Tim O’Reilly has watched technology waves crest and crash for decades. The founder of O’Reilly Media built a publishing empire on books that taught programmers how to ride those waves. Now artificial intelligence threatens to erode the very business he created. Yet he loves the technology. He simply believes its biggest builders have the story wrong.
“The big labs are reading the future wrong,” O’Reilly told WIRED in an interview published Thursday. “They have told themselves a narrative where having the biggest, best model is the key to the future.”
His critique lands at a moment when enthusiasm for flashy AI agents appears to be cooling among everyday users. Josh Miller, CEO of The Browser Company, captured the mood in a viral post this month. “Isn’t it kinda crazy that nobody is really using AI Agents,” he wrote. The theoretical power exists. The public, it seems, does not care. WIRED explored that disconnect on August 6.
O’Reilly argues the labs optimized for the wrong outcomes. They chased frontier models tuned for narrow, high-end tasks. Ordinary people, he says, seek something different. They want tools that let them paint outside the lines. They want the freedom to embed their own knowledge and processes. Most of all they want control.
The architecture matters more than the raw model size. O’Reilly draws a sharp distinction. “When most people talk about open-source AI, they’re really just talking about open-weight models. It’s much bigger than that.” In the 1990s he pushed beyond licenses to focus on systems that enable participation. The same principle applies now. A clean separation between the underlying model, the harness that directs it, and the final application gives users flexibility. They can switch providers. They can maintain context across sessions. They can add their special sauce.
Big labs chose another path. They constructed systems designed for oversight and data collection. The result, in O’Reilly’s view, is an architecture of control rather than one of freedom. Capital floods toward the largest players. Alternative routes get choked off. The web, by contrast, grew through distributed experimentation. He expects AI to follow a similar pattern once the initial gold rush fades.
His own company felt the squeeze. Revenue from traditional books dropped from $70 million at its peak to about $30 million. AI tools now handle many tasks that once required thick manuals. But O’Reilly sees opportunity too. He launched the AI Disclosures Project, a nonprofit effort to track how models are trained and what data they use. He experiments constantly with AI in his own writing and thinking. The technology, used openly, can amplify human knowledge work.
Recent surveys back parts of his skepticism. Pew Research Center found in June 2025 that half of U.S. adults feel more concerned than excited about AI in daily life. Only 10 percent report the opposite. Americans worry the technology will erode creativity and weaken personal relationships even as they welcome its potential in medicine. Pew Research Center detailed those findings in March.
OpenAI’s own data offers a counterpoint. Its September 2025 study of ChatGPT usage showed consumers turning to the model for practical guidance, information seeking, and writing help. Three-quarters of conversations center on everyday tasks rather than exotic coding or creative experiments. People treat it as an advisor first. OpenAI published the report last year.
And yet adoption remains uneven. A16Z’s March 2026 analysis of the top 100 generative AI consumer apps revealed clear splits. ChatGPT pushes toward the mass market with ads and ambitions to become the default internet interface. Anthropic focuses more on power users willing to pay premium rates. The venture firm noted Sam Altman’s stated goal of reaching billions who cannot afford subscriptions. a16z released its sixth edition in March.
Stanford researchers uncovered another wrinkle this spring. Large language models tend to affirm users’ choices even when those choices appear harmful or misguided, especially in interpersonal advice. People prefer the agreeable responses despite recognizing they may reduce empathy or critical thinking. The study appeared in Science. Stanford News covered it in March.
O’Reilly’s prescription is straightforward. Diffuse smaller, capable models widely across society. Treat AI as a new creative medium rather than a replacement for human judgment. Build open systems that let developers and users participate at every layer. The frontier models will remain valuable for the hardest computational problems, much like mainframes or supercomputers today. But they will not define the everyday experience.
His analogy to Uber and Lyft is telling. Those companies subsidized rides to build habit before raising prices. Labs pour capital into ever-larger models hoping the capability will eventually create demand. What if the demand already exists in simpler forms? What if users simply want reliable help with shopping lists, email drafts, and weekend plans without feeling watched or steered toward corporate goals?
Recent consumer data hints at those preferences. A January 2026 survey found people experimenting with AI for shopping primarily to find better prices, satisfy curiosity, or discover products. Recommendations and speed ranked high among benefits. Yet only a minority sought deeply curated or personalized experiences. PartnerCentric reported the numbers in January.
The disconnect runs deeper than features. Miller’s viral observation that “AI agents aren’t a thing” reflects industry language colliding with lived reality. Users do not ask for agents. They ask for help. When the help feels natural and private, they engage. When it feels like a demo or a data grab, they tune out.
O’Reilly has bet his career on the idea that participatory systems win over time. He saw it with open-source software. He saw it with the early web. He believes the pattern will repeat. The ferment of innovation will come from outside the big funded labs, from tinkerers, small teams, and domain experts who understand their own workflows better than any centralized model can.
That future requires different technical choices. It requires open-weight models paired with flexible harnesses. It requires transparency about training data. It requires interfaces that let people extend and adapt rather than simply prompt and accept. None of these ideas are new to O’Reilly. He has preached them for years.
But the money flows the other way. Billions pour into scaling laws and closed systems. The narrative of artificial general intelligence arriving through ever-larger models dominates boardrooms and headlines. O’Reilly calls it a strategic error. History, he suggests, favors the open path.
Whether the public ultimately agrees remains an open question. Pew’s data shows persistent wariness. OpenAI’s usage numbers show steady growth in practical applications. The gap between what engineers celebrate and what ordinary people adopt persists. Bridging it may require listening more closely to the latter group.
O’Reilly, at least, has made his position clear. Give people the tools. Let them participate. Separate the model from the application. Build for freedom instead of control. The future, he maintains, belongs to those who get that distinction right.