OpenAI unveils ChatGPT Health—personalized advice that can tap medical records, raising privacy and accuracy concerns. NousCoder‑14B shows open-source code models can compete with rapid RL training. NVIDIA's Nemotron slashes ASR latency, and Google’s Gemini turns lessons into podcast-style audio. Accuracy matters.

Lenovo's Qira ambient AI and Motorola’s wearable pin promise seamless context handoffs. Boston Dynamics is adding DeepMind’s Gemini to Atlas for smarter factory work. Liquid AI’s LFM 2.5 brings multimodal edge models to Hugging Face, while the push toward proactive agentic AI raises privacy and reliability questions.

At CES Nvidia unveiled a full‑stack robotics platform aiming to be the OS for physical AI. Boris Cherny’s viral thread shows multi‑agent coding workflows (Opus 4.5, CLAUDE.md) reshaping developer work. Plus: Jerry Tworek exits OpenAI and Amazon repurposes Alexa.com as a web agent.

A study exposes LLMs’ “expert blind spot” — they ace exams but struggle to teach. We unpack rumors of OpenAI’s rumored Project Strawberry reasoning engine, explore Google’s new customizable Gemini “Gems,” and break down California’s controversial SB 1047 kill-switch debate and industry fallout.

OpenAI merges audio teams to close accuracy and latency gaps for rumored ChatGPT hardware. TechCrunch's "War on Screens" asks whether voice will replace displays. A new study shows LLMs can act as world models, enabling massive synthetic training environments for AI agents.

Meta reportedly spent $3B to acquire Manus AI as it races to catch up on autonomous agents. Chinese firms like Zhipu and Deepseek are rushing to IPOs. Plus: transactional agent design with LangGraph (two-phase commit) and Cloudflare open-source tokio-quiche for QUIC/HTTP3.

Meta’s acquisition of Manus AI — built on competitor models — highlights urgency in the agent race. MIT found 59 scientific models converging on the same internal ‘physics’ of matter. Alibaba’s MAI-UI boosts mobile GUI agents, and open-source LLMRouter smartly routes queries to cut inference cost.

ChatGPT now links apps like DoorDash, Spotify, Uber and Canva—seamless but centralizes sensitive data. Some “privacy” browser extensions are scraping and selling AI chat logs, YouTube is clogged with AI-generated low-quality Shorts, and new reasoning models often overthink simple tasks.

Nvidia quietly scoops up Groq talent via a license deal, consolidating AI hardware power. They also unveiled NitroGen — a vision-action model trained on 40,000 hours of gameplay — while OpenAI’s Operator gains real mouse/keyboard control. The EU AI Act now forces stricter copyright transparency.

A 1,200-line Waymo system prompt reveals heavy guardrails; Salesforce loses faith in LLMs; Google may put AI answers behind a paywall; OpenAI's voice engine can clone voices from 15s of audio — huge potential and big risks.