Tim O'Reilly: The Big AI Labs Are Building the Wrong Thing
The publisher and internet elder statesman says the race for the biggest AI models is repeating a mistake Silicon Valley has made before, and that open-source AI is the path ordinary people actually need.
Key points
- Tim O'Reilly, publisher and tech commentator, argues that today's major AI companies are repeating the strategic error Microsoft made in the 1990s by locking users into their products.
- O'Reilly says open-source AI, meaning AI whose underlying components are freely available and modifiable, matters far more than simply releasing model weights.
- His nonprofit, the AI Disclosures Project, is working on an "open-memory consortium" that would let users carry their personal AI context across different providers.
- O'Reilly believes the highest-capability "frontier" AI models are becoming less useful for ordinary people, not more.
- He uses AI himself as a brainstorming and note-processing tool but draws a clear line between that and AI replacing human judgment in writing.
Tim O'Reilly has spent decades watching technology cycles eat themselves. The founder of O'Reilly Media, whose books taught a generation to code, sees the current AI moment as a replay of a familiar story: a handful of companies grab control, users get locked in, and the real innovation happens somewhere nobody is watching.
In an interview with Wired AI, O'Reilly laid out why he thinks the biggest AI labs are making a strategic mistake, and why the open-source path matters far more than which company has the most powerful model.
What does O'Reilly actually mean by "open-source AI"?
He means something bigger than most people assume. Open-weight models, where the underlying mathematical parameters of an AI are published for anyone to inspect and modify, are only part of it. O'Reilly wants the full stack open: the model itself, the "harness" (the software framework that connects the model to real-world tasks), and the application layer on top.
His argument is structural. Without clean separation between those layers, companies can track users, push them toward specific products, and dictate how AI gets used. He calls this "an architecture of control rather than an architecture of freedom."
Are the biggest AI models actually getting worse for ordinary people?
O'Reilly thinks so, at least for everyday tasks. He points to reports that the most advanced AI models, the so-called frontier systems, are producing weaker creative writing than smaller, cheaper ones. He also argues that China's competitive edge may not come from building a single world-beating model but from spreading lower-cost, more accessible AI widely across its economy.
His analogy: mainframe computers were extraordinary machines, but the technology that changed daily life was the personal computer, which nobody thought was serious competition at the time.
| Concept | O'Reilly's view |
|---|---|
| Frontier models (highest-capability AI) | Useful for hard, specialist problems; poor fit for mass adoption |
| Open-weight models (AI with published internals) | Cheaper, more adaptable, closer to what most people need |
| Full open-source stack | Lets users switch providers and keep their own data |
| Meta's personalisation strategy | Locks users in through accumulated personal context |
What does this mean for someone who uses AI tools today?
Practically, it shapes what you can and cannot do with AI products you already pay for. If your AI assistant learns your preferences, your writing style, and your work history inside a closed system, you may not be able to take that context with you if you switch tools. O'Reilly's open-memory consortium idea would change that: your AI context would belong to you, not the platform.
He is also candid about his own use. He feeds hour-long interview recordings to AI and asks it to turn them into usable notes. He calls it brainstorming support, not ghostwriting. The distinction matters to him: the judgment stays human.
The honest takeaway: Before you commit deeply to any single AI platform, ask whether your data, preferences, and conversation history can travel with you if you decide to leave. Right now, for most major tools, the answer is no.
