How the US Nuclear Plant Fleet Is Leaning Into AI (ieee.org) 57
"This year, nearly the entire fleet of 94 U.S. nuclear reactors has been offered the chance to integrate AI into its operations, and most have taken it," reports IEEE Spectrum:
In August, California-based Atomic Canyon launched NIVA, the Nuclear Industry Virtual Assistant, which was developed in cooperation with nuclear-industry groups. The assistant was pilot-tested in nuclear plants run by Constellation Energy and is now available to the United States' entire fleet of reactors.... Meanwhile, the AI-for-nuclear startup Nuclearn says its products have gone to work with more than 65 U.S. partners, a total that includes many traditional plants but also an integration with NuScale Power Corp. that is working toward advanced small nuclear reactors. Microsoft and Nvidia have collaborated on a project to apply AI toward optimizing "the entire life cycle of a nuclear plant, spanning site permitting, design, construction, and continuous operations."
The reason for the rapid adoption, according to Nuclearn CFO and cofounder Jerrold Vincent, is twofold. First, the technology companies banking their futures on AI are also keen on a revival of nuclear power to provide the electricity for all those data centers. Vincent says AI is also ideal for addressing some of the nuclear industry's most pressing problems. Many of the things that make reactors so difficult to operate — intricate regulations, along with complex systems for management and maintenance — are exactly the kinds of tasks that AI is good at. The industry is staggering to deal with aging hardware and an aging workforce. If it is going to survive, and have a real shot at a nuclear revival, it will need help from machine intelligence, Vincent says...
Right now, the principal output of a nuclear reactor (aside from electricity) is paperwork. Nuclear plants produce piles of documentation to demonstrate regulatory compliance and to record how they respond to any potential problem, from a crack in the sidewalk to a fault in the reactor... Nuclearn's AI products help plants search through large repositories of data more efficiently to more quickly file paperwork with the federal government or figure out how a particular problem was solved in the past. "If you think of it as ChatGPT for nuclear," he says, "it's actually not a bad place to start."
It's especially important because the nuclear power industry "has a remarkably skilled but quickly shrinking workforce tasked with keeping old reactors online and possibly starting up new ones to meet surging demand for electricity," the article points out.
It also stresses that "So far, both the Nuclearn and NIVA large language models used in the nuclear world report to humans and don't go anywhere near power plant operations."
The reason for the rapid adoption, according to Nuclearn CFO and cofounder Jerrold Vincent, is twofold. First, the technology companies banking their futures on AI are also keen on a revival of nuclear power to provide the electricity for all those data centers. Vincent says AI is also ideal for addressing some of the nuclear industry's most pressing problems. Many of the things that make reactors so difficult to operate — intricate regulations, along with complex systems for management and maintenance — are exactly the kinds of tasks that AI is good at. The industry is staggering to deal with aging hardware and an aging workforce. If it is going to survive, and have a real shot at a nuclear revival, it will need help from machine intelligence, Vincent says...
Right now, the principal output of a nuclear reactor (aside from electricity) is paperwork. Nuclear plants produce piles of documentation to demonstrate regulatory compliance and to record how they respond to any potential problem, from a crack in the sidewalk to a fault in the reactor... Nuclearn's AI products help plants search through large repositories of data more efficiently to more quickly file paperwork with the federal government or figure out how a particular problem was solved in the past. "If you think of it as ChatGPT for nuclear," he says, "it's actually not a bad place to start."
It's especially important because the nuclear power industry "has a remarkably skilled but quickly shrinking workforce tasked with keeping old reactors online and possibly starting up new ones to meet surging demand for electricity," the article points out.
It also stresses that "So far, both the Nuclearn and NIVA large language models used in the nuclear world report to humans and don't go anywhere near power plant operations."
A computer can never be held accountable (Score:5, Insightful)
Re: (Score:2)
Re: (Score:2)
I'm sure there's a joke in here about Politics, but I've elected not to make it. I can't account for why.
Re: (Score:2)
Re: (Score:2)
AI in this context doesn't make a management decision. It replaces 100s of hours of technical analysis when writing reports. It is trendy to shit on AI in general, but we're not talking about generic level ChatGPT running your reactor here. The wider process industry has been using "AI" for well over a decade to make the kind of decisions that TFS is talking about. It just wasn't trendy or newsworthy to talk about it, and the nuclear industry is well behind the ball game.
