In an interview with WIRED's Steven Levy, Tim O'Reilly—publisher, internet pioneer, venture capitalist, and conference organizer—warns that the biggest AI labs are "reading the future wrong." His measure for any company or society has long been to "create more value than you capture," and he now applies that test to artificial intelligence. The result: he worries hyperscalers are repeating Microsoft's 1990s playbook of locking users into their products, and he is pushing for open-source AI that unlocks the whole stack, not just neural-net weights. (During the interview Levy and O'Reilly also found themselves disagreeing about AI's role in producing original content.)
The case against frontier models
According to O'Reilly, the big labs "have told themselves a narrative where having the biggest, best model is the key to the future." But models like Claude, he says, are "optimizing for particular use cases, but they aren't necessarily the use cases that people want." He argues that frontier AI breakthroughs are "pushing us further away from what ordinary people are going to need."
O'Reilly also warns of a geopolitical consequence: "We could win frontier AI here in the US, and China will kick our ass because they have lower-level models diffused widely through society." He says increasingly the latest and greatest models are "better for some things and worse for others," citing industry discussion that Fable and Sol are worse writers than lower-level models—though Anthropic and OpenAI, the companies behind those models, would disagree.
Open source is more than open weights
When most people talk about open-source AI, O'Reilly says, they mean open-weight models. But in the 1990s, he recalls, "when everybody else was focused on open-source licenses, I was like, 'No, no, it's about the architecture of the system. Does it enable participation?'" That conviction shapes his current demands: a "clean separation between the model, the harness, and the application." Instead, he charges, the big labs have "built an architecture of control rather than an architecture of freedom and participation, so they have the ability to track you."
For enterprise users, the consequence is lock-in. O'Reilly notes that Mark Zuckerberg's thesis is that Meta will give users "the AI that knows you best" and lock them in; the open-source vision, he says, needs to say no and give users the ability to switch models and providers.
Security debate: frontier models, not open weights
O'Reilly turns the standard security concern about open-source AI on its head. "All of the cybersecurity incidents we've seen are from the frontier models," he told Levy. Risks such as cybersecurity and the ability to develop pathogens are, in his view, "an argument for slowing down the frontier models more than an argument for restricting open-weight models."
A path beyond the frontier
O'Reilly doesn't predict the future, he said, but "the world is proceeding the way that I hoped it would." He suggests frontier models might end up like mainframes or supercomputers—"aimed at really hard problems, which are not actually the thing that gets diffused throughout society." He points to efforts such as Pi, an open-source agentic harness, and the work of his nonprofit, the AI Disclosures Project, which is developing the idea of an "open-memory consortium" to prevent permanent lock-in.
Two AI architectures compared
| Dimension | Frontier-lab approach | Open-source approach (O'Reilly) |
|---|---|---|
| Core focus | Biggest, best model | Architecture that enables participation |
| Model access | Open-weight only, control of stack | Full stack, remove control and tracking |
| User agency | Lock-in (e.g., Meta's "knows you best" AI) | Ability to switch models and providers |
| Security record | Incidents "from the frontier models" | Open-weight models not the cited risk |
| Societal impact | Hard problems, not diffused | Lower-level models diffused widely |
What it means for enterprise buyers
O'Reilly's argument reframes the AI procurement question: are you buying into an architecture of control or an architecture of freedom? He says the most important thing is the ability for users to "embed my own special sauce," and he sees AI as a new creative medium he uses extensively—even blogging about his chats with the AI. For technology leaders, the decision involves whether their AI investments can switch models, harnesses, and applications—or whether the frontier labs' architecture, and their ability to track you, will define the relationship.
Sources:
