AI at a Nuclear Power Plant - DTNS Weekend

0:00This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. Hey, it's Paige from Giggly Squad, and I want to let you know that if you want the newest and hottest beauty products, you have to go to Sephora.

0:36I'm obsessed with the One Size Liquid Blotting Paper Spray. It's the first-of-its-kind blotting spray that mattifies for up to eight hours. It's unreal. There's also the Summer Friday Shade Drops SPF 50, which basically lives in my bag at this point. Oh, and the Kayali Eden Plush Pear smells so good. It's sweet, warm, and addictive. like people literally stop me and ask what perfume I'm wearing, which you know I love. So if you want the makeup, skincare, hair, and fragrance products everyone's about to be obsessed with, shop only at Sephora. You want to get your backyard summer ready, but you don't want to break the bank?

1:11Wayfair gets it. Planning on dining al fresco or relaxing poolside? Wayfair has everything you need to prep your space. Shop now and save up to 70 % off during Wayfair's 4th of July clearance. Score huge deals on outdoor furniture, area rugs, and more. We're talking thousands of products for every style and budget. Plus, surprise flash deals July 6th. Don't wait. Shop Wayfair's 4th of July clearance now through July 6th at Wayfair.com. Wayfair, every style, every home.

1:48welcome to the dtns weekend edition gratefully provided to you and thanks for your support of this show and we are going to do a follow-up today you may recall a few months ago we talked to Tanner about using AI at a nuclear power plant and some of the ways that that turned out well, not a problem for Tanner, but he's been able to use it for a few more months. And so we thought we would check back in and find out what he has learned. We don't have video of Tanner, not because he's in an undisclosed location or anything. But Tanner Goodman, of course, is actually a little different title than the last time we talked to you.

2:32You're now a trainee for Senior Reactor Operator. Congratulations. Thank you. Yeah, it's been fun. It's a lot of learnings. Fantastic. You'll be represented in the video version by this man in an orange suit. I don't know if you wear an orange suit though. I could, but I appreciate it. No, absolutely. So catch us up for those who may have not heard the other episode. Just the short version of what you are using AI for at a nuclear plant. So last time we kind of talked about Diablo Canyon nuclear plant kind of created their own, not really large language model, but model that they could use to search and retrieve documents and stuff like that.

3:22And we kind of discussed whether AI belongs in nuclear at all, because it kind of freaks people out when you say that. And then since then, my plant actually developed our own, I think they call it retrieval augmentation generator or something, RAG model. and what we end up using it for is you can use natural language queries to find any document that's ever been published even like handwritten stuff from the 70s and 80s and provide you with these references and answers to your questions so that's mostly what we've been using it for which is what I had hoped the kind of model that we would hopefully get which we did would do.

4:16But yeah, we have, the bot is called NOVA, N-O-V-A, stands for Nuclear Operator Virtual Assistant, which is an amazing name. Yeah, it's perfect. Yeah. We use it just, like I said, a natural language, plant-specific research assistant. So it's not trying to do any operations, not control of anything. It's essentially a better search engine? Correct. Yeah. Effectively. And I know that is kind of the concern that a lot of people that I've talked to have. It is not tied into any plant systems. It doesn't even know what the current plant state is. It is just a very, very fancy search engine that when you ask it a question, it provides you an answer.

5:09But more importantly, it provides documentation and citations to how it got to the answer. And then you can click into the citation and it brings you right to the source document that it pulled from. And it even highlights like, hey, this is what I pulled from to answer this question for you. Yeah, that feels like the best case use of large language models, which is training them on a specific set of data. In this case, all of the manuals that you're talking about. And then when it gives you an answer, I guess that would be the next concern people have. What if it hallucinates and tells you to pull this switch and that's exactly the wrong thing to do?

5:51And the safety net there is you shouldn't just rely on the answer. You're going to look at the citation. Yeah, that's right. And I think we've kind of self-policed ourselves and we kind of talked last time about like how do you control this the nuclear industry is very good about self-policing and we've kind of come to the conclusion that verification is non-negotiable when it comes to the answers that it provides so we treat these answers as like a starting point and then the cited uh the cited documentation as the authority um to verify everything that it is telling us. So yeah, we don't just say, hey, AI, make this decision for us, and we're just going to take that answer and run with it.

