Idle Macs, Programmable Gateways, and the Work of Judgment

Different ways to provision AI compute

Apple’s newly detailed Minions framework pools idle Macs for AI inference and training, making intermittent consumer devices part of the compute pool. That approach could reduce dedicated GPU requirements for some workloads, but scheduling, thermals, battery wear, and reliable job completion remain engineering questions. Its usefulness depends partly on how well workloads tolerate machines becoming unavailable.

Meta’s Muse Spark Open is a released multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. For teams considering self-hosting, the distinction between total and active parameters matters: the smaller active footprint is relevant to inference work, while the full model’s size remains part of deployment planning. Apple’s approach concerns where jobs run; Meta’s model offers a different choice about what to run.

Changing where software behavior lives

Cloudflare’s secure web gateway supports custom DLP and traffic-inspection logic, allowing security behavior to move into code that teams can version and test. At a smaller scale, TinyLittleOffice makes the shareable URL the document, avoiding a backend database row for its experimental office suite. These are separate design choices with concrete consequences: one makes policy programmable, while the other removes a storage dependency from lightweight document sharing.

Google is bringing Gmail Live’s real-time AI assistance to paid Workspace users and rolling out AI-organized inbox views through AI Inbox. Those changes place assistance directly in the inbox workflow. Microsoft’s Windows 11 “everyday experiences” updates address core OS interactions through incremental refinements. The Gmail rollout changes how messages can be handled; the Windows announcement describes a more modest set of interaction updates.

Preserving judgment as workflows change

Tools that reshape work still leave organizations deciding how to think through consequential choices. A former Amazon engineer’s account of its declining narrative-writing culture argues that six-page memos have given way to slide decks and shortcuts, weakening a practice intended to force rigorous thinking before decisions. Separately, Steering AI Beyond Automation argues for directing models toward judgment-intensive work rather than treating task automation as the main objective. Both essays concern how work is framed: one defends a discipline for developing decisions, while the other proposes a broader role for AI.

ETF growth still below its peak

BlackRock’s crypto ETF assets rose 63% but remained far below their 2025 peak. The increase and the remaining gap both matter: substantial growth does not by itself establish a return to the earlier high. This is a narrower signal about one asset manager’s crypto ETF assets, rather than evidence of a broad recovery across crypto markets.