Claude’s new skill starts with “watch me”
Claude Cowork can turn a recorded desktop demonstration into a reusable skill, but its announcement leaves control, privacy and reliability questions unanswered.
THE LOG — TRANSMISSIONS FROM ALTITUDE
What we learn running AI systems every day — workflows, growth loops, tools, and the occasional hard lesson. Written down so you can reuse it.
Claude Cowork can turn a recorded desktop demonstration into a reusable skill, but its announcement leaves control, privacy and reliability questions unanswered.
OpenRouter has added xAI’s image-to-video model, promising motion, camera direction and sound in one generation—but the announcement is not a quality test.
OpenAI says Sites can turn an idea into a live, publishable site, but its European expansion announcement confirms access only for Plus and Pro accounts.
Google Gemma presents Gemma 4 31B as the reasoning layer inside an open cascaded voice stack, but supplies no latency figure for its speed claim.
NVIDIA Agent Toolkit now includes Omniverse libraries that help AI agents add physical-AI capabilities to existing applications and prepare 3D content for simulation.
Claude says Fable 5 will reach more subscribers from July 20, with included access and usage-credit terms split by plan.
Google’s renamed research notebook can now run code against supplied sources, but access remains tiered and Search integration is still forthcoming.
Moonshot says its 2.8-trillion-parameter Kimi K3 can build compilers, design simulated chips and refine playable software, but deployment and control come with serious catches.
Qwen is teasing a 2.4-trillion-parameter open-weight model, while builders currently have access only to a limited Max Preview through named Alibaba products.
OpenAI demonstrates a four-step path for starting a Codex Security scan, while leaving access, coverage, privacy and scan quality unanswered.
Manus says Auto-Publish lets WebDev users opt in to automatic publishing so each completed successful build reaches the public URL without another manual publish step.
OpenAI's GPT-5.6 family pushes the cost-vs-performance frontier — Sol, Terra, Luna trading price for score, ahead of Fable 5 and Opus 4.8 on the coding-agent index.
Anthropic's Reflect beta turns Claude use into a self-check surface — what we hand off, what stays human, and whether our habits still match our goals.
xAI is pitching Grok 4.5 around speed, token efficiency, Cursor-linked training, and low API pricing for repeat engineering workflows. What that means for coding agents.
Anthropic is putting Mythos-class capability into public reach — but the release is capped, routed, and split from restricted Mythos access.