6:41Yeah. You're not just, what do I do? Yeah. Yeah. Yeah. Yeah. For sure. And I think if you keep that frame of mind as a person listening to us talk here out there, I think that may help you understand it. And Tanner, I'm not trying to put words in your mouth, but it sounds to me like when you're using it, you're not asking it to answer the question for you. You're saying, well, I could search for this. The old way would have been try to find the manual page through a bunch of things to find the answer. The secondary way would do a search engine and it'll give me a bunch of lists. And then I have to search through the list to see if it's got the right citation.

7:20This is going to get you to the page in the manual you need faster. And you're not trying to get it to answer the question for you. You're trying to get that. You're trying to get that relevant piece of documentation. Correct. And even sometimes, though, some things can fail that we've maybe, since me being there, have never seen before. Uh-huh. Or something breaks and we need to know the technical design behind something. I can go into Nova, type a question like, hey, this is what we're seeing. Give it as much detail as I can. And then it goes and finds any relevant documents and relationships between various documents and then provides it to you and says, here is what I think is going on.

8:06Here's what I think the answer might be. And it'll source documents that we didn't even know existed. so uh we could do all this manually before but it would take teams of people and sometimes days going through combing through millions of documents like i kind of know that something like that talks about this should exist i just don't know where it is and that is where this has been invaluable where it it would take days before it could take 30 seconds to provide back an answer And at least I hope most of the time, the problems you've never seen before are seriously non-critical problems. Oh, yeah, yeah.

8:47In any nuclear reactor situation, you don't want to let anything go unattended any longer than possible, I'm sure. Right. And we fall back a lot on our training. So when something is going wrong, which thankfully hasn't and hopefully won't, we rely on our training to deal with what is going on at that present time. So when I say like something breaks, like any of these mechanical components over time, stuff wears down and may, you know, spring an oil leak or something. So using this natural language to then ask questions about, hey, some of these, these are some parameters that we're seeing, something going wonky, what are some potential solutions?

9:35And then having it, you know, provide all this information, information, documentation has just been really useful. Yeah. And I can see where, you know, something like an oil leak, I'm sure you could have a workaround that is relatively safe, but you'd rather not have the oil leak at all. And the faster you get that plug, the safer it is. No. Yeah, absolutely.

10:11Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade-in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. You want to get your backyard summer ready, but you don't want to break the bank? Wayfair gets it. Planning on dining al fresco or relaxing poolside? Wayfair has everything you need to prep your space. Shop now and save up to 70 % off during Wayfair's 4th of July clearance. Score huge deals on outdoor furniture, area rugs, and more.

10:45We're talking thousands of products for every style and budget. Plus, surprise flash deals July 6th. Don't wait. Shop Wayfair's 4th of July clearance now through July 6th at Wayfair.com. Wayfair, every style, every home. Hey, it's Paige from Giggly Squad, and I want to let you know that if you want the newest and hottest beauty products, you have to go to Sephora. I'm obsessed with the One Size Liquid Blotting Paper Spray. It's the first of its kind blotting spray that mattifies for up to eight hours. It's unreal. There's also the Summer Friday Shade Drops SPF 50, which basically lives in my bag at this point.

