OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux

Often I look at what people are building and think: you’re not quite getting it. If you really imagined models becoming roughly ten times better within a year—cheaper, faster, seamlessly multimodal—you would build very differently. Most activity online will be carried out by agents, so products need to handle that scale and economics, while we also invest much more in delightful human experiences.

I don’t think the future is everyone fiddling with loops and graphs forever. You want an extremely capable agent that works around the clock, understands your goals and preferences, learns from feedback, and is there through whichever screen—or no screen—you happen to be using. Walk into a room, pick it up in email, send it a text; it should help when needed and disappear when it isn’t. Carrying a laptop around like a brick is not the end state.

The shape of my agent team expands and contracts. As I push the frontier, I create more specialized agents; then a model breakthrough arrives, a larger agent can hold more context and learn, and suddenly I need fewer. We’re starting with one primary dot that knows you deeply, then you’ll be able to create a virtual team with specific roles. I have one for Twitter because that can genuinely be the work of an entire dot.

This should make work more natural, not turn every task into a solitary prompting adventure. I want an agent in the physical space with people: we’re talking, jotting something on a shared surface, it starts building quietly, and we keep having ideas together. We have to reduce configuration fatigue and the pressure to do more just because we can. A week fully disconnected changes how creatively I think; maybe fewer meetings and better rest make you more productive, not less.

The sleeper hit is the open ecosystem. A useful plug-in should earn discovery through retention and quality, not just the right words around it, and partners whose products see real use should share in the economics. We’re opening sign-in, extensions, discovery, and distribution because people will build things we never anticipated. That community feedback is humbling, inspiring, and necessary if we want technology that is actually useful for humans.

The technology is still clunky until, suddenly, it isn’t. Voice has changed my own behavior more than I expected; I dictate and call my agent constantly. I still care deeply about human creativity, taste, kindness, collaboration, and learning fast—roles are blurring, and this is a moment to build. But we also have a responsibility to harden safety, security, guardrails, and monitoring before giving powerful systems broad access. I want the app to reach this essence of simplicity. No model picker, no reasoning-effort homework, no needing a PhD in model pickers. Just talk to it.