This week in AI, I covered a few stories that actually matter if you’re building real automation + agent systems (not just reading hype). Google dropped Gemini Embedding 2, which is their first natively multimodal embedding model—text, images, video, audio, and documents all mapped into one embedding space. That’s a big deal for retrieval and classification because you can push mixed media into a vector DB without doing awkward conversions first. The specs are solid too: up to 8192 tokens for text, six images per request, up to 120 seconds of video, native audio embedding, and PDFs up to six pages. On the Anthropic side, I talked about Claude Code Review—basically an AI reviewer for the flood of AI-generated “vibe code.” I was against vibe coding for a long time because of bugs, security risk, and poorly understood code, but the last 6 months have been improving fast. Still, $15–$25 per PR is a real price tag, so you’ve got to weigh that against engineer time and risk. Anthropic also launched the Claude Partner Network with a $100M commitment, and Claude Opus 4.6 reportedly found 22 Firefox vulnerabilities in two weeks with Mozilla. Then we hit the weird one: Meta acquiring Moltbook, a Reddit-like social network made entirely of AI bots. I also covered Microsoft (plus Google, Amazon, Apple, and even OpenAI) backing Anthropic against a Pentagon designation, Yann LeCun raising $1B+ for AMI (physical-world intelligence over language), and a key Department of Commerce deadline under the new federal AI executive order.