11:19Oh, and the Kay Alley Eden Plush Pear smells so good. It's sweet, warm, and addictive. Like, people literally stop me and ask what perfume I'm wearing, which you know I love. So if you want the makeup, skincare, hair, and fragrance products, everyone's about to be obsessed with shop only at sephora yeah uh you were fairly optimistic about this becoming useful when we talked last time uh has your opinion changed at all over the course of actually using it uh i was optimistic i was i would say cautiously optimistic yeah because like many people I do fear the over reliance on tools like this and replacing not necessarily replacing people but the knowledge that people have uh I would say I'm still pretty optimistic about it just seeing the the powerfulness of how much time it saves being able to to find all this information um i i really hope in the future it doesn't get like tied into systems to be able to control things like that i'm still not not a fan of thankfully we're not anywhere close to that uh but yeah i am i'm pretty optimistic about it having used it more and even like you had mentioned i'm in a training program now i can use it to generate my plant specific like study guides and test questions and stuff which has been really helpful so i said hey i'm learning about this system this week uh generate a 10 question quiz and it's not just some generic nuclear quiz that it spits out it's very specific to our plant which has been really helpful too So this brings up something that I think is a bit of a parallel.

13:19When navigation systems first came around and everybody's using Waze and Google Maps and all of that, there was a fear that people wouldn't know how to get around anymore because you're using the navigation and you don't need to know where you're going. What I have found personally is that I know more ways to get around because the navigation app has taken me some routes I would have never known about, would have never tried. And now even if I'm not using the navigation app, I know, oh, wait, I can go up here, take a right here, and that'll loop me around some of that traffic. It's increased my knowledge.

13:53Do you feel like there's any chance of that happening with this kind of system? I do think so. kind of like I said when some of the questions that we may ask it it'll provide documentation that I'm sure somebody knew existed at some point in time but is over the 30 40 years of operation has been you know buried under all the other documents so the fact that it knows everything and it's like having an intern that you say hey your job is to go read and memorize all millions of pages of documents and whenever I need something you know exactly where it's in the library the fact that it can source us info and i'm like oh man i'm learning new stuff about whatever xyz uh and it's very helpful one engineers love it because when we're on nights and weekends at two o 'clock in the morning if we have a question we don't have to call and wake somebody up yeah like if it's not uh super serious we can just say hey i'm curious about learning about this and it's made the whole learning aspect and teaching aspect really good and we kind of talked a little bit about hallucinations uh it does hallucinate and we have seen that before um but what really is really helpful and the feature that we all really like is that it like i said cites the sources gives you the source document and then you can go in and verify yourself which is what we all default to doing uh yeah well at first when we first got it we're saying we're going to ask it questions that we already 100 know the answer to and see how good it is and it was pretty good and then there were some things that would spit back we're like this is not right at all yeah it has these uh little feedback buttons on it like thumb up thumb down and then you can tell it why it was wrong and then we have an on-site team of people that look at all this real-time feedback that we're providing it to then go back and retrain the model, which is pretty cool.

15:59Yeah, I think there are three really important points to the story you're telling is one, it is trained on a limited set of verified information. A lot of times these LLMs get a reputation because they've been trained on whatever was available on the internet, which may or may not be very reliable and certainly isn't a limited set of information. So narrowing the scope means its expertise is going to hallucinate less because it's not pulling in from irrelevant sources. And it means that you can verify that it is trained on things that we know are true. The second thing is that you've kind of self-trained yourselves to say, I'm not going to just take responses at face value.

16:42I'm going to look at the sources and using it that way, I think is also important to say, I'm not relying on it to give me the answer. I'm relying on it to help me understand the answer or help me find the documentation that would have taken me longer to find it before. I think those two points in particular are not just important for understanding how it's helping you, but I think important for understanding how to make better use of LLMs in general. No, absolutely. And it's funny. This is a small story. So it is trained, like I said, on all of our internal plant documentation, but it's also trained on the Nuclear Regulatory Commission, NRC, all of the public information that they have available, which I think publicly, if you just search on Google, it's like several million documents, but it's like 50 some on million pages of information that's publicly available to anyone.

17:36You can go on their website and you can search up anything. But it's also trained on all that. So when we were first training the model, we sent it to the NRC and said, okay, we're going to download all of this information to then train the model on it. We actually crashed the NRC servers because we were pulling so much data so fast that they kind of freaked out a little bit. And we're like, no, this is what we're doing. We're just training this LLM on all this, uh, on all this information. But yeah, that was, that was pretty funny. So yeah. Yeah. Explain yourselves. That's right. Why are you doing, which honestly, that's, that's similar to companies seeing all of the spiders from, from AI companies coming to their website and like, what are you doing?

