You're listening to a very spatial podcast, episode 768, August 31st, 2025.

Hello and welcome to a very spatial podcast.

I'm Jesse.

I'm Sue.

I'm Barb.

And I'm Frank.

And this week we're gonna talk about stuff, but first, some news.

The USGS has unveiled a new national geologic map.

If you haven't seeded, it's a comprehensive web tool for understanding the nation's geology.

And as always, it is one of the prettiest maps I've seen, the geologic map of the US.

This new map, um, was created from more than a hundred preexisting geologic maps.

Um, and it is the first nationwide map to provide these multiple layers of data in one location.

That surprised me 'cause I guess I'm used to seeing our state's geologic interactive map, so I just assumed we, you know, had that everywhere.

Yeah, but I agree with you.

I could have sworn I've already seen one, so, but they're telling us that it's new.

Um, it's, I mean, if you play with it, it's pretty cool.

It tell, it gives you a lot of information and I don't know the first thing about geology really, but there's a lot of interesting things and I'm just clicking around West Virginia.

I have some conceptualization of what's going on there, but it's, it's a pretty fascinating map and it even tells you where it's getting the information from and stuff so you can see the actual sources and where you can get more information.

I'm assuming nine times outta 10, it's gonna be some state geologic survey of some sort, but that's just a guess on my part.

The European Space Agency has developed a AI project to address satellite collisions and space debris, which are one of the biggest issues facing, you know, the geospatial community when it comes to satellites, and it is called the collision Risk Estimation and Automated mitigation or cream project.

There's a lot of stuff floating around up there, and so having, uh, a good mapping of it, is critical.

And oftentimes we just don't know what's up there very well.

I am trying very hard not to get a copyright strike by making a reference to one of several songs by the band cream.

But you go ahead and do that yourself.

Well, as long as we don't actually play their songs, you're okay.

Yeah.

Well, I, I could sing it in a hummed and that's why, but there's lots of cream songs I Anyway, that are appropriate for AI or for space.

Which one?

Yes, exactly.

And, um.

Uh, I also think that we should, um, as a collective body of humanity realize that acronyms are not always the answer for all things.

Yeah.

What's interesting about their approach with AI is that it's about the active satellites communicating with each other rather than the focus on the debris, which is what has usually been the approach.

Um, so rather than focusing on how do we handle space debris by the debris, they're actually, how do we handle space debris by, um, working on satellites, not causing it.

Recently, the Norwegian University of Science and Technology had, uh, a research project where they focused on the curricula across a number of different countries, looking at how geography was approached, and a number of things were found about how we think geographically, how we kind of begin to put this into the curriculum across countries.

What kind of things did they find?

One of the things they found is that there was more focus on.

Cramming or rote memorization than there was on thinking geographically and within the geospatial community.

We hear that all the time.

We need to be spatial thinkers.

But you know, this showed across this, this study that there does need to be a focus more on how do you create pedagogy that leads people to understanding geography in an applied way across their lives.

Yeah.

And one of the things they pointed out right about what, uh, the study, uh, researchers felt is a deficiency, right?

Is the concept of place.

And to us, that just seems like absolutely is part of geography and a part of what you should know when you understand our relationship between the world around us.

But they felt that, for example, was, uh, not really emphasized in curricula.

And the countries, by the way that they looked at, are Australia, China, the Czech Republic, Denmark, France, Norway, Slovakia, South Africa, and the United States.

So there are a couple of critical regions, maybe that, that weren't looked at, but some some pretty well-developed education systems.

And so, uh, I, for one m is looking, you know.

Disappointed to, to hear that.

You know, not a lot is thought about place.

However you could find it woven into other, other parts of the curriculum.

Just not emphasized that it's geographic, especially from the US perspective, which we're, that's the one we can talk about most directly is that geography's not in the curriculum.

It's in the curriculum, but it's not there as geography a lot of times even though it may be for that state listed as these are geography standards.

It's still taught as part of history or part of politic, not political science, social studies.

Yeah.

It's, it's part of that larger social studies class, um, that we have in K to seven, K to eight.

And so yeah, it's, I think that is where we're seeing a lot of that inability to make that next step, because it's not really taught as geography is taught as you need to know these place names or you need to know where these events happened as opposed to.

These are ideas of geography.

Now, you do get to that if you do human geography at the AP level in high school in the us, but if you can, if you can, if it's even available.

I mean, one of the things that that we don't talk about, we don't use this at all as a concept, at least maybe in K through 12 teaching they do.

But this idea of education deserts that that.

There are places that don't have that as an option.

So not even if you do, it's, you don't even have it on the table as an option.

And you know, we should think about, in the United States at least, should think about some of the implications of that.

Yeah.

And it comes back to the fact that there's not necessarily a geographer, anybody who's taken, you know, more than one or two basic geography classes in a school district, let alone in a particular school.

So it's, it's, it's a thing.

I think Barbara has an anecdote she can share today.

That just happened this week.

Oh yeah.

