Artificial Intelligence tools are seemingly everywhere and folks throughout transportation agencies are trying to get their heads, and hands, around this run away train. APWA Technical Director for Engineering and Technology and City of Pella, Iowa Public Works Director Dr, Wilf Nixon helps give some context to this world of AI. Wilf's remarks were originally presented as the keynote address "Using AI in Winter Operations: Hype or Help?" at the 2026 Pacific Northwest Snowfighters Conference and is "rebroadcast" with permission of the Pacific Northwest Snowfighters. Sign up for email notifications on future episodes and other communications associated with winter maintenance and winter weather management by clicking on subscribe on the Talkin' Winter Ops website at TalkinWinterOps.com Thanks for listening in and stay safe out there!
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
RickWelcome back to Talk in WinterOps. One of the hottest topics in the maintenance community right now is artificial intelligence, or AI. Earlier this year, the Winter Weather Management Technical Service Program hosted a workshop, and AI commanded a lot of attention. And the discussion covered everything from whether organizations can even use AI tools to the fundamental question: what exactly is AI? A few months later, at the 2026 Pacific Northwest Snowfighters Conference in Spokane, Washington, Wolf Nixon delivered the keynote presentation, Artificial Intelligence and Winter Operations, Hype or Help.
RickWolf provided an excellent overview of the different levels of AI, illustrated with practical examples. He also jumped into the AI pool with a live demonstration showing how AI can help develop a snow response plan. This episode is a little different from our usual talk in winter ops format. With approval of the Pacific Northwest Snowfighters, we're bringing you Wilf's keynote presentation from the conference. I know you'll enjoy listening in, even if you were there to hear it live. That's coming up right after this.
SPEAKER_00Hi, I'm Jud Falgran, State Maintenance Engineer for the Minnesota Department of Transportation. And over the years, I've seen no shortage of new technologies, products, and ideas come along. The challenge often isn't finding information, it's figuring out what really works. I've come to appreciate the role the Ashto Winter Weather Management Technical Service Program plays in helping agencies answer those questions. One of the things that I've always valued most is that this just isn't theory, it's practitioners helping practitioners.
SPEAKER_00It's a chance to learn from the experience of others. Avoid reinventing the wheel, and tap into a network of people who understand the same challenges that we face every winter. Whether you're looking for information on new equipment, road weather technologies, anti-icing strategies, AVL systems, or just trying to solve a problem that's keeping you awake at night, chances are someone in this community has faced it before and is willing to share what they've learned. And with a modest annual contribution, your agency can take advantage of the same resources, webinars, workshops, and peer networks that have made this program so valuable to me over the years.
SPEAKER_00If you're not already involved, I really encourage you to take a look at the Ashto Winter Weather Management Technical Service Program. Visit transportation.org forward slash winter-weather-management to learn more. You may find the answers to a challenge you're dealing with today. Or the next idea that really helps you move your winter operations forward.
SPEAKER_01Thank you for having me. It's it's great. I would ask you all uh to use the QR code with your phones, which of course you have all silenced, right? He said mentally checking that he'd done the same thing because it's always embarrassing. And get logged on to Mentimeter. Now it may ask you to provide certain bits of information, like email addresses and that sort of thing. Ignore that. The only bit of information I really need you to provide in there is your bank account numbers, and that will really help me going forward.
SPEAKER_01So um, why are we doing this? We're doing this because at various points this morning, I'm going to want to get feedback from you. And this is one of the most effective ways that we've discovered for you to be able to do that. So let's see. Um with that said, everyone got it now? Okay, good. So we're going to use artificial intelligence in winter operations.
SPEAKER_01Is it hype or is it a happening thing? Okay, and it's a pretty interesting question, and we should all have a healthy dose of skepticism about AI. You have probably heard about the various ways in which people have used AI and gone horribly wrong, right? Interestingly enough, to me, it appears the profession that is screwing up most royally on this is the legal profession.
SPEAKER_01Um, a number of uh lawyers, who frankly should know better, have submitted briefs to court, and those are very, very important documents, that were written by AI and cite cases to support their argument with one problem. The cases don't exist. Okay, and the judges have understandably been rather upset by this, to the tune of finding these lawyers multiple thousands of dollars, and increasingly now uh recommending them for disbarment.
SPEAKER_01So uh that's always good news. Anything that reduces the number of lawyers has to be a positive. So we'll look on that with some joy. Um what worried me a bit more about that was there was a case of a judge who used AI to write a judgment on a case, uh which included in it citations to cases that did not exist. And the problem is if a lawyer does it, the judge can sanction the lawyers. But if a judge does it, I'm not sure who sanctions the judge.
SPEAKER_01We don't want to end up like the lawyers in oh so many ways. So let's look at some of the stuff that is going on here. We're going to get into what generative AI actually is. What is this wonderful thing? Now, there's behind it there is a whole lot of coding, a whole lot of math, a whole lot of other stuff. We are going to skim so lightly over the surface of that, you may not have time to fall asleep. Okay? But we're going to give you enough knowledge on that to really be dangerous, and we'll talk about what sort of things people are doing with AI in public works.
