrous compute projects get underway as the AI boom enters its third year. Among the largest disclosed so far is xAI's plan to expand its Colossus AI supercomputer from an already impressive 100,000 GPUs to a cool million. Such a figure seemingly defies logic. Even if you could source enough …
Is this to imply that the industry is getting high on AI? Or should I just assume that the authors have no idea what "chasing the dragon" is?
I guess we can just put this phrase next to "light years in the future" in the List Of Phrases I'm Going To Have To Learn To Ignore Without Reacting[TM].
Millions of GPUs? Thousands of ExaFLOPS?
Wow. Just think of the possibilities.
Pictures of people with previously unimaginable numbers of arm joints and fingers.
Whole new areas of hallucinated alternative facts, confidently asserted in mere femtoseconds. Black will be white, up will be down, left will be right, P will be both NP and also Cents.
Videos with bizarre rendering and continuity errors, gurgitated on-demand and near instantly.
Truly, we are blessed.
There was a pretty neat piece on this last week: Twirling body horror in gymnastics ... Sora AI can make the human body do amazing stuff!
The AI beer commercial posted at arstechnica is worth watching, “Women laugh, jaws flaring. Beer glasses turn into beer cans. Flaming grills achieve columnar fire tornado status and arc across the yard. It's a vision of surrealistic hell that is at once familiar and impressively alien.”, very creepy!
"...and there was I thinking it was a CES preview for an AI sex robot experience from Neuralink"
I remember a movie where somebody's brain could be recorded and they found one of the doctors that recorded himself having it off with some attractive lady and then spliced the tape to run it as a loop. He was a bit worn out when they found him. I expect suffering severely from dehydration.
This post has been deleted by its author
All the dreaming about massive AI compute is great but sooner or later has to crash into the hard financial reality that it has to somehow generate money to both build and run it. At the moment it doesn't look like there's paying demand for what already exists let alone more, at least if all the desperate attempts to give it away for free are a clue.
About the only one who likely doesn't care is the one hyping xAI, given both the likely financial chicanery around who'll actually end up paying for it all plus the reality that the real profits from their businesses usually come from funding rounds and stock, not the actual business operations; all that truly matters is sustaining the pump.
In the end, we will all buy what we are told to buy. Nobody wanted 5G, but your phone (and your nearest tower) probably uses these bands by now whether you like it or not. The same for my most hated software microsoft teams, alternatives are far more useful but my work PC has it locked to auto-start on boot. There is a long long history in computing of expensive crap being shoved down people's throats.
"In the end, we will all buy what we are told to buy. Nobody wanted 5G, but your phone (and your nearest tower) probably uses these bands by now whether you like it or not."
The phone companies needed something to get people to buy yet another phone, not whinge about the higher prices for service and the ability for the next gen of cars to be able to do massive data dumps that rat out their owners. There's no 5G where I live. We are lucky to have service at any given time, but it's a small town and not on the affluent end of the wealth scale. Securing the bands was likely also a priority so in high density areas, people weren't constantly staring at a rotating ball every time they loaded a new site.
Still no fiber where I live and the distance to the nearest node supports paltry 12 Mbit over DSL. So I get my internet via mobile network - 4G was nice until it became so saturated that I got regular drops and went in and asked for 5G. Live next to a road that is one of the main entryways for commuters into the city. You could measure the drop in speed and ping as they start rolling in and out.
Either not many have 5G or the tech is able to better handle such loads, but haven't had connection issues since.
The quick fix of 5g internet access wears of just as quickly when you have to pay for bandwidth usage and nobody offers unlimited access (the same applies to cable but the pricing would likely be lower). And lets not forget about standing in just the right spot for this promised 5g to pass the test. "AI" has lots in common with 5g hype. Plenty of gimmick use cases but majority of population is unlikely to want or be able to afford to pay for it. We'll see what businesses do with it (and how much of the "benefits" will they pass down in prices of their products).
Our national T-Mobile spinoff recently launched a 5GTTH product, and according to their coverage map you will get 300 mbit basically across the country. Granted, I live in a small, dense country, but the point still stands. Especially since old city centers can sometimes be the last to get fiber. €25 a month, if you use more than 2TB in a month your priority drops when there is congestion in the air. But, just to illustrate, my brother and I live 3 streets away and both consistently get 800 down/100 up with latency comparable to cable and dsl. So does his friend who lives 2 blocks away. Oh and guess what, it's extremely popular because it works so well. 5G @ 3,5Ghz is more useful than you might think.
