OpenAI's President Says: The Best Product Is No Product
- OpenAI President Greg Brockman sat down for an exclusive interview at the Big Technology AI Summit in San Francisco: ChatGPT is on the verge of one billion users, the company is pursuing an IPO at a trillion-dollar valuation, and recently closed a $122 billion round — the largest venture round in history.
- His central thesis is that interfaces will eventually dissolve: in the future, users won't need apps or products — they'll speak directly to a persistent AGI entity, give it goals, and let it act.
- Codex has evolved from a coding tool into a general-purpose Agent framework; non-technical teams like PR and ops are using it heavily, and its internal adoption is approaching Slack-level penetration.
- He asserts that scaling laws remain valid — every time a "model hits a wall," investigation reveals an engineering bug, not a real ceiling. OpenAI's compute strategy is "buy everything," and the company has been building its own chips for years.
- OpenAI is developing multiple hardware devices and building an end-to-end bidirectional voice model — aiming to replace turn-taking with a natural, interruptible conversation like talking to a real person.
Eleven Years In: How Big Is OpenAI Now
OpenAI President and co-founder Greg Brockman recently sat down for a headline interview with Alex Kantrowitz at the Big Technology AI Summit (Commonwealth Club) in San Francisco — their fourth conversation, covering the Agent era, interface evolution, the compute race, and the product roadmap.
In 2015, he co-founded OpenAI alongside Elon Musk and Sam Altman in pursuit of AGI. Eleven years later, the company is on track to go public at a trillion-dollar valuation within the next year. What makes this interview matter is who's speaking — someone at the very top of that scale, talking about what the product should become next.
Why it matters: This is the first public confirmation from OpenAI's president that multiple hardware devices are in development. Codex Agent has reached Slack-level internal adoption. The context window has grown from 2K tokens in 2022 to 5–12 million tokens today. And 230 million people use ChatGPT to ask health questions every week.
He Says: The Best Product Is No Product
The host floated a framing: OpenAI is merging Codex, the browser, and ChatGPT into a "super app" where everything starts from a prompt, and the technology then calls your browser or computer to get things done. Brockman agreed with the view — but wanted to zoom out further.
What he actually wants to build is AGI. When ChatGPT launched in 2022, it had no memory, no tools, no context — conversational intelligence was just a tiny slice of what people need to get things done. What OpenAI is moving toward is an AI that genuinely looks out for you: you give it goals, and it continuously thinks about "what can I do for you today" — solving hard problems and handling the mundane, so you wake up with your inbox sorted and your medical options researched and followed up.
His conclusion: you almost don't want an "interface" or a "product." What you want is to talk to a persistent entity the way you'd talk to another person — and have it help you reach your goals. The long-term direction is simplification and unification: fewer buttons to click, switches to flip, modes to switch between. The interface gradually dissolves.
He made a pointed remark about the relationship between machines and people: the beauty of AI is that it brings the machine closer to the human — rather than forcing people to adapt to files, folders, and other details that are about how machines work, not how humans think.
Don't Believe It? Codex Already Works This Way Inside OpenAI
"Interface dissolution" sounds like a distant vision — but Brockman says it's already happening. Despite having "code" in the name, Codex has little to do with code. It's a general-purpose execution framework that can use tools — an Agent. You can connect it to Slack, Gmail, and calendars. Many non-technical employees at OpenAI are already using it.
A PR team member organizing an event had Codex reach out to every attendee to ask about dietary preferences, then automatically arrange the entire seating chart — handling all the busywork so she could focus on what actually mattered.
Codex started with a software engineering focus, but non-engineering usage has exploded. Inside OpenAI, it has essentially reached the same penetration as Slack — the whole company runs on Slack and barely uses email. Codex is starting to feel the same way.