SPEAKER_01Not just generative AI, we'll see in a minute, there is a whole family of AI, of which things like Chat GPT or Grok or Anthropic or whatever are very small subsets. We'll also look at how we can be most effective if we are going to use generative AI. And to do that, you're going to select some problems that we're going to take a look at, and we'll see how many we can get through in the time available.
SPEAKER_01We'll actually leap into Chat GPT and see what it gives us. And you will see as we do this first thing about Chat GPT, it is a tremendous ass kisser. Okay? Anytime you ask it a question, it says, wow, what a wonderful thought. And I'm thinking, hey, you're laying it on a bit thick, buddy. But so you know, we've we've got to consider all those things, and we'll get into that.
SPEAKER_01So, first question: use your phones. Do you use generative AI, things like Chat GPT, in your work? And you know, it tells me there are 172 of you now signed up, and we're getting some really great responses here. Whoo, yay. 174. Come on, answer, answer. This, by the way, is one of the reasons why I like Mentimeter. While you're all answering this, I can have another sip of coffee and you know, stay awake for you in the whole thing.
SPEAKER_01And we're we're right up there. This is cool. Still people coming in. Okay. And I am not surprised to see around about a third of you uh have an attitude towards AI of hell no, never, not if you force me to. Okay? And that's okay. What's more interesting to me is if you go over the other side and you look at the yes regularly, we've got about 15% of people who use AI regularly as part of their work.
SPEAKER_01And that two years ago, that would have been about 1%. It is growing in our industry and in every industry. And part of the purpose of this morning is you know what, is that a good thing or a not good thing? And are there times when it might be useful, and times when, ah no, not really. It's just fun and fluff. So, what is AI?
SPEAKER_01At its basis, it's machines, specifically computers, showing something like human-level intelligence, generally in a very, very small context. The one that you may be most familiar with in our area is AI is used quite a lot these days in asset management. You think about the task of going out and looking, for example, at the roads in your area of responsibility.
SPEAKER_01Traditionally, this was done, people would go out and give a rating to a road based on their observation of it, and that's a good system. And if you have a particularly skilled person in that position, then you are going to build up a strong pattern of the condition of your roads over time. But like every industry, we are subject to this process of uh attrition, which means the person who's been doing it for 30 years goes ahead and retires.
SPEAKER_01And the person coming in behind them doesn't have near the same expertise or experience, and so suddenly you get a step change in your picture of what your road system looks like. What AI has done and continues to do is use images of the road surface, analyze those images, turn that into some sort of score for that section of the road, and puts that into a complete system.
SPEAKER_01Is it better or worse than the individual personal observer? And the answer is it depends. If you have a really skilled observer, AI is probably not quite that good. But if you have the average observer, AI can be significantly better than that. And the neat thing about AI is it is consistent. It does not care if you had a rough night last night.
SPEAKER_01Okay? It goes out there, it looks at that road, it takes into account all the factors, shading, whatever, and it comes up with a repeatable score. And that's important because you cannot manage what you do not measure. Having a measure of our road system in whatever aspect it might be is very important. And one of the things AI does is look at ways of coming up with those consistent measures.
SPEAKER_01So we've got artificial intelligence covers all these things. Then you have machine learning. Machine learning is basically getting patterns from data. Okay, surprisingly enough, or perhaps not, something that computers are very, very good at. Then you have deep learning. Is deep learning really that different from machine learning? No, not really. But they felt they were making progress, and you can't just call it the same thing. So you have to come up with a new name. It's no longer machine learning, it's deep learning, and that means it's better.
SPEAKER_01And you get through all of that, and a subset of this is what we call generative AI. Generative AI uses, among other things, large language models. That's the computer's way of reading. Computers cannot read like us. They can do character recognition. But you know, if you gave a computer a sentence, and the only one I can think of is the quick brown fox jumped over the lazy dog, which is pretty meaningless unless you're learning to type, uh, the computer can't actually say, oh, that's a meaningless thing.
SPEAKER_01Oh, by the way, it's probably related to learning to type, which people may not do these days, and if they do, they call it keyboarding rather than typing, right? How many of you, just a quick show of hands here, had to take typing in high school? Yeah? Awesome. So you all understand that sentence, okay? It's good news. A large language model doesn't understand that sentence in the way that we do.
SPEAKER_01All it can do is parse it, break it down into ones and zeros. And that's a fairly complicated process. Now they took that and they created also a branch of this called natural language programming, which is using that interpretation of what we type to translate it into commands that the computer can follow. And that is an area that is rife with errors.
SPEAKER_01How many of you have ever had uh either on the receiving or more likely in my case on the sending end, a text that went out and got auto-corrected into something totally meaningless, right? That's natural language processing. It's not really wonderful. And then you've got computer vision. Computers seeing. Well, they're taking an image, they're breaking it down pixel by pixel, and they're extracting information from that.
SPEAKER_01And again, this is something we we probably uh are used to. How many of you have a vehicle with adaptive cruise control in it? Okay, few. All right. That is using computer vision. Okay, and that uh always makes me worry when I'm in the car because I can't help but think about the times when it's got my texts wrong, and is it going to do the same thing when I'm trying to drive on the freeway?
