In 2016, Kevin Kelly published The Inevitable, a book-length argument that a handful of macro technological trends — cognification chief among them — would reshape daily life in ways we'd barely notice until they were load-bearing infrastructure. A decade later, a retrospective of the book's AI predictions reads less like futurism and more like a first draft of the present. The headline prediction — "By 2026, Google's main product will not be search but AI" — landed almost exactly. Google rebranded its entire product strategy around Gemini, restructured its org chart around DeepMind, and now treats search as a distribution channel for AI-generated answers rather than a standalone product. Kelly's reasoning was structural, not vibes: he traced the feedback loop from search queries to training data to better models back to more queries. That loop is now the core business logic of every major AI company. The oligopoly call was equally sharp. Kelly predicted "an oligarchy of two or three large, general-purpose cloud-based commercial intelligences" driven by network effects. As of 2026, OpenAI, Google DeepMind, and Anthropic dominate general-purpose AI, with Meta's Llama as the open-weight wildcard Kelly simultaneously advocated for. The mechanism he identified — more users makes the model smarter, which attracts more users — is exactly how frontier model economics work today. The "Amazon Web Services" framing was the most conceptually precise prediction in the book. Kelly argued AI would become "cheap, reliable, industrial-grade digital smartness running behind everything, and almost invisible except when it blinks off." This is a near-perfect description of how AI APIs are consumed in 2026: developers plug into foundation model endpoints the way they plug into cloud compute. The utility metaphor — AI as electricity — has become the default pitch deck analogy for the entire industry. Where Kelly missed, he missed for fascinating reasons. His prediction that "99 percent of the artificial intelligence that you will interact with" would be "nerdly narrow, supersmart specialists" inverted what actually happened. The transformer architecture — published by Google researchers one year after The Inevitable — enabled general-purpose models that do translation, driving assistance, video recall, and conversation within a single system. Kelly didn't foresee what Rich Sutton later codified as "the bitter lesson": that general methods leveraging computation beat specialized approaches almost every time. Kelly's mental model was a toolbox of bespoke instruments; the reality became a single Swiss Army knife that keeps adding blades. The conversational interface prediction — asking the cloud "in conversational tones, any question at all" — is ChatGPT, Claude, and Gemini described with perfect fidelity six years before any of them existed. Kelly even anticipated the clarification dialogue, the willingness to surface "deep, obscure, complex factual knowledge," and the trajectory toward hundreds of questions per day at nanocent marginal cost. He got the UX right before the architecture existed to deliver it. What makes the retrospective genuinely useful isn't the hit rate — it's the method. Kelly's predictions held up where he tracked structural incentives and feedback loops. They failed where he assumed architectural constraints would persist. The lesson for reading any technology forecast: trust the economics, doubt the engineering assumptions. The forces Kelly identified were real; the implementation path was unknowable. That's a more honest batting average than most futurists achieve, and it tells you something about which kinds of predictions are worth making at all.
