Sam Altman on ChatGPT Dots, 1.2B weekly users, OpenAI's chip program, and why alignment is still unsolved science

Sam Altman at OpenAI DevDay

1.2 billion weekly users. That number, which surfaced the day of the event, is the backdrop against which Sam Altman is framing OpenAI's next phase — one he describes as moving from a pull-based product into something that reaches out to you.

ChatGPT Dots and the always-on shift

Dots, OpenAI's ambient agent layer, is the product Altman is most focused on right now. The pivot it represents is straightforward: ChatGPT has always required a prompt to do anything; Dots can initiate. Altman says his own experience over the past few weeks has been that OpenAI has calibrated it too conservatively — everything his Dot sends him is useful, which suggests there's still value being left on the table. He acknowledges the obvious risk, that proactive pings become a retention mechanic, and says OpenAI will try hard not to go that way. But the demand question he poses is telling: if users love one Dot, will they want 10, or 100? That's not rhetorical. He treats it as a genuine infrastructure planning assumption.

“I think it's so important to build an open ecosystem here and we want people to take well, to bring whatever they want into ChatGPT... One thing that OpenAI has that doesn't get as much attention is our chip program which will start to come online in the first half of twenty twenty seven and will scale a lot in future years... Anyone who says we have solved the science of alignment, I believe is wrong in a very dangerous way.”

Open ecosystem and portability

Altman is direct that OpenAI's subscription should be portable. Users should be able to bring their ChatGPT compute into third-party apps, and developers building on top of the platform should be able to serve users through it without absorbing the cost of free-tier scale. He describes seeing other AI companies move in the opposite direction — locking subscriptions to their own surfaces — and says OpenAI intends to push the open model further.

Model picker fatigue

The model picker, a staple of developer interfaces, is something Altman wants to retire for most users. The goal is a single model that makes routing decisions automatically — defaulting to the right cost and speed trade-off without asking. Developers will keep fine-grained controls, but for everyone else, Altman wants the underlying model selection to be invisible.

Chip program

OpenAI's custom chip program, which hasn't received much public attention, is expected to come online in the first half of 2027 and scale from there. Altman frames it as an inference efficiency play — more useful output per watt — rather than a training compute story. Given the demand growth across consumer, developer, and enterprise, he argues more efficient chips are a necessity, not a hedge.

Startup culture at scale

Altman repeats advice he heard recently from a new OpenAI hire who had come from a slower public company: the way large organizations rot is that too many people have too little real work, and they start competing over artificial hierarchy instead. The antidote is keeping real workload per person growing. It's the kind of thing that's easy to say and hard to operationalize, but it's his stated model.

He's equally pointed about AI-generated communication bloat. His rule: if you're writing bullet points, having AI expand them into a document, and your recipient is going to summarize it back to bullet points, just send the bullet points.

Alignment

On alignment, Altman is unambiguous that anyone claiming the science is solved is wrong, and dangerously so. The engineering work — better sandboxes, better monitoring — is ongoing but insufficient. Building models that substantially exceed human intelligence requires actually solving the science of alignment first, not just the tooling around it. He doesn't offer a timeline or a framework, only the position that more research progress is needed and that OpenAI expects to make it.

Forecasting and sequence

Altman is candid that the order in which capabilities show up is harder to predict than the fact that they will. He expected Navier-Stokes-level scientific breakthroughs to precede AI booking haircuts. The reverse happened, and he thinks it reflects something real: more people want their lives to be less hectic than want to benefit from frontier science, at least in ways they can directly feel. His working view is that OpenAI will be right about the broad trajectory — continued capability gains, accelerating scientific discovery, durable human values — and wrong about the specific sequence.