The week AI left the chat window

Perplexity's new Mac app touches your file system. Grok ships speech-to-text in 25 languages. Codex agents skip ads on your behalf. The interface is fragmenting — and that's the point.

The AI interface is finally escaping the chat window. This week brought a Perplexity desktop launch that touches your file system, a Grok speech-to-text API spanning 25 languages, and growing developer chatter about Codex agents that operate browsers like a junior employee. The shift from "type a prompt, read a response" to "give the AI a job and let it run" is no longer theoretical — it's shipping.

Perplexity launched its Personal Computer app for Mac on April 17, 2026, and the framing matters: per AIbase News, the app "directly accesses the file system and native applications," moving beyond browser-tabbed search into something closer to an OS-level assistant. xAI moved on a different axis, announcing that Grok's Speech to Text API is now generally available with multi-speaker transcription across 25 languages — voice as input is becoming table stakes, not a feature. And on the agentic side, one developer's reaction to demoing Codex computer use — "showed someone codex computer use and it clicked skipped ad" — captures the visceral moment when an autonomous agent's behavior crosses from impressive to uncanny.

These three threads point in the same direction. The AI product surface is fragmenting beyond chat: voice in, file system out, agents that click ads on your behalf. For builders, the design problem is no longer "what should the prompt template look like" — it's "what should the AI be allowed to do when nobody's watching." For buyers, the privacy and audit questions get serious fast: an assistant with file system access is a different threat model than one that only sees what you paste.

The visible-reasoning trend is the same story from another angle. As SitePoint describes, DeepSeek R1 exposes reasoning tokens as a first-class output, letting developers render chain-of-thought in real time. That's not a new model capability — it's a new API contract, and it tells you something about where the field is heading: more streams, more side-channels, more ways for the AI's work to be inspected and acted on.

The bottom line: If your product still lives entirely inside a text box, you're building for an interface that's already being replaced.

Sources