Um, I was like, what is he talking about?

I had a student that I taught who's now teaching in the, um, upper levels of junior high high school, and he approached me and he said, can you share all the stuff you shared with me for class?

Because our state's now focusing on world geography.

Which is really cool.

So I was gonna say, I think some things might be changing.

He, well, he was, he's a, um, a K through 12 education, social studies teacher, so it is very much in the social studies program, and he was sort of built around the notion of history.

Mm-hmm.

As.

Course social studies.

And then suddenly, I guess last year they said, no, you need to bring the geography in.

And so he, he walked up to us.

We were at a coffee shop just getting coffee after work one day and said, Hey, you know, could you share this information?

Yeah.

Help.

I don't have the, I took your class, but I don't have the background for this.

I, I, I believe this topic would be a really good main topic for us at some point, and I think we should put it on the list of things we need to talk about.

Because there's a lot to talk about here, but this is a really interesting news item.

Well, I, I love, I, I love the title where it says We need to have less fact based and more thinkers.

If we do have that conversation, we need to have at least one social studies.

Person in that conversation as well.

Yeah.

Yeah.

And I think it goes along with the, I, I can't remember if it was the American Association of Science or a a g that recently talked about the need to create a bridge between high school geography teaching and college university teaching for Geograph.

And that's it for the news.

In the web corner, if you love the True Sizes app, if you've used the True Sizes app where you could overlay a country over another one to see the true size or the projection of it.

ESRI has developed their own, of course, being ESRI it's a lot more intensive.

And it's fun to play around with.

I thought that this was something they had earlier, but I, they've updated it.

In August, 2025, the world Imagery.

For the sizes, this is clearly using a, an A scene, an arc scene.

And it's pretty cool because you essentially, you click on a boundary and go, okay, I want to use that and move it around and slide it to where I want to go.

And you know.

It's, it's, it's neat.

It starts in the United States, obviously because, well, US centric co, uh, company and, and also we screw up boundary sizes a lot in the United States.

But it is pretty fascinating to play with this, and I think in some ways it's a little more powerful of a tool to give to students than some of the images we've traditionally done where somebody's done that for them.

So I, this is gonna make it into our cu it's gonna make it into my version of, of our curriculum At the minimum.

I suspect it'll end up in all of our curriculum.

'cause it's, it is kind of a cool way to explain to people the McCat isn't the end all, be all the universe.

And like many things with ESRI apps, this might have been around a while, and it was just that news item brought it to my attention and was like, Hey, you know, let's put it in the web corner.

And like a number of them, there's like.

No instructions.

No, no.

You said you're gonna use it for something.

Make sure you play around with it first.

Figure out how it works.

Yeah.

Click and do, don't read.

Yeah, that's yeah.

Fun part is Greenland.

This week we're gonna talk about geography and ai.

We threatened it, I think last week in, in the episode.

'cause there were a couple of AI things in, uh, the last episode, including a discussion of, well, we talked about Gian splats, but really was the natural language search and map apps that kind of led, I think, to this conversation today.

But I, I think.

So do we wanna start with definitions?

Yeah, I think you have to.

I think you have to start with defining ai.

Well, I, I, I, I'm not gonna try to define AI because there's so many different definitions of it, but I'll define the different parts of AI so that whenever we're talking about it, we, we will use those terms.

We have machine der, machine learning, and deep learning, which are the traditional ways that we have used AI and what they're currently being called.

We can talk about expert systems, neural networks.

Many of those are now built into deep learning algorithms, and there we're, we're doing more with image based for the most part.

Is that safe?

Yeah.

Well, I, I, I'm gonna go just a little bit, if I, if I can sort of exclude pieces of the definition of AI is, um, going back to at least, the three laws of robotics in science fiction.

We've been talking about AI in science fiction for a really long time.

It predates that even.

But that's, that's usually the place where a lot of people, you know, think of as a delineation.

And what we talk about in fiction is AI is really, I wanted to say it's really fantasy, but the point is, is that we're not there.

Uh, a lot of things that Jesse just mentioned is really mechanical pieces of like how it happens, and I would liken this to a per perpetual motion machine.

Someone essentially saying, well, we have machines.

Yes, we have machines, but we don't have a perpetual motion machine because it's a lot harder Get, it may be actually impossible, we just don't know.

So I, I wanna have in our definition an exclusion to say what you think about in science fiction that we're nowhere near that.

So don't, don't have that as your frame of reference when you're thinking about AI at all.

Much less in geo.

Uh, I would, I would counter because as whenever we talk to Elvin a lot, you know, all of the things that became a lot of the mobile GIS were influenced because of science fiction.

So people are reaching for those things that they saw.

No, I'm not saying that it's not relevant.

I'm just saying that's not, we're not there.

That's the North star.

We we're never, arguably, we'll never get there in our lifetime.

But it may be what helps you navigate the, the, the landscape.

But it's just, we're not there.

That's not what we're talking about when we talk about AI in the year 2025.

Okay.