18:22Why are, why are you crawling my site? What are you taking? At least in this case, you had a justifiable reason for doing it. But before we wrap up, any lessons that you've learned from using this that might be helpful for others who are looking to put this in place in their workplace? I think, one, don't just, and I think hopefully this is common sense, but don't just trust anything that an LLM or a similar model spits out at you. I think if you are creating something like this for your workspace, the citation feature is really what sells this for me. And then I think to kind of place ease on people, which it was when it was framed this way to me, it made me feel a lot better about having AI in nuclear is what before took us hours or days to search for information, which is then time that I'm not helping provide oversight, digging through tech manuals and stuff, trying to find the answer.

19:30now takes 30 seconds to retrieve this information which then frees us up to either oversee other people or you know just monitor things uh more better i guess we always have people dedicated to monitoring but having freeing other people up to also do the same thing um has been really helpful but i think lessons learned wise uh yeah one never trust it and two if you can uh have it source, the actual source document that it was trained on to then provide to the end user and then have them make the ultimate decision on whether the answer it is providing is accurate or not. Yeah. Well, Tanner, thank you so much for sharing your experience with us because I think this is really illustrative of that dividing line that gets lost in a lot of the arguments about LLMs.

20:25And it's, it's, it's wonderful that you're willing to share that with us. Yeah, no, it was, it was good coming back and, and talking about it. I'm actually not all sold completely on LLMs in general, like be it chat GPT or anything like I do have, there are some use cases that I think are useful, but then once it was framed in a context that actually benefit my every day to day. Like it was like a light switch in my head flipped and I'm like, Oh, I kind of get it now. But, uh, yeah. Yeah. It's, it's not good for absolutely everything. Right. And this is an example of like, Oh, here, here is a thing that it's good for.

21:08It's good for some other things too, but you know, let's, let's, let's figure out what those are. Uh, Tanner, if, If folks want to follow you or do anything, do you have anything you want to tell folks about? I know you're probably not out there, you know, putting as much things, getting as many people to follow us, like if we had a journalist on or something. But here's your chance to tell people about whatever you want to tell them about. Yeah. Oh, I am on threads at Straven, S-T-R-A-V-Y-N. Just post random thoughts throughout the day. and then also part of the the frog pants community the core podcast is a video game podcast we have a world of warcraft guild so i guess if you like world of warcraft and you follow frog pants related shows you can pop on over to the frog pants discord and ask some info come on come join us kill some dragons and stuff yeah fantastic uh well tanner thanks again man i I appreciate it.

22:07No, thank you for having me. It was great. And thank you folks for being along with us on the weekend. If you're in the United States, hope you're having a fabulous holiday weekend. If you're not, I hope you're keeping cool or having fun doing whatever you're doing. Don't forget that we have new patron perks. And in celebration of that, with Experiment Week coming on Monday, we are offering 26 % off your first pledge for new patrons monthly or yearly, use the code experiment. That's patreon.com slash DTNS. And starting Monday, it's experiment week. All new podcasts launching right here on the network.

22:50We'll hand over the feed to them. Check them out. Let us know what you think of them. And we'll see you soon.

23:00The DTNS family of podcasts. helping each other understand. Diamond Club hopes you have enjoyed this broker.

23:13This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade-in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. Hey, it's Paige from Giggly Squad, and I want to let you know that if you want the newest and hottest beauty products, you have to go to Sephora.

23:49I'm obsessed with the One Size Liquid Blotting Paper Spray. It's the first-of-its-kind blotting spray that mattifies for up to eight hours. It's unreal. There's also the Summer Friday Shade Drops SPF 50, which basically lives in my bag at this point. Oh, and the Kayali Eden Plush Pear smells so good. It's sweet, warm, and addictive. Like, people literally stop me and ask what perfume I'm wearing, which you know I love. So if you want the makeup, skincare, hair, and fragrance products everyone's about to be obsessed with, shop only at Sephora.