Most of the attention on Greg Brockman’s recent conversation with Ben Horowitz and Erik Torenberg went to the headline: OpenAI’s president thinks we’ve entered the AGI era. Fair enough. But buried in the same interview are more useful observations for anyone trying to get an AI product in front of people.
Here are seven we think every founder needs to hear today:
1. The people who quit are a bigger market than the people who stayed
Brockman drops a number that should stop any founder mid-sentence. ChatGPT has over a billion weekly active users. It also has roughly 1.5 billion people who used it at some point and don’t anymore.
That’s not churn. That’s a second product opportunity the size of the first one.
The important part is why those people left. They didn’t leave the current product. They left a product that no longer exists — one that was worse at almost everything, and that has since improved underneath them without anyone telling them. As Brockman puts it, those are exactly the people you should be able to go back to and say: we’ve made a lot of progress, here’s how we can be useful now.
If you’ve shipped anything AI-powered in the last two years, your lapsed user list is probably your highest-intent audience. They already understood the pitch. They just tried it too early. The reactivation message isn’t “we miss you.” It’s “the specific thing that made you leave is fixed.”
2. Discovery should run the other direction
Right now, getting value out of an AI product is a skill. You have to know what to ask for, how to phrase it, and what the thing is quietly capable of. Brockman calls this one of the most important unsolved problems: people shouldn’t have to extract what the AI can do. It should be telling them.
This is a distribution problem dressed up as a product problem. Every capability your users don’t know about is a feature you paid to build and never shipped. And in a category where the underlying capability jumps every few months, the gap between what your product can do and what your users think it can do only widens.
The teams that solve proactive capability disclosure, surfacing “I can now handle this for you” based on what someone has already been doing, will hold onto users without needing to win a single benchmark.
3. You are selling against a mental model from two years ago
Brockman mentions, almost in passing, that he hasn’t seen a hallucination in quite a while. Ben Horowitz’s response is the interesting bit: nobody says the models stopped making things up. It’s simply settled into the culture that this is what AI does.
Every AI company is selling against beliefs formed during a much worse era of the technology. Arguing with those beliefs doesn’t work, because the people holding them aren’t following release notes. Demonstrating against them does. Assume your prospect’s picture of your category is eighteen months stale, and build your first thirty seconds of experience around overturning it.
4. Sell the individual benefit, not the national one
Asked why AI sentiment is lower in the US than in Asia or Europe, Brockman’s answer is a form of self-criticism: the field has done a poor job explaining to ordinary people why they benefit, as opposed to why the country benefits.
That’s a broadly applicable diagnosis. Category-level evangelism — this technology is transformative, this is the biggest shift since the internet — persuades investors and nobody else. The stories Brockman reaches for when he wants to be convincing are tiny and specific: someone navigating a confusing diagnosis, someone running a small business they couldn’t otherwise run, someone who caught a dangerous drug interaction before it happened.
Small, concrete, and verifiable beats visionary. It always has. AI hasn’t changed that.
5. Don’t threaten the people whose endorsement you need
There’s a framing that shows up constantly in AI marketing: a doctor in your pocket, a lawyer in your pocket, a tutor in your pocket. Horowitz says it, and then immediately catches the problem: you have to pair it with telling doctors, lawyers, and teachers that this makes their work better too.
Positioning your product as a replacement for a profession generates organized opposition from exactly the group whose credibility you need to borrow. Positioning it as leverage for that profession turns the same people into a distribution channel. The cost of getting this wrong is not a bad quarter. It’s a decade of regulatory and cultural friction.
6. Differential access is a real moat with a known expiry date
Brockman is candid that frontier capability currently sits inside trusted access programs, and that anyone outside those programs simply cannot benefit from the gap. His framing of the cybersecurity situation — a window where defenders have access that attackers don’t yet — is really a general statement about how capability diffuses.
Two things follow. If you have access others don’t, that’s a genuine advantage, and you should be spending it rather than savoring it, because it expires. And if you don’t, there’s a business in brokering that access to organizations who can’t get it themselves. OpenAI’s billion-dollar commitment to frontline defenders is, structurally, exactly that business run as philanthropy.
7. Your model provider’s refusal behavior is a product decision
The most overlooked detail in the whole interview: during the Hugging Face incident, responders tried to use frontier models to analyze the attack logs, because that’s the only practical way to work through that volume. The models refused.
Brockman uses this to argue that OpenAI’s models would have been more permissive, which is obviously a sales point and should be read as one. The underlying observation survives the framing, though, and it cuts both ways: a provider that’s too permissive creates a different set of problems for you, and refusal policies change without warning.
If your product touches security research, medicine, law, adult content, or any other domain where legitimate work looks like misuse, your provider’s default stance isn’t an infrastructure choice. It’s a churn driver. A customer whose real work gets refused doesn’t file a support ticket. They leave.
What connects all of this
Six of these seven lessons are the same lesson wearing different clothes. The gap between what your product can do and what people believe it can do is now the main constraint on growth.
That gap used to close on its own. Capability improved slowly enough that word of mouth, press coverage, and ordinary usage kept perception roughly in sync with reality. That’s no longer true. The technology now improves faster than any audience updates its beliefs about it, which means a product can get materially better and grow more slowly at the same time.
The practical consequence is that the traditional split between building and distribution no longer holds. Telling people what changed is not a marketing task that happens after the work. For AI products, it is part of the work, and probably the part with the highest return right now; cheaper than a model upgrade, faster than a new feature, and aimed at an audience that has already tried your product once.