And so within that machine learning, deep learning, the other side is generative ai, and these are just the ones we're talking about.

There's different aspects of AI that we're not even too far into in the geography geo area that we won't go down.

But with the generative ai, we're generally talking about using large language models, maybe visual language models to basically interact with, um, more often, not right now, they're being used as.

Half of a chat, you're having one side of the conversation, it's having the other side of the conversation, and that's kind of what's going on with it.

So those are two things, and maybe it would help it if, if you could, to kind of briefly explain just what a large language model is.

I find that's a sticking point, right?

When you try to say, you know, when you're working with these things, they're, they're not truly artificial intelligence.

They're these large language models, and then you just get a blank stare.

Like people can't evaluate what that means.

In terms of a cautionary, I'll give it.

I'll give a lay definition of that.

All right, so essentially what it is, is that there's all these algorithms that were developed largely in the eighties.

Some predate that a little bit, but they essentially said, you know what, if we know everything that was ever written, and at that time, obviously it wasn't in digital format, most of it was in in analog format.

We're very used to this in geography.

Most of the information we have is in a paper map somewhere in a drawer somewhere.

But if we had everything digitally and we could look through it simultaneously, and we had ways of understanding how one thing written in one book in one paragraph relates to another thing written.

In a white paper that was published by government agency in a completely unrelated paragraph, we could go, you know what?

These two things are related and therefore they matter, the connectivity.

And we can then infer in theory that other pieces of that are in fact related.

And we start saying, well, maybe these connections relate to this third thing, fourth thing, 10th thing, 6000th thing down the line.

So it's essentially, you can think of it like.

Not dissimilar to a, a research, uh, um, a paralegal that goes through and researches laws and say, well, this law depends upon that, depends upon this, upon the other thing.

Or you can also think of it like a search engine that we'll go through and just go, let's go find all those connectivity over space and time.

And I think this is where we have to go into another term that's, um, which is, you know, open source, open science, open data, because where it's being trained is from a lot of sources, and the spatial community generates a lot of this, um, which is openly available, created by volunteers, and then openly available to the world.

So if you're thinking about the, the pot that it's.

Taking from then, a lot of that is going to be that open source, open data, so I'm gonna follow up on both of those first.

Of course, we can get into the whole copyright issues and the fact that a lot of the large language model companies are sometimes paying, sometimes not paying for access to non-open data.

So the internet's generally safe unless you market as non readable by robots, which should, in theory, keep it from being read.

But there's also, uh, yeah, the other part of that to Frank's, I, I, he's, he's kind of tiptoeing towards, he's not getting too agentic, which we'll talk about later on, but he's tip touring to closer to it.

To simplify it a little bit, it's not necessarily so much about relationships between different documents, it's the relationships between words, because it's not coming up with big ideas.

It's actually usually generating everything that you're doing word by word, pixel by pixel.

So it's saying, okay, well commonly we find that these two words are in this same order, in this context.

And so that's how it's writing.

It's.

Going through and doing this word by word usually I was, I was trying to simplify the concept a little bit, but yeah, you're right.

And at the at, at the rubber hits the road piece of it and the history involved here, essentially what happened was the conceptualization is what if we get these concepts?

It was very, it was at concepts, and then at some point they said, well, how do we figure out a concept?

A concept is nothing more than a string of words put together.

Right?

That's a concept which I would dispute that that is.

That is one, and I'm not sure even the best way to talk about a concept, but that was one way you can talk about concept.

If I take these words in a sentence and I disaggregate them and I go, well, you don't see the, the, the the in, in, in, in ever.

So that's not relevant.

I need to look at words that are together and say, well, that is an idea.

I can link it to other words that are similarly together and say that is also an idea.

The two things that, again, this has been around for 40 or 50 years, conceptually, maybe even longer, but certainly in terms of the algorithms, the things that it was missing was one, a really big data set that was digitally connected.

We like to call that now the internet.

So we have that, we have that data set that we can get to, whether you include or don't include everything legally or not legally, ethically, or unethically.

That's a whole other, like you said, minefield ball, wax.

We have to contend with, the second thing is, is that we need really, really, really powerful CPUs to pull this off.

We needed all these CPUs.

Well, invention of the GPU allowed for amazing amount of computing power using a different set of architecture.

So now we have all these GPUs out there that we can in a very small amount of physical space, we're not even talking about, you know.

Uh, it, it's literally physical space.

We can get astounding amount of computing power in a, in a lot less space than we ever could before.

So therefore, we can build these really, really, really, really big computing engines, these CPUs that act together and parallel and work together.

So we have these two bits of technology that exist that allow us to do it.

So now we're doing it.

What I don't think we've done that we think we've done is actually link concepts like Jesse's saying, we've just linked words and typical articulations of an idea and saying, well, that seems to be a consistent idea, and that also seems to be consistent idea.

And you say, okay, ai.

What's it about?

You know, I don't know.

I don't think in ideas.

