Tim O'Reilly warns large AI labs’ control architectures will cost them dearly

Tim O'Reilly, longtime technology-industry commentator, venture-capital investor and conference organizer, proposes a simple yardstick for AI companies: do they generate more value than they capture for themselves. He says the major AI labs fail this test and cautions against attempts to lock users into their products, just as Microsoft did in the nineties, while advocating full openness of AI technology.

O'Reilly stresses that most current discussion of “open source AI” actually refers only to open weights—the numerical parameters of neural networks—and that this is a mistake. In the nineties, while many focused on software licenses, he insisted the issue lay in system architecture and whether it permitted genuine participation. Today he calls for a clean separation between the model, the harness layer, and the application so that users and developers can embed their own logic without dependence on a single vendor.

His central claim is that the large labs read the future incorrectly. They are built on a narrative that the biggest, best model is the key to success and they fine-tune models such as Claude for specific use cases. He points to community debates in which models like Fable and Sol are said to outperform larger models on writing tasks—a point on which Anthropic and OpenAI disagree. Breakthroughs in frontier AI push the industry farther from the needs of ordinary people, and he believes the United States could lose to China, which spreads small, cheap models throughout society and enables free innovation.

Responding to the usual concern that open source could facilitate security circumvention, O'Reilly reverses the argument. He notes that every security incident observed so far has stemmed from the most closed and advanced models. Consequently, risks such as pathogen design or cyber attacks are, in his view, an argument for slowing down large models rather than restricting open-weight models.

O'Reilly declines to forecast the future but observes that the world is moving in the direction he hopes. He envisions a scenario where large models become mainframes—tools for exceptionally hard problems—without permeating other parts of society. At the same time he mentions projects such as Pi, an open harness for autonomous agents, and work by his nonprofit AI Disclosures Project on an open-memory consortium. The aim is to counter Meta’s strategy of locking users into its ecosystem by claiming its AI knows them best; O'Reilly sees open memory as the answer to that business model.