What if generative AI, but nucular?

March 12, 2024 5:38 PM

Tech firms and Silicon Valley billionaires have been pouring money into nuclear energy for years, pitching the sustainable power source as crucial to the green transition. Now they have another incentive to promote it: artificial intelligence.
posted by cupcakeninja (49 comments total)

Honestly, if AI is the excuse that gets us cheap, safe, and ubiquitous nuclear, that wouldn't be the worst outcome.
posted by Joakim Ziegler at 5:54 PM on March 12 [10 favorites]

Technically, Skynet irradiating the entire surface of the Earth qualifies as ubiquitous.
posted by notoriety public at 6:01 PM on March 12 [16 favorites]

We're not going to be so lucky as to get Skynet though. We'd get Marvin, and he'd hook himself up to the Internet and all of our technological infrastructure would just give up and crash.
posted by notoriety public at 6:03 PM on March 12 [7 favorites]

I guess the bills are coming in for the loss leader and they’re beginning to freak out.
posted by Artw at 6:05 PM on March 12 [7 favorites]

It's a better use of their money than this nonsense, at least.
posted by AdamCSnider at 6:13 PM on March 12 [1 favorite]

Sam Altman explains why he's helping to take nuclear microreactor company Oklo public.
posted by clavdivs at 6:15 PM on March 12

Is anyone making money using LLM-style AI? I see lots of cool demos and scams and spam. Dunno if I've seen any concrete business uses, though.

Seems a bit early in the hype/money cycle to start building nuclear reactors just for AI.
posted by ryanrs at 6:49 PM on March 12 [5 favorites]

>if AI is the excuse that gets us cheap, safe, and ubiquitous nuclear,

Is it really necessary to point out that techbros will manage nuclear power exactly the way the OceanGate clowns managed their submarine? The jackasses in the article are already crying about how those dumb regulations are preventing them from innovatin' like they need to. It's EXACTLY the same rap as Oceangate.

Technologists always want to tell you about how nuclear power can be made safe. But it can't be made safe by technologists because the problem isn't technological. We have created a civilization where absolutely nothing can be prioritized above the making of more money by people who are already rich. Even the question of whether we're going to keep destroying the biosphere has to be reframed as "how can we make it more profitable for rich people to STOP destroying the biosphere than it is for them to keep destroying it?"

Nuclear power plants in our civilization can only be constructed and run by people expecting to make a profit by doing it. Those people will always regard demands for regulation and transparency as nuisances which can fortunately be got around by paying fines, scratching backs, and greasing palms.

Sorry, but it's never going to be safe. Maybe a different civilization run by a different animal that doesn't happily murder and enslave its own brethren to get rich could do it, but we can't.
posted by Sing Or Swim at 6:50 PM on March 12 [55 favorites]

Is anyone making money using LLM-style AI? I see lots of cool demos and scams and spam. Dunno if I've seen any concrete business uses, though.

I sit beside an exec at the company who has been using whatever the Microsoft write-a-meeting-summary product is.

The other day it made up a whole paragraph from nothing about how the meeting participants had been talking about being drunk, and it editorialized that they might be struggling with their work because of their alcohol consumption. And then it sent the summary to the CEO.

You could argue that's a concrete business use.
posted by clawsoon at 7:21 PM on March 12 [36 favorites]

France has been safely generating the majority of its electricity from nuclear power since the 1970s.
posted by gwint at 7:23 PM on March 12 [12 favorites]

... France does that by throwing its best engineers and a money-no-object budget at nuclear power, not techbros.

Also, AI control of nuclear might go as well as Therac-25 did.
posted by scruss at 7:30 PM on March 12 [23 favorites]

Micronuclear just seems like a poor idea. You have all the safety and security concerns of nuclear, but you need to have hundreds or thousands of sites because you're only generating... *squints* 15MW? Are you shitting me? A full scale nuclear plant pumps out over a gigawatt. You're telling me it's more effective to build one thousand small scale nuclear plants than just to build another ordinary plant? That sounds insane.

The fusion investment at least sounds interesting, and I wish their great-grandchildren the best of luck in ironing out the final wrinkles.

or you could just build more renewable power generation and use less power, you absolute tool
posted by phooky at 7:41 PM on March 12 [29 favorites]

How many large datacenters can you power with 15MW? Less than one, right?
posted by ryanrs at 7:59 PM on March 12 [1 favorite]

... France does that by throwing its best engineers and a money-no-object budget at nuclear power, not techbros.

