A startup with five founders, each owning 20% of the company, has a simple advantage: every one of them wins only if the company wins. OpenAI researcher Noam Brown thinks AI could, in principle, give that advantage to a company with 10,000 workers. The condition is that the AI workers are aligned, meaning their goals reliably match what the company actually wants.
Brown made the argument on the Dwarkesh Podcast, in an episode published on 17 September 2026. His point runs against a common expectation about AI and business: that tools which let one person do the work of many will mainly help small, fast newcomers against established firms, known as incumbents.
Why startups beat big companies
Brown agrees that AI helps startups. "It's much easier than ever before for one person to step in and be like, I'm going to make a multimillion dollar company," he said. "The AIs amplify an individual so much."
But he pointed to a second reason newcomers disrupt incumbents, alongside their greater willingness to take risks. As organizations grow, he said, the interests of the people inside them drift away from the interests of the organization. In a 10,000-person company, he sees "a lot more instances where people are territorial" or care mainly about getting headcount for their team, "building their fiefdoms," and gathering resources so they can publish impressive work and get promoted. "This is actually a real detriment," he said. "I think this explains a lot of why startups are able to disrupt incumbents."
The conditional case for incumbents
That drag is where Brown thinks AI workers could change the balance. "If the alignment problem is solved, then you don't have the issue of misalignment between individuals and the company," he said, before narrowing it: "At least that's mitigated." AI agents that are aligned well "could just be aligned to the interest of the company. And you can have 10,000 of them. And they're all going to be working as hard as if they were like a 20% share co-founder."
The argument depends entirely on that "if". Brown did not claim the alignment problem is solved, and the episode also took up how anyone would know whether models are actually aligned. His proposal describes what a large firm could gain once that question is answered, not what companies can do with today's agents.
It is also a different mechanism from the one in the host's own writing on AI firms, which stresses that cheaper coordination inside a company could let firms grow larger. Brown's point is about incentives: an employee's own career can pull against the company's success, and an aligned agent would have no such pull.
Copying workers and forking context
The exchange grew out of a question from the host, Dwarkesh Patel. In a January 2025 essay, Patel imagined fully automated firms staffed by human-level AIs and asked what would make them different from human organizations. On the podcast he summarized some answers: AIs can share context, the working information a model holds while doing a task, far more seamlessly than people. They can merge what they have learned. And a firm could spin up or shut down any number of copies. Instead of the slog of recruiting, "your best talent, you can just make an infinite copy of them," and a company could replicate its most effective teams, or whole effective organizations. The essay presents this as speculation about future, broadly capable AI organizations rather than a description of any existing company.
Brown picked out the copying point. "If you have a person and you want just two copies of them, you can't just clone the person," he said. With AIs, "it's actually really easy to just say, okay, well, just fork yourself and then have those copies work on this thing and then merge back together."
Part of this, he said, already exists. When the multi-agent systems for the models he called Astra and 5.6 Sol spin up subagents (helper agents given part of a larger job), "the context is just forked. So it has all the context that's relevant." A new helper starts with what its parent knew instead of needing a briefing. That is the piece Brown described as working today. Merging many copies' knowledge back into one system and cloning a company's best teams remain part of the hypothetical picture. OpenAI's developer documentation for its multi-agent API describes a related pattern: a root agent hands bounded tasks to subagents, each keeps its own context, and the root agent combines their results.
Agents that talk too fast to follow
The discussion also turned to speed. Before Brown spoke, the conversation raised the prospect of AI systems thinking more than ten times faster than people, working without sleep and collaborating more intensely than humans can. In that picture, a "shadow organization" inside a company might do in a week what a human organization does in a year.
Brown said working with today's models has been "surprisingly natural," but that this could change. He described "ultra fast modes" that make sampling, the process by which a model generates its output, roughly 10 or 15 times faster, and said "it's going to be pretty hard to keep up with these things." His expectation is that agents will move very fast when they talk to each other, but will recognize when they are talking to a person instead of another agent and behave differently.
What hasn't been measured
Brown was more cautious when the discussion turned to whether 10,000 AIs could work together on a hard problem such as the Navier–Stokes equations better than 10,000 newly hired mathematicians, who would struggle to cooperate at first. "I want to be conservative here because we haven't measured how effective the 10,000 agents are at coordinating," he said. "We think it helped," but there are no good measurements showing, for example, that 10,000 agents gave a twofold speedup over 2,000.
OpenAI's documentation lists practical limits of its own: running agents in parallel uses more tokens, the units of text models process, and helps less when steps have to happen in sequence or when several agents keep changing the same shared resource.
Brown went further than the caveat about measurement. "I think it is very possible that 10,000 humans are better at coordinating than 10,000 agents right now," he said.