I just think about consistent linkages of things, and these are consistently linked, therefore, they must be the same or similar things.

So taking that.

Well, let's kind of take a turn then towards the geography.

How do we then use these things that are linking things which we would like to think are linking concepts, but eh, and use them to help us in geography?

And there's whole discussions on this as spatial ai, as spatial worlds.

It's GOI.

AI and geography, all of these terms are being bandied about.

So what does it, in, what does any of it mean for your day-to-day?

Well, I was gonna say, if you go back into the timeline of ai, and you were mentioning machine learning surprisingly in or not surprisingly, because it deals with big data and geospatial big data, and it deals with linkages and connections.

The um, geospatial community was in there very early on.

Because we could use it to handle all of the data that we have that's being generated in order to handle things at different scales.

So you see countries using it as we saw in a one of the episodes where you had in the UK them using it to identify the bog.

You had whole countries using it mainly for things like land cover.

Um, so it, it shortens the time.

Um.

You have to do things and then you do have to go back in and ground truth, but you don't have to take so long to do the work that you would normally do.

Uh, and if you saw the plenary from ESRI, you saw the whole big vision that ESRI had that went from, you know, earth rise to at the end, this giant ai, you know, is the future because of its ability to, to help us do our work.

More efficiently, faster, and in some cases to be more accurate then, then you're stepping, just to make it clear, you're stepping into those conversations about generative ai.

Correct.

'cause that's where you're talking to it and having that chat on how do I do this?

Right.

You're moving from that machine learning into that generative ai.

Yeah.

And those of you have had an intro to remote sensing class or had geos stats classes whenever you're looking at.

The results of a supervised or unsupervised classification, that's ai.

If you're looking at, um, some of the correlation models, those are built essentially in different types of ai or actually the same type of ai, but just different algorithms that they're focusing on.

And they've been, you know, pre-trained and those type of things, which is another key to either the machine learning side of things or the generative side of things.

Is that they have been trained and a large part of the large language model is that it was trained on these large amounts of text or video or audio or whatever it's building on.

And I'll hand it over to Frank to finish that idea.

And not to finish that idea, I was just realized we have need to have another definition of training, right?

What do we mean when we say training?

And one of the reasons I think that this notion of inference is an important piece is because the idea behind training is that I give you all the things that I know about.

That's the big internet, all the knowledge of humanity piece, right?

I go and I say, this is everything we know and we, these are all the connections we know are true.

That that is definitely that, um, hear and fireplace.

This house are the same thing, that that's what we're doing when we're training.

We're telling you those connections.

And in theory, in the ideal, it's the machine then says, well, if this is this and that's that, then this is that.

And it can infer that those connections must also exist.

And to some extent there's a truth to it that it can do that.

It works so much better on ways that we've been doing for, I don't know, how long have we've been doing remote sensing, you know, classification trained, what, 40 years?

Something like that or, yeah.

We'll, we'll call it, uh.

60 for our first digital attempts.

Yeah, so we've been doing this for decades in the geospatial realm and it's very easy because it's kind of mathematically is or isn't.

So it is pretty easy for it to go well, you know, in all probabilistic model.

If you say, if you're 90% certain I'm okay with it, then yeah, okay.

We've been doing this for a really long time.

It gets a lot harder to do that whenever you're thinking about, I love my dog, does that mean I love any dog?

That's a harder.

Conceptual thing when we're talking about human emotions, stuff like that.

This is where large large language models are attempting to move towards.

Is these conceptual ideas, are they, are they, are they connected?

Are they not connected?

Uh, and we certainly debate whether they work or not.

And whenever we were first, that's what training and whenever we were first a podcast, we talked a lot about things like semantic search and, and those ideas, and those were earlier.

Iterations of what we're talking about now with the LLMs and whenever we've talked about you know, pulling out things like emotion from a certain text.

Those are LLMs that were being used on specific texts or specific group of texts.

So we've talked about these things in the past, it's just in the last few years.

The te the training data sets got so much larger that we just jumped a few yeah.

Iterations or scales of magnitude in terms of what they could do and what we were using them for.

Well, and I think related to that too, is one thing, and, and I, I mean I admit to, to being super surprised by right, is how quickly not just.

Not just the evolution of the, the models themselves and, uh, you know, the applications that, but just how much it would explode.

And, and we could, I mean, we could probably give this parallel right to the, the seminal event of the release of Google Maps, right?

Is that I was surprised just simply by just a massive explosion and uptake, um, because of an interface that made almost anyone be able to access.

Access the model and use it for things.

And so that, I think, has been that fast explosion, right?

Of the, the chat GPTs, the geminis, the clauses, whatever it is.

It's the interaction mode, right?

Making it this chat.

Uh, and, and a few others like that.

But, but that's where all the stuff behind it kind of disappears unless I guess you interrogate the model itself and say, well, how do you work?

Or something like that.

But, but that's something that, that I'm still wrapping my head around.

I think magic and a lot of people, even the people who are making them are like, it kind of is because there are certain parts of LLMs that nobody actually understands how they're working.

