Desk
Broad artificial intelligence coverage.
Google DeepMind introduced Gemini 3.8 Flash TTS and Flash-Lite TTS, which let creators prompt new voices, choose from a library of more than 2,000 or replicate a voice from a 30-second sample with consent verification. Every clip carries SynthID watermarking and C2PA credentials, and replication is limited by region.
An Nvidia profile of validation engineer Sakeena Fiza describes the work of proving new data center hardware works, from first power-on through customer deployment: bring-up, integration, firmware and software fixes, then stress testing until failure modes appear. The work starts before products are public, including Rubin-generation systems.
At Made on YouTube, YouTube added AI features to its Studio app, including Ask Studio on mobile, feedback on drafts' pacing and structure, a research feed of what is working on the platform, generated titles and thumbnails, and dynamic thumbnails that show different images to different audiences. YouTube says creators have run more than 40 million title and thumbnail tests.
YouTube will let viewers prompt Gemini to build custom feeds, such as commute-length podcasts or relaxing commentary videos, with include and exclude rules, pinned as their own home-page tabs. TechCrunch reports multiple custom feeds roll out on web and mobile next month, while the main recommendation feed remains.
Nvidia detailed Nemotron 3 Diarization, an open-weight model of about 100 million parameters that identifies who spoke when, supporting up to eight speakers in offline and streaming modes. It ranked first on VoiceArena's initial Diarization-Bench with a 14.72% diarization error rate across 139 English conversations.
OpenAI says it will give Ukraine's government access to Daybreak, its program offering advanced AI for authorized security work, to help defend civilian infrastructure. Working with the Ministry of Digital Transformation, Ukrainian teams will use the tools to find software vulnerabilities and develop fixes faster. CERT-UA handled nearly 6,000 incidents in 2025.
Security researcher Patrick Wardle found a zero-day in Meta's Muse for macOS that let any locally installed app or terminal command grab the token authenticating a user's Muse account and take full control of the agent's privileges. Meta issued a hotfix more than 12 hours after the report. About 12 hours before the disclosure, Amazon had begun blocking Muse from shopping on its site.
Ema, a Mountain View startup that coordinates teams of AI agents across HR, IT and finance work, raised a $77 million Series B led by Bengaluru's Creaegis, bringing total funding to $140 million. TechCrunch reports the company says the round more than quadruples its 2024 valuation, though it would not name the figure.
In remarks to the UN Security Council, OpenAI CEO Sam Altman framed AI as a choice between a Renaissance of human agency and an Industrial Revolution of upheaval. He named two failure modes, losing control of systems and concentrating power, and called for national and international frontier AI standards, incident reporting and secure vulnerability channels.
In an OpenAI customer story, legal AI company Harvey says GPT-6 Astra lets it pass more matter context, including court filings, firm documents and case law, into drafting and get back more structured legal documents, so lawyers spend less time formatting. The claims are from Harvey and OpenAI and have not been independently evaluated.
In an OpenAI customer story, video platform invideo says GPT-6 Astra roughly tripled success on color grading and correction, planned edits with frame-level precision and helped editors ship about 50 custom effects in one day. The company says Astra uses fewer reasoning tokens on complex plans. The figures are from invideo and OpenAI.
In an OpenAI customer story, Indian startup Ringg says its voice and chat agents, built on GPT-5.6 and related models, handle more than 7 million connected calls a month and resolve up to 65% of requests across voice, chat, WhatsApp and web. It reports about a 90% model-cost reduction on selected workloads compared with GPT-4.1.