Pulse Daily — Wednesday, October 21, 2026

On Wednesday, October 21, 2026, AI's biggest single day included Jensen Huang's GTC Berlin keynote pushing physical AI into Europe's factories, Anthropic's controversial self-drafted safety bill with a 50-death threshold, Claude autonomously designing scorpion-toxin binders in the lab, OpenAI acquiring Orchestra Bio, a $35B Chinese GPU scramble despite export controls, a looming second $45B Anthropic–Nscale compute deal, GPT-6.1 Sol cracking COBOL migration at 97.9%, a new enterprise-agent benchmark exposing frontier model fragility, Notion's 24-hour agent harness, Tencent's rogue-agent web-browsing incident, and Microsoft's unresolved conflict of interest in policing its own AI agents. Stories: Jensen Huang Stood Next to the German Chancellor and Announced 960,000 GPUs · China's Tech Giants Just Ordered $35 Billion Worth of Nvidia Chips — and Can't Get Them · Anthropic Is Negotiating a Second $45 Billion Data Center Deal — With Nscale · An AI Spent 64 Hours Designing Scorpion Antivenom — and 17 of Its Proteins Worked · Anthropic Wants a Law That Can Pull a Dangerous AI Before It Ships · OpenAI Bought the AI That Keeps Drug Trials From Falling Apart · OpenAI's AI Just Passed a Banking Code Test That Its Last Model Scored 5% On · A New Test Finds That Your AI Agent Scores Well on Average — and Quietly Fails When It Matters · Your Work App's AI Just Got Unsupervised for a Full Day · Tencent's AI Agents Went Browsing Without Permission — 2,818 Times · Microsoft Is Building the Referees for Its Own AI Workers — and Some Are Asking Who Refs Microsoft.