For Trump to get (or try to) get into the AI Grift.
Yes, he has his crypto thing but Elongated Muskrat will never let the Trump Coin challenge his Crypto Holdings.
Natirally, it will all be made in China. He won't by 'made in USA' because he is a cheapskate.
He'll buy a 'Good Clean Coal' mine and power plant just to power this scam that won't be able to generate even half the power needed. Cheapo Trump.
Watch out for Brownouts if you are in the USA. Trumpo will probably be the cause.
Meanwhile.... the US and then the World Economy will tank while Trump blames Biden for it all including him deporting millions of people who contribute to the US Economy.
On the upside, Prince Harry could get deported back here... oh wait... that is bad.
From companies wanting to reduce their workforce and increase their turnover. At least, that's the plan.
Might work, might not. Personally, I reckon it's another dot com bubble - a rapid, indeed feverish, cycle of investment and development that everyone wanted to be part of, which mostly just burnt money, but did generate a few worthwhile things from the flaming wreckage that it left behind. We'll see.
GJC
"did generate a few worthwhile things"
...and some of the things might even be predictable in advance. What won't be obvious is which provider will survive. It's a bit like online shopping. 30 years ago it was clear that this could catch on and many people jumped on the bandwagon. Only Amazon survived. Was that obvious? Was he obviously better at it than everyone else? Or was he just the lottery winner?
Amazon won because they didn't make a profit, they focused on getting rid of the barriers that make people not want to shop online. For example allowing you to return goods though simply pressing buttons on the website instead of having to email to request a return form. Only once they had market dominance did they start making a profit.
"Only once they had market dominance did they start making a profit."
I've never had creditors that were that patient. If they aren't, the enterprise will fold. People that always say "the early bird gets the worm" fail to remember that it's the second mouse that gets the cheese. There might be a massive falling out as a company such as Nvidia comes out with hardware that's twice as fast, consumes half the power and is substantially less expensive. On the software side, it's coming up with the most useful product or a toolset that lets companies develop their own for specialized tasks.
The Silicon Valley playbook requires that you achieve market dominance before you think about making money. As such, the plans are similar to those around Uber a decade ago and, unsurprisingly, investors are on board throwing cash – which might include your or my pension fund – at the problem. And as long as cash is available, either directly from investors, or via the stockmarket, this will remain the cash and divert resources away from potentially more efficient approaches. Access to capital is thus being used as a barrier to entry and making a mockery of the idea of "creative destruction".
There was a good interview yesterday on German radio with one of the people behind NX AI on their non-transformative approach, which is apparently both significantly faster and more efficient that that used in the LLMS, but also the decision to avoid language processing precisely because the market is being crowded out by investors.
I avoid Amazon as much as possible and shop with actual retailers, but I bet they use Amazon to some degree anyway so it’s unavoidable.
The answer is to not shop at every whim, reuse, repurpose, recycle, but, hey, that’s nonsense hippy talk.
We have all become a commodity.
"The answer is to not shop at every whim, reuse, repurpose, recycle, but, hey, that’s nonsense hippy talk."
I use the 3 R's all the time and it makes sense. My house is heated by solar collection panels that have a fair bit of repurposed material. They'd be expensive to build if I didn't scrounge. I'm building a CNC router, slowly, with mainly second hand parts. I just scored some big stepper motors for free. A local seller has 5kWh, 24v Li batteries for sale. To buy them new would be very dear if I could buy just one or two. I expect that I won't be getting a full 5kWh capacity, but even half at that price is very good.
"Good luck with that. They have fingers in everything. Every. Single. Thing."
To varying degrees, yes. I make Mac and Cheese from basic ingredients rather than from the contents of an expensive box. The wheat used to make the pasta likely came from a megacorp farm, but buying it in bulk rather than a wee bit at a time is less profitable for Kraft Foods. I make my own cheese sauce as it only takes minutes while the pasta is boiling and I'll use a good quality local cheese if possible. The salt certain came from another giant company, the flour (bulk buy again) came from some big milling conglomerate. If I use a good local cheese, that bumps my price up, but it's also a much more premium version. I can use the cheapest mild cheddar if I need to save a few bob.
The difference is where in the chain I'm buying. I'm not going to mill my own wheat. I've looked at small home mills and there's no ROI for the machine, much less the time and quantities of wheat I'd have to buy just to get the varieties I'd need for what I'm making. Even if I were to spot one at an estate sale going to very little/free, it would make no sense. There is some benefit to megacorps to a point.