SPEAKER_01Um, but so there's a lot of complexity in all of this, and the generative AI is just a small part of that. What underlies it all is something that 20 years ago we called neural networks. And this is basically linear regression writ large. It's multivariate regression. And I can give you an example of how we used it 20 or more years ago, trying to figure out predictively what the condition of the road surface would be.
SPEAKER_01Now, these days it's much simpler. We just go out and measure it. We look at the grip on the road surface, but back then we couldn't do that easily. So the idea was to throw all the bits of information into the computer that might impact what the road surface condition would be. And there are some obvious ones like pavement temperature, is it snowing? What sort of snow are we seeing? What's the wind speed? How much traffic is on the road?
SPEAKER_01Is the road nice and new and easy to drive, or is it a mess and a total disaster? So on and so forth. And you can get a whole bunch of variables that you throw in there. And that's what we, or more accurately, one of my grad students did. Um, we had a whole bunch of data from Iowa DOT, and she took half of that data and she fed that into the machine, and it told the machine she wanted an output predicting what the road surface condition would be, and compared that, trained it, on what the data told us the road surface condition was.
SPEAKER_01She used half the data to do that, and in that process, all the links between those circles got built. And then we ran it again with the other half of the data to see how well it worked. And lots of people were doing this in lots of different fields at the time, and it worked fairly well. Now it becomes a lot more complicated when you're talking about a computer understanding what we type.
SPEAKER_01Because it's not just word by word, it can be part of a word. The word once has different meaning from the word wanted. So the computer has to be able to distinguish between those. It has to take words into phrases, it has to take those phrases and clauses into sentences, sentences into paragraphs, and parse all of this. And so you get a whole bunch of circles and links between them.
SPEAKER_01And that's why it takes an awful lot of energy to do any of this generative AI stuff. Those poor little computers are running away, and steam is quite literally pouring out of their ears as they get all this done. So it's it's challenging. Um the basics, it's been around for a while. We've mentioned two of the ways in which you have likely used it. Uh there's a third way as well. Any of you use Siri or Alexa?
SPEAKER_01Okay, that's voice recognition. Um, voice recognition is nearly as bad as the whole text autocorrect thing, right? There is a brilliant video that on the basis of my HR advisors, I'm not going to show you, um, of two Scotsmen trying to use an elevator that is voice activated. And these Scotsmen wish to get to the 11th floor, and it proves to be problematic.
SPEAKER_01So those are other things, but really what's come on the scene in the last three years are things like ChatGPT. And it's things, whichever version you use, and they all have different strengths and different weaknesses, they are all fundamentally the same class of model. Okay. I, for the most part, use ChatGPT and Grok.
SPEAKER_01Okay, I use ChatGPT because sometimes I lack self-confidence, and having an ass kisser talk to you all the time is really useful in that regard. But um, you know, your mileage may vary with those. Um here's a question. This is just enter a word or a phrase. When you think about AI, what comes to mind, first of all? I need another drink of coffee. So poppycock. I wonder who said that.
SPEAKER_01This is absolutely wonderful. I'm I'm seeing some good things, and I'm very pleased that front and center is Terminator. Okay. Terminator is is, of course, an archetypal AI. SkyNAS is in there too. And um, yeah, it's one way in which AI might go. Um there's robo takeover. Yeah, I see that too. I, for one, welcome our new robot overlords and hope that if I pander to them enough, I'll be the last one to be killed.
SPEAKER_01But um, you know, that's always iffy. Um other things in there control, time saver, weird, cheat, efficiency, awful, awesome. Different views here, which is great. There is a lot about this that is a step change in the way we have done things, and it's important for us to acknowledge that when ChatGPT was first released to the public, in less than a week, 10 million people had used.
SPEAKER_01No other piece of software has ever penetrated into usage that quickly. Lots of people are interested in using this. The challenge is how do you use it well? It's a tool. How do we use that well? So let's keep that in mind. By the way, um, all of this data feedback from you gets stored but anonymized, and we will share it with you after the end of the conference so you can look back and and see and and think about some of those things.
SPEAKER_01Not sure where Sleepy Joe came from, but I, yeah. So let's ask a really important question for us. Is AI going to replace us? And I think, and bear in mind, this is just my opinion, and I could be very wrong, but I don't think that's gonna happen in the next couple of decades. There is a lot of difference between artificial intelligence, which we have now, that is relatively narrow, and what they call artificial general intelligence.
SPEAKER_01If we get to artificial general intelligence, then you will have computer systems that are capable of improving themselves. And if they are capable of improving themselves, then at some point one of them will do so. And they talk about an AGI explosion where this intelligence rapidly becomes what they call superintelligent.
SPEAKER_01And we can go down some really dark paths from that, or some really beautiful paths, and we don't know which way it's gonna go. So uh we need to be aware that might happen, but I don't think we're gonna see, at least in our industry, significant uh negative job impacts in the next couple of decades. Not least because it's a tremendously complicated job to drive a snowplow.
