Vision2025 · Vision · 2 min read · Google DeepMind

ShieldGemma 2 takes vision-language safety classifiers out of the lab

An open, fine-tunable multimodal safety model shipped as a production component rather than a research demo — a template for scene-understanding safety work like ours.

Read the original↗

Karakalpak text-to-speech with voice cloning is live at karakalpakvoice.uz

Any product that needs to speak Karakalpak can now call an API instead of building a voice from nothing.

Research note · 2026

Why we are not launching Karakalpak speech recognition yet

The model works. It does not work well enough for the things people would immediately use it for, so it stays in research.

Deployment · 2026

AvtoTax enters service with the Nukus City Administration

Debt notifications are now sent automatically, through the right channel, with a record of what was sent and when.