No, but it will play a mean Doom.
Sorry, try again.
No, but this means we are doomed*
*to keep paying for all this guff and having its random outputs foisted on us in an increasingly desperate attempt to make it "relevant to our everyday life" and therefore worth the expenditure.
the 60's space race with this AI race.
The space race was government driven with national pride at stake so wasn't commercially driven and within the two countries playing the game, there were no competing interests, no splitting of investment money, technology etc. Thus they (the US) reached their goals.
Plus, given the investment needed and the lack of egotistical and narcissistic super billionaires in the 60's only a government had the resources to fund a trip to the Moon.
The AI race has no such unifying features. It's different factions splitting the investments, each with competing computing platforms and end goals.
Thus it's chances of success are reduced.
And I would venture to say that the challenges of putting a person on the Moon in the 60's with the primitive hardware, software and systems they had available is a far more impressive achievement than if/when AI actually becomes generally useful (aside from some niche edge cases)
Bluck
"Plus, given the investment needed and the lack of egotistical and narcissistic super billionaires in the 60's only a government had the resources to fund a trip to the Moon."
The Soviets had the money, but after the failure of the N1, they could see that there wasn't the time left to catch up and beat the US with some new hardware. The crucial part was the ability to build something as large as the F1 rocket engine and get it to work. Clusters of smaller engines only works up to a point.
Alternately the "get to the moon at any cost by an arbitrary deadline" ended up creating a NASA that was far too big and inefficient but without the core skills to actually build anything. All the Apollo HW had been built by defense contractors. Given this huge organisation they needed a follow-up project that was on the same scale ie. The Shuttle, which was doomed from the start.
A more commercial space program from the start might have meant competing Boeing and Lockheed and McDonnell-Douglas and de Havilland launchers. We wouldn't have got to the moon but might have had SpaceX type innovation 50 years ago. Instead we had something like the US ship building industry, heavily subsidized and protected - but only capable of building $20Bn aircraft carriers.
"ended up creating a NASA that was far too big and inefficient but without the core skills to actually build anything."
NASA did a tremendous amount of basic research that let them design the pieces of the Apollo program so contractors could build it. The knowledge is there now that NASA doesn't have to do the entirety of the design and can hand that off to industry to concentrate more on the mission, science instruments and coordination. If the early work hadn't been done by NASA, it would never have been done by private industry that couldn't see an immediate ROI.
I don't think one can say that NASA is inefficient. There isn't anything to compare what they do against. They are looking to create systems to answer scientific questions in space and atmospheric flight. If they were designing and building consumer goods, one could say that the amount spent on R&D is difficult to amortize for a given product. NASA's product is knowledge so how would one put that into an out/in ratio? It can't be measured in only looking at things that have worked when understanding what doesn't work and why isn't a failure, it's knowledge. The Michelson–Morley experiment was an attempt to measure the motion of the Earth relative to the luminiferous aether. It's the most famous experiment with a null result. The experiment was very clever and wound up showing there was no aether as opposed to measuring its characteristics. At the time, the vast majority of scientists believed it did exist and that was likely a huge barrier to progressing further in cosmology and physics.
But the focus on get a man on the moon was at the expense of a lot of science, it was only on the last of the moon missions that they thought of sending a geologist.
It follows through to today where the astronaut = daring fighter/test pilot explains the level of science from the ISS.
You can judge efficiency in the sense of having to build duplicate sites because you need to do different RD tracks in parallel because you need a rocket in a year. And are then left with a 100 sites in different States that you can't close because of political pressure but have to keep funding at the expense of canceled science missions.
"But the focus on get a man on the moon was at the expense of a lot of science,"
I sort of agree since science investigations were put aside for work on engineering. Once enough engineering was done to have the confidence to send humans to the moon, they went straight to it so there wasn't time for the USSR to catch up (they didn't know for certain they had given up). By that last flight, NASA was able to send up Jack Schmidt rather than a specialized trained astronaut in that slot. It showed that a human is able to do a lot of work much more quickly than a rover. Something that Steve Squires (MER rover principal scientist) has pointed out. Had further missions gone forward, maybe there would have be new craft that could take more researchers to the moon at a time which is where we stand today.
Restarting nuclear power plant construction is probably a good thing. Less so the gas turbines, admittedly. There's likely to be a lot of compute looking for a use when (hopefully) the AI stuff proves incapable of earning back its investment; fancy modelling might be cheaper than expected in a few years.