Not to mention sending French paratroopers to Niger when necessary to keep cheap uranium flowing.
posted by clawsoon at 8:04 PM on March 12 [10 favorites]

Even the question of whether we're going to keep destroying the biosphere has to be reframed as "how can we make it more profitable for rich people to STOP destroying the biosphere than it is for them to keep destroying it?"
[...]
Sorry, but it's never going to be safe. Maybe a different civilization run by a different animal that doesn't happily murder and enslave its own brethren to get rich could do it, but we can't.
Sing or Swim

sadly agree with this. businesses will cut whatever corners they can on the way to implementing this technology.

and thank you for linking that wikipedia entry gwint. there it mentions l'autorité de sûreté nucléaire as the independent agency designed to regulate france's nuclear safety. I don't know enough to say if the usa's nuclear regulatory commission is exactly identical in mission, but I know it (the nrc) has been criticized as too industry friendly. yet according to the nbcnews article, even the nrc denied oklo's "aurora" design application in 2022 for want of more safety data about the powerhouse.

quite honestly, I don't put it past any of these fuckers to cause some kind of costly disaster that requires more resources to clean up despite (the ounce of) prevention having a cheaper price tag in the long run. if the nrc is raising caution, I think it's a good idea to listen. the usa is most definitely not france in terms of holding businesses accountable for when they (inevitably) fuck up.
posted by neitherly at 8:06 PM on March 12 [4 favorites]

The thing about nuclear power is that we have not ever really had a worst-case nuclear accident. Even Chernobyl and Fukushima - without good luck, exceptional personal heroism and sacrifice, expensive and prompt disaster response - both could have been much worse.

But pretty much every day that a coal or gas power plant runs is about as bad for us all as they can be. Correct me if I'm wrong but it seems that the only way they could really be worse is if the fuel burned without generating power at all?

The price of solar power and storage continues to plummet, while nuclear gets more expensive. There's an observation that stuff made in factories tends to get cheaper, while stuff constructed in the world gets more expensive. I think this partly explains the drive for SMRs - instead of needing a custom design for every site, you just build them in a factory and ship them out to where they will used, hopefully enabling the same "learning" dynamics in the process.

SMRs potentially that could be used in many existing sites of coal and gas plants, swapping the heat of a furnace or burner with one or more reactors, utilizing the installed generation equipment, grid connections, etc.
posted by rustcrumb at 8:33 PM on March 12 [3 favorites]

Nuclear power, if something goes wrong, renders a significant area inhospitable to human life for hundreds of years.

Fossil fuel power, even if nothing goes wrong, renders the entire globe inhospitable to human life for many multiples of "long enough" to wipe us out.

And we are not going to get nuclear power right.
posted by notoriety public at 8:41 PM on March 12 [2 favorites]

US coal power plants killed at least 460,000 people in past 20 years

Luckily, as of 2023, more electricity is generated by nuclear than by coal in the US (20 years ago coal generated 3x more than nuclear. Total nuclear generation has essentially remained the same over that time. Natural gas and renewables have filled in the gap)
posted by gwint at 8:45 PM on March 12 [6 favorites]

Why do these tech people try SO hard to seem like cartoon supervillains?

Also, come on, they can't do marshmallows right. Don't trust them with nuclear power!
posted by mmoncur at 9:02 PM on March 12 [4 favorites]

The fact that neither governments nor private industry have been able to run nuclear plants without disasters should be warning enough.
posted by They sucked his brains out! at 10:07 PM on March 12 [4 favorites]

How many large datacenters can you power with 15MW? Less than one, right?

Depends on what you mean by "large". Calculate maybe 2000 watts per server, all included, and 15MW is about 7500 servers. That's not the largest data center in the world, but it's not small either. There are data centers now that house more than 100k servers, but I think that's still fairly uncommon. A small data center is maybe 1000 servers. So, a medium one?
posted by Joakim Ziegler at 11:28 PM on March 12

Micronuclear just seems like a poor idea. You have all the safety and security concerns of nuclear, but you need to have hundreds or thousands of sites because you're only generating... *squints* 15MW? Are you shitting me? A full scale nuclear plant pumps out over a gigawatt. You're telling me it's more effective to build one thousand small scale nuclear plants than just to build another ordinary plant? That sounds insane.

I don't know where you get "one thousand small scale nuclear plants" from. If you want a Gigawatt, you need 67 of these 15MW plants. I also assume the first one they build is the smallest they can build, so I'd expect the design to be able to scale to, say, 10 times that, at least. Naval reactors , probably the most efficient small reactors in current use, put out somewhere north of 150MW, and I see that the typical range cites for SMRs is 5-300MW.