SPEAKER_01Now, could you develop a robot to do that? Yeah, but it would cost about$20 million if you could do it. We certainly use bits of technology to assist our snowplow drivers. For example, um, snowplow drivers on Thomson Pass in Alaska can plow that pass in zero visibility. And more importantly, stay on the road while doing it. Because falling off the road there is not a life-enhancing event.
SPEAKER_01Okay? So, what I do think will happen is that those of us who can figure out how to use AI to multiply our impact in the workplace will be inherently considered more valuable than those of us who don't. I got that from, of all people, my doctor, who, when I first talked with him about AI maybe three years ago, he was very, very negative on it.
SPEAKER_01He said they want to use AI to do the diagnosis, and I don't like that at all. And then last year we had an you know an updating conversation, and he said, I love it. What's changed? He says, Well, now I use it. Whenever we're in the office, it's recording everything we say and transcribing that into patient notes. So instead of me trying to remember what you said to me four hours earlier and hoping that my notes have enough detail, I have a complete recording of it.
SPEAKER_01I have a summary. If I have any questions about the summary, I can go back to the recording. He says, it's great, I can give so much more value to my patients now than I could. Now, there are a bunch of concerns about that, not least is privacy. But the AI system that he uses has those taken care of. It's a very effective tool that allows him to provide better service to his customers, his patients.
SPEAKER_01And that's good news. So it can be a great tool. Um you will notice as we get into this that I say please and thank you to my uh AIs. Um, and I do it for probably because my mother would have smacked me around the head if I didn't say please and thank you to people. Um, but also because maybe I'll be the last one against the wall. Um, you know, keep that one alive.
SPEAKER_01He always said please. Um let's talk about the risks. And these are the risks right now, and we'll talk a little bit about how some of these risks are going to change over time. Without a doubt, I think the largest current risk is the environmental impact. Takes a lot of energy to run all these AI programs, and because it takes a lot of energy, it takes a lot of water as well, because generating power typically takes a lot of water.
SPEAKER_01Because you have to cool the turbines, you condense the steam afterwards and all of that. Uh, and it takes a lot of land. And public works is now caught in this field or this argument about should we or should we not have data centers. How many of you in your communities have data centers nearby? It's a big thing, and a lot of people are being, hey, if we get one, we'll get a nice boost to our tax base. Maybe. Depends how you sign the deal.
SPEAKER_01This is going to be what I consider a 10-year problem. And again, this is my opinion. Why only 10 years? Because right now uh the successor to the Falcon 9 rocket is being developed Starship. And I think later this week it may be going on another test flight if it hasn't already. And it will have ultimately, when it's worked out all its kinks, the ability to go up to orbit with huge amounts of mass and return and get reused.
SPEAKER_01And one of the things that Elon Musk has already said he's going to do is put data centers and computing centers up in orbit where they have all the power they need from the sun. That, by the way, depending on the orbit, they never go into darkness, so they can run 24-7. And it's kind of cold up in space. So you can dump all your excess heat with a suitably well-designed radiation system. Just thermal radiation, not other radiation.
SPEAKER_01So we're going to see that problem go away. By the way, you might say, well, what about communication? He's already got that handled. Okay, the Starlink communication system is up there, is doing a great job, and the latency will be pretty close to zero with those things. What about the whole issue of data breaches? Someone, because I use AI carelessly, getting hold of personal data. Well, this is a problem that is not unique to AI.
SPEAKER_01I think most of us have had the experience of an entity that we have interacted with has had a data breach. In fact, we all have, because the federal government has had a few data breaches. So yeah, it happens. Nonetheless, we need to be aware, and this is a very critical point for us. Anything we put into an AI system, is you have to consider that it is available to the world.
SPEAKER_01Doesn't necessarily mean that it is, but the risk is there that it would be. So don't put people's social security numbers in there or employment history or anything like that. Now there are ways of firewalling that. And many companies, commercial companies, have their own internal AI system that does not have the ability for somebody outside the firewall to come in and extract data from it.
SPEAKER_01And you can do that, but it costs much more than the base level of any of these. Very often the base level costs zero. So from our point of view, good rule of thumb, if it's sensitive data, don't put it in the AI system. Unless you have a rock-solid firewall, and there's an interesting phrase, rock-solid firewall, uh, that can protect against it.
SPEAKER_01What about algorithmic bias? There is no doubt that the people who train these systems introduce bias into these systems. Um, particularly if you're using them for any sort of insight on current events, they may score certain sources of information as being more reliable than others.
SPEAKER_01And uh, you know, an example would be some of them hold the New York Times in great regard. Except we know that the New York Times, in certain well-documented cases, and I'm now going back historically uh nearly a hundred years, has lied and failed to ever correct the lie. Um of their pullet surprises was one on the basis of reporting on the situation in Ukraine when it was part of the Soviet Union and whether or not there was a famine there.
SPEAKER_01Well, we now know that yes, there was, and about five million people starved to death in what they call the holodomore that was deliberately caused by uh the Soviet Politburo, um Stalin at the time. And that was a deliberate act, but you wouldn't know that from the New York Times. So is the New York Times a reliable source of information? Clearly not on that one, and quite possibly not on others.