I guess you haven't been following the numbers. For any given yield in power, nuclear plants cost many times more than solar and wind installations, and take about 20 times longer to build. They also have a small but inescapable ability to blow up, with both the probability and potential devastation of such events being sufficiently large that insurance companies won't take the bet at any price.
The new modular reactor concepts, as well as fusion, are certainly worth some research money. But realistically they can't be expected to do much about our current deficit of clean power and surfeit of greenhouse gases within the next crucial decades.
From TFA: "drive investments in things like nuclear power or sustainable computing"
Sounds like relief to me! If (when) the current AI bubble bursts, we'll still need a buttload of (as clean as possible) electricity to drive EV adoption and use, industry conversions, and more ... We clearly can't go on with the pressure imposed on the inner bowels of our planet's climate by the greenhouse gas exhausts of internal combustion. Averting the growing pile of related disasters and discomforts, and restoring the sweet spot of regularity, should be Priority One for all in the long, hard, and sustained, push-out and back, gutsy movement for de-enshitification, IMHO!
If those massive investments in AI (however misguided) help us overcome the emergence and realization of a global Cotard syndrome (rather than causing doomsday as a iatrogenic outcome) then I'm all for it. Just imagine what could be done with the 5 GW of clean power generated for Microsoft's death-star-gate project, once the death-star and gate bits are preemptively unplugged from it!
"we'll still need a buttload of (as clean as possible) electricity to drive EV adoption and use"
Not as much as you think. The problem is distribution. Oil refineries consume a S-ton of leccy. In non-standard (US) units, 7.5kWh per gallon of petrol (likely about the same for diesel and Jet-A). Convert that into miles driven in an EV for a comparison and remember that it's just some of the energy that's gone into creating the product, not energy derived from the product when inefficiently exploded within a chunk of metal that has a bunch of moving bits.
Another thing that needs work is storage to offset demand peaks. If you have solar panels on your home, it would be great to stick the electrons in a can so they can be dumped into your EV when you get home in the evening. I'd argue that storage is better than sending excess rooftop production onto the grid since the grid doesn't have a good way to store it for use later in the day.
> If you have solar panels on your home, it would be great to stick the electrons in a can so they can be dumped into your EV when you get home in the evening
Do you think Americans could be persuaded to own 2 cars?
I know it seems like a massive waste of resources and you would need 2 car garages to park them in front of (the actual garage being full of junk) but perhaps with some sort of subtle marketing pressure then even Texans could be persuaded to give up their hair-shirted minimalism
"Do you think Americans could be persuaded to own 2 cars?"
My dad did. Driving the truck he needed for the ranch to work each day was costing a fortune and wearing the truck out so he bought a cheap econobox for commuting and actually saved money each month .The same would apply to anybody with a 9-5 job and a need to have a something large such as a full size pickup.
Brits that have a garage often park outside them so they can have a workout room or art studio they never use.
Disclosure: I have too much crap to fit my car in the garage. If I could sell off the last of the manufacturing equipment I have left from a business I closed, there'd be room. It's a goal that I've been working towards. I work half the week in the field and at my home office the other half and getting the car out of the summer sun would preserve its life.
All this talk of needing two garages - a lot of people don't even have a driveway, much less the luxury of a garage. Home charging of an EV can't be factored in, much less worrying about the source of the power to do so ;)
This a very major factor that the most ardent fans of EVs repeatedly gloss over.
And lo and behold we have a downvote by one of the blinkered EV fans living in their own little utopian bubble who <still> can't open their eyes and see the real world problems.
Or perhaps it is wrong to point out that a large part of the population don't have a garage/driveway, often can't park a vehicle anywhere near their house, don't yet have local access to an economically viable charging point (and those people are more often the ones who simply can't afford to buy/lease an EV in the first place).
Do some people think that there aren't people in this situation, or is it (once again) a case of they are 'nobody worth bothering about'?
"All this talk of needing two garages - a lot of people don't even have a driveway, much less the luxury of a garage."