The host described this as OpenAI becoming an "operating system," but Brockman pushed back. He said an operating system is almost a concept from another era — a layer in the tech stack — while what he's thinking about is "what's the ideal interface to AGI?" It might be called a kind of "personal AGI." That Agent would have its own computer and access permissions, like an ideal colleague who comes over and types on your machine with a set of delegated access rights — not fundamentally different from building a trust relationship with a human assistant.
Why Old-School "AI Assistants" Simply Didn't Work
OpenAI's first attempt at tool use inside ChatGPT was the plugins launched in March–April 2023. Brockman's verdict: "completely didn't work." But he says the problem wasn't the idea — the underlying conditions weren't ready; the model wasn't prepared. The timing is right now not because AI suddenly got smarter, but because the foundations changed.
- Only 3 connectors could be exposed at once — add more and it started forgetting
- Context capped at 2K–4K tokens — virtually no memory
- Using tools caused it to forget things — like early computers in the '60s and '70s with tiny memory
- Models can use hundreds of tools simultaneously, connected to entire file systems
- Context reaches 5–12 million tokens — nearly the whole internet at your fingertips
- Can control entire browsers and computers; already solving unsolved math and physics problems
The context window is the amount of text an AI can "remember" and process at once — think of it as the size of your work desk. The bigger the desk, the more material you can spread out and multitask with. In 2022, that desk was a single sheet of paper. Today it's an entire room.
He also pointed to an earlier failure: OpenAI had tried letting you call an Uber from inside ChatGPT. Lots of companies attempted in-chat actions, but nothing ever really caught on. The difference now is that a chatbot can take over your actual browser or computer — you don't have to worry about whether a plugin works; it gets things done by controlling your machine directly.
How Much Do You Trust the AI? That's the Real Product Design
Brockman kept returning to one word: trust. The defining differentiator in the Agent era isn't technology — it's the trust boundary: how much autonomy the AI is actually allowed. He says trust must be earned, not granted directly; the way to earn it is to give operators plenty of tools, controls, oversight, and governance. He illustrated three levels of autonomy using the example of a single email — toggle through them yourself.
"Can I send this?"
The AI uses a Gmail connector to find the right people in your inbox and drafts the email — then stops to ask for your permission. Safest mode: it asks before every action, and you retain full final control.
"I've drafted it — you send it."
The connector is configured to block sending, so it handles everything up to the draft and hands the final send back to you. Useful when permissions are hard-capped at the product level — no judgment call needed in the moment.
"I drafted it, and it's already sent."
Once you've built enough trust with the system, it completes the whole task and reports back afterward. Maximum autonomy — on the condition that trust has been earned through a track record.
Same AI, same email — but different trust levels produce completely different products. Brockman says this idea of "giving operators sufficient oversight and governance" is exactly what differentiates Agent products. Whoever designs the trust boundary well is the one whose Agent actually gets used.
80 Years, No Ceiling: What Is the Scaling Law
A few years ago, a popular claim emerged: large language models were about to hit a wall. Brockman says that's wrong. His answer has two parts, and the first is foundational science: scaling laws remain valid — one of the most mysterious and important empirical observations he knows of. Every time something "didn't scale as expected," investigation turned up a bug, a math error, or a flawed implementation — not an actual ceiling.
Give a neural network more compute, more data, and better architecture, and it gets smarter — like a factory scaling up: more lines, better machines, more output. No one has ever seen a point where more investment stops paying off. This pattern has been observed for about 80 years.
He says he looked into it: neural networks were designed in the 1940s, before computers existed, as a model of how the brain processes information. The first hardware implementation was the perceptron in 1959. The field's milestone achievements follow an extremely smooth, predictable path — one defined by more compute being invested.
So the fundamentals allow it — the hard part is engineering. Building these massive supercomputers is expensive and difficult. You have to develop your own network protocols and have people who understand every layer of the stack, because neural networks have no abstraction layers — a small error anywhere cascades and surfaces somewhere else entirely. Assemble the right team, put the mission in front of them, and grind — the result is achievable.
