Anthropic launches Claude Tag — an agentic AI coworker for Slack: Claude now lives in Slack channels, learning context and providing suggestions as a persistent team member. This marks agentic AI’s entry as a first-class participant in enterprise collaboration tools. Anthropic · ZDNet
OpenAI unveils DayBreak (GPT-5.5-Cyber), a cybersecurity-specialized frontier model: The first major frontier model purpose-built for a specific domain — cyber defense — signals a shift from general-purpose chatbots to domain-specialized AI. OpenAI positions the model as a defender tool for security teams. OpenAI
Mistral OCR 4 sets a new standard for structured document extraction: With bounding boxes, block classification, and multilingual support, Mistral’s latest OCR model demonstrates European AI infrastructure competitiveness at the frontier, particularly for document-intelligence use cases. Mistral
Meta unveils $299 AI smart glasses ‘Adventurer’ and ‘Fury’ — its first own-brand AI glasses: Powered by Meta AI and Muse Spark, these are Meta’s first glasses without a partner brand. The Starfire model comes in at $399. Consumer-priced AI hardware is becoming a real category. Reuters · Meta
AI inference startup Baseten hits $13B valuation — infrastructure gold rush continues: Serving GLM-5.2 at 280 tokens per second, Baseten secured Australia’s Blackbird’s largest bet ever with a $1.5B round. AI inference infrastructure is solidifying as its own platform layer. Baseten · Tekedia
VibeThinker-3B: A 3B-parameter model that beats Opus 4.5 on reasoning: Using novel SFT+GRPO training, researchers push the boundaries of verifiable reasoning in small models. A strong signal for on-device and edge inference becoming competitive with massive cloud models. arXiv
AI’s Affordability Crisis — the broken economics of hyperscale AI: A rigorous analysis of how AI companies burn capital on training while struggling to monetize. If the math continues to fail, a structural correction in the AI industry is inevitable. DSHR Blog
The Coming Loop — AI models training on AI-generated data: Armin Ronacher (creator of Flask) examines the technical and philosophical implications of AI models increasingly trained on synthetic data from other AIs. We have crossed a threshold where human-generated data alone can no longer sustain AI progress. Armin Ronacher
Stanford HAI warns on algorithmic monocultures in AI hiring tools: New research finds AI hiring tools systematically exclude 26% of Black and 15% of Asian applicants due to “algorithmic monoculture” — the widespread adoption of a narrow set of AI screening approaches. A critical risk for any company deploying AI in hiring pipelines. Stanford HAI
AWS Lambda MicroVMs enable safe execution of AI-generated code: AWS introduces micro-VM isolation in Lambda, providing infrastructure to run AI-generated or user-submitted code safely in production. A building block for teams connecting AI agent code generation to trusted execution pipelines. AWS
YOLO26 continues the evolution of real-time vision AI: The standard framework for object detection reaches version 26, with unified end-to-end vision capabilities. Edge deployment of computer vision keeps getting more practical — worth experimenting with in manufacturing, retail, and security domains. Roboflow · arXiv
Anthropic introduces age and identity verification for Claude — a new front in AI access control: One of the first major AI platforms to require identity verification, moving away from the default anonymity of AI services. The privacy-versus-safety framing is now reaching AI platforms. Anthropic Legal
Polymarket’s fake betting videos erode trust in prediction markets: Viral videos showing huge wins were revealed as fabricated bets, raising questions about integrity verification in AI-era prediction markets. As AI makes synthetic media easier to produce, verification mechanisms become critical infrastructure. Ars Technica