Hello readers,
Welcome to the AI For All newsletter! Today, we’re talking about a leap forward in cyclone forecasting, how AI agents are changing connectivity management, and more!
AI in Action: AI knows how bad it’s going to get
Cyclone forecasting has always forced a trade-off. A storm's track gets steered by planet-scale atmospheric currents, best captured by coarse global models. Its intensity comes from tight, localized physics near the core, which needed specialized high-resolution models to catch. Forecasters ran both and stitched the results together. Google DeepMind's WeatherNext Cyclones, detailed in a Nature paper published August 6, folds track, intensity, and wind structure into one model — and does it at 28x28km resolution, roughly 100 times coarser than what intensity forecasting was assumed to require.
The payoff shows up directly in lead time. Tested against historical storms from 2023 and 2024, WeatherNext's three-day forecasts matched the accuracy older models needed a full extra day to reach — DeepMind puts the jump at roughly a decade of typical meteorological progress compressed into one model release. The system trained end-to-end on nearly 20 terabytes of atmospheric data plus the IBTrACS record of almost 5,000 historical storms, and it now runs 1,000-member ensembles per forecast in under a minute on a single TPU, up from 50 members a year ago. That scale lets it flag low-probability, high-consequence outcomes like rapid intensification before they happen, rather than after.
It already has a track record: during the 2025 season, the model helped the National Hurricane Center anticipate Hurricane Melissa's rapid intensification and landfall in Jamaica, buying evacuation teams extra time on the ground. DeepMind built this with forecasters at the NHC, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office, and it's now open-sourcing the code and weights on GitHub alongside a public interface, Weather Lab, for exploring live predictions. Even DeepMind admits it doesn't fully understand why the model performs this well at such low resolution — an open question they're inviting outside researchers to help answer.
🔥 Rapid Fire
AI companies may need $10 trillion in annual revenue to justify capex
“Barring a massive increase in productivity growth, this will be very difficult to achieve.” To be clear, it will be impossible.
AI suppliers are preparing for the AI data center bubble to burst
Siemens executive: “We are actively working to not be wholly dependent on data centers … We understand at some point there could be a bubble and we’re racing to pay back the investments as quickly as possible.”
Commentary: The AI Hater’s Manifesto
More executives depart OpenAI: key sales executive, head of data centers
Anthropic’s best AI model struggles to attract users as cheaper tools thrive
Meta tried to replace staff with AI but failed after AI was proven ineffective
Waymo says there’s no AI shortcut to self-driving cars
Gemini has a branding problem, and so does the rest of AI
Designer creates shirt that confuses AI surveillance cameras
📖 What We’re Reading
APIs make connectivity data accessible. AI-assisted workflows make it easier to act on. Access and usability are not the same thing. While a developer can pull raw device and usage data through an API, turning that data into an operational summary, identifying the exceptions that matter, or preparing a briefing for a non-technical stakeholder still takes time. AI agents can help close that gap when they have structured, governed access to trusted data. That last part matters. AI tools operating on vague context or stale exports are not operationally useful. They need live, structured access to real systems.
