Tim O'Reilly says the frontier AI labs are building the wrong product. In an interview with Steven Levy, the publisher and longtime open-source advocate argues that OpenAI, Anthropic, and Meta have optimized around an architecture of control that will lose to a diffuse open-source ecosystem, and he points to his own collapsing book business — down from a $70M peak to $30M over 25 years — as evidence that betting against openness is a career-ending trade.
O'Reilly's core claim is that the industry conflates open-source AI with open-weight models, and that the distinction matters. Open weights alone, he says, do not deliver what mattered about open source in the 1990s: a clean separation between the model, the harness, and the application, so that any developer can slot in their own logic. The current stack from the major labs bundles all three, which O'Reilly frames as a deliberate lock-in strategy rather than a technical necessity.
“The big labs are reading the future wrong. They have told themselves a narrative where having the biggest, best model is the key to the future.”— Tim O'Reilly, Founder of O'Reilly Media
He is blunt about the strategic bet the frontier labs are making. Bigger is not automatically better for the use cases users actually reach for, he argues, citing user complaints that OpenAI's Fable and Sol produce worse writing than earlier, smaller models. Anthropic and OpenAI would dispute that read, but O'Reilly's point is directional: capability gains at the top of the curve are diverging from utility gains at the middle.
Key facts
- 01O'Reilly's book business has fallen from a $70M peak to $30M today, a 57% decline over 25 years.
- 02O'Reilly argues frontier models like Claude and OpenAI's Fable and Sol optimize for use cases that ordinary users don't want.
- 03His AI Disclosures Project is building an open-memory consortium to let users switch models and providers without losing context.
- 04O'Reilly predicts China will out-diffuse the US in practical AI by spreading lower-level models widely through society.
- 05He argues every major cybersecurity incident to date has come from frontier models, not open-weight ones.
The lock-in critique extends to memory and personalization. O'Reilly reads Mark Zuckerberg's Meta strategy as an explicit bet that the AI that knows a user best will keep that user captive, and he wants the open-source community to reject that model outright. His nonprofit, the AI Disclosures Project, is working on what he calls an open-memory consortium — a shared context layer that would let users port their history across models and providers.
“They built an architecture of control rather than an architecture of freedom and participation, so they have the ability to track you.”— Tim O'Reilly, Founder of O'Reilly Media
On security, O'Reilly inverts the standard argument against open weights. Every cybersecurity incident to date, he says, has traced back to frontier models rather than smaller open-source ones. He treats that as an argument for slowing frontier development, not for restricting the open-weight releases that regulators most often target.
The geopolitical version of the argument is sharper. O'Reilly predicts the US could win the frontier race and still lose the diffusion race to China, where lower-level models are being embedded broadly across the economy. His framing echoes the debate over whether raw capability or deployment breadth determines strategic AI advantage — a question that has become louder as US export controls tighten and Chinese labs continue shipping capable open-weight releases.
O'Reilly also connects the AI capital structure to a broader complaint about Silicon Valley. He argues the sector became anti-capitalist around 2010, when VCs used the Uber and Lyft playbook to spend billions subsidizing rides and pick winners the market would not have chosen. He sees the same dynamic in AI, with capital funneling into a handful of labs that, in his view, do not actually have a moat.
The personal stakes are visible. O'Reilly's publishing house exists to compensate experts for sharing knowledge, and generative AI hoovers that knowledge up without payment. He is not arguing for prohibition — he uses AI himself for brainstorming and turning hour-long interviews into usable drafts — but he wants tooling that pays the people whose expertise trained the systems. He runs a personal blog documenting his AI chats and treats the model as a medium comparable to paint or the camera.
The counterweight is that O'Reilly is describing a preferred future rather than a measured one. Anthropic and OpenAI have concrete revenue, enterprise contracts, and a widening deployment lead; the open-source projects he points to, like the agentic harness Pi, are still small. His own analogy to the mid-1990s web boom cuts both ways — the web did emerge outside VC funding, but the eventual winners were still concentrated platforms, not the distributed ferment he envisions.
For the AI market, O'Reilly's argument is a useful stress test on the frontier-first thesis that has driven roughly all of 2025 and 2026's capital allocation. If mid-tier and open-weight models keep closing the gap on the tasks users actually run — writing, summarization, agentic workflows — the pricing power of the top labs erodes even without a single dramatic disruption. That is the scenario the recent OpenAI and Anthropic price cuts already hint at, and it is the one open-source proponents will keep pressing until either the diffusion argument or the frontier argument wins on the numbers.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