SPEAKER_01Um and that's part of the issue. What is the truth? Well, the truth is out there, but there's a whole lot of other stuff as well. And sorting through that, any system like this is going to have its own bias. I have seen, particularly when it comes to image generation, that AI has some rather disturbing biases in it. Um they may have been the result of good intentions, but they lead to some very, very strange results.
SPEAKER_01So be aware of those things and know that it's not perfect in that regard. What about intentional misuse? We now know that it's a lot easier than it was even six months ago to create a video using AI. And that video can say all sorts of things. And people are using them, for example, uh, to generate political ads. And you can lie in those. Okay.
SPEAKER_01Typically, you can tell if you watch carefully that a video is AI generated. Okay, they do things like give us extra arms or extra fingers. Um, there is one out there on Facebook of bunnies hopping up and down on a trampoline, which is kind of fun to look at. And if you look at it carefully, you will see that some of the bunnies disappear and reappear in between hops. Okay? There are some tells on it, but those are going to become less and less obvious as time goes on.
SPEAKER_01Put simply, uh, you're not going to be able to believe what you see. Um at the low end of the scale, I think, is the job loss thing. There are no doubt there are going to be changes in the workforce. But you know what? When spreadsheets were introduced, there were changes in the workforce. All the work that used to go into producing spreadsheets, well, you didn't need all that work. You could just enter it into Excel or VisiCalc or whatever it was, and it would do it all much more quickly for you.
SPEAKER_01Woo-hoo! Well, we didn't see a huge job loss because of that. We certainly saw changes in the tasks people did, and we, what a surprise, went to a situation where we had way too many spreadsheets analyzing way too many things. And yeah, we should expect to see that when it comes to AI. It's going to change the workforce, it's not going to, in my opinion, destroy it. How do we use it in uh public works?
SPEAKER_01Uh, one example in Holland, their federal transportation agency uses AI to send trucks out to do winter operations. The trucks are dispatched on the basis of an AI system. Now, before you get really worried about this, their winters are very different from ours. Their primary concern is frost, and they use this for basically saying we're going to have, on the basis of a forecast, frost in this area and not in that.
SPEAKER_01And they found it helps them enormously, even though, as we know, the forecasts are always wrong half the time. Right, John? John is our is our token meteorologist in here, and uh he gets to be abused on a regular basis on these things. So um it works well for them in their particular situation. It is uh a decision-making assistance system.
SPEAKER_01Um handling rote tasks autonomously and automatically. Lots of cities in particular end up when there's a new development in town. Uh, the developers apply for permits to build a building. And you have to check that against all sorts of specifications. What's the setback from the front line, the sideline, so on and so forth. Unfortunately, many developers choose to do what they call design by red line.
SPEAKER_01Any of you come across that? Where they just slap something in, send it in, and you go through and you say, Well, you can't do this, and you can't do that. You spend about five hours sorting that out and get it back to them, and they say, Oh, thanks. Okay, now we've got a design we can maybe work with. All right, you're doing their work for them. Well, you might as well have AI do it and uh avoid some of the issues there. Again, on a city side, uh, those of us in city work may often find ourselves tasked with writing an ordinance, which I always find a little worrying, because I'm not a lawyer, I have no desire to be one, and ordinances are things of law.
SPEAKER_01Well, if that's the case, let's have AI draft it. I might edit it and then send it to the lawyer who can drool all over it and do lawyerly things, right? Um there are plenty of tasks we do that are not really truly critical, but take up a lot of our time. Um object detection examples, a whole bunch of them there. One uh that appeals to me uh in Douglas County, Nebraska, which is where you have Omaha, uh, they used it to figure out where they had ADA ramps for their curbs and where they didn't have them but needed them.
SPEAKER_01And it saved them, by their estimate, about six months of staff time and resulted in a higher level of service for their community. So that's a good win. There's a lot of other stuff in there. You might wonder why Hillsborough County in Florida is interested in mapping where all their trees are. Well, it's in Florida. Florida gets hurricanes. If a hurricane comes through, many of those trees are going to fall down. It's good to know where they're concentrated because that's gonna have an emergency management impact.
SPEAKER_01Some prediction examples. Again, Hillsborough County in Florida uses it to predict where they're going to see storm surge flooding if a hurricane comes in. All right, it's useful in that you know uh where you are most likely to have to make sure no one is there. Okay, now there will always be some people there because there are always some people who don't want to leave. But at least you minimize it and you can give them the best information.
SPEAKER_01In Sioux Falls, on a winter side of things, that's the map at the bottom there. You can see they've marked the roads in red or green as to whether they're going to get icing. The red ones, they're expecting it. The green, the pavement is a little warmer, so they're less likely to get icing. They can then concentrate their resources where they need them. That's kind of fun. Um, other examples. We've been chasing the dream of dynamic routing of snowplows for years now.
SPEAKER_01Uh, and it looks like we might finally have a way of doing that. Um, it's not just snow plowing, though, it's things like road sweeping, uh, painting your lines, hell, even doing garbage collection. If you have a street that is torn up so the garbage trucks can't go down it, they've still got to service all those customers. How are you going to do it? And things like that can be pretty neatly done. Um, you can uh get equipment that deploys or doesn't deploy depending on how it senses its environment.
