Publisher, internet pioneer and venture investor Tim O'Reilly argues in an interview with Wired that the big AI labs do not understand what people actually want. His analogy is direct: these companies are trying to lock users into their own products the way Microsoft did in the 1990s.

What "open source" means

O'Reilly's most concrete objection is to the term itself. In his view open-source AI means something far broader than most people assume, and publishing neural-net weights does not on its own satisfy it.

What is required is a clean separation between the layers of the system: the model, the harness that runs it, and the application the user sees. When those three can be pulled apart, a user can change each layer independently — keep the model and swap the application, or keep the application and swap the model.

The big labs, O'Reilly argues, do exactly the opposite: they weld the three layers together and present them as a single whole. Even when weights are published, a closed harness around them leaves the user without a real alternative.

The claim about security

The most contested part of the interview concerns security. Against the common argument that open models are dangerous, O'Reilly notes that the largest cybersecurity incidents have come from frontier models.

That claim has recent support: the year's most discussed security cases involved not small open-weight models but the newest systems from the largest labs. Risk rises with capability, and the most capable models are the closed ones.

Where innovation will come from

O'Reilly's forecast is that real innovation will come from an open-source movement not funded by venture capital. His reasoning is historical: the web, Linux and most of modern software infrastructure emerged that way — not from a return-on-investment calculation but from people wanting to solve a problem.

He adds a complementary framing: seeing AI not as a product but as a new creative medium, like writing, paint or music. That frame puts at the centre not what the technology will do but what people will do with it.

Why it is worth hearing

The weight of O'Reilly's words comes from having lived his own analogy. He is among the people who named and popularised the open-source software movement; he saw the lock Microsoft built on the desktop, and what broke it, at first hand.

Whether his forecast holds is a separate question. Nobody in that era's open-source movement needed billions of dollars of compute; the capital required to train a frontier model today is not on a scale a volunteer community can reach. That is where the argument is weakest.

The distinction between training and use softens that objection somewhat, though. Training a model from scratch takes billions of dollars, but adapting an existing open model to a specific job is far cheaper. The tooling ecosystem that has grown around open-weight models recently develops in exactly that space — in use rather than in training.

O'Reilly's three objections

  • Architecture — model, harness and application are welded together; publishing weights alone does not suffice.
  • Security — the largest incidents come from frontier models, not open ones.
  • Innovation — the real breakthroughs will come from communities solving problems, not calculating returns.