Zoom's CEO Eric Yuan has unveiled plans to deploy AI avatars as meeting proxies — agents trained on individual decision-making patterns that would attend calls on behalf of absent employees, read and respond to emails, and participate in what amounts to machine-to-machine meetings. The vision is ambitious but reveals growing anxiety: Zoom's pandemic-era surge has flatlined as Microsoft Teams and Google Meet capture workspace users by bundling communications with docs, spreadsheets, and other productivity tools. Yuan's strategy is defensive — layer AI agents and a suite of workplace applications onto Zoom's core video infrastructure to remain relevant. Yet the bet carries real friction. Training avatars to reliably make decisions in high-variability meeting contexts is orders of magnitude harder than automating email triage, where responses typically follow predictable patterns and written records persist. The company hasn't addressed how these proxies learn or whose data trains them, and there's genuine uncertainty about whether enterprises will trust machines to negotiate on their behalf.
Meanwhile, two Portuguese companies are making concrete waves in different directions. Apptoid — a Portuguese app store — becomes the first alternative iOS marketplace under EU digital regulation, launching this week with a waiting list and staggered access. The mechanics matter: Apple charges alternative app stores 50 cents per annual installation but permits them to run in-app purchase ecosystems using Apple's infrastructure. Apptoid absorbs this fee into its commission on in-app transactions, undercutting Apple's standard 30% cut. Developers gain leverage; apps can migrate to Apptoid at 15% commission, a long-standing industry grievance epitomized by Epic's Fortnite lawsuit. For a Portuguese startup to pioneer this category signals how regulation can reshape market structure faster than competition alone.
Sword Health announced Phoenix, a physiotherapy AI that layers computer vision and conversational agents into a single interface. Where the previous generation used wearable sensors to track movement, Phoenix relies on camera-based pose detection combined with real-time dialogue — patients ask questions, receive feedback, and see workout protocols adapt mid-session. The system mimics what a human physiotherapist provides: responsiveness, adjustment, continuity. Sword has raised $340 million at a $3 billion valuation, liquidity that partly reflects how venture capital now rewards not just deployment scale but architectural credibility in regulated healthcare domains. The company's shift from sensor-based to vision-based systems also signals a consolidation in how AI commoditizes expertise: fewer specialized inputs, more general-purpose models, greater density of intelligence per deployment.
Finally, Microsoft's Aurora model for weather forecasting demonstrates how transformers — the architecture underlying LLMs — apply directly to sequential, temporal prediction at 5,000x cheaper compute cost than traditional meteorological models. Aurora predicts temperature, humidity, wind, pollution, and greenhouse gas concentration at 0.1-degree resolution (roughly 11 km at the equator) days ahead. This matters not as a novelty but as proof that the transformer paradigm generalizes beyond language. Time-series problems across agriculture, energy, infrastructure, and finance are now tractable candidates for the same architectural primitives that power chatbots. The economic implication is stark: entire modeling domains built on expensive numerical simulation may see rapid displacement by cheaper, faster learned models.