WindBorne: a better feedback loop, not just a better model
A team that started by building better weather balloons now out-forecasts agencies that have owned the field for decades, because the balloons and the AI feed each other.
I stumbled on one of those stories that make me think we sometimes use AI really well.
A small group of ex-Stanford students started out building better weather balloons. Now they're beating institutions that have been forecasting weather for decades.
WindBorne keeps around 400 self-flying balloons in the air at any given time, launched from 15 sites worldwide. These aren't passive balloons: each one carries a microcontroller that vents gas or drops ballast to change altitude on its own, steering toward wherever the forecast has the least data.
Here's the clever part: the balloons and the AI feed each other. The model, WeatherMesh, tells each balloon where to fly to cover its blind spots; the fresh readings then sharpen the next forecast.
The result? An hourly forecast at 3 km resolution across the continental US, as accurate five days out as a traditional one is the day before. They've already beaten the US National Hurricane Center on Hurricane Milton's track.
Almost every "AI beats the incumbent" story is about a better model. This one is about a better loop between hardware and software. And to me that's far more interesting.
Sources
- WindBorne / IEEE Spectrum: https://spectrum.ieee.org/ai-weather-forecasting
- WindBorne / TechCrunch: https://techcrunch.com/2026/06/01/this-ai-weather-startup-is-out-forecasting-government-agencies/