Tim O’Reilly has a simple test for technology: it should create more value than it captures. In a conversation with Steven Levy, he applies that idea to artificial intelligence and argues that the future should not be controlled only by the biggest AI labs and their largest models.
His central claim is not just that AI should be more open. It is that the useful version of open-source AI must give designers and users room to participate, customize, switch tools, and keep control of the systems they depend on.
Open-source AI means more than open weights
O’Reilly says the common discussion around open-source AI is too narrow. In his view, many people use the phrase when they are really talking about open-weight models, where model weights are made available but the rest of the system may still be controlled elsewhere.
He argues that the important question is broader: does the system’s architecture let people participate? That idea, he says, goes back to debates in the ’90s, when many focused on licenses while he focused on whether technology opened a path for others to build.
For AI, O’Reilly wants a clean separation between three pieces:
- The model, which produces the AI output.
- The harness, which shapes how the model is used.
- The application, where people actually work with the system.
He believes today’s dominant AI architecture tends to concentrate control instead of distributing it. That matters because, in his view, the structure of the system determines who can add value, who can customize the experience, and who can track the user.
The big-model race may miss what people need
O’Reilly’s critique of the big AI labs is strategic as much as philosophical. He says they have built their plans around the belief that the biggest and best model will define the future. But he argues that this assumption is starting to break down.
For a time, he says, the newest and strongest models were broadly better across tasks. Now, in his view, they can be better for some work and worse for other work. He points to discussion around Fable and Sol being worse writers than lower-level models, while noting that Anthropic and OpenAI would disagree.
The larger issue is fit. O’Reilly believes ordinary people may not need frontier AI for many practical uses. They may need AI systems that are flexible, widely available, and easy to adapt to local needs. In that framing, a lower-level model spread broadly through society could matter more than a single powerful model locked inside a controlled product.
He also raises a geopolitical concern. The US could win the frontier AI race, he says, while China gains advantage from lower-level models diffused widely through society. The point is not only who has the most advanced model. It is who enables the most people to experiment, build, and use AI freely.
Security fears cut both ways
A frequent objection to open-source AI is risk. If models or tools are more available, critics worry that bad actors could get around guardrails, especially in areas such as cybersecurity or the ability to develop pathogens.
O’Reilly turns that argument around. He says the cybersecurity incidents seen so far have come from frontier models. From his perspective, risks like cybersecurity and pathogen development are stronger reasons to slow frontier models than reasons to restrict open-weight models.
That does not remove the risk question. It reframes it. O’Reilly is arguing that concentrating capability in the largest systems is not automatically safer, and that policy or industry caution should not focus only on open models while assuming closed frontier systems are the safer path.
Memory, switching, and the fight against lock-in
One of O’Reilly’s most practical concerns is lock-in. He compares today’s hyperscalers to Microsoft in the 1990s and worries that users may be pushed into AI products they cannot easily leave.
He points to Mark Zuckerberg’s thesis that Meta can keep users because it will provide the AI that knows them best. O’Reilly’s answer is an open-source vision where users can switch models and providers while maintaining the context the AI needs.
That is where the AI Disclosures Project, his nonprofit, enters the discussion. O’Reilly says one effort there is the idea of an open-memory consortium. The goal is to prevent personal AI context from becoming the thing that traps a person inside one company’s system.
He also mentions Pi, an open-source agentic harness, as an example of work happening beyond the largest frontier models. In his view, frontier systems may end up more like mainframes or supercomputers: important for difficult problems, but not necessarily the technology that spreads most widely through daily life.
AI as a creative medium
O’Reilly also sees AI changing creative work. He uses AI extensively and has a blog about his chats with it. In the interview, he and Levy disagree about AI’s role in producing original content.
Levy says he would edit the interview himself and that the introduction would be all his own, adding that this is policy. O’Reilly replies that this may be a legacy position that will fade. He describes AI as a powerful thought partner when writing.
He does not frame that as having AI do all the writing. He says he uses it for brainstorming and functional writing, including turning an hour-long interview into something usable when he does not have time to write up all those interviews.
His broader claim is that AI is a medium, comparable to written words, paint, music, or a camera. Different people will get different results from it, and some will become skilled at expressing ideas through LLMs. In his view, the idea that serious work cannot be made with AI may eventually seem outdated.
That argument connects back to his open-source AI position. If AI is a medium, then the structure around it matters. A medium controlled by a few companies offers one kind of future. A medium with open models, open harnesses, portable memory, and broad participation offers another.