Japan exploring whether AI could help inspect its nuclear power plants Japan’s Nuclear Regulation Authority has requested extra funds to experiment with AI-powered nuclear plant inspectors. Japanese media report that the authority wants to explore AI inspection because many nuclear plants operated by Japanese energy companies are already old and will likely need more oversight as they continue …
I downvoted this because AI might actually have spotted two major design issues at the Fukushima-daiichi facility :
1. The Fukushima facility was designed at a time when the potential for huge earthquakes associated with geological subduction zones was just beginning to be understood. As a result the potential strength of tsunami at the site was underestimated, The seawall protecting the facility was not high enough, AI reviews of the facility in the decades after construction when the geologic risk was better understood might well have spotlighted that problem.
2. There was probably adequate backup diesel generation on site. But it was located in an area that flooded and it was not hardened against flooding. If it had been better sited and/or properly hardened, the loss of grid power to the facility probably would not have led to the catastrophic failure of 3 of the six reactors and severe damage to another. Again, AI might have identified that problem.
I am not suggesting that AI can/should replace human inspection and review of nuclear facility design, construction, and operation. It's hard to imagine a worse idea, But AI might well be a useful supplement to human review.
> AI might have identified that problem
That's some pretty stinky weed you're smoking.
Today I Googled how to find out the model number of my Pixel 6 Pro. (For some reason, it's hidden under regulatory info instead of model number)
Google's AI was completely wrong. About Google's own damned phone. On a technical question which had a simple uncomplicated answer 4 results down.
In summary: On a sample size of 1, with a completely unrelated problem, using a type of ML (presumably an LLM) that would almost certainly not be the type of ML to be deployed for the kind of problems mentioned in the post you responded to - and given that your expert human oversight spotted the error instantly - you arrived at an entirely meaningless conclusion.
Seems to me that you've missed the obvious. Sure AI makes mistakes. Lots of them. Some really bizarre. You wouldn't want AI doing jobs where mistakes are costly and/or lethal -- Air Traffic Control for example. But mistakes in the design of nuclear power plants and other infrastructure items like dams and bridges were/are/and will in the future be made by humans. Asking AI to review the human generated design adds very little cost and might occasionally save money and lives.
I might add that there look to be a few other applications where being wrong sometimes isn't a big deal and AI might be useful. All niche applications as far as I can see. Nowhere near enough to justify the staggering investments that are being made. Seems to me that AI is a tool. And not a tool that a sensible person would want to use all that often. But still, something one might want to have in the toolkit.
That's a bit of a knee-jerk reaction. It is not implausible that AI (ML, if you prefer) could help flag up issues that might potentially be missed by human experts. The human experts would still, of course, investigate any issue flagged up by AI.
Whether that could be at all helpful in the particular scenarios described by the article is another story:
It is unclear what role the regulator imagines AI will play in nuclear inspection, but highly-regulated industries such as nuclear power produce huge quantities of data and documentation. AI could conceivably spot anomalies in that information and advise human inspectors where best to focus their attention.so some small clues there.
From a broader perspective, however, it's becoming clear that AI can potentially play a useful role in the science/engineering arena - but not in the form of the straw-man cartoon caricature which imagines that scientists/engineers will simply let AI make decisions without expert oversight.
This post has been deleted by its author
"mighty Godzilla, the radiation-created monster that gave anxiety about nuclear energy a prominent place in Japanese popular culture."
Other way around. After being on the receiving end of the first nukes used in warfare, japan's population understandingly had anxiety about any form of radiation.
Godzilla was born of that anxiety.
There aren't enough quotes. What do they mean with "AI"?
If it's to review documentation - it might beat the alternative, if the alternative is not looking at it at all. But the proper solution is to fund the inspection agency properly.
If it's to analyze scans of structural elements in search of material stress and microfractures - go ahead, a specialized ML model would be good at this. You don't need to call it "AI", though.
I thought the lesson of Fukushima was not to build in a known tsunami risk zone, at a height above sea level which would be easily topped in the event of one. Just like the lesson of Chernobyl was not to f*** around with a known dangerous design, and the lesson of TMI was not to turn off all those annoying warning signals. Silly me.
