Future Technology[Daily Briefing] 🌧️ Hourly weather AI, 🤖 robots beat humans, 🔓 7 exploited flaws
Google's weather AI now updates every hour, at 5km
Google DeepMind launched WeatherNext 3 on Wednesday, and it is already running inside Search, Maps and the Gemini app. It produces 15 day global forecasts, re-initialised every hour, across 64 ensemble members. Temperature and moisture come out at 5km resolution, other surface variables at 10km, wind and the atmospheric layer at 25km.
The headline number is up to 50 percent more accurate precipitation forecasting when you are planning a day or more ahead. The interesting part is where those gains land. Traditional physics-based forecasting needs a supercomputer and a dense observation network, so the regions with the thinnest sensor coverage historically got the worst forecasts. A model that has learned global patterns does not care where the observation stations are.
If you want the raw data rather than the app, it is available through BigQuery and Earth Engine, or as a bulk Zarr download from Cloud Storage. There is also a Maps Platform Weather API if you are building something on top of it.
Why it matters: weather is the cleanest test AI has. There is no benchmark to argue about and no evaluation methodology to pick apart. Either it rained on Tuesday or it did not, and the model gets marked against reality every single day. Very little else in this field gets graded that honestly.
Quick hits
- Seven exploited flaws land on CISA's list before Patch Tuesday. Attackers are actively hunting AI infrastructure now, including stealing LLM API keys straight out of exposed RAGFlow servers. See the list
- Four frontier model launches in 72 hours. The release cadence has gone from quarterly to weekly, and the benchmark gaps between the top labs are now inside the margin of error. The rundown
- Nvidia is raising AI server prices by more than 15 percent. Memory is the reason. HBM and DRAM costs have gone vertical, and that flows through to what your cloud bill looks like next year. Why AI infrastructure costs what it does
- Apple's first foldable arrives Wednesday. No standard iPhone 18 this year, just the 18 Pro, Pro Max and the folding Ultra. Price talk starts near 2,000 dollars, and the crease is the only spec that decides whether this is a product or a demo.
- A humanoid robot ran a half marathon in 50 minutes 26 seconds. The human world record is 56:42. It was built by Honor, a phone company, which tells you something about where robotics talent is pooling. Running models locally, if you would rather not rent a robot
Tool of the day
A personal weather station. Honest take: WeatherNext will beat your back garden sensor on forecasting, every time. What it will not do is tell you the actual conditions where you are standing, which is a surprisingly different number to the one your phone shows for your postcode. If you garden, cycle, run or care about damp in an old house, a station that logs local temperature, humidity and rainfall is genuinely useful. If you just want to know whether to take a coat, the free app is now better than it has ever been and you should save your money.
Home weather stations on Amazon
Quick question, and I do read the replies: what is the one thing your weather app still gets wrong where you live? I want to know whether the 5km resolution actually fixes it.
Nath, Future Technology
Forwarded? Get your own free briefing at futuretechnologyhq.com/newsletter
Some links may be affiliate links.