"Buy All the Compute": What's the Logic Behind It
The host recalled asking Brockman a few months ago how much compute to buy. His answer: "all of it." "No, seriously" — "all of it." It sounds like a reckless bet, but he says it's a fundamental judgment call. The first pillar: compute is a genuinely scarce resource. The world simply will not have enough compute to meet all demand. We are heading toward a compute-driven economy.
The most concrete evidence he offered was the massive gap between two user numbers.
Only about 10–20 million people currently use Agents — not yet a global scale. ChatGPT has roughly one billion users, but Agent capabilities haven't reached that level, and usage depth is tiny compared to where it needs to go. That ~50x gap represents untapped market, signaling that mainstream Agent adoption is still in its earliest days.
The second pillar is counterintuitive: the Wall Street Journal reported that OpenAI's upcoming models may be drastically cheaper — but Brockman says lower prices won't compress demand. Instead, like the Jevons Paradox, demand will surge. Frontier intelligence is always the most expensive, but in a year, today's capabilities will feel ordinary and widely accessible — while something better will have emerged that makes you not want to use the old thing.
When the efficiency of using a resource improves, demand doesn't fall — it explodes, because lower costs bring more people and more use cases into range. When electricity got cheaper, people bought more appliances. When oil extraction became more efficient, global oil consumption rose. AI becoming cheaper follows the same logic.
↻ The flywheel spins — which is why lower prices mean you should still "buy everything"
See More: H100 Price Inversion and Custom Chips
Responding to Satya Nadella's claim that "models are being commoditized," Brockman pushed back: no layer of the tech stack gets eliminated — these layers multiply together. As for compute at the bottom layer, the value is evident from what people are willing to pay for H100s. Hopper is a prior-gen chip that normally nobody would want, yet its market price has actually risen relative to newer hardware — a price inversion — because everyone is facing avalanche-like demand.
He also noted that OpenAI's custom chip project has been in development for years with exciting progress — more announcements coming soon. Achieving full vertical supply chain integration is rare. He also observed a shift on the customer side: people used to think "get AI fast or fall behind"; now they're asking "does this actually deliver ROI?" — so OpenAI released a spend-control feature that very day.
He Closed with Healthcare: His Most Personal Answer
Brockman returned to health multiple times throughout the interview, and closed with it. He says applying AI to medicine is the personal motivation behind everything he's built at OpenAI. This isn't vision — it's something he's watching happen in real time.
His friend Sid Sijbrandij, CEO of GitLab, was diagnosed with cancer. After exhausting all available diagnostic tests, he fed the data into ChatGPT; a small team also built a dedicated application for him. Together, they managed to bring the cancer under control. There's also a dog named Rosie in Australia whose owner ran a biopsy on the cancer, put the mutation data through AlphaFold, and with a chatbot's help designed an mRNA vaccine that was injected into the dog — the tumor shrank, and Rosie can jump over tables again.
Brockman says this will absolutely become the norm, not the exception. He pointed out that patients have historically been disempowered — forced to become their own doctors. You're the one making decisions and bearing the consequences; if a doctor makes a mistake, it's the rest of your life that pays. That's a very different set of incentives. This is also personal to him: his wife has several health conditions, and he says that without ChatGPT, he doesn't know how he'd manage many of her situations.
Among the evidence he cited was a development OpenAI announced that same day: in peer-reviewed literature, physicians used O3 — one of OpenAI's earliest reasoning models — to find diagnoses for patients who had gone years without answers from doctors. One case involved a mystery illness that had gone undiagnosed for 20 years. Given that healthcare accounts for an enormous share of the economy and that physician and nurse burnout is a real crisis, he says if AI can help people prevent health problems before they occur, it can relieve enormous burden.
Without ChatGPT, I honestly don't know how I'd manage many of my wife's health situations. Personalized medicine is happening right now — it's not theoretical. Greg Brockman · Big Technology AI Summit