You're not missing anything. It's good when it has code to train on so... great at making tutorial matplotlib bar charts and D3 line charts and it's absolutely terrible at anything else. It has no diagrammatic conceptualization so it can neither properly design or read anything but simple charts.
The conclusion we came to in a recent grad seminar class on cartography & AI was that it would be easier to learn how to do cartographic design than it is to contort current AI systems into making even the most primitive map.
I assume eventually someone will make a diagrammatical model that understands topological relationships along with cartesian ones but until then unless you can map it to structured text then you're wasting your time. To be clear: There are better ways to do that taking advantage of stuff like JSON.
Some folks working on a thing they called MapGPT had to develop 68 separate modules in order to make a really primitive map - and that's before we might actually try to do something creative: www.tandfonline.com/doi/full/10.... It's not too surprising that a viz may be worth thousands of words...
The thing is, maps are already structured data so you don't need to go through the transformer process, you can use more traditional ML approaches to identify anomalies and insights. I can't even imagine what GenAI does in that case other than translate natural language into commands.
Yeah, there is just a huge mismatch in type it seems. I also think a chat interface as a means to create a graphical output is always going to be a terrible fit. It's like driving a car with voice commands - beyond awkward. I don't think in sentences when I am designing a map... maybe others do?
Well... when you hail a taxi you're driving via voice commands so I think there's some application for it but not in the way people are envisioning given its current limitations.
With a taxi, you describe where, not how. A bit like declarative vs procedural distinction. This is why turn directions from satnavs are so awkward and unlike human. (Disclaimer, my PhD topic years back).
But with AI, I can describe where not how, too. The examples we had with Noteable were very loose and non-expert-oriented and worked out very well. www.youtube.com/watch?v=Myea...