Even the LLM, whenever you ask it doesn't exactly know how it's doing certain things, but.

It kind of is doing it, whatever.

It's, and, and if we get back to the, the geography heart of it, it the generative AI is a very hard time with geography for the very reason in that national geography curriculum study that we have a hard time really explaining place outside of geography.

Um, yet, you know, places.

Everything.

But it's something that's very difficult to be captured outside of generative ai.

So why is it not surprising that it would be difficult to be captured?

Yeah.

Or even, and this goes back to our, our previous discussions.

Was it, uh, meta card or we were talking back in those days about these types of search, right?

How do you define near something very simple?

And, and I sometimes use this to, to kind of, you know, make students think about.

The uncritical acceptance of results of anything.

Right?

But in this case, like a, you know, a, a chat with a a generative AI model, you know, to say, well, it's gonna tell you something, but if you don't even know or cannot define it, and we can't right?

Everyone's definition of near, then you have to, you have to critically look at what you're getting back as results, right?

So, so it's just a very, very simple.

Not even a, a collection of words, right?

But just one word that does not have a single definition.

And yet it's a huge, you know, a hugely important concept when we're looking at space, right?

And we're trying to understand everything from, you know, neighborhoods to connections between ecosystems and all that, you know, or spa.

Lot of correlation for that matter.

Right?

What is near?

So I think it's, I think it's an, it's almost a, an overwhelming dilemma sometimes to think about.

How to move forward with a critical use of it.

I think one of the, the tricky bit here is that generative AI hasn't figured out how to ask questions insightfully and ask questions, but they're usually, I need just a slight semantic understanding.

And then you go, oh, I mean this.

I go, okay, I'm wrong with it.

So if you ask where do you live?

It's a very simple question.

A lot of people ask it every, I mean, probably millions of people are asking at any given hour.

Where do you live?

It's a actually a lot harder question to answer if you ask me Where do I live?

I live in Morgantown, West Virginia.

No, you don't.

I don't live in Morgantown, West Virginia.

If you look at the political boundaries of Morgantown, West Virginia, I'm nowhere near the political boundary of Morgantown, West Virginia.

I don't live in the town, but my address says I live in Morgantown, West Virginia, and I conceptually think that I live in Morgantown, West Virginia.

So do I live in Morgantown, West Virginia?

That's an an objective question that we could, you know, look at because you look at the boundaries, you go, you're in, you're not in discussion.

But even that doesn't get generative AI to go, okay.

What do you think about when you think about where you live?

Like what, how are you framing that statement?

And since it's not, it's what we do in academia naturally, right?

We say.

So I'm gonna make this statement, this thing is true, but by that I mean blah.

And by that I mean blah, that's something we're used to having to do in academia, that it feels like the generative AI now hasn't quite, because it's designed for a, a, uh, a general audience and the general audience isn't, doesn't think that way, and they don't wanna communicate that way.

And it's really annoying and it annoys me when we do it in academia, so I totally get it, but sometimes it's necessary.

To say, well, what do you mean by place?

What do you mean by near?

What do you think of as near?

Where do you live?

I live near, do you live near Charleston, West Virginia?

If I'm in West Virginia, no, I don't live anywhere near Charleston, West Virginia.

If I'm in Sweden, yeah, close enough.

That's about right.

You know, that's.

The distance where I'm physically standing when I'm asked the question is going to impact the answer to the question, is this near there?

So it, it's a really complicated thing and I feel like AI hasn't gotten to the point of being able to start interrogating back in a meaningful way to say, I need to understand how you understand this.

I need to understand how you want to think about this.

And then we can start talking about how we're tying AI to data.

And there it's usually, you know, it takes and finds one and runs with it.

It's not, it's not giving you options of which data you wanna use because are you using the metropolitan statistical area in which you are in?

Are they using the specific political boundary which you're not in?

Are you in one of the zip codes?

Yes, you are.

That's, that's not even getting into perceptual questions.

That's just straight.

I have multiple data sources I can choose from to determine whether or not you are, and should I check one of them?

All of them.

It's not asking, and that's where we can kind of come back around to, it's a puppy, right?

It wants to make you happy, it wants to give you an answer.

It wants to give you a simple answer.

Um.

And so it's not always going to necessarily get things right because it just wants to make sure that you have what you want.

Sometimes that's not correct.

And, and this is an interesting thing that it also does is back to the semantic web.

You, you brought that up and to me, this is the, okay, this is the, for geography, it's the, you know, brass ring.

If we can get that working, then AI becomes a hell of a lot more useful.

The problem is, is it's almost impossible to get that working air quotes.

You can't see it.

It's a podcast, but I'm doing air quotes around the word working because it depends on who's asking and their point of view.

What purposes they're asking.

That's gonna have a lot to do with how you semantically interpret a connectivity within a geography.

Part of the challenge there is that the Geos mantic web doesn't exist and it's hard to construct.

