Lenny's Podcast: Product | Career | Growth
OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux
This debrief was written and voiced by AI from the episode's audio.
🎧 AI Debrief
🎙️ Full Episode
Key Moments
Tibo builds bigger teams of agents as he pushes the frontier, then collapses them back into one agent after each model leap. He argues that hand-tuned loops and graphs aren't the future; systems that learn are.
The vision for Dots is a persistent intelligence available through any screen or channel, from meetings to email to text, that gets out of the way when you don't need it.
The open ecosystem is Tibo's underrated launch. It started as a handshake deal with OpenCode and Pi creators and now includes plugin distribution to 1.2B users with shared economics for popular plugins.
Plugin discovery inside ChatGPT is driven by retention and real utility, not keyword optimization. Plugins that stop delivering stop getting recommended.
Tibo's Dot connected the dots between Dev Day, a live demo and a production outage, then warned him minutes before he went on stage. It also offered to fix the outage itself.
The Dots harness doesn't run on your machine. It has its own computer, can connect to many devices at once, and specialist Dots run with extra guardrails.
The skills trending up are taste, user empathy and the founder mindset. OpenAI has more than 120 ex-founders on staff and calls itself a mega startup.
Ahmed Ibrahim went from new grad to owning OpenAI's compute fleet and much of the Codex harness. Tibo credits his kindness, his collaboration and his sponge-like learning speed.
Most internet actions will be taken by agents, models will get cheaper and faster quickly, and modalities will merge. Builders should design for tech that is 10x better a year from now.
OpenAI spends growing compute on secondary monitors that watch primary agents for risky actions and prompt injection. Tibo says most of the API stack investment goes to safety, and they held back a more capable model.
Search Episode
Summary
Recorded live at OpenAI's DevDay, hours after a launch of more than twenty products, Lenny talks with Tibo Sottiaux, who leads ChatGPT and Codex. The centerpiece is Dots, a personal agent that runs around the clock, learns your goals and preferences, works across channels like text, email and meetings, and has no model picker. Users currently get one primary Dot, with multiple role-specific Dots coming soon. Tibo argues that hand-tuning agent loops and graphs is a passing phase that will give way to systems that simply learn. He also says Chat and Work modes will merge, and that Dots' capabilities will eventually ship to ChatGPT's roughly one point two billion users. Tibo calls the open ecosystem the launch's 'sleeper hit.' It includes sign-in with ChatGPT for partner tools, plugin discovery, and a revenue share for popular plugins. Plugins are recommended based on retention and quality rather than marketing. His biggest 'not priced in' claim is that most actions on the internet will be taken by agents, while models get dramatically cheaper and faster and modalities merge. His advice to builders is to imagine everything ten times better within a year and to design for agent-scale traffic. On people and work, the skills gaining value are taste, user empathy and a founder mindset. OpenAI has more than 120 ex-founders. Roles are blurring, and new grads who learn fast and go all in on AI have an edge. Tibo addresses loneliness and pressure to do more, and argues AI should reduce noise rather than add 'one more prompt per second.' On safety, he points to heavy investment in secondary monitoring and to withholding a more capable model. His biggest frustration is complexity: users shouldn't need 'a PhD in model pickers.'
Key Takeaways
Most internet actions will soon be taken by agents. Build products and infrastructure (APIs, MCP servers, pricing) that can handle agent-scale traffic, as Notion learned when its MCP launched.
highOpenAI is converging on a single, always-on personal agent (Dots) with no configuration or model picker. Chat and Work are merging, and Dots' capabilities will eventually come to all ChatGPT users.
highHand-tuning agent loops and graphs is a transitional phase. Expect systems that learn from feedback and goals instead of requiring orchestration.
highThe open ecosystem is the underrated launch. It includes sign-in with ChatGPT for partners, plugin discovery to a huge user base, and revenue share for popular plugins. Ranking is driven by retention and quality, not keywords.
highBuild as if models will be roughly ten times better, cheaper and faster within a year, with all modalities integrated. Most builders are underestimating this.
mediumTaste, user empathy and a founder mindset are the skills trending up. Roles are blurring, and fast-learning juniors who go all in on AI have an advantage.
mediumFaster models restore flow. Ultra-fast inference has let Tibo run fewer parallel agents and get back into a creative, voice-driven flow state.
lowNotable Quotes
“If you were just pushing yourself and really imagining all of this being roughly ten times better than it is today in a year, you would build in a different way.”
— Tibo Sottiaux
His core critique of what builders are getting wrong about the pace of AI progress.
“It still feels a bit clunky, and I think it will feel clunky until it isn't.”
— Tibo Sottiaux
A crisp framing of how AI interfaces will suddenly cross a threshold into feeling natural.
“We may not have coders anymore, but we have more builders than ever.”
— Tibo Sottiaux
His answer to where humans remain valuable: taste, creativity and building.
“You kind of need a PhD in model pickers. And I just want to get rid of all of that.”
— Tibo Sottiaux
His biggest frustration with the current ChatGPT experience, which signals a push toward radical simplicity.
“That's the promise of AI, and it's not just one more prompt per second.”
— Tibo Sottiaux
His response to the pressure to run ever more agents and ship ever more.
“Build a good plugin.”
— Tibo Sottiaux
His blunt advice on plugin discovery. Recommendations are driven by retention and quality.
Predictions
The majority of actions on the internet will be taken by agents.
— Tibo Sottiaux · Confidence: high · Near future, unspecified
Models will become roughly ten times better, and much cheaper and faster, with all modalities seamlessly integrated.
— Tibo Sottiaux · Confidence: high · Within about a year
Users will be able to create multiple Dots with specific roles, forming a virtual team.
— Tibo Sottiaux · Confidence: high · Soon
Dots' capabilities will ship directly into ChatGPT for its full user base, and Chat and Work modes will merge.
— Tibo Sottiaux · Confidence: high · Eventually, unspecified
How people work will continue to change radically, with agents present in physical spaces and collaborative surfaces beyond chat.
— Tibo Sottiaux · Confidence: medium · Next two years
The partner and plugin ecosystem will grow quickly, with revenue share for popular plugins.
— Tibo Sottiaux · Confidence: medium · Coming months
Action Items
Audit your product for agent readiness: expose an API or MCP server, and plan capacity and pricing for agent-driven traffic.
Tibo expects most internet actions to come from agents. Notion saw a surge in traffic after launching its MCP.
Stress-test your roadmap by assuming models are ten times better, cheaper and faster a year from now.
Tibo believes most builders aren't pricing in the pace of improvement.
If you build a ChatGPT plugin, optimize for retention and real utility rather than discoverability tricks.
OpenAI recommends plugins based on retention and quality, and popular plugins get a revenue share.
Stop over-investing in hand-tuned agent loops and graphs. Favor flexible setups that can ride the next model upgrade.
Tibo sees orchestration fiddling as a passing phase that each model breakthrough undoes.
Try voice and dictation as your primary interface with agents.
Tibo changed his mind on voice and now does much of his work through dictation and calling his agent.
Invest in taste, user empathy and rapid learning. If early-career, behave like a sponge and be kind and collaborative.
These are the skills trending up in OpenAI's hiring, exemplified by new-grad-turned-leader Ahmed Ibrahim.
Use AI to cut noise from your work, for example a bot that audits your calendar for energy-draining meetings.
Both Lenny and Tibo argue AI should create focus and rest, not pressure to do more.