Peter Yang says he now gets real work done on a walk, without touching a keyboard.
In his post “How We Use Computers Is About to Change Forever,” Yang describes turning on ChatGPT Voice while out walking and rambling instructions at it: check his inbox, follow up on an open thread, move a task forward. He writes that directing an AI agent this way feels “magical compared to tracking everything myself with a keyboard and mouse.”
From that habit, he predicts a bigger shift is coming for how people use computers at all.
Who Yang is
Yang is a product lead at Roblox and writes Creator Economy (published as “Behind the Craft”), a newsletter on AI and product work. He has previously led product teams at Reddit, Twitch, and Meta.
The five shifts he predicts
Yang lays out five changes he thinks follow from voice becoming the main way people direct AI:
- Voice becomes the orchestration layer. Talking replaces typing as the primary way to direct AI agents through a day’s work.
- Personal computers move to the cloud. Yang argues: “Your laptop’s specs will matter less when your real computer is in the cloud,” pointing to always-on cloud agents like Grok Bot, which BuilderWithin covered two days ago, as an early version of this.
- Products get built for agents first. Software should work inside “your favorite AI harness” (the app that hosts and runs an AI agent, giving it tools to act with), with an agent as the primary user instead of a human clicking through a UI, the visual screen of buttons and menus people normally use.
- Software gets commoditized. Yang writes that “software has become easier to build but much harder to monetize,” so the products that win pair software with something an agent can’t easily copy, like a proprietary dataset or network.
- Trust decides the winners. He calls trust “the biggest barrier to broad agent adoption” and says the products that win will make transparency about what an agent can access a core feature, not an afterthought.
What holds up, and what’s still a bet
The voice-as-input shift is the part with the most evidence behind it already. Voice dictation tools and voice-driven AI assistants have gone from novelty to daily habit for a meaningful slice of builders in the past two years, and Yang’s own workflow is a real example, not a hypothetical.
The cloud-computer and agent-first-product predictions are further out. They depend on always-on cloud agents like Grok Bot becoming reliable and cheap enough for daily use, which is not yet established. Yang’s own post treats these as bets he’s making, not conclusions he’s already confirmed. His trust point is the most testable near-term: builders can already watch which AI products ask for narrow, explained permissions versus broad, unexplained account access.
Why it’s notable
This is a builder describing his own workflow shift, then generalizing it into a five-part thesis about where computing goes next. The individual claims aren’t new on their own. Voice input, cloud compute, and agent-native products have each been discussed elsewhere. Yang’s contribution is tying them together as one connected shift, driven by the same underlying cause: once an agent can act on your behalf, the interface for directing it stops needing to be a keyboard and screen.
What it means for builders
Treat this as a framework for questions to ask about your own product, not a roadmap to build against yet.
If you’re building a tool today, ask whether it works when the “user” is an AI agent acting on someone’s behalf, not just a human clicking through your UI. Ask whether your product’s value survives if a competitor can reproduce the AI-generated parts of it in a weekend, and what non-software asset, data, relationships, credentials, would still make yours the one people choose. And ask what you tell users about what an agent working on your product can see and touch, before they have to ask.
Voice-first, agent-first computing is a real direction, not a shipped reality. Yang is betting early. Builders don’t need to make the same bet today, but the questions underneath it are worth answering now.
End of article