Tech Visionary Says the Big AI Labs Don’t Get What People Want

Details

In a wide-ranging Wired interview, publisher and tech strategist Tim O'Reilly argues that today's hyperscalers are misreading the AI market — much as Microsoft did in the 1990s by locking users into products rather than enabling an open ecosystem. His core critique is architectural: 'open-source AI' must mean more than releasing model weights. It requires clean separation between model, harness, and application layers, giving developers genuine control without vendor tracking. O'Reilly's sharpest observation: the US could 'win' frontier AI while China wins the broader market by diffusing capable lower-level models widely through society — giving Chinese developers an innovation advantage at the application layer. He cites Anthropic's Fable and Sol as examples where frontier model optimization may actually degrade performance for ordinary writing tasks compared to smaller, more accessible models. (Anthropic and OpenAI publicly dispute this characterization.) He draws a direct parallel to the internet: the web's architecture of participation enabled mass innovation; current proprietary AI stacks centralize control and create lock-in. His prescription is not just open weights but a fully open stack that enables users to embed their own logic. The interview also surfaces a genuine disagreement with interviewer Steven Levy about AI's role in producing original content.