AI assistants enter the interfaces users already inhabit
OpenAI brought voice-driven multi-step work to desktop, Cognition bought messaging assistant Poke, and Attie added social research questions. Assistants are entering existing interfaces before independent evidence shows how reliably they work there.
Artificial Intelligence··Morning
Voice starts work on the desktop
ChatGPT Voice on desktop works with ChatGPT Work and Codex subscriptions, allowing multi-step spoken commands and control of several agents at once. On macOS it can read screen content as alt text through Appshots. OpenAI's demonstration shows one command opening a thread, creating a pull request and investigating a bug's root cause. That is a product demo, not a measured task-completion or error rate.[1]
Messaging and social networks become research surfaces
Poke, reportedly acquired by Cognition at a low-nine-figure value, operates through iMessage, SMS, Telegram and WhatsApp. Users exchanged more than 100 million messages in three months, but the product was expensive to run and integration experiments are deferred until next year. Attie's Quests feature instead aims to answer open-ended questions over Bluesky and the AT Protocol; its beta opens through a waitlist. The cited 45.6 million accounts measure registrations, not active use.[2], [3]
The common mechanism is the distribution channel
What joins the three developments is not a proven new capability but a distribution mechanism: the assistant is moving through desktop voice, private messaging and an open social network. Each interface adds a different constraint on access, context, sourcing and human hand-off; reliability is not yet known. Comparing the same tasks across channels after general release, using completion, error, citation and hand-off measures, would test reliability. For now, demonstrations, message volume and account totals describe reach rather than work quality.[1], [2], [3]
Related columns
For more information on this topic, you can read the related columns.