DeepMind's WeatherNext AI gains 24 hours of hurricane forecast accuracy

·3 views

Google DeepMind published results of its WeatherNext AI model in Nature, showing its three-day hurricane forecasts now match the accuracy previously only achieved by two-day forecasts — a gain DeepMind equates to a decade of meteorological progress. The model jointly predicts a cyclone's track, intensity, and wind structure in a single system, replacing the need for separate models for each variable. During the 2025 hurricane season, WeatherNext accurately predicted the rapid intensification of Hurricane Melissa, giving Jamaica and the National Hurricane Center extra time to issue warnings and prepare. The model generates 1,000 forecast scenarios per cyclone, was trained on nearly 20 terabytes of atmospheric data, and can produce a full 15-day forecast in under one minute on a TPU. DeepMind has open-sourced WeatherNext 2 and WeatherNext Cyclones so meteorologists, universities, and under-resourced agencies worldwide can run comparable forecasts without supercomputing infrastructure.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Mautic 5 users can swap SMTP for Symfony Mailer DSN to improve email deliverability

Mautic 5, the open-source marketing automation platform, has moved from SwiftMailer to Symfony Mailer, enabling users to configure email delivery via a DSN string instead of traditional SMTP. MailKite, an email delivery service, offers a first-party Symfony Mailer transport that provides clearer error messages compared to generic SMTP responses. Users can install the transport via Composer, set an API key in the MAILER_DSN environment variable, and clear Mautic's cache to activate the new configuration. For managed Mautic hosts without Composer access, a standard SMTP fallback using MailKite credentials on port 587 with TLS is also supported. Successful delivery requires that all campaign sender addresses belong to domains with SPF and DKIM records properly configured.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Automating a Legacy Windows Desktop App Required a Real Interactive Session

A development team attempting to automate a Windows-only, GUI-first desktop application initially tried running it inside a Windows compatibility layer within a Linux container to stay aligned with their cloud-native stack. The approach produced intermittent, hard-to-diagnose failures — the automation SDK would randomly fail to connect to the running application, with no consistent trigger or clear root cause. After months of patching symptoms, the team abandoned the compatibility layer and moved the application to a native Windows machine, which resolved most issues. However, a second requirement then emerged: the application's automation layer only worked reliably inside a real, interactive, logged-in desktop session — not a headless or background service context, even on genuine Windows. This reflects a broader class of legacy software — including engineering tools, financial systems, and CAD applications — built with the assumption that a human is actively logged in, making headless server automation fundamentally incompatible by design.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Stop AI Tools From Generating Generic Web Designs

Developer Maneshwar, creator of the open-source AI code reviewer git-lrc, argues that AI-generated web designs tend to look identical because models are trained on the statistical average of the public internet, producing what he calls 'AI slop.' He contends the root problem is not model quality but a lack of deliberate design taste on the part of the user. His proposed fix begins with actively collecting design inspiration from platforms like Dribbble, Pinterest, and X, rather than relying on vague prompts like 'build me a modern landing page.' He recommends organising those references into a personal inspiration library — even using Claude Code to build one — so that prompts are grounded in specific, curated aesthetics rather than generic defaults. The broader argument is that better outputs require users to develop and supply their own taste, since improvements in model capability alone will simply raise the baseline of what qualifies as generic.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineer Builds Full AWS Network in 90 Lines of Terraform Code

A developer has shared how a small Terraform project can provision an entire AWS network — including a VPC, two subnets, an internet gateway, a security group, and an EC2 instance — using just two commands. The setup replaces a tedious manual process of clicking through roughly fifteen AWS console screens, which often led to misconfigured route tables and connectivity failures. The project is organized across five files covering provider config, variables, core infrastructure, and outputs, making the setup fully reproducible. One subnet is configured as public with internet access via an internet gateway, while the second remains private and isolated. The author also highlights two specific pitfalls encountered during the build, aiming to help others avoid the same mistakes.

0 comments Read more at DEV Community