AI Updates on 2026-02-21

Anthropic releases Claude Code desktop app with live app preview, code review, and background CI/PR handling capabilities @claudeai
Anthropic launches Claude in PowerPoint with connector support for Pro plan users, beating Microsoft to market @claudeai
Google releases Gemini 3.1 Pro Preview topping multiple benchmarks and matching Claude Opus 4.6 at less than half the price @WolframRvnwlf
Google Labs launches Photoshoot feature in Pomelli for generating campaign-ready product visuals with templates and flexible editing @GoogleAI
Google releases Lyria 3 music generation model in Gemini App @OfficialLoganK
NVIDIA announces Cosmos Policy for physical AI, turning world foundation model into unified robot brain without separate action heads @NVIDIAAIDev

AWS engineers using Kiro AI tool caused outage, proving inadequate verification systems rather than AI capability issues @GergelyOrosz
Microsoft reportedly in "code red" after Anthropic shipped Claude Cowork desktop app before Microsoft's own agentic Office integration @GergelyOrosz
Mac Mini sales surge as developers buy hardware to run OpenClaw and similar agent systems locally @karpathy
DeepSeek achieves 1.53x performance improvement over GB200 on NVIDIA GB300 NVL72 with 226 TPS/GPU on long-context inference @lmsysorg
Human genome sequencing cost drops to $100, down from $500M-$1B in 2000 and $600 two years ago @EricTopol
Phil Spencer retires from Microsoft after 38 years, new Xbox CEO vows not to flood ecosystem with "endless AI slop" @XboxP3

OpenAI debated calling police about suspected Canadian shooter's chats, raising questions about AI platform safety responsibilities @TechCrunch
Users report fatigue from AI-written content on X, citing predictable structures and overused phrases making platform "increasingly boring" @emollick
Security concerns emerge around OpenClaw with reports of exposed instances, RCE vulnerabilities, and supply chain poisoning attacks @karpathy

Anthropic hackathon winner CrossBeam speeds California permitting process with AI-powered code compliance and plan review tools @claudeai
Postvisit.ai built by cardiologist turns medical visit transcripts into personalized ongoing health guidance for patients @claudeai
TARA system converts dashcam road footage into infrastructure investment recommendations, tested on actual Uganda construction project @claudeai
Linear Slack agent achieves product-market fit with engineers managing all tasks through agent without opening dashboard @hahnbeelee
Pika launches AI Selves feature allowing users to create persistent AI agents with memory and personality traits @pika_labs

Research shows small language models improve as judges using "backwards" approach where models predict instruction from response @cwolferesearch
MIT physicists reduce quantum noise in optical atomic clocks, improving fundamental measurement stability @MIT
Anthropic research finds AI models become incoherent rather than systematically misaligned with extended reasoning @AnthropicAI
"Claw" emerges as term for OpenClaw-like agent systems running on personal hardware with messaging protocols and task scheduling @simonw