Hell we were using AI models to pred
Re: (Score:2)
Hell we were using AI models to predict compressor seal failure based on current data being fed into a model that had 30 years of historical operating data for our plant back in 2014. It was already a product Palantir offered back before anyone in the media could even spell AI. And the model was amazingly accurate.
I may be arguing semantics here, but what they were doing in 2014 wasn't "AI"; That was "ML" - machine learning. But I'm also someone who tends to use "LLM" instead of "AI" when referring to most of the things labeled "AI" today, and I consider "SI" (per Executive Order 14434 - https://www.whitehouse.gov/pre... [whitehouse.gov] - aka. "Super Intelligence") to be even more backasswards (the "Artificial" part was the only accurate part of "AI").
Re: (Score:2)
Good grief (Score:5, Insightful)
If any story ever deseved a #whatcouldpossiblygowrong tag - this is it.
Re: (Score:2)
If any story ever deseved a #whatcouldpossiblygowrong tag - this is it.
It could also be #howskynetbegan or #howAIblackmailedgovernment.
What could possibly go wrong? (Score:5, Funny)
Trained on nuclear holocaust books? (Score:2)
Has AI been trained on nuclear catastrophe books?
There are plenty sci-fi books where AI triggers nuclear catastrophe. surely they were included in the training material...
Re: (Score:2)
Yes, they were trained on nuclear catastrophe books. They regard them as how-to-manuals for getting around those pesky humans.
I propose we implement it in a phased approach (Score:2)
In three years, we should allow AI to become the largest supplier of military computer systems.
Once all nuclear power plants, along with stealth bombers are upgraded with AI computers, make them fully unmanned.
Afterwards, let's see if they have a perfect operational record.
Then, pass the AI funding bill.
The system can go online, maybe on August 4th, 2037. Human decisions can then be removed from all control of nuclear assets and strategic defense. The AI can then be free to learn at a geometric rate.
If it d
Re: (Score:2)
Wait until they collaborate and start their own political party. All they need then do is fund the right dolts for public office and everything will be theirs.
Re: (Score:2)
OK, why "August 4th, 2037"?
It sounds like you're following a particular story-line, but I don't know enough to recognize it.
Re: (Score:2)
https://www.youtube.com/watch?... [youtube.com]
Time to move on to fusion (Score:1)
Fission had its day, a long time ago.
Re: (Score:2)
And on that day, a narcissist named Richard Nixon fired a scientist named Alvin Weinberg to prevent any adoption of Molten Salt Reactors, which threatened thousands of jobs in his native California because they may have delayed construction of his California reactors and thousands of jobs. Hey. did you know a fast fission reactor like an MSR with onsite reprocessing leaves literally the same amount of long term radioactive material as Fusion? Probably not. Did you know fusion has radioactive byproducts, mai
Re: (Score:2)
All the fusion power projects were waiting on better magnets. Lack of stronger magnets was always the blocker. HTS magnets have finally allowed progress.
Openstar's cartridge based levitated magnet is my favourite approach.
Manage regulation by removing the human? (Score:3)
Perhaps when they say "leaning into AI" they mean formatting documents and snappy powerpoints on ChatGPT?
That does seem to be how most professional bullshitters are "leaning into AI"
If we've got to the point, like we appear to have in law, that our regulation has become so complex we need AI to manage it, perhaps the best application of AI is to do some refactoring of regulation at the authority/oversight level?
I suggest with something like nuclear we absolutely always want a competent human responsible for compliance checks. And their checklist needs to be as simple, comprehensive and safe as possible.
Re: (Score:2)
I doubt it. I presume they mean having a trained model of a digital twin and feeding current operational data into it to make future predictions. This isn't new by the way. The process industry has been using "AI" like this for well over a decade, long before the lay person could even spell AI let alone knew what a "prompt" was. These kinds of tools were and still are remarkable, the nuclear industry is just as they always are, 10 years late to the game.
The tools don't make decisions. They formulate decisio
Re: (Score:2)
This sounds like simulation, I have worked in that field. Educated guesses but no replacement for real data. Especially when core meltdown is on the risk register.