There are significant advantages to being able to generate power closer to the users. 5-7% of the power generated in the US is lost to transmission loss, for one, but also if your goal isn't to generate power for domestic use, but for, say, industry, localized spot production makes a lot of sense, you can put the reactor on the factory grounds, and if you have power to spare, you can sell it out to the grid like is common with solar. I imagine aluminum smelters, for instance, would be quite interested in having a reactor on-site.

And regarding safety, one of the main ideas behind new smaller reactor technologies is much better safety, mainly passive safety, where, if you just stop doing anything to the reactor, and/or the control systems shut off, it just slowly turns itself off on its own. I don't know too much about the details of Oklo's design specifically, but there's a lot of research on thorium-fueled reactors, for instance, which both produce much less waste, and whose fuel is entirely unsuitable for nuclear weapons, and is also extremely abundant in nature.

So, in short, don't write off small nuclear. Almost every technology we've created has initially been large, unsafe, custom-built, required highly specialized maintenance and care, and posed hazards, and we've managed to get pretty much all of them down to small, near maintenance free, efficient units produced in factories. Yes, nuclear poses some unique hazards, but there's no real reason they can't be managed, with proper oversight.

We desperately need to get away from things like coal-fired power plants, which not only are hugely polluting, but also actually release much more radioactive contamination into the environment than nuclear power plants ever have, even including the large, well-known nuclear disasters.
posted by Joakim Ziegler at 11:46 PM on March 12 [12 favorites]

Radiation accidents are not actually the only or even statistically the most threatening scenario of nuclear power.

What seems to be often forgotten is that there is NO way of storing the waste safely. Maybe for a generation or two. But not for the thousands of years necessary. No cave system, pit, old coal mine, no man made facility will be leak proof for that long.

Hundreds of generations will have to care for that shit to keep it from contaminating the planet.

Nuclear accidents are a statistical risk that can be mitigated. Waste leakage is certain.
posted by uncle harold at 11:46 PM on March 12 [8 favorites]

What seems to be often forgotten is that there is NO way of storing the waste safely. Maybe for a generation or two. But not for the thousands of year necessary. No cave system, pit, old coal mine, no man made facility will be leak proof for that long.

Thorium cycle reactor nuclear waste only needs to be stored for a century or two, and only produce a few percent of the waste mass that current reactors produce. It's not ideal, but we should be able to manage that.

Also, nuclear waste is not a liquid or other material that can "leak" as such. It's converted into a glass-like ceramic material, so if, for instance, you bury it in "old coal mines" up to a kilometer under ground, and then fill the rest of that mine up with rock, concrete, etc. to seal it off, the only thing that's going to get it out of there is drilling for it and manually bringing it up. There's no natural process that'll bring it to the surface on a hundreds or thousands of years time scale.
posted by Joakim Ziegler at 11:53 PM on March 12 [8 favorites]

AI as it currently stands isn’t much reason to grow the power grid: numbers are all “leaks” because OpenAI once heard of transparency and nearly fainted from shock, but it’s all in line with what we’d expect from other, known models: GPT-4 required 25,000 A100s for 90~100 days, for a total training power consumption of ~50 GWHr. Once trained, inference runs 0.001 to 0.01 kWHr depending on complexity, vs a Google Search’s typical 0.0003 kWHr. So: between 3x and 30x with the often asserted but never formally sourced claim that, on average, like-for-like query complexity, it’s usually 15x.

Perspective: I’m almost always redlining my work PC or gaming PC to the tune of ~0.8 kWHr every hour, 16 hours a day. My carbon footprint is incredibly low otherwise (for the US) so that’s like 50% of my bill. Point is: my typical hour working remote is equivalent to 100~1000 ChatGPT-4 queries.

GPT-5 is now a month into its training run assuming the sudden flurry of “damn it feels good to be a gangsta kick off a huge training run” on Twitter by senior researchers was for that. Widespread rumor is 100k A100s for the same time period, total power consumed ~200 GWHr. This does not mean inference will likewise jump 4x: multi-modal is the big thing for this generation (text/images/video/speech all informing each other) so there’s a lot more to chew through in training, and they’re starting to see diminishing returns on increased parameter counts / training tokens, so line-goes-up = find entirely new methods before long (already in progress since GPT-3 was transformer-at-scale and GPT-4 went sparse mixture of experts).

Putting some context on the training numbers: every single year Bitcoin consumes 127 TWHr, or in equivalent units: 127,000 GWHr. OpenAI’s major training runs are a rounding error by comparison. Average annual US household consumption is 10,600 kWHr = 10.6 MWHr = 0.0106 GWHr.