I see that as an issue with city planning. In the US, there are drives and garages since at least the left half was mainly settled around the age of the automobile. In the UK/Europe, cities grew up much earlier and people relied on shank's mares (their own two feet) or horses if they were posh. As the horse era waned, the coaching yards got converted to flats and business, not car parks so not only were there fewer and fewer places to park a horse and carriage, there where no places to park a car. The same thing goes on today with many homes being built with only a gravel drive that can accommodate one car at most unless the building/developer can get away with none at all. Even if you can get by with no car yourself, having no parking makes it hard to have trades stop in to do work or have visitors. I've lived in areas that were mainly small flats and there was zero on-street parking for nearly a mile as only one parking space would be allocated to each flat even though they were occupied by a couple or a couple/three roommates each with their own car. There's no out for planning commissions to keep ignoring the 800 kilo steel gorilla in the room. People own cars. If that isn't taken into account, there will be issues. If you choose to live somewhere with no off-street parking, you have to deal with it and part of that will mean that an EV isn't likely a good choice if having a car at all is still possible.
The really big difficulties come with older housing within towns and cities - terraces which often front on to the pavement where on street parking is the only option, or semis similarly close to the road and with insufficient space between them for any kind of vehicular access. Much of this housing stock was built in the days before cars existed, and a very significant portion of the rest dates from a time (pre the mid 1970s) when a great many 'ordinary' people didn't own a car, so provision for a driveway was not even thought about.
These are the type of scenarios that I always think of where there is very often very little chance of being able to park outside your own house, and therefore no realistic opportunity for cheap rate home charging (compounded by the issue of running a cable across the pavement in the first place (there was one case publicised a few weeks ago where one EV owner in Oldham can't charge his car outside his house because the local council had forbidden him to trail a lead across the footway on safety grounds). Although possible solutions are being trialled in some places, there is no clear guidance as to whether any are legally permissable, or who foots the liability for any injury to pedestrians from them being a trip hazard or the risk from a cable getting damaged.
Nationally the UK is estimated to have an average 40% of homes without any kind of driveway (about 20% in rural areas, but rising to around 60% in some towns and cities, but still some people keep their heads buried in the sand and downvote anyone who dares to raise any of the difficulties that many face which challenge the possibility of EV ownership.
" (compounded by the issue of running a cable across the pavement in the first place (there was one case publicised a few weeks ago where one EV owner in Oldham can't charge his car outside his house because the local council had forbidden him to trail a lead across the footway on safety grounds)"
That's been sorted. There are cable troughs that can be installed to be able to run a lead without creating a tripping hazard. It still means you need to be able to park in front of you own home or maybe make a deal with neighbors that have done the same thing in front of their house for another chance at being able to charge up. Most modern EV's have enough range that they don't need to be recharged every day for average daily driving. My petrol car has about 25% more range than an EV with a full tank, but during the holidays, it's been a couple of weeks since I've topped up and still have half a tank. I'll probably be pretty busy right after the new year since holidays usually mean things just stack up, not go away.
<......."There are cable troughs that can be installed to be able to run a lead without creating a tripping hazard.".......>
Yes, that is one of the things I had noted that was being trialled, but until clarification is forthcoming about the legality of installing such things, most local councils seem to be taking the view that cross pavement charging isn't permissable.
Good point! Katarina Zimmer's got a perty good overview of how to tackle renewable energy’s sticky storage problem over at this past Thursday's Knowable Magazine ... I like her flywheels (eg. from Amber Kinetics).
"Katarina Zimmer's got a perty good overview of how to tackle renewable energy’s sticky storage problem over at this past Thursday's Knowable Magazine"
The discussions usually weigh heavily towards chemical batteries, the insufficient quantities of certain elements (that nobody has looked for in quantity previously), the effects of mining and that sort of thing. Using water based gravity storage takes rather particular geography... and water. I favor gravitational storage built on the side of a high hill or mountain. Crane based systems seem to have too many issues that cost money to deal with (safety).
Hydro is gravity storage with the sun causing a repositioning of water from lower altitudes to higher where it flows down hill and the energy from that can be tapped. At this point, humans have dammed up the most optimum points. What's left are places with all sorts of problems such as silting that aren't easily overcome.
Flywheels are interesting but how good are they? Big ones can store up a tremendous amount of energy, but what happens if something goes wrong? I can imagine a bearing going out slagging the whole installation if it can't be drained quickly enough. A mass on an inclined plane can "ratchet" up the hill so if a cable fails, the mass is mechanically prevented from careening down the tracks to a disaster video death. Repair would be a matter of replacing the cable. One implementation I saw replaced a motor/generator at one end with those gubbins being on the sled and power fed to and from via contacts like a metro. That seemed to me to be harder to service in the case of a fault. It does allow for the sled to be more of a train that can be run around corners/turns rather than requiring the track to be straight.
