Brief — March 28, 2026

01

Anthropic's Claude Can Now Control Your Mac

Anthropic shipped a Mac computer-use research preview for paying subscribers, transforming Claude from conversational assistant to "remote digital operator." The update lands inside Claude Cowork and Claude Code, with Dispatch (assign tasks from mobile, return to finished work) now extended to Claude Code for the first time. End-to-end pipeline: issue instructions from anywhere, collect finished deliverables. This is the agentic pivot in full mot

02

Anthropic Security Leak: Next Model Is "Mythos"

An unsecured internal data store exposed Anthropic's next model name ("Mythos"), details of an invite-only CEO event, and other internal materials. The leak arrived via Fortune. The name signals continued mythology branding (Claude, Prometheus, Capybara — now Mythos). The security lapse itself is the real story: at the frontier, operational security is as critical as model capability. 🔗 https://fortune.com/2026/03/26/anthropic-leaked-unreleased-m

03

OpenAI Shuts Down Sora — Abruptly

OpenAI is shutting down Sora (its AI video model, app, and API) with little warning — despite releasing updates to the product as recently as this week. No official explanation given. Either internal consolidation ahead of a major new push, or quiet acknowledgment it couldn't compete with Runway, Kling, and others. The video AI race is getting culled. 🔗 https://venturebeat.com/technology/openai-is-shutting-down-sora-its-powerful-ai-video-app

04

Google TurboQuant: 8x AI Memory Speed, 50% Cost Cut

Google's TurboQuant algorithm was open-sourced and the community immediately began porting it to MLX (Apple Silicon) and llama.cpp. Memory is the new compute bottleneck. An algorithm that compresses LLM data 6-8x with zero accuracy loss changes the economics of local and edge AI deployment. Research Lab angle: this is critical infrastructure for the "AI everywhere" thesis. 🔗 https://venturebeat.com/infrastructure/googles-new-turboquant-algorithm-

05

xMemory: Cutting Agent Token Costs Nearly in Half

New research technique xMemory replaces flat RAG with a four-level semantic hierarchy, cutting token usage nearly 50% for multi-session AI agents. As agent systems scale, context management is the cost moat. Tools like xMemory could be the difference between sustainable and unsustainable agentic pipelines. 🔗 https://venturebeat.com/orchestration/how-xmemory-cuts-token-costs-and-context-bloat-in-ai-agents

06

Cloudflare Dynamic Workers: AI Agent Code 100x Faster

Cloudflare launched Dynamic Workers — ditching containers entirely for AI agent execution. 100x faster cold starts at $0.002 per unique Worker loaded per day. Edge-native AI infrastructure is becoming real. The "agent-everywhere" architecture is quietly being built at the infra layer. 🔗 https://venturebeat.com/infrastructure/cloudflares-new-dynamic-workers-ditch-containers-to-run-ai-agent-code-100x