And this, I think, could help us start to get at that, but only if it can also collect data and sort of store it.

And it doesn't seem to necessarily do that very well.

I could be wrong, but.

It stores information about the project like that.

If I, if I set a project in chat, GPT as I'm making a thing for a this and I go, well what is this?

I go, well, what is this is the organization.

I go, okay, I got it.

And it will remember the context of the organization, but it doesn't seem to create new, I don't wanna call it logic, I'm wanna call it facts.

That it can go, oh, okay, wait a minute.

You asked on this date, this, and you said these two things are connected and.

7,000 other people said the same thing.

So maybe those two things are in fact connected.

I'll store that knowledge over here.

It doesn't feel like it does that efficiently.

It seems like it goes back to sources more than not that more often than not, one of the reasons you can't think about AI in the science fiction terms is because it's not creating new information for itself.

Not even knowledge didn't even get there, new information for itself to say, oh, okay.

A lot of people think these two things are.

Close enough connected, even though it's not in any of those sources I can find on the internet.

Okay, let's just assume they are and then make decisions and inferences and, and responses based upon that being a quote unquote fact.

But it doesn't do that.

Well, I actually for for when you, when you said you, you thought, you know, you kinda lost your train of thought for a second though, but what you were talking about there is the thing that, that, first of all I'm gonna preface by saying is terrifying to think about, but the idea that working with these models and their interfaces, right, this is back to the interface, is if the interface can discern.

A little bit about your perspective and who you are that would be a huge leap.

Again, a terrifying leap, but a huge leap because the questions that, that we would ask in using it right as the user the perspective that we bring to that is super important to what comes out of it.

And so here's the thing that's kind of, I guess, a little scary, and I've seen it mentioned a number of times now, right?

Is that as we're adopting this.

Those of us who are already experts in a field or have a, you know, a more mature, say, professional life or interaction with, with things will not just frame our questions differently to the, the system maybe, but also be able to more critically interpret the results because we have knowledge that predates the use of the system.

But as more and more people who are, let's say, are younger or are coming through, are using it earlier on, they aren't gonna have that critical knowledge.

'cause you know, like a lot of us get on the internet and they, we look at things and we see stuff from, from, you know, the, the chat bots coming back and we're like, ha ha, that's funny.

Right?

That's stupid.

Uh, you're, you're an idiot.

You know, like, you don't know this because we have the ability to discern that.

Uh, but those who are coming in and using it more, uncritically just because they don't have the previous knowledge background, it's gonna get harder and harder for them to be able to interrogate it.

Right?

And that's, that's adding to the training and the learning of the model is when it gets interrogated back and like, no, that's garbage.

Look for this.

Or, you know, this is how you need to approach this.

Right?

So expert users all over are, are helping to refine the training.

But, but, but the thing that it needs is to be.

Better in interacting.

I think, and again, this is interface, this is not what some models are doing behind the scenes.

This is just talking about interface.

But that improvement interface is gonna require continued interaction from humans that can say, because this is what we did with our, to go back to, uh, the remote sensing example from Barb, right?

Is that we did the super five classification, we did all this stuff, but we trained the models to be better at it because a human intervened.

To say, actually no, that cluster pixels is not an agricultural field or whatever it is we did, right?

So we, we intervened, or even when we did supervised classification, it's more, more overt, right?

We intervened and said, you're, you're correct or you're not correct.

So it's kind of that writ large and, and you know, Sue, when you were talking, I was thinking this was something I'd been thinking about, that when you use geo ai, that there is a level of expertise involved, but you know, there are.

Usually geographers involved.

But when you're talking about the generative ai, sometimes it's not just people, it's organizations and companies that are trying to solve a, a spatial problem, but they're, again, not involving the spatial experts in it because they're getting answers that they think are correct and they don't know that the answers they're getting are, you know, gonna cause more problems than they're solving.

Um, but they're out there selling things that are a solution that, you know, might not be a solution because they haven't had to reach out to the experts within the spatial community.

And, and as a slight counterpoint to both of that, what, what I would say is, is that one of the dangers that runs in, even if you do have an AI that's smart enough to go, oh.

Sue Bergron has a PhD in geography and she focuses on the idea of place and she says, this is a truism, therefore it must be a truism.

I'll accept that as a fact as opposed to Bob's, you know, fish bait and tow shop saying that this is the expert.

But what you do run into the fact that, you know, think about.

And these are for, if you don't know the history of geography as a thought process.

So basically up until mid sixties, somewhere in the sixties the dominant model after World War II of thinking about the way geography was correlations, let's just correlate the hell outta everything.

And if it's correlated, then it's a true, it's that's true.

And if it doesn't, then it isn't.

And that changed.

But what if we were doing generative AI then using that?

Sort of, this is just how we think about the world.

And that's just true.

And somebody who comes along later and says, well, I think that there's a little more nuance.

There's something else going on here.

We're not capturing with that.

And yeah, I think you have to ask the question as much as you want.

You should respect experts.