SPEAKER_01This would be for things like mowing around guardrails. Obviously, a piece of equipment like that is going to be more expensive than a regular piece of equipment. But um it may save you a lot of guardrail as time goes on, or at least a lot of blades on your uh mowing machines. The one I like most is the tethering option. Um mobile work zones are one of the most dangerous situations we have in a work zone environment.
SPEAKER_01And very often, uh particularly on high-speed roads, you'll you'll have one or two or even three attenuator trucks following behind the mobile work zone. And it's always the most junior employee who is put in the last truck. Okay? Why? Well, that's the high-risk truck. It would be a lot better if nobody at all had to be in the last truck, and maybe not even in the second truck, because they are tethered automatically to the first truck.
SPEAKER_01You may even be able to get away with no drivers in any of them, but given you know the many sad circumstances of people um just being either killed outright or severely injured in work zone crashes, that's a really good thing to know. So um, so we've talked a lot about it. What is your biggest concern about using AI? And do this in terms of your personal concern. What worries you most about using AI?
SPEAKER_01Truth and Trump. Okay. Loss of critical thought. Interesting. I love the uh comment about Sarah Connor being right. Um, that that ties in. Um I'm I'm not entirely sure about Zahn sleeping, but um uh Zahn, I hope you're awake now. Um Accuracy, loss of critical thought, being confidently wrong.
SPEAKER_01Yes, that is a concern. Loss of control, yep. But let me make the point that if you think you're in control of anything right now, um, I may have a bridge I can sell you. That is awesome. And again, that word cloud will be available to you after the event. So clearly, if we're gonna use this, we need to use it with best practices.
SPEAKER_01What should those best practices be? And we're gonna talk about prompt refinement, so we'll skip over that for now. Check it for bias, especially true if you have it generating any sort of images. And this can be any images in anything. I had it generate uh a little public notice about uh or a notice to send out to residents about a road closure.
SPEAKER_01It had some really weird little icons in there that had no bearing on a road being closed. Um, just very, very strange. So um, continuous learning, these things are changing pretty much on a weekly basis. Each week, one of the models out there issues a new version. And that new version may have strengths and weaknesses different from the one that is currently in place. You have to stay up to date on these.
SPEAKER_01Sometimes what the people who are creating these models think is an improvement, is not actually an improvement. So be aware of that. Verify the output. This is is, you know, there's that old phrase, trust but verify. Um I would say don't trust and verify anyway. Um, you should, if you're using anything, this generating something that might end up in a report, ask it any information it gives you to provide a reference.
SPEAKER_01And then check the reference. Go and see whether A, the link they've given you even exists, and B, if it does, is it relevant? Because I've not yet had a link that didn't exist, but I have certainly had links that really weren't relevant. Um, so be aware of that issue. Um don't believe it. Work on your prompts, we'll talk about that. Key thing here, it's a tool.
SPEAKER_01It is just a tool. Like any tool, you can learn to use it well, but a key thing about any tool is that you should only use it where it's going to work. A hammer is a wonderful tool if you are driving nails. It is less effective if you are trying to change light bulbs. Okay? You've got to use the tool in the right way.
SPEAKER_01So the prompts. This is where we're asking it a question. And there is a process that has come to be known as prompt engineering. And this is how do you ask the question in the right way? Um it's been somewhat glamorized, but really what we're trying to do here is to get our initial query in a form that gets us the best possible initial response.
SPEAKER_01It's an initial query. You don't ask one question and then you're done. There will always be follow-ups in this. And we're going to take uh a look at some of these prompts and suggestions. If you're doing a grant proposal, you might use RASEF and so on down the list there. Uh the prompts are all acronyms. Of course they are. Okay? These are geeks who put these together.
SPEAKER_01They don't actually tell you what something means until you ask. So what do those mean? RASEF means you structure your question in the format of role, action, content, examples, and how you want the output formatted. Clear, it's context, length, examples, audience, and response, and so on. Okay? So we can dig into those. We are going to have five examples of which we will we're gonna rank them.
SPEAKER_01Well, I say we, you are gonna rank these. Okay. First one, develop a winter storm response plan. Tag, go and look at the following web page and provide information related to AI. Okay. Which pages? It's the Clear Roads page. I know that Clearroads folk have been talking about AI quite a bit. So let's see what AI they already have. Craft, create a training brief for your snow plow operators before a big storm.
SPEAKER_01Clear, draft a resident communication during a snow event. And smart, set performance goals for snow operations. Are you all clear on those? And they're not in the right format yet, but that's where we can begin. Okay. Now we vote. Which of those five would you like to start with? And by the way, I think this is set up so it is not Chicago voting. You can't vote early, vote often.
SPEAKER_01Um I think it's one vote only counts. Um, but I think we're up there and oh yeah. So it looks like we're gonna begin with our winter storm response plan, and then maybe it's the training brief and setting the performance goals, and we'll we'll see where we go on those. Okay. So first up will be the winter storm response plan. Let's go to the AI and see.
SPEAKER_01All right. Now I've already got Chat GPT up and running, and I am going to go here. Oops, come on. We'll go to Chat GPT, and it says here, ask anything. So we'll put this in. Um develop a winter storm response plan, but first put that request into the RASEF prompt format.