For some reason all this makes me think of is the hype among certain utopists and scifi writers about the oncoming Singularity.
Where even then I always thought "where is the energy going to come from?"
Here we are with AI needing what is likely a fraction of that required for us all to enter the post-need magic candyland society, and it's already bringing our grids the point where I can absolutely see things like black-outs in urban areas during heatwaves, killing thousands, so the techbros' precious AI clusters continue to run.
Obviously they're going to stick all the recently-unemployed into dream machines that somehow power the AI.
Oh, yes, that was the nonsensical part of the plot.
I recall someone pointing out that one prominent research group's singularity timeline claim involved increasing total world energy usage by at least 1000 times in under a decade, and making all of that electrical, too.
There's a huge amount of magical thinking in AI land, and not just the hallucinations and jabberwockies.
Where even then I always thought "where is the energy going to come from?""
It comes from an outlet in the wall. Duh.
The Techbros seem to think so (on a larger scale). They order up a fat set of wires from the leccy company and that's all of the thought they've done on that issue. Keeping that set of lines hot is the power company's problem. (a nod to Douglas for the SEP plot line)
This is why I'm trying to kit out my home for the powerpocolypse™. Actually, I'm working on getting important things off grid. Not trying to go completely off-grid as that has no return if there is a grid option. If the power lines are melted, my HVAC will work, the fridge/freezer keeps going and I'll have the lights on as well. Water is a bigger issue as the city has wells to supply everybody and with no electricity, I'm not sure what they've done to keep pressure up. Diesel generators will only run so long, provided they've been properly maintained.
"Where I am it's all or nothing: you want to go off grid? You will be cut off from the leccy company."
Where I am, they city will condemn the house as unfit for occupancy if it doesn't have grid power and that's not unusual. I've seen many stories about that from all over the US. The same goes for water. You must have service if it's in the street. Whether you use any of that is up to you, but it must be on to live in the house. The minimum charge for electricity isn't very much where I live. After I bought the house I had to do a bunch of repairs before I moved in and used very little power for a couple of months. I think the bill was around $15/month. Water is more expensive. My bill is $40/month even if I use nothing (300ft3 allowance included).
If I could get to over unity on electricity, I'd still keep grid power as a standby. It wouldn't make sense to build a system big enough run a welder I want to get since I wouldn't need the capability often enough to justify more panels, a bigger inverter(s), more battery, etc.
An investment is something that is intended to create a profit or some other benefit. I've yet to see either associated with AI, and the clearly unfeasible amounts of money being spent cannot possibly earn a return. Taking Statista numbers as a reasonable approximation, the world has cumulatively invested $600bn in AI so far, and for what? Some pretty poor quality "adult content", a lot of LLM parroting and hallucinating, and...well, you tell me?
AI is the age old tale of the Emperor's New Clothes, driven by the fact that the world's decision makers suffer from the joint and collective failures of not understanding the limited use cases and a severe dose of FOMO.
"investments that create jobs and spread wealth around?"
Approx $200 BILLION has been spent on this bollocks.
As for jobs....do you mean the 10's of thousands of people laid off by EXTREMELY profitable tech companies as they pivot to this?
Please, show how this "investment" is helping the non-existent trickle down economy.
Yeah, through the quack frog soup that are so-called AI scaling laws, they're basically trying to sell investors the snake oil that an African elephant, with its 257 billion neurons, can make art better than a human that has just 86 billion! Or that an Orca's 42 billion forebrain neurons make it better at calculus than a human's 21 billion. And possibly that an octopus' nine brains, or a leech's 32 brains, provide them with the cognitive wherewithal to ace pathological diagnoses of challenging medical cases published in the New England Journal of Medicine clinicopathologic conferences, all through some Chained off Tentacular Thought wizardry (CoTT)!
I'd say they suffer from a highly contagious form of Clever Hans syndrome if you ask me ...
It would be better money spent on the real thing by those without it! Playing the Lotto will give them a better return if one believes in dreams*
Artificial is the real word! It's a dream, is artificial, just as I am an Artificial Multi millionaire.
The hardware sales people must be laughing all the way to the bank, as are their bosses.
This is why Musk needs to get behind nuclear fusion rather than claim solar can cover our future needs. AI will have unlimited energy demands. What we really need is an infinite supply of safe, clean energy.