You also have to ask the question of, is that just how we think about it now?

Is that, is that an a?

Where we're at in our intellectual thought processes, and we haven't had the ability as of yet, to run into something and we go, well, wait a minute.

We haven't quite got it right.

And it's really group think is what I'm, what I'm talking about here.

I, I'm thinking in particular that I'm so outta my depths on this.

This is, I'm, I'm, this is a bounce this to various space.

The south, as we were talking about archeological stuff, there's, you know.

A large body of literature that tends to say, oh, this is generally true.

But then you'll find something and go, wait a minute, that doesn't jive.

And somebody go, well, maybe this, these facts could mean this instead.

And it takes a while for that thought process to be accepted, interrogated confirmation evidence to happen, all that sort of stuff.

And I feel like somewhat, the ai, even if it did pay attention to experts, it would run the risk of.

Stating that is absolute truth.

When in reality is, is that we know as academics that look from this perspective, we think this is the most likely scenario is verbiage.

That means we don't know.

No, it's not an absolute fact of the universe.

It's our best guess given the evidence that we have and we feel fairly confident about it, but we're open to change.

Yeah.

No, I, I, I would totally agree.

Right.

And that subtlety, that subtlety of language sometimes that, that allows, that Right.

Is just one example of, but our ability to read the subtext right.

As, as maybe not experts in the field.

Right.

But the, the subtleties of language, the large lightning models we have are, are not quite there yet.

And as you would expect, right.

That process, right.

This, this time period of quick adoption and you know, and, and I liked how you went over some of the technology, right.

That has converged to, to let it be this moment in these last couple years.

'cause you're like, why suddenly now?

But it, it, it's, it's all of that, right?

It's connectivity reach a certain point where we can have internet speeds and internet reach, right.

And all of those things and, and.

And a significant portion of populations using devices that can interface with it, right?

So all of these things kind of came together.

And, and so, but it's just a flood of information going back and forth.

The rapid adoption of these, and, and I don't know numbers for some of this, but just like, you know, we talked about this with mapping too.

Like it was like amazing, everybody used Google Maps, but, but.

There are these caveats, right?

That now you gotta worry about actually these kinds of things, right?

So, so the usual sort of pushback, but, but I think that it's hard to see in the moment where we go next, I guess.

And that's for me, right?

Because I'm not, I, I'll, I will, uh, give the caveat that I'm not much of an AI user.

In the current, you know, large language model interface, sort of shattered face.

I don't really do it outside of coding.

What's that?

Outside of coding?

Yeah, outside of coding.

So I'll ask some questions to get coding and, and the best part is they'll tell me the six steps ahead of, you know, using my script and I'm like, you know what?

I could have written the script myself.

What I really needed to know is, I don't know how to do step number three in your preface to put this script on your object.

Right.

So it's interesting that I find it.

Less than useful, uh, for that reason.

But but I think that that just the, the flood of things going through and not really knowing what the, the language models are.

Learn learning, right on this just flood all at once from all kinds of users.

Uh, it doesn't, you know, a anyway, so it's, it's, I think.

This is kind of a rambling response 'cause you could tell I haven't organized my thoughts on it, but to your very good point, Frank, about, you know, sometimes what we know now isn't what we're gonna know later.

And even saying, well this is expert knowledge now isn't necessarily gonna mean that it's infallible.

So.

Okay.

I'm just gonna, I'm gonna cut you off and just take over now.

Okay.

Because you don't know where you're going.

Yeah, I did, but, huh?

I did not know.

Yeah, so I think one of the things that we have to keep in mind is the large language models are built on now decades, centuries of text and, and content.

And so they have.

Disagreeing information in what they're building from.

And so this is how we end up with different responses and different systems and different responses from the same system.

They are as confused as we are about the world, and so we have to be careful with how we use them.

And that's of course where the big question that Sue is really going for is, is that people aren't necessarily being careful with the systems.

So, so I just wanna give an end to, it has nothing to do with geography really, but it, I think it does a really good job of highlighting how you have to think about, um, that, that contradictions and how that can play out.

One of the things I recently said in, uh, a professional development about AI that we did on our campus, I said, have a, what you should tell your students to do is have ai, ask the sa, ask the same question of AI a different way, and figure it out if you get the same answer.

Because that's gonna tell you a lot about what you can trust or not trust.

It's a very simple method.

I was playing a, a tabletop role playing game with some friends of mine and the game is called Traveler.

It's a science fiction game.

It's like Dungeons and Dragons for those who don't know what those type of things are.

And it has a thing in there for sis.

It's a little known rule that's not particularly well developed and it just so happens that one of our friends wanted to do it.

So okay, we said, let's do it.

And so we asked AI because the rules as written in the version we had were a little ambiguous 'cause they hadn't really thought out that part of the game.

And a little bit confusing.

So we asked AI to interpret a thing that we were having trouble interpreting, and it said one thing based upon page blah, blah, blah.

Well, that's not the page in the book that we're using.

So we said, okay, we're gonna use this edition.

