"Open-source AI" is mostly open weights, and O'Reilly says the lock-in sits in the rest

Tim O'Reilly told Wired's Steven Levy that when most people say "open-source AI" they mean only open-weight models, and that the real question is whether the surrounding architecture enables participation ^[3][4]. Wired frames his worry plainly: like Microsoft in the 1990s, today's hyperscalers are trying to lock users into their products ^[2]. For anyone signing a contract on the strength of the word "open," that distinction decides whether leaving is possible.

O'Reilly's stated test is structural, not legal. In the 1990s, he says, while everyone else argued about open-source licenses, his position was that what mattered was the architecture of the system and whether it enabled participation ^[4]. Applied to AI, that means a clean separation between the model, the harness, and the application, and he says buyers are not getting it today ^[5]. If the open release covers weights only, then two of the three layers he names stay outside the buyer's control ^[6]. What the labs built instead, in his words, is an architecture of control rather than one of freedom and participation, which gives them the ability to track you ^[7].

The sharpest procurement point is about memory rather than models. O'Reilly characterises Mark Zuckerberg's thesis as lock-in through familiarity: Meta will give you the AI that knows you best ^[8]. His counter-position is that open source should let you switch models and providers while keeping the context the system needs ^[9]. At his nonprofit, the AI Disclosures Project, he says one work item is an open-memory consortium ^[10], and he points to Pi, an open-source agentic harness, as the kind of component being built outside the labs ^[11]. Weights you can download are not portability if the accumulated context stays on someone else's account.

He concedes the obvious objection. Giving up control is totally against the big companies' interests, he says, but that does not make it the right strategic decision ^[12]. His reasoning is that frontier models used to be better at everything and are now better at some things and worse at others ^[13]. He cites talk that Fable and Sol are worse writers than lower-level models; Wired notes that Anthropic and OpenAI would disagree ^[14][15]. He goes further, arguing that frontier breakthroughs are pushing away from what ordinary people need, and that the United States could win frontier AI while China wins on diffusion of lower-level models through society ^[16][17]. That is a bet, not a measurement, and it is his alone.

On safety, O'Reilly inverts the usual argument: he says every cybersecurity incident seen so far has come from frontier models, so the risk case argues for slowing frontier development more than for restricting open weights ^[18]. Wired presents no independent verification of that claim.

Three things to watch. Whether any vendor ships a documented, testable separation between model, harness, and application, since that is the only version of "open" that survives a migration ^[5]. Whether the open-memory consortium produces a specification or stays an intention ^[10]. And whether frontier models settle into the role O'Reilly floats for them, mainframes and supercomputers pointed at hard problems rather than the layer that diffuses through society ^[19]. He declines to predict, saying only that the world is proceeding the way he hoped ^[20], and his general yardstick remains create more value than you capture ^[1].