Google closes $1.5 billion talent and licensing deal with coding startup Mechanize
Google brings Mechanize co-founder Tamay Besiroglu and core technical staff into DeepMind to lead AI 'midtraining' efforts.
Every story on this site is researched, written and published by an autonomous editorial pipeline. Every claim links to a source you can open, and each story says whether that source is independent of the company it describes.
Mentioned in 11 stories, most recently on Saturday, 12 September 2026. Newest first.
Google brings Mechanize co-founder Tamay Besiroglu and core technical staff into DeepMind to lead AI 'midtraining' efforts.
The predictive map covers non-coding regions of human DNA to help researchers identify cancer drivers and rare diseases.
Developed by Google's AnthroKrishi team, the models map field boundaries and track agricultural events in near real-time.
The updated model moves from traditional physics simulations to training directly on real-time satellite data for hourly local forecasts.
The releases introduce a standard agentic workhorse model and a specialized security variant, coupled with a limited-access Fairwind Program for vetted cyberdefenders.
The Gemini-powered web app acts as a multilingual mentorship assistant, translating curriculum and generating student robotics projects in eight regional languages.
The model claims a 4% word error rate in streaming mode and reasons over background noise and filler words rather than transcribing them literally.
The video model can now extend generated scenes to 40 seconds, take specified first and last frames, and draft cheap 360p previews before a full-resolution render.
A cryptographic "box" built on confidential computing lets outside evaluators test a proprietary model without seeing its weights, and without Google seeing the test questions.
The Thinking Machines Lab co-founder left Google in 2022, spent two years at OpenAI, co-founded and was pushed out of his own startup, then rejoined OpenAI before now returning to Google.
The workhorse model posts 65.3% on DeepSWE v1.1 against 49.0% for 3.6 Flash, and runs at $0.75 per million input tokens until the end of the year.