Technology publisher and investor Tim O’Reilly believes the AI industry is using an incomplete definition of open source. Making model weights downloadable is useful, he argues, but it does not give users meaningful freedom if one company still controls the surrounding application, memory, data flows and service relationship.
In an interview with WIRED, O’Reilly described openness as a property of the whole system architecture. His preferred design separates the model, the software harness and the application, allowing people to change providers while preserving their own context and customisation.
That vision challenges the frontier-lab assumption that the largest and most capable general model will define the market. O’Reilly’s case is an argument about strategy and participation, not a benchmark showing that smaller or open systems outperform frontier products in every task.
Open weights are only one layer
O’Reilly’s experience with the open-source software movement shapes his distinction. In the 1990s, he came to see licences as only part of the story; the more important question was whether an architecture enabled others to participate, modify components and build outside a vendor’s preferred path.
He applies the same test to AI. A downloadable model may still sit inside a stack that makes it difficult to move personal memory, workflows or application logic elsewhere. The result can look open at the model layer while leaving the user dependent on a platform for the most valuable parts of the experience.
O’Reilly supports an “open-memory consortium” through his nonprofit AI Disclosures Project. The idea is to make accumulated context portable so a user could switch models or providers without abandoning the information that makes an assistant useful.
The proposal is still an initiative rather than an established technical standard. Interoperable memory would have to address data formats, permissions, privacy, security and differences in how models interpret stored context. Portability can reduce lock-in, but poorly designed exchange mechanisms could also move sensitive histories into new risk zones.
Frontier capability may not match everyday demand
O’Reilly argues that the largest labs are optimising for a narrow idea of progress: building the most capable model and assuming applications will follow. He believes ordinary users may instead value systems they can adapt, understand and combine with specialised knowledge.
In his analogy, frontier models could become like mainframes or supercomputers—essential for the hardest problems but not the layer through which most innovation spreads. Smaller models and open agent harnesses could support a wider variety of local or specialised applications.
He also links concentration to venture capital. Large pools of funding can select a few winners and crowd out alternative paths before the market has tested them. O’Reilly expects more important experimentation to emerge from open communities, much as the web disrupted a computing debate centred on control of the personal computer.
Those comparisons are provocative, not forecasts with guaranteed outcomes. Frontier models continue to improve and can support applications that smaller systems cannot. Open development also carries costs: maintenance, evaluation and security work still require resources, even when the code or weights are available.
Publishing exposes both the promise and conflict
O’Reilly Media was built around books and expert knowledge, and its book revenue has fallen from a reported peak near $70 million to about $30 million. O’Reilly acknowledges that AI systems absorb material from publishers while weakening parts of the business that produced it. He nevertheless uses AI extensively as a brainstorming and functional-writing partner.
His position differs from WIRED interviewer Steven Levy’s stated policy of writing and editing without AI. O’Reilly sees generative systems as a medium comparable to words, paint, music or photography: the operator’s choices can shape distinct results rather than merely outsourcing authorship.
That analogy does not resolve compensation or originality. A creative medium built partly from other people’s work raises questions that cameras and paint do not answer directly. O’Reilly’s own company is searching for tools that let experts be invoked and rewarded as traditional publishing declines.
His wider message is less about declaring one model “open” than distributing control across the stack. Whether that becomes practical depends on standards, sustainable economics and usable tools. But it offers a clearer test than a licence label: can a user leave the provider and take the intelligence of the relationship with them?
