Mapping Hilaire Belloc's Path to Rome

• #

In Hilaire Belloc's travelogue The Path to Rome, he embarks on a trek from Toul (in northeast France) to Rome, a pilgrimage of sorts. The book documents his journey as he interacts with locals and runs into hardships and challenges walks a beeline (or as close as he can manage) over the Alps.

As a geography guy, when I read any book like this that refers heavily to places — towns, gorges, mountain peaks, rivers — it takes me forever to read when in parallel I have a map open tracing the mentioned locales. When Belloc talks about his attempt to scale Gries Pass, or crossing Lake Lugano on a long mole, I have to spend just as much time looking up these places and panning around on maps as I do actually reading.

In parallel to this reading, my experimentation with agentic software tools continues. I've been flexing more and more the limits of where these models can go, not just in domains like code and writing, but internet search, data processing, computer use. If it can be done with a computer, Opus and Astra can probably take care of any ambitious ask.

Mapping, with AI

So I asked myself: how far can you go with mapmaking using only agents and prompts? No GIS tools, no databases, no typing code, not even searching Google for public sources or opening up GIS datasets. Just let the agents take the sticks and drive purely through prompts and live previews.

The subject of this experiment was Belloc's journey. I wanted to see how far I could take it, to build a map that could act as a geographic companion to reading.

Here's the result: Belloc's Path to Rome, visualized in an interactive map, historically (mostly) accurate to the time period.

The Path to RomeOpen atlas separately

Above is an embed, but you can view the full map here.

The map traces his journey between all the noteworthy places he visited along the way. Some spots are mountains or river fords, some small to mid-size villages. If you click on a place, you get a quote or two from the book with his musings on it.

At the bottom you can pull up his complete pilgrimage itinerary, to search for any place he visited. Because the book is in the public domain, the entire text is available here, too. Any quote you pull up, you can follow a link to that specific passage in the text. At any point you can open the full text in a drawer, replete with his drawings and sketches. In the text are also many linked names, places mentioned that you can instantly pan to for context. As you scroll and read, the map follows his path.

Image

If you zoom in far enough, the smaller scale map give

Image

s way to detailed survey maps from the era. As you pan around, the map legend dynamically shows you the relevant features in your view.

Image

If you select "Explore in 3D", you get a tiltable 3D view where you can actually follow the journey from above, like you're flying behind him. Watch the elevation profile change as he traverses the Jura, the Alps, the Apennines.

Data, processing, cartography

Everything you see on the map came from Astra finding what I asked for and building its own processing pipelines to give me the result I described. A thing about these frontier models is that if you can describe in detail your desired outcome, and give the model a long enough leash to pursue the goal, they can do it. There's hardly a barrier they can't surmount at this point. The gap now is having a clear idea of what you want, and has little to do with how it should be built. The agent already knows that part.

I started out having it generate a simple basemap, the rough journey path, and some place names. Then from there it's about iteration — layering it on. It found DEMs to generate shaded relief with color ramps. Historic boundary data to represent the 1901 national borders. Rivers and lakes. Mountain range locations and labels. Almost all the feature categories I added didn't happen in one shot. It'd find a dataset or two for something like rivers and drop them in. Those datasets would have gaps or missing lake bodies, or not enough detail in streams or smaller creeks. I just kept asking for more and it kept finding more.

If you zoom in close you'll see these historic raster survey maps. These it found from sources like IGM, the Siegfried Map, and the Austrian military of the 1880s. When it initially dropped them in, there were heavy seams between disjointed map sheets. A few prompts later Astra had clipped them up and restitched them into a more seamless canvas. Absolutely incredible what it can do. The drudgery of the processing to find, clean, and organize all the data for a map like this would take weeks. This whole thing was probably less than 4 or 5 hours of active work.

I also did a heavy amount of specific cartographic design prompting. Make the river stroke widths variable by size. Add bathymetric data for the lakes with a light/dark blue ramp. Adjust the typefaces to use more/less letter spacing. Curve the labels on splines representing the mountain ranges, if there are no splines, create some. There's pretty much no limit to what it'll do, though there's probably some bound if your goals are specific enough. Then you might be better off popping the hood on the data or taking it out as a static map in Illustrator to refine the last mile.

Every source used in the map came from open datasets published online, and Astra found them all, downloaded, merged, and processed them. What appear to be single feature layers in the map are actually composites of sometimes five or six different datasets. This is a perfect use case for AI as processor. When I'm building finished maps, I just want to present a coherent, unified result, not disjointed separate layers. Sources are important, but secondary to the map's function and presentation. We can cite sources all day without encumbering the user with sourcing information. With AI, it turns out munging data (once the GIS Analysts most-time-job) is a thing of the past.

This was mostly an experiment to see how far I could push the models to automate the full design and publishing flow. We're doing a handful of other experiments at Liminal Lab (my product design studio) and for Endovar (our fishing equipment and software business) that we're looking forward to sharing more about soon. I want to keep pushing the boundaries to see where "agent-native mapmaking" can go.

Check out the full Path to Rome Atlas →

1,157 words • 6 min read