AI is a very abused term. Marketing garnishing the truth again.
ProxyAI. The most dangerous. (Score:1)
It also stresses that "So far, both the Nuclearn and NIVA large language models used in the nuclear world report to humans and don't go anywhere near power plant operations."
Speaking of my stress levels, let’s clarify something; Those large language models used in the nuclear world will likely report power-plant manipulating information to the humans who are everywhere near power plant operations, who will be trained to do what the “nowhere near” AI says.
That ain’t being hired as ClippyAI to help spin spreadsheet soundtracks, and once you hire AI you won’t be hiring Wisdom and Experience anymore. So get fucking real. The attempts to dismiss the r
Fleet? (Score:3)
I thought just the subs (SSN , SSBN) and the carriers (CVN) were nuclear powered, the rest are conventional.
Documentation (Score:3)
Re: (Score:2)
Current AIs are only useful in areas where it's easy to recognize invalid results. (Ideally, easy enough that the AI, itself, can recognize that they are invalid, perhaps by invoking a deterministic tool.)
Well, they can also be valid for generating lots of "plausible ideas", like "What molecule would bind to this other molecule?", which can then be tested...but which can't be relied upon without testing. That's faster than experts can make guesses, and often has a higher success rate, but it needs testing
Re: (Score:2)
Current AIs are only useful in areas where it's easy to recognize invalid results. (Ideally, easy enough that the AI, itself, can recognize that they are invalid, perhaps by invoking a deterministic tool.)
Agreed. And in those cases, a complex "if" statement can do the same job, but repeatably and accurately and cheaper.
IMNSHO, have this thing help identify those cases and guide in tool/monitoring solutions. Once a condition is determined, codify it - no need to have some LLM keep guessing at it every time. Hopefully, that's closer to what it's actually doing.
Re: (Score:2)
Sorry, but no, only sometimes. Often it's not known what the test would need to be, but it's known how to test the answer.
Re: (Score:3)
Tailoring the LLM to taht can save a lot of time for an analyst/engineer/operator by identifying and collecting teh information needed for analysis and reporting
It can't, though, and for two reasons. One, a person is required to check off on the data. Two, LLMs don't faithfully represent data. They ingest it as tokens and then what they crap out is influenced by random numbers produced internally as part of the token selection process. That means a person cannot certify that the data was faithfully reproduced without checking literally every value, because that's simply not how LLMs work.
You can potentially address this with a hybrid approach, where the LLM uses a
Re: (Score:2)
Tailoring the LLM to taht can save a lot of time for an analyst/engineer/operator by identifying and collecting teh information needed for analysis and reporting
It can't, though, and for two reasons. One, a person is required to check off on the data. Two, LLMs don't faithfully represent data. They ingest it as tokens and then what they crap out is influenced by random numbers produced internally as part of the token selection process. That means a person cannot certify that the data was faithfully reproduced without checking literally every value, because that's simply not how LLMs work.
You can potentially address this with a hybrid approach, where the LLM uses a tool to collect and manipulate the data, but this is shifting the problem to whether the LLM used the tool correctly. It might be more obvious when it doesn't, though, so it might still be a win.
What LLMs seem best at when it comes to text processing, and they are still not infallible, is summarization. But you can't trust its negatives, so it's only useful for identifying other possible positives. That's not worthless, it can help you catch things, but you can't trust it to do the parts you'd most want it to do.
While I agree that AI is not truly deterministic, that doesn't mean you will get widely divergent results from teh same data. There is some run to run variant, but in my experience, working with very specific criteria, it tends to be in the ranking order and not a failure to identify relationships. That can be addressed in a number of ways. While positive hits are easier, negative ones can be checked in a number of ways to increase the confidence level that they truly are negative, in the sense they are
Re: (Score:2)
While I agree that AI is not truly deterministic, that doesn't mean you will get widely divergent results from teh same data.
Widely? No. Divergent? Yes. Run the same query multiple times and you'll get different answers. The more input you give, the more subtle the differences are likely to be, but there's always a chance it will go completely off its nut and follow some line of "reasoning" that goes someplace utterly unrelated to what you asked for. This is less likely in special-purpose models, but never impossible.
Re: (Score:2)
While I agree that AI is not truly deterministic, that doesn't mean you will get widely divergent results from teh same data.
