Note the author of that article also runs theeyewall.com and has a meteorology degree, so he’s not just a run of the mill science reporter.
Just last year, the AI models weren’t very good. Looks like extra year of data and training has helped a lot. The GFS model has gotten worse over time but still a solid tool. Doing better than that is a big achievement.
I wonder how much of a factor this is. An awful lot of weather forecasting is based on data collected by the National Weather Service and NOAA, and the quality and quantity of that data has been seriously degraded lately. If Google has other sources of data that the standard models aren’t currently using, that alone could make their forecasts much better than the standard models.
Eric Berger touches on that a bit in the article. Hard to say, really.
Though, the other standard models haven’t degraded much. So, it’s just left to be seen if the Google AI just had a particularly good year or if this is the standard we can expect from now on.
While true, that degraded quality of current meteorological data in the U.S. would likely affect any model. My understanding is that the GFS model (which is run by the NWS) has been regularly updated and upgraded over the past years. However, despite these upgrades, GFS’s accuracy has been generally falling behind the European EMCWF model.
Right, which is why I’m wondering if Google has access to other data sources. For instance, there’s a very widely dispersed collection of a great many barometers out there, which Google could potentially be harvesting data from: Do they use data from phone barometers?
TIL that many smartphones have built-in barometers. 
If Google’s paper is to be believed, they use the same data available to traditional ensemble models.
There are other bits of meteorological data you can gather from phones, too. For instance, a widely-distributed array of antennas is a great way to pinpoint the locations of lightning strikes.
Temperature measurements would be tricky, though, since phones spend a lot of time in close proximity to body heat.
I suppose those might help a bit but one of the traditional problems of hurricane forecasting (and general weather forecasting to a lesser extent) is limited data in the vicinity of nascent and active storms. There’s a lot of ocean out there