'cause there's multiple editions of the game and went, oh, well, okay, that is blah, blah, blah, based upon this page.

Well, no, that's actually a different edition that it was referencing yet again.

And we said, no, no, no.

You need to reference this specific page on this specific, uh, thing.

And anybody who's talking about it went, okay, blah, blah, blah.

Based upon this form we found.

Okay.

But that form specifically references an earlier edition that we've already told you that's not what we're interested in.

It ended up being effectively useless and we had to make our own interpretation.

But you know, it's a non geography thing, but it is, it, it, when it doesn't find exactly what it's looking for, it fails back to a point where it goes, well, this is what I can absolutely find.

But it doesn't have, it doesn't, it didn't present anything as a caveat.

Look, I know you asked for this, but this is all I can find.

But it references that, do what you will with that information.

We didn't get that.

We got, look, this is how you interpret that.

You say, well, how do you know that?

Oh, I got it from here.

That's not how you interpret it.

It's the wrong location.

It's the wrong thing.

It, you have to know that going into, to dealing with this, like, how do you know this?

And many times it won't necessarily tell you how.

It knows it in a, I'm gonna use the word honest.

Maybe that's not the right word way because it will get conflicting information.

It'll go, uh, that one.

'cause the probabilistic model tells me that's more likely, so I'll pick it.

So what you're asking it to produce is its metadata in a, in a, in a very real way.

Yeah.

Your sources.

But yes, your metadata.

So the conversation I intended to have included things like esri's added, uh, generative.

LLM capabilities into documentation and into coding.

And you can find lots of tools available, uh, not just from machine learning, but generative as well for QGIS and other applications.

But we didn't make it into that part.

Go out, look at 'em.

Maybe we'll talk about 'em someday.

But what it comes down to is that we have apparently a lot to say about ai.

Even those people who don't use AI a lot well, you have to confront it.

That's the thing you do.

Uh, and I, I see it as confronting.

Well, it's, it's, in some ways it's not, I don't think it's the same scale.

I don't think it's the same scope, but I do think it's not dissimilar to the internet.

When the internet came on board, you couldn't just go, ah, I'm not a should.

It just, it's, it's too big and too pervasive for you to not at least think about it a little bit, figure.

To confront it, if you will, but, or incorporate it how, whatever verb you want to use there, you've got to, you can't just go, ah, don't care.

I, I think I would use, you've gotta get your, your hands on it.

You, you've got to, to try it out, use it and inform your own approaches and opinions on its, um, there's a lot of potential there, but it's, it's gonna be dependent on people exploring.

I can't remember if it was in the episode or not in the episode last time, but Frank mentioned that really where we're at right now is at the same place where the GIS and society debates developed in GIS.

Uh, it's a shorter timeline, of course.

Well, not really.

No.

Not but 40 years after it started.

Yeah.

Yeah.

Um, but we're, we're to that point where we.

We need to get to these issues on how this works and how it's going to impact society.

'cause it is having a much broader impact, I think than, I mean, maps had a, a fairly significant impact as GIS and remote sensing came online.

It changed the ways we use maps dramatically, but I think this is having even more of an impact on literally how people think.

So we'll see how it goes.

So I'm, I'm gonna propose that we, we treat this ai discussion as a series, and this is part two.

Either the part one of the series or part two.

I, I think there's, like you said, there's a lot to talk about and there's a lot more I want to talk about.

So hopefully.

For everyone else.

We'll do more on AI in the coming months and year.

And by then I, I have broken down and signed up for our university's AI training, just so I can hear what they're saying.

So it's, it's by then I will, I will say different.

We, we've had more in-depth discussion today in this conversation than you will have the entire Oh, but I will have done it.

Five minutes of this conversation, done it more in depth than.

The, all four of those.

But again, they're meant to help those people who are, I'm not, I'm not truly an, I don't think I'm anti AI or anything like that, truly, I just, in my, my life and workflow, it hasn't, I haven't seen it as something that I potentially wanna explore using.

So I just have nobody, nobody's, you don't have to be, yes, I know an ai cheerleader.

I'm not an AI cheerleader.

Um, but I, I have to talk about in class.

So, as I was saying, no events this week because we've now already talked too long.

But if you do want us to add your event to the podcast, send us an email to podcast@veryspatial.com.

If you'd like to reach us, individ individually, I can be reached at sue@veryspatial.com.

I'm barb@veryspatial.com.

You can reach me@frankenveryspatial.com, and I am still on the social medias at NoJa par and I do in fact do things.

So if you wanna follow me, you can.

I'm available at kind of spatial, and of course, you can find all of our contact information over at very spatial.com/contact.

As always.

We're the folks.

I'm very spatial.

Thanks for listening.

We'll see you in a couple weeks.

I know.

Why do I doubt?

And you say it's all for love.

Do.

Still feel so lost.

Why do I cry when, when?

And you say There's room for me in the broken, we free.

And you say you'll always be.

Loving arms that carry me.