Thomas Ptacek, one of the most respected voices in software security and infrastructure, has left Fly.io to co-found a startup with Kurt (last name not disclosed) building what he describes as a phone designed from scratch for AI-generated software. The thesis: the era of prefabricated fixed-function applications is ending, and the operating systems designed to manage them are architectural relics. Ptacek's argument runs in two stages. First, AI has dissolved the boundary between programmers and users. He reports his own Mac is cluttered with native macOS applications he built himself using natural language, and asserts this capability is now available to any power user. Second — and this is the swing — if most software will soon be conjured by its own users for audiences of one or two people, the entire purpose of a modern OS (partitioning untrusted stranger-code into sandboxed fiefdoms) becomes vestigial. The historical framing is pointed. Ptacek traces a line from Egghead Software shelves to CD-ROMs to internet distribution, positioning each as a collapsed assumption about how software reaches users. The next collapse, he argues, is the assumption that software is something you receive rather than something you generate. In this world, applications pull in fragments of existing capabilities — "1/7th of the guts of a word processor" — to solve idiosyncratic problems no professional developer would have addressed. The phone itself is deliberately vague at this stage. Ptacek acknowledges the startup-announcement credibility tax and withholds technical details. What he does commit to: it won't be designed around running AAA games or TikTok. It will build apps on-device, on demand, from user intent. The implicit bet is that frontier model capabilities will continue their current trajectory of cost collapse and capability expansion. The structural bet here is genuinely large. Every mainstream phone in 2026 was planned in 2023, before the current AI capability explosion. Ptacek is wagering that the 2023 assumptions baked into those product cycles — that users consume apps built by professionals, that OS security models must sandbox untrusted code — are already obsolete. If he's right, the phone is early to a platform shift. If he's wrong, it's a device nobody asked for. Notably absent: any discussion of what happens to trust, security, and reliability when users run AI-generated code they don't understand. Ptacek built his career on the insight that software is insecure by default. The essay waves away the OS sandboxing model as unnecessary when code has "the same provenance," but provenance in an AI-generated world is a deeply unsolved problem. An LLM that generates a weather app and an LLM that generates a keylogger have identical provenance. This is a vision piece attached to a company announcement, and should be read as such. The core observation — that AI is compressing the gap between intent and implementation — is real and already visible. The leap to "therefore we need a new phone" requires you to accept that this compression will be total, fast, and irreversible. That's a bet, not a prediction.