GPT-4 training cost = 5,000 US households’ annual consumption.
GPT-5 training cost = 20,000 US households’ annual consumption.
Bitcoin continues to consume, year after year, as much power as 12,700,000 US households. JFC.

My actual concern is that the Q* proposal (assuming the OpenAI STaR paper actually is the method and I’m reading it right) consists of solving ChatGPT’s inability to follow chain of reasoning or common-sense sanity check its hallucinations by spooling up millions or even billions of small Reinforcement Learning models and continuously culling underperformers until a consensus is reached.

Point is: if things just continued as-is, we’re not in trouble. I’m not convinced this remains true for GPT-6 in 2026. I trust OpenAI to do the ugly, wasteful, brute-force thing if they think it gets them there first. Every goddamn time.

Calculate maybe 2000 watts per server, all included, and 15MW is about 7500 servers. That's not the largest data center in the world, but

Typical 8xA100 / H100 DGX or HGX draws 6.5~10 kW during training depending on load. GPT-5 is almost certainly training with 12,500 of these - or functional equivalent - for 90~100 days. No such thing as perfect efficiency but you can save a lot of waste if the datacenter is located next to the power station.
posted by Ryvar at 12:12 AM on March 13 [9 favorites]

Also, nuclear waste is not a liquid or other material that can "leak" as such. It's converted into a glass-like ceramic material

Not in the US. You have to do nuclear reprocessing to get something you can vitrify. In the US, we use dry cask storage, which means stacking the used fuel rods out in the back parking lot of the nuclear power plant (not a joke).
posted by ryanrs at 12:18 AM on March 13 [6 favorites]

Not in the US. You have to do nuclear reprocessing to get something you can vitrify. In the US, we use dry cask storage, which means stacking the used fuel rods out in the back parking lot of the nuclear power plant (not a joke).

The handling of spent nuclear fuel and misc nuclear waste (ion exchange filters, rags etc.) are essentially two separate nuclear waste industries. Vitrification is still done, just not on fuel.
posted by groda at 1:55 AM on March 13 [2 favorites]

Micronuclear just seems like a poor idea. You have all the safety and security concerns of nuclear, but you need to have hundreds or thousands of sites because you're only generating... *squints* 15MW? Are you shitting me? A full scale nuclear plant pumps out over a gigawatt. You're telling me it's more effective to build one thousand small scale nuclear plants than just to build another ordinary plant? That sounds insane.

The nuclear industry is trying to pivot to small reactors because nobody has been able to build a large nuclear plant that came in anywhere near the original time and budget for decades now. Too much knowledge has been lost and the complexity of large plant designs that meet current safety standards are beyond what the current industry can effectively build.

The hope is that you can get economies of scale by going smaller, constructing parts in centralised factories, and also dodge some safety regulations.

I guess it's possible, but the forecasts I've seen have small reactors actually being more expensive per gigawatt hour generated than current plants and there's no way the public will be any more willing to allow them to be built next to their suburb than they would be a new large plant.

Given the lack of public tolerance they'll probably end up being built next to existing nuclear plants anyway, in which case they're really just an expensive way to expand existing plant capacity that might be more likely to arrive on schedule if they can get regulators to ease up on the safety regulations.
posted by zymil at 2:41 AM on March 13 [4 favorites]

France has been safely generating the majority of its electricity from nuclear power since the 1970s.

And yet somehow all the new generation French built reactors come in massively over budget and very late.
posted by biffa at 5:18 AM on March 13 [2 favorites]

I am dubious of the long term stability of the social and logistical infrastrusture necessary to keep nuclear power viable.
posted by The Manwich Horror at 6:47 AM on March 13 [2 favorites]

I'm more worried about the idea of handing over nuclear material to idiot failsons in Silicon Valley whose business models are built atop the principle of literally breaking laws.

One can imagine that inconvenient safety and regulatory measures are going to get ignored once the VC runway is spent, or the billionaires otherwise need to cut costs.

Nevermind that AI technology is in its infancy and that there are practically no laws regulating its usage. There are no do-overs when nuclear technology fails.
posted by They sucked his brains out! at 7:12 AM on March 13 [7 favorites]

It is kind of ridiculous to praise the French power system as uniquely French and not relevant to an American context. The vast majority of France’s nuclear power plants are Westinghouse designs. France deployed the nuclear power system the US developed and then abandoned.

The US scaled up coal and natural gas instead.
posted by Headfullofair at 7:47 AM on March 13 [4 favorites]

Yeah there's a reason he's Gun Jesus and not Nuclear Reactor Jesus.
posted by GCU Sweet and Full of Grace at 8:22 AM on March 13 [1 favorite]

This smells like the crypto bros who claim they can run their bitcoin mines by capturing methane from animal waste. I suppose you could, but why not use the power you generate to run something people actually want and need?
posted by The Ardship of Cambry at 9:17 AM on March 13 [2 favorites]

They would basically just like us to continue to ignore their externalities, everything else is hit air.