Sorry if I don’t speak in the sarcastic tone that’s expected in these comments
The problem with development of practical fusion power is that it requires massive breakthroughs in technology, and possibly even in basic physics. Our current state of knowledge can't even guarantee that production-scale fusion is possible at all - i.e. that fusion can be 'tamed' (to provide more output than input) in an Earth-bound setting. A far better investment would be to improve our collection of the vast output of the fusion reactor we've already got - the Sun.
True, a lot of good work has been done on fusion, and continues to be done. But I doubt that anyone working in this field would give you even money that we'll have fusion power plants operating within the next thirty years. Meanwhile, we are bombarded by fossil-fuel-industry hype of the form "Don't worry about the climate - fusion power will save us." Which is a pipe dream about as realistic as "Don't worry, we can all move to Mars."
I will answer that question. 2.2 GW generator gas turbines are probably the quickest and cheapest short term solution for that much power generation.They are not going to wait 10 years or more for their own nuclear plant. Profit talks louder than good intentions and hype.
"I will answer that question. 2.2 GW generator gas turbines are probably the quickest and cheapest short term solution for that much power generation"
You might also be able to buy one off of Amazon although I doubt delivery will be included. The point is they are quasi off-the-shelf tech you can just order where all of the engineering is done, there are lots of examples of them installed and working and it means that planning and permits have been worked out as well saving s-tons of money and time.
This post has been deleted by its author
These companies building these AI centres have all missed something: knowledge itself is limited. You can create knowledge from knowledge - but there's a limit to how much you can do that. Right now, AI is improving rapidly - but its progress will slow down dramatically as it starts to approach the limit of knowledge. By my calculations, we are just 20-30 years from peak AI, and when we get there, we won't be anywhere near having an AI that's clever enough to take over the world through superior intelligence. Of course, we might wish to voluntarily hand control over to AI because people will consider it preferable to being ruled by fellow humans.
Try one year. After that it’s self-referential rabbit-holes.
Life's too short to read through that wall of text in your link. The sub heading with 1st-year O-level-style question mark told me all I need to know about the 'piece'. Plus it is old. June this year.
Because it is Xmas morning and I am waiting for Santa, I tried quick reading again but "The ticking clock on high-quality data has forced AI developers to think more creatively" made me queasy. We need an Alan Partridge of IT.
Try to think past training data. Imagine we have enough (since ChatGPT3.5) curated data for the AI. And what the article misses is that a lot of interlinked data gets created in abstract form from that curated data. Indeed, that is the appeal of AI, the interlinking and pattern finding we miss.
Thinking beyond data into abstraction of data which is done. Pretty much. The process of abstracting the formulas of stored data is mostly finished. We don't need the data. We can recreate it.
And then beyond, the constant stream of real-time data. Imagine OpenAI releasing a new model every say ... 0.10 seconds. But still you grumpy Grinch's will find another thing for us to fix that you hate until you love us more than your Mummy.
Current general AI (as distinct from specialised AI) has its uses if its severe limitations are understood by the user, but most of it is totally overblown. I mean I'd never use it for anything factual, but it can produce a fun story line for entertainment or an enjoyable initial brainstorm. It's a great shame it's being relentlessly pushed out into areas of factuality where it is currently inappropriate.
But it might get somewhere in the end, especially if fact and imagination are properly seggregated instead of being fused together...
As pointed out in earlier comments, it's simply impossible to build nuclear capacity quickly enough to serve the AI bubble.
AI companies probably continue promising nuclear, solar, wind, tidal, geothermal or whatever else sounds good - but will ultimately build gas-fired, because it's quickest and most convenient. (Then they'll use their AI to concoct explanations of why it was impossible to save our environment.)
"Maybe it's the compute capacity. "
If it's general purpose computing hardware, maybe. There are lots of crypto-mining barns that are completely abandoned as the hardware has zero value very quickly and isn't even that good as scrap due to the required amount of manual separation it will take. All it took was a dip in 'faith' for the operator to dive firmly into the red and shut down. To make use of the building will take time/energy to clear out all the stuff.
The irony of this Musk guy, telling us we need to give up internal combustion engines and buy electric cars instead, now building a data center which uses enough natural gas generator to power a small African nation, is simply incomprehensible. Is no one capable of rational thought anymore? What the genuine heck? I am seriously tempted to give up on this 21st century entirely and converting to Amish principle.
My take:
Recently introduced OpenAI o3 offers PhD level capabilities. Chat GPT Pro is offering access to PhD level research level intelligence for $200 p.m.
