Weekly AI Brief

Vendors widening their reach (Anthropic, GitHub, Ant) at the same week DeepMind and Thinking Machines push the interaction surface — and builders quietly start pricing the agent infra and the «AI productivity» narrative.. From the Labs is the wide section this week: Anthropic ships Opus 4.7 Fast as a same-model fast tier and runs a distribution week alongside it (Claude for Small Business, $200M Gates Foundation partnership); DeepMind rethinks the mouse pointer for the AI era; Thinking Machines Lab names «interaction models» as a non-turn-based class; Qwen Image 2.0 lands; IBM Granite refreshes 32K-context multilingual embeddings; ByteDance proposes a third visual-gen route; Ant open-sources the LingBot-VLA real-robot post-training pipeline; GitHub wires secret scanning into MCP. Builder Signals is shorter and harder: Peter Yang anchors the «AI changed how we work» layoff template against the Q1 80k+ number; nikunj names «consensus capital»; InfoQ reads the RAG retreat from the teams who built the wave; qbitai reads the agent-era infra shape as «a database per user»; GitHub Trending shows three repos converging on spec / memory / skills. Long-form: Suno's Mikey Shulman on Training Data, on building a consumer AI product where 90% of users create.