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

Host Lenny Rachitsky speaks with Tibo Sottiaux, who leads ChatGPT and Codex at OpenAI, about a shift from conversational AI toward always-on agents that act on users’ behalf. The discussion centers on Dots, described in the notes as OpenAI’s new personal agent platform, and the prospect that agents will soon perform most actions on the internet. The core thesis is a change in how people interact with software: less direct operation of individual tools and more delegation to AI systems.

The episode connects that vision to product architecture, developer ecosystems and workplace behavior. Topics include bringing Codex, ChatGPT and Dots together, plugin discovery and revenue sharing, and the argument that loops, graphs and fine-tuning agent workflows are a passing phase. Its most concrete operational example is Sottiaux’s account of his Dot warning him about a production outage five minutes before a launch. That anecdote illustrates proactive monitoring rather than merely answering prompts, but the notes provide no broader reliability data or detailed explanation of how the warning worked.

The conversation took place at DevDay, hours after Sottiaux’s team launched more than 20 products, giving it the context of a broad product expansion rather than an isolated feature announcement. Alongside the benefits of delegation, the episode explicitly addresses agent fatigue, loneliness, context switching and pressure to accomplish more with AI. It also covers where humans remain valuable, changing career skills, and autonomy, trust and learning from mistakes inside OpenAI. These subjects frame agent adoption as an organizational and human challenge as well as a technical one.

Important questions remain unresolved in the available notes: how agent guardrails work across connected devices, how plugin recommendations and revenue sharing operate, and what evidence supports the forecast of agent-dominated internet activity. Safety and alignment receive a dedicated discussion, as does moving beyond model pickers toward simpler interfaces, but the notes do not establish specific safeguards or conclusions. The episode therefore presents an executive’s product direction and forecasts, supported by a concrete usage anecdote, rather than a demonstrated timetable for the transition.