Why this is in the vault
Thompson's framing of "trust" as two separable questions — do you trust Apple to ship integrated, private hardware vs. do you trust Siri AI specifically — is a reusable lens for evaluating any vendor that bundles a trusted platform with an unproven AI layer.
The core argument
Apple's new HomeView smart-display and the accompanying Home Collection (cameras, thermostats, locks built with LG and Schneider Electric) revive the AirPort-era playbook: Apple doesn't invent the underlying tech, it marries partner hardware to its own "ease-of-use" layer and privacy promise. Thompson argues Apple has earned genuine trust on integration and privacy — but HomeView's entire value proposition was designed around a new Siri that could handle personal context and multi-user recognition by voice, and that Siri slipped nearly two years waiting for its AI rebuild. The Siri AI that finally shipped is slower and more error-prone on basic tasks than the old deterministic assistant, while competing agents (Muse, Instinct, Dots, Grok Bots) already act on information rather than merely retrieving it. Committing a one-time hardware purchase (a hole cut in your wall) to Siri AI as primary interface means forgoing every other AI ecosystem — an opportunity cost Thompson calls "too expensive" for himself, and the real question consumers must answer isn't about Apple's integration or privacy track record (both strong) but whether Siri AI specifically will keep improving enough to justify that lock-in. He ties this to Tim Cook's record: the "Cook Doctrine" of owning core technology clashed with Apple's reliance on licensed/partnered AI models, and services-revenue pressure (iCloud+ gating for camera history) complicates the trust pitch further.
Mapping against Ray Data Co
The separable-trust framing maps directly onto the founder's own all-in bet on Claude/Opus as Ray's substrate (feedback_delegation_model_effort_pairing): the founder trusts Anthropic's integration and reliability track record the way Thompson trusts Apple's hardware/privacy track record, but that's a distinct question from whether any specific model generation will keep improving fast enough to justify the lock-in Ray's architecture now has (single-vendor, single-model-family, everything-on-Opus). Thompson's "opportunity cost of committing to one AI" is the same risk calculus the founder is implicitly accepting by not hedging across model providers. Second mapping: phData's sales motion rests on being credible about "agents in production" (project_credibility_for_phdata_sales) — Thompson's point that the real diligence question for an AI-bundled product is "does the AI actually work reliably on basic tasks," not "is the vendor trustworthy in the abstract," is exactly the question phData's enterprise clients are asking about Cortex/Claude agent deployments, and it's a sharper framing than the demo-vs-production distinction Ray has used so far.
Related
- [[2026-10-14-stratechery-apple-homeview-tv-app-story]]
- [[2026-10-07-stratechery-apple-lg-smart-home-agent-standards]]
- [[2026-06-09-stratechery-iphone-last-stand-siri-good-enough]]
- [[feedback_delegation_model_effort_pairing]]
- [[project_credibility_for_phdata_sales]]