SPEAKER_01Okay, and we'll see what it does with that. So it tells us what it's doing, first of all. And I I know this is not easy to see at the back. When it's finished, I'm gonna cut and paste this out so that we can um see the whole thing, okay? And I think I can go. So uh there's quite a lot down there. Um let's go to there. And we'll come over to here.
SPEAKER_01So here's what it's told us. Here's how the request could be structured. For this, he's using role, action, context, expectations format. He must have been listening to me because he did not begin this by saying, wow, what a wonderful idea. Okay, um, so that's good news. Um here's the role. You're an experienced public works winter operations superintendent with expertise in these things, and municipal winter maintenance planning for mid-sized communities.
SPEAKER_01Okay, so I'm asking it to be that. Action, develop a comprehensive winter storm response plan for a municipal public works department responsible for snow and ice control. Context, we've got, as you can see, 870 lane miles of street, 13 single-axle plow trucks, and five tandem axle. We've got to address multiple storm types, uh, mixed precip events.
SPEAKER_01The community includes residential neighborhoods, arterial roads, emergency routes. The plan should consider staffing and shift scheduling, so on and so forth, escalation triggers, coordination with police, fire, and emergency management, because while they call themselves first responders, they can't first respond to anything until we've cleared the roads, right? Point of order there, we are first responders. Um and resource limitations, expectations.
SPEAKER_01We want to be realistic, we want to include decision-making criteria, uh clear differential operation or differentiate operational responses, so on and so forth, include both strategic guidance and tactical actions, and then provide an executive summary, a storm classification matrix, operational response levels, so on down there. Gosh, it's providing a lot. It says it works well because it gives AI a clear professional identity, a specific task, and a defined deliverable structure.
SPEAKER_01Any and I realize this came at you quite quickly. Any things missing in there? Any questions? Seen some is that a hand in the back with a question or just a gesticulation? Okay, no worries. Yes, go ahead.
unknownI didn't know how many trucks you had. I put that in already.
SPEAKER_01Nope, it it guessed. Okay. Now, to be fair, I use it a lot. That is way more than I have in my city. I only have about 150 lane miles and I have five single-axle trucks, but it's a good number. Um, so it was probably saying, let's make this medium size and see. It will, and that raises a good point. It will remember and use what you have done before with it, unless you specifically tell it not to.
SPEAKER_01Okay, so very good point there. All right, let's dive back in and so what I'm telling it to do now is all right, you've developed a great set of prompts. Go ahead and use those prompts to develop the plan. So it'll go ahead and do this. One thing that I think is very useful about this, I did not have to come up with the specific formatting that it was that it developed.
SPEAKER_01I just said, I want a plan, format it, go ahead. At this point, if I were doing something serious, I would put in the details about my city. I'd say, no, in this section, you know, the RASF, the context part, let's use this context instead. But everything else, keep it the same and see what you come up with. Okay. Um so we now have, and yeah, it's as you can see, it's done a lot.
SPEAKER_01So let's put that there. And again, I'm gonna uh cut and paste that. Now, it's gonna be kind of tough to go through this in detail, so we're gonna skim it. Um, but it's already got a fair amount of good stuff for us. Here we go. Winter storm response plan for this municipal public works department, executive summary, it's a structured operational forecast, yada yada, gives us storm classification, response activation, staffing and shift procedures, equipment deployment, route prioritization, escalation decision making, recovery.
SPEAKER_01Now, how many of you have had to develop and write a plan like this for your agency? Any of you done that? I'm not saying, by the way, that this is in its final form. I would never say that. But what we now have is a first draft from which we can work and see how we go. Okay. I think some things are interesting.
SPEAKER_01The overall operational philosophy preserve life and safety, maintain critical transportation routes, restore mobility systematically, communicate clearly with the public, protect employment safety and equipment longevity. Are there other things that you would include in there? Let's just see if that's the whole list. Yeah, that's the whole list. What else might you add, or is that fairly complete?
SPEAKER_01I would say it's a good place to begin. After I'd gone through this and chewed through it, I might look at it in more detail, but it's definitely a good place to start. Then it comes down, winter storm classification. Okay, this is not gonna look good in this formatting, but it's light snow, the wind is low, operational risk is low, and so on and so forth. And you, you know, there's gonna be a lot of this in in a lousy table.
SPEAKER_01Umperational priorities. Priority routes must remain passable at all times. I wouldn't express it that way. Because that opens you up to liability. If you get the storm from hell and it's not passable at all times and somebody crashes, well, it's your fault. You said you would do it. Okay? So I would do some tweaking there.
SPEAKER_01Um, but my priority one routes and you know what? School bus routes don't need to be open at all times. They need to be open when the school buses are moving, not at all times. Priority two routes, collectors, transit routes, industrial access, and then priority three is everybody else. Umperational response levels, and we won't go through all of this, but you can you can see there is a lot of information in here, much of which is useful.
SPEAKER_01Okay, I wouldn't for a minute say all of it is useful, but it's again, it's a good starting point. Now, we're at the point where we should be thinking about wrapping this up. So before we dive into any more of that, are there any questions at this point about anything that we've done? Okay. They're all kind of shy.
