Gregory's ServerSo now that we see how AI fails to accomplish anything you throw at it, unless the task is "do an AI thing with no consequences", you might understand why the @openstreetmap community has always felt negative towards bulk-importing AI-detected roads and buildings and stuff onto their hand-crafted map.
 7 comments
I agree with the negative feels, too much of that in OSM :) I'd say Meta benefits from Rapid because it enables bulk-importing of ML and third-party data onto OSM under a guise of manual verification. And corporations need our map to be perfect pronto, no time to wait for mappers. The thing with OSM, when you have spent 15 years mapping the world by hand, and then somebody offers to double the data, you can't but notice how bad the ML is compared to skill and knowledge of mappers. @djh What changed is, couple years ago I assumed the tech might get better, with the proper pre- and post-processing, increased power, and applied geospatial knowledge. Now we see it's all dazzle, MVP with no VP in sight. No people with geo experience do ML, it's all twiddling settings of models invented last century. Because people are not profitable, hype is. And there are no two identical buildings out there, stats won't do. Fixing ML output would mean rewriting detection with heuristics.  @zverik That's a very grim outlook and doesn't reflect what I've been seeing. I'm sorry you feel this way. There's lots of ML at work in the geo domain out there: improving above ground biomass estimations, allowing us to interpret satellite radar imagery for geospatial understanding, cloud removal, improving validation and vandalism detection, spotting outliers and mistakes in the map, better terrain models, better global landcover, and on and on the list goes. @djh That's a good long list of successful ML applications. I think half of those could be done without ML, but still. My point is with OpenStreetMap, you're not working in a vacuum, but alongside human mappers, so the net should be at least half as good. And HOT for some reason chose the one particular application of ML that has been proven to produce bad mapping — it's visible even on their promos.  @djh @zverik @djh @zverik TBF Rapid is only what it is *because* of the “negative OSM community”. Facebook started just silently dumping their AI data into OSM, and bad-mouthing anyone who opposed it as “risking dooming OSM to irrelevance“. It took *a lot* of push back from this “always negative” OSM community to get what exists now. So when I learned about @hotosm "fAIr" AI-assisted mapping service, I was like, uh-oh. We all know now that AI takes too much resources, is an environmental threat, and is never profitable. Apart from also being wrong. That a humanitarian org turns its employees and hardware to replicate things Microsoft and Meta failed to do... Well, won't end good. https://www.hotosm.org/updates/fAIr_2024/ (On the other hand, their FMTM is shaping up to be better than I expected.) |
@zverik @openstreetmap To be fair the vocal OpenStreetMap community feels negative about pretty much anything 😛
The Rapid editor and folks around it seem to benefit from AI assisted mapping, so I wouldn't say it's binary good / bad; it's more complex and nuanced as most of the time.
I also can't see much of an overlap between the current LLM hype and the previous efforts in geo to build machine learning tools and workflows.
Disclaimer: I have built such workflows before.