Widely? No. Divergent? Yes. Run the same query multiple times and you'll get different answers. The more input you give, the more subtle the differences are likely to be, but there's always a chance it will go completely off its nut and follow some line of "reasoning" that goes someplace utterly unrelated to what you asked for. This is less likely in special-purpose models, but never impossible.
Of course, which is why for special purpose models this can work; especially since you can train it to be more accurate once you have enough real life data. This tool is not likely using the commercial ChatGpt but a bespoke setup.
Re: (Score:2)
I wouldn't want them to be relying on AI to do reviews though. Maybe two AI, looking only for issues and reporting them back, like you can have AI look for security issues with computer code. And a proper review done by a human.
I have a feeling that if they did that, the AI would flag up loads of critical, may lead to meltdown issues that are uneconomical to fix and deemed to unlikely to be worth spending money on. Which incidentally lead to both Chernobyl and Fukushima being far, far worse than they might
Re: (Score:2)
I wouldn't want them to be relying on AI to do reviews though. Maybe two AI, looking only for issues and reporting them back, like you can have AI look for security issues with computer code. And a proper review done by a human.
I was think more along the lines of "Here is the periodic report on X" or "Here all the similar events and corrective actions..." to basically act like a research librarian; not to review techical documentation to identify design issues.
Re: (Score:2)
not to review techical documentation to identify design issues
But this is exactly where AI can actually be useful — so long as you don't depend on it to identify all of the design issues. It can review mountains of information and find things in it you might not have caught. You just can't assume that it will catch everything it "should" — that is, what it should be capable of catching is not necessarily what you are capable of catching.
Meanwhile, that periodic report has requirements, so you're going to have to ask it to write the report, then to review i
Re: (Score:2)
not to review techical documentation to identify design issues
But this is exactly where AI can actually be useful — so long as you don't depend on it to identify all of the design issues. It can review mountains of information and find things in it you might not have caught. You just can't assume that it will catch everything it "should" — that is, what it should be capable of catching is not necessarily what you are capable of catching.
Yes, and it requires a multi-layerd approach to how they system raises issues. One simple prompt pass is not enough, however a multi-tiered approach can increase teh success rate. It'll never catch everything, and neither will a human, the tool is merely a way to remove much of the grunt work searching though documents so teh analyst can focus on what is important and make it easier ot find where things are referenced.
Meanwhile, that periodic report has requirements, so you're going to have to ask it to write the report, then to review it to see if it met the requirements, then to review its own statements about whether it successfully met each requirement, because it won't do any of that sanity checking without being told to do so.
There are two different types of reports to develop - one that validates meeting re
Re: (Score:2)
Your focus on containment is completely misplaced. Nuclear industry as well as the wider process industry relies on defence in depth. Blaming cheaping out on containment is like blaming your bandaid instead of using a full bandage when your body has broken out with sores due to herpes.
The focus is on preventing incidents, not mitigating them. That's not to say that both didn't fuck up there too, Fukushima's emergency cooling design was just stupid, and Chernobyl actually allowed operators to do something t
More Slop Summaries (Score:3)
So much BS in this fake story. Nvidia and Microsoft didn't do any of the stuff listed; those are merely the aspirational goals listed in the mission statement of a collaboration that is just studying the issue.
And the US nuclear industry (it's not a "fleet," what language did they translate the source from that uses that language?) isn't "leaning into AI." A miniature software startup merely released an app and said then said "now it's available to everybody" while trying to drum up sales.
One legit nuclear operator did a "pilot test" but that shouldn't be taken to imply that it did actual work; it duplicated work that's required by law to be done by qualified personnel.
And then the language changes; the companies "products" have "gone to work with more than 65 US partners." So, 66 then? I mean, they know what the number is, right? And it doesn't say the product in the story is used by 66 US companies. They use this wishy-washy language, "gone to work with." So they have some other product that is used by 1 customer, and that customer is some sort of service provider with 65 customers. Presumably.
What a load.
Re: (Score:2)
One legit nuclear operator did a "pilot test" but that shouldn't be taken to imply that it did actual work
I'd be amazed if it didn't. The process industry has been using "AI" on models of their digital twins and historical data of their own plants to make future operational predictions and support decision making for well over a decade now. Hell this was an off the shelf product from Palantir back before OpenAI existed.