SPEAKER_01So I'll give you one role. Okay.
RickWell, you know, one of my one of my concerns about AI is this idea of AI hallucinations. Yep. And I think you talk about it a little bit, but when the AI comes back and and it's it seems very confident in what it's telling you, and uh it makes a very compelling case, but in fact, it's not right.
SPEAKER_01Yep.
RickAnd um it's kind of, in my mind, it's kind of worse than you know Siri messing up or autocorrect messing up.
SPEAKER_01Yeah.
RickBecause if you don't have the experience to know or to differentiate what is giving you a good story or giving you something real, I think that's uh that that opens you up for a uh kind of a dangerous circumstance. You talk about with the court cases and so on, but yeah, but I think in your example where you're developing plans and things, you know, I I think that's a and the deeper you go in the rabbit hole, the more convinced it is that it's right.
SPEAKER_01Oh, absolutely. Yes. Um they very cleverly uh labeled um errors or lies of AI systems as hallucinations, because that makes it sound a lot better. It's not wrong, it's just off in its own little world somewhere. Um and it's a problem. And I view this from um my status, if you will, as an engineer. I have a responsibility, anything I produce, it is mine.
SPEAKER_01I have to guarantee it. So too do lawyers, by the way, but you know, that's on the lawyers. So I have to check it. What I also have to do is if I'm using it in an area where I genuinely don't know enough, then one, I should be very, very careful about using it, because I have no way of verifying the data properly. I can't tell if it's BSing me or poppycocking me or not, because I don't know enough.
SPEAKER_01So that's the first thing. Don't use it outside your area of expertise. Okay? Or if you do, bring someone else onto your team who has that area of expertise. Okay. Second thing, this is only ever a starting point. You can take that output, and then it's up to you to refine it. And don't have it refining the whole document. Work on little bits and pieces separately.
SPEAKER_01If you think, well, this bit is pretty much okay, put that in the Word document and keep it there, but then okay, I need to rethink this. All right, why did you say this? What makes you think that that is such a good answer? You know, and there are a number of memes uh out there about AI being wrong. Uh the one I like is somebody using their phone with AI. Is this mushroom edible? Oh, yes, it's edible, and they eat it and die.
SPEAKER_01Um and uh the AI comes back. Would you like more information about fungi? Um yeah, it's a little late for that. So, yes, we do need to use it carefully. It's a tool, it can be useful. I'll be honest, I find it very helpful for moving me off the blank page. You have to check it. You never trust it without checking.
SPEAKER_01And even so, it can be a great force multiplier, can get things going. There is another concern, this tends to be more educational, that it stops people from thinking critically. And that may be true, uh, but I think partly that's, and that was particularly in response to education. And people have asked me, well, would you like to be a professor now with AI where they can write things for you? And my answer is very simple.
SPEAKER_01If I were a professor now, I would not set them essays to write. I would give them a problem and say, show me how you research this using ChatGPT or whatever it might be. Because this tool is out the box now. We have a responsibility to learn to use it uh successfully. So I think there was another question in the back. Yeah.
SPEAKER_00If you were to ask me the same question three times, you don't get the same response.
SPEAKER_01That's a really good question. If I ask the same question three times, will I get the same response? How much time are you going to leave in between? Um and and I told you it it remembers everything you've dealt with. Okay, so if I ask it a question today, and then I'm doing a bunch of other stuff, and I come back in three weeks' time and ask that same question, I would expect to get a different answer. Um and that should make us pause a little bit.
SPEAKER_01On the other hand, if you were having a conversation with someone about something and you ask them a question, and then three leak weeks later you ask them the same question, would you expect to get the same answer? It might be generally the same, but I bet there'd be a little difference. But yeah, very good question. Okay, I think we're done. Whoops! Yep.
RickIf you like this episode, don't keep it a secret. Share it with your friends and co-workers. And while you're at it, click subscribe and leave us a review on your favorite podcast app. It really does help. You can also sign up to get all the Talkin' WinterOps news direct to your inbox by clicking on subscribe on our podcast website, talkingwinterops.com. Sign up today so you can get all the news firsthand. I'm your host, Rick Nelson. Until next time, thanks for tuning in and stay safe out there.
On this episode we spend some time visiting with Mike Chapman, Colorado DOT about the state of the practice when it comes to weather forecasting for highway operations, some new technology, and pondering the future....
The AASHTO Committee on Maintenance meets annually to discuss highway maintenance, emerging trends, new technologies, and compare notes. In 2026 this meeting took place in Westminster, Colorado and we mic'ed up Chair...
This National Briefing segment is presented by Dr. Brittany Welsh of the National Center For Atmospheric Research and Chair of the AMS Road Weather and Technology Committee. Check out their work products and get more...
This National Briefing segment is presented by Anna Arvidsson Swedish National Road and Transport Research Institute and features the Transportation Research Board (TRB) Standing Technical Committee on Winter...
This National Briefing segment is presented by Peter Nutz, City of Vienna Austria and features the World Road Association’s Winter Service Technical Committee. Check out their work products and get more information...