So the geospatial revolution continues, a new day is dawning in GIS, boring tasks such

as looking up latitude and longitude and applying drop shadows to every bar in your Adobe Illustrator document. Those days are numbered. So yeah, let's go for a round of applause.

So also, the days of giving up while making a map are gone too. So raise of hands, who here has given up, it's okay, don't be shy, I've given up too. It can take a long time and we're here to change that, I think we're all here to

change that. We're just playing a small role in that. Millions of user interactions are currently untapped. It's sort of the new gold or oil or whatever they call it these days, sort of like big data. And there's millions of interactions every day through GIS software, but many of those

aren't tapped. They're not captured, they're not learned from. And one of the things that first draft does is it learns from all those user interactions. Mountains of data. So unlike Smog's mountain or the Lonely Mountain in Lord of the Rings, our data is open.

We're open. And there's just so many great data sources out there. Just to name a few geonames, Wikipedia, OSM, OSM names from OSM. And then, I mean, all the geonodes and I'm probably missing a few, so sorry.

But there's so many out there. And so what is AI? I know this is a big question. Are they doable robots? I think you guys all know they're not. So this is AI. It's a little disappointing, but it's better than it looks at first glance.

So this here is the table used for linear classification. It's just a bunch of weights, but there's a lot of models out there. There's random forest. There's night bays.

There's recurrent neural networks. I know you guys are all here to hear about recurrent neural networks. So we can guess that in the Q&A. So what can AI do? And I'm sure this is super controversial. So if you want to slug me in the face, I'd appreciate it if you would until after the presentation.

So what can AI do? So right now, these are things that are done with like higher than 90% accuracy. So geotagging tweaks, finding places mentioned in newspaper articles or reports, sort of cleaning Excel spreadsheets depending on the situation.

Like if someone has a typo, you should be able to catch that. And autostyling maps. That was a reference cardo. I saw a blog post on autostyling. And then as soon as things are coming in the near future, and we should all be prepared for, is a speech recognition, talking to your maps, creating maps by speech.

And there's like this speech where you talk into the API, synthesize voice, and those are likely in your Chrome browser already. So we just have to tap into those. And detecting features in imagery. I know there's already work in that, but with the space net challenge and all that, there's so much more work to be done in order to detect imagery at the same level

that sort of like OSM has it. And then like way far into the future, I don't know when this is going to happen. I'm trying to understand user requirements. I don't even think we can do that now.

Building community. I mean robot communities aren't that strong right now, but I mean bots are coming, so maybe. And then making the final draft, sharing the map, understanding the audience. Those are all things I don't think AI can do right now. But let's talk about it in Q&A.

So how can we help our first draft GIS? So with first draft, you can give it like a disgusting spreadsheet. Sorry for my language. And then you'll get a map in return of the places mentioned in the spreadsheet. So I was working for the Red Cross back in the day, an intern under Dale. So shout out to him. And we got survey data and had a big map split.

All right. So then there's PDFs, you know, just like a PDF report. I don't think PDFs are dying anytime soon. So you can make a map of places mentioned in that.

And then like New York Times articles, Washington Post, Ars Technica, whatever you read, you can make articles. All right. So this is the big picture. The old timeline of making maps is on top and the new timeline is on the bottom.

Just like the boring stuff always takes the longest, just looking up latitude and longitude, just correcting typos. I mean, it really isn't fun. And we're just trying to play the boring stuff in so everyone can focus on the fun stuff.

Okay. So now I'm sort of diving into the deep end here. We use multiple classifiers. I'd love to talk to people about this. We found that we couldn't do everything with one classifier because different users will use GIS differently.

Some people are focused more on local stuff like where are the bars around me. And some people are more like where are weak warehouses in Somalia. But they both might be in Boston at this time. So on the left-hand side, if you want to look for local stuff, you're going to want to make sure it's in your same time zone.

And it's probably the most important thing to probably explain. If there's two places called Lucky's Irish Bar, you're going to want to get the one that's most popular.

So that's sort of like the most important. And then on the right side, that's the global scope. There's more columns than what's displayed here. The importance, the third to bottom, that's using the importance from OSM names. So great job. They really did a great job there.

And then look at the standard population and stuff like that. Okay, so how do you actually use this stuff? You just plug in, download it. It's experimental. You install it, and then... So if you were making a map for Gary Johnson,

he was running for president a while ago. I mean, but not if he knew where it was, he'd win. I don't think so, but we'll never know. So if you just type in where is Aleppo, and then you click run, and then you'll see in the center of the screen the red dot. That's where Aleppo is. And it can take in text or links to articles or documents.