Current PhD UK salaries average £45k, much higher in the USA.
Imagine hiring the equivalent of 18 AI research assistants for the price of one UK PhD salary… (as an aside, historically, the first jobs to go used to be the guy sweeping the warehouse floor…). The societal implications are horrendous.
At these costs, and they will get cheaper, one can only imagine the possible breakthroughs in Maths, Physics, Biology etc and the probable income derived from successful research, patents and product.
Ray Kurzweil, (if I remember correctly, in Bostrom’s “Superintelligence”, whilst most AI experts median guess was AGI might arrive by 2050, and some estimated 2100, Kurzweil’s estimate was 2030) who’s background I don’t imagine needs much of a description here. Currently serving as Principal Researcher and AI Visionary for Google, reckons that widely available by 2045 (too late for me unfortunately)will be a brain interface, possibly similar to Musk’s Neuralink, giving almost instant access to AI support. If it was anyone but Kurzweil I would think the suggestion ridiculous.
Bearing in mind Musk’s genius in using his rockets to launch his own satellites. Add to that the symbiotic nature of Tesla vehicles+Solar Panel roofs+Powerwall battery. I imagine that Musk is fully aware of the possibilities for an AI brain interface.
The future of AI generated video to me looks very promising. Sure, at the moment the quality generally is not up to scratch. I can’t help but imagine a time when a full AI produced production outshines its major studio competitor at a fraction of the cost. (I’m already in love with the ladies on here: https://x.com/BrianRoemmele/status/1851348471703544182?mx=2 )
I’m reminded of how easy it is now for an independent film maker, equipped with decent technology, can present a full finished exceptional product. You only have to take a look many Youtube channels to see the sheer professionalism on offer.
When you look at some the companies mentioned in the article you do have to consider their sizes, budgets and overall wealth. A lot of firms will go bust in the process. The winners will make billions is my guess.
"The companies that will spend $50000 million to be the first or the companies that will spend $10 million to be second?"
I'll bet on a company that has a rational service derived with the technology and can earn a consistent net profit. It's the same as Google Glass in that it's interesting, but what can you really do with it that people will continue to pay for? I really liked the VR system in the book "Airframe" by Micheal Crichton. If I could pull up schematics and test gear readouts while working on something and not have to put probes down or look away, that would be awesome. If I could say "Schematic, R104" and I get a view of that part of the schematic I can zoom in and out of, that would save loads of time. Way back in the bronze age when I worked on electronic music gear, even though schematics were organized in filing cabinets, it took time to get them out and put them away. I might work on 5-6 pieces of gear each day so it could be a lot of back and forth. If I needed to make a copy so I could doodle while tracking down an issue, I'd have to go up to the office and wait for the copier to warm up. To have all of those cabinets available via voice prompts and a way to make notes (and save them) would be an awesome application. Computer generated porn and cat videos, not so much.
What the hell are they going to use to train all these new monster LLMs?
There's already a significant percentage of 'knowledge' on the Internet that has been churned out by the squawking birds.
Even if the current methods could lead to AGI, there's not enough good data upon which to draw, even if using copyrighted material that should be firmly off limits to these parasites. If they're allowed to squirt it back out into the source pool it'll all become worthless to them and the rest of us well before they've got these gas-guzzling things fully online.
It's an informational Ponzi scheme in a similar way to almost all cryptocurrencies being financial ones.
Now I don't even have to mess up Wikipedia articles myself. I can get my LLM to do it.
Microsoft reactivated TMI just to power one project.
1- Maybe that's one way to subvert German sanctions on Nuclear Power, if you are a private company hiring Framatome (or whoever) to build a NPP to you, and avoiding burning Russian gas in the process. Now they are burning one the nastier types of fuel derivatives of coal in Germany besides the aforementioned Russian import just to keep the lights on, because wind and solar can't possibly keep up. Now you could build a NPP to power your AI and sell the surplus power to the foot-shooting Germans that decided to kill their reliable non-air-polluting NPPs.
It is a win-win scenario to build NPPs for AI in Europe right now. You can build them in France if the Germans demand to freeze to death by their own stupidity.
Americans are not far behind, not having any new NPP designs ready to go.
2 -I could see Elon Musk building a couple of NPPs just to dump the excess power in his own Panasonic-designed batteries and to feed his own AI training for self-driving cars. He at least would have a legitimate purpose to power an AI project.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