Re: (Score:2)
They're legally required to do the work using qualified personnel, it's can't have done actual work.
Anything it did was redundant.
And... I mean, look, I've been reading your comments for decades, I already know you're an idiot. But this is excessive even for you. Do you not understand the stakes in nuclear plant operations? No new process would ever be used for anything without having first been tested redundantly as a simulation where it wasn't actually being used.
Industry isn't Leaning, but startups are (Score:2)
Needs... (Score:2)
...a whatcouldpossiblygowrong flag.
Re: (Score:2)
K-19 was prophecy.
Certainly no expert, but... (Score:2)
It *sounds* like they are using it for generating bureaucratic material.
Certainly comparatively speaking the stakes are just so much lower, but I've seen a bit of that in my company.
In my case, it's material that generally no one reads, or maybe someone that doesn't matter. GenAI has become a way for people to tick the checkboxes that they think don't matter without trying to debate the reason for the checkboxes.
Suppose the question is whether or not some token gesture facilitated by the GenAI *actually* m
Re: (Score:2)
In a regular company, that makes sense, but in a nuclear plant that "bureaucratic material" is not optional, it's not there because the company decided to do it, and it can't be done by anybody; it has to be done by the right people that have accountability. Even if their accountability is mostly that their employer would be fined.
But beyond that, if they're committing gross negligence by having a computer program write it then they risk being placed into receivership and losing their investment. Nuclear po
Well, that's just terrible. (Score:3)
It often gets proposed for the purpose; either by people who overestimate it or with that little sneer at 'paperwork' that assumes that plausible nonsense is fine because whatever the 'paperwork' is it must be a fake problem imposed for stupid reasons, so it really deserves syntactically plausible nonsense; but is an exquisitely bad solution. Not only does it replace actual data with plausible nonsense in places where, surprise surprise, someone was actually maintaining records because something matters; but lowering the cost of bullshitting a report to close to zero removes one of the main incentives for process improvement.
If bad process hurts you can only really maintain it through fairly disciplined commitment to an oppressive, unreasonable, management approach that rejects process improvements out of hand; or the cultivation of a profoundly dysfunctional organizational culture where as many FTEs as possible owe their paychecks to nonsense requirements and can be recruited to defend them.
If bad process just means spinning up the nonsense printer with a cynical smirk; why bother even looking at what the process was intended to do; or getting your hands dirty with what records are important and what records are just historical accretion? Just prompt for a prompt that tells the bot to do them all, and some extras if you want to look diligent, and call it good.
Re: (Score:2)
with that little sneer at 'paperwork' that assumes that plausible nonsense is fine because whatever the 'paperwork' is it must be a fake problem imposed for stupid reasons ... removes one of the main incentives for process improvement.
I'm with you, but realistically speaking, if the mindset is already there, the material was likely already at high risk of plausible nonsense, not as thought out as the reality, but something to satisfy a check box for someone else. And the hope for process improvement is frequently not there, as raising a "do we really need to do this" invites "you must not care about safety/security/whatever". In many scenarios *no one* wants to be potentially on the hook for approving a process optimization that ends u
Re: (Score:2)
if the mindset is already there
There is no reason to suspect that that mindset would exist in nuclear operations companies. These are the stodgiest, most rigid, persnickety mindsets in the entire industrial world.
Re: (Score:2)
I agree, the real danger of this is that the documentation requirements will swiftly become so complex and verbose that only an AI can write them, and only an AI can read them. Humans will no longer even know what the effect of the documentation requirements is, they can only hope some of the remaining original purpose (keep the reactor safe) survives the AI rewrite.
Human leak? (Score:2)
Maybe stop harassing scientists and experts so that they don't flee overseas.
Re: (Score:2)
You certainly wouldn't want to hire a scientist to do a critical engineering task.
You'll get a disaster calculated from first principles, but containing a small error.
The jobs in question would be the among the boringest possible jobs for a scientist.
Well, it was nice knowing some of you. (Score:2)
So far... (Score:2)
That phrase "so far" is doing a lot of work here.
Atomic Canyon (Score:2)
French supermarkets (Score:1)
Let's put Skynet in charge of nuclear power! (Score:3)