So that's the URL. And I guess these are reported, so you can get it. The important one to learn is the first strategy I asked. That's the ORP. You can see all the refills under that.

And then there's a pet Python API that's installed. And then return a key symbol. So if you want to know bars near Cambridge, and then just get a GeoJSON. It's not going to get the bars. You're going to want a special current for that in other sort of platforms.

But it will pick out place names like Cambridge and return a GeoJSON. So this is a special current. They're one of our users, and they use it to resolve place names so the user's going to ask for bars near Cambridge.

And then special currents, we're going to use first draft to figure out which Cambridge they're talking about. This is the link to the Python API. Sorry, I'm sort of blazing through this. I know you guys want to do your journey. Oh, yeah. So UFOs, or supposed UFOs, I guess.

So there's this website, National UFO Reporting Center. And one of my friends was doing a GIS class, and she had to make a map of this. But they don't really have a DB dump you can use. It's sort of like HTML tables. So she got the URL for that

and put it into first draft. And she went to the website first draft GIS. Create a map. And then in the far right, you can add a link, file, or text. Put in the link. And these are the supposed UFO sites from Massachusetts.

So I know a lot of people are asking, like, what's to do here while you're out? There you go. I mean, trust me, there's a lot to do there. So you generate the map. Wow, okay. I mean, I don't think it's like a fun city, so you don't know exactly where it was.

But, yeah, so if you're into that stuff, I won't be joining you. But if you are. So that's sort of the presentation. I'd have to take your questions. Thank you.

Yeah, for sure. That's a great question. So you can just send it, like, as a document. So you can do the URL to the New York Times article. And it sort of uses newspaper.

I'm not sure if you're familiar with that. That's like a Python package that identifies where the article is on the page. So it will just focus on the text of the article. But, yeah, it can totally handle documents. I think it did 100 pages for, like,

some human rights report. It did that one too. So it can totally handle it. But it does use some sort of, like, sense of what the article's about. So you wouldn't want to mash two articles together and send that together, because it's not going to understand the...

It's not going to cluster and understand what locations you're talking about. So, yeah, you just send over just an article.

So the sort of formats that it goes through right now are tabular and just sort of unstructured text. So if it's in tabular format,

it's going to look for the column that represents place names. And then if it's sort of in unstructured text, if the text is super long, it's going to use rule-based entity abstraction to get the place names. But if the text is, like, less than 1,000 words, it's just going to look up just every word

in the database. And, I mean, it can just crush it. It's, like, amazing. I'm sure people here have worked on Postgres, so thank you.

Yeah, that's a really great question. So it does some sort of, like, looking at the PDF table. You probably know that's pretty difficult. It'll look at extract the text from a PDF. It doesn't keep it in the tabular format.

It will add, like, new lines for each row or, like, put, like, space between the different columns. And so it can identify if the table is big enough that it has a repeating pattern. But it has to be, like, five or more lines or long.

But then sometimes... Sorry, just real quick. Sometimes you get, like, a PDF from a country, and they don't really like sharing their data, but they kind of feel like they have to. So they're going to send it in sort of the ugly, somewhat tabular PDF format, in which case it's also going to just

try to crunch every single...

Yeah, that's really cool. That's a great question. So you can turn on the experimental features

for the first draft website, and it includes, like, all the base maps, I believe, from... I believe it's leaflet. The tile layer is the plug-in for leaflet. And so you can just, like, use whatever base maps you want. But currently the AI doesn't choose that for you.

If that's, like, what people want, let's do it. Not yet. Great question.

Sorry. So it's... You know, AI, that's a great question. It's something I've thought about a lot. It really just relies on the database.

So whatever is in ger names, all of a sudden there would be... If those aliases are in there, it'll find them. But if they're not...

So for the Python API, the map format, you can do CSV or TSV, or I think PNG, you get all the image formats. So you can get it back, get it back in whatever format you want.

That's a great question. So I've ended up implementing my own, and that was mostly because I wanted to do stuff in Arabic,

and that's not widely supported. So it also... For the short text, I honestly just don't even really geo-parse. I just look up every single word in, like, a diagram, which is, like, two words together. Just look them all up in the database, because it's so fast.

If you just put everything in one custom function in Postgres, you guys know there's not that transaction cost of looking up sort of each place back and forth. You just look them all up at once.

Feel free to, like, keep questioning me after. Thank you very much.