Which is ironically the only kind of air we will have thanks to them.
posted by Artw at 9:24 AM on March 13 [2 favorites]

Positing that nuclear energy will never be safe for our civilization is specious without a comparative analysis of how safe it will be for that same civilization—with all its peccadilloes—NOT to use nuclear power.
posted by lastobelus at 11:54 AM on March 13 [3 favorites]

Yeah there's a reason he's Gun Jesus and not Nuclear Reactor Jesus.

Clearly the splitting of loaves and fishes is an allegory for fission. How else did he do all that stuff?
posted by credulous at 12:01 PM on March 13 [1 favorite]

→ Thorium cycle reactor nuclear waste ...

There are precisely 0 operating generation stations that use the thorium cycle. There's one operational research reactor in India, and one "almost there" in China. It's not part of any climate solution right now.
posted by scruss at 12:56 PM on March 13 [3 favorites]

There are precisely 0 operating generation stations that use the thorium cycle. There's one operational research reactor in India, and one "almost there" in China. It's not part of any climate solution right now.

There are also precisely 0 operating generation stations made by Oklo. I thought we were talking about future developments in nuclear power generation here. And yes, India was first because India has a shitton of Thorium natively, and they see it as essential to their energy generation needs in the future, which are of course non-trivial.
posted by Joakim Ziegler at 1:02 PM on March 13 [1 favorite]

I was under the impression that there is no such thing as a thorium reactor that isn’t also an excellent resource for making isotopes of various elements useful in nuclear weapons, but this isn’t a topic I know much about.
posted by Ryvar at 1:51 PM on March 13

That would make them no different than your bog standard (and actually operational) current nuclear fission reactors, which are very much encouraged as a key step towards nuclear weapons by those who want them.
posted by biffa at 2:49 PM on March 13

I think you can use a thorium reactor to make U-233, which could theoretically make a bomb, although no one has ever tried.

But a thorium reactor isn't going to create U-235 or Pu-239, which are the usual bomb-making isotopes.
posted by ryanrs at 4:26 PM on March 13 [2 favorites]

Thorium reactors are pie-in-the-sky, like fusion reactors. It's kind of pointless to mention them in the context of real-world energy production in 2024 for the very same reason.
posted by They sucked his brains out! at 5:16 PM on March 13

Correct, thorium reactors do not make any isotopes that we know can be used for nuclear weapons, it's one of the several advantages it has.

Thorium reactors are most certainly not "pie-in-the-sky, like fusion reactors". The reactor running right now in India produces 40MW thermal at full power, right now, not in the future. That's a usable, practical amount of energy for power generation, indeed, it's more than twice the power of the one Oklo is building as their first operational reactor.
posted by Joakim Ziegler at 5:26 PM on March 13 [1 favorite]

Also, India's Prototype Fast Breeder Reactor, is 500MW, and is finally, after many delays, scheduled to be commissioned in December this year. That reactor is capable of running both on Uranium-238 and Thorium, and it's unclear to me which they're starting out with, but point is, these reactors exist and can generate useful amounts of power.

Fusion, while potentially very cool, is perpetually 20 years away from that, as the saying goes.
posted by Joakim Ziegler at 5:31 PM on March 13

That reactor is capable of running both on Uranium-238 and Thorium

TBH that sounds purpose-built to make plutonium for weapon, ha ha.
posted by ryanrs at 5:35 PM on March 13

TBH that sounds purpose-built to make plutonium for weapon, ha ha.

It's not impossible that India wants a dual-use fast breeder reactor for use in their nuclear weapons program as well, no. But this is extra work in construction, you can build thorium-only reactors more easily, KAMINI, the 40MW thermal Indian thorium reactor operating today that I mentioned is thorium-only.

Here's an overview of India's three-stage nuclear programme. They're currently comissioning reactors in phase two, and phase three is moving all their power generation to thorium only (again, India has vast reserves, so their motivation is basically energy independence until the end of time). That's planned for 2050-2060, if all goes according to plan, although phase 2 was delayed quite a bit.
posted by Joakim Ziegler at 5:47 PM on March 13

9 out of 10 coders on the Internet agree: Thorium is the future of nuclear engineering.
posted by groda at 4:14 AM on March 14 [1 favorite]

You gotta laugh.
posted by They sucked his brains out! at 5:58 AM on March 14