Ten years on, Kevin Kelly's 2016 AI predictions get a split verdict: AI-as-utility arrived, 'Google's main product' didn't

Kevin Kelly's AI predictions from his 2016 book The Inevitable score a split verdict ten years on, and the split has a shape: every call that a price or a count can settle landed, and the one call that required Google to relabel its main product did not 1. Developer Taylor Troesh lined the book's AI passages up against 2026 reality 1, and Hacker News readers grading the same excerpts landed between three hits out of five 2 and eight out of sixteen on the most generous split 2.

The pattern inside the grades is ours, not the scorecard's: every hit is a forecast a price sheet or a count can settle, and the miss required a company to change its story about itself. Kelly wrote that AI would arrive like electricity from a grid, utility-grade intelligence running invisibly behind products 1, and one commenter called that outcome settled by "tokens and APIs" 2. He wrote that the industry would consolidate into an oligarchy of two or three large, general-purpose, cloud-based commercial intelligences 1, a network-effect argument that comes due as a countable number. Forecasts about the cost of intelligence survived the decade because they came due as invoices.

Graded call by call, ours is the scorecard that survives the audit:

• AI as a metered utility you plug into: earned hit. Intelligence is bought per request through cloud APIs 2.
• Any question answered in conversational tones: hit, and no commenter in the thread disputed it 2.
• An oligarchy of two or three commercial intelligences: hit by the scorecard's read, with the wrinkle Troesh flags himself, that Kelly also wanted AI free and open 1.
• Google's main product becoming AI rather than search by 2026: earned miss 2.
• The bulk of AI as narrow, special-purpose brains: contested, adjudicated below.

The miss that needed Google to change its story

The boldest call carries its deadline inside it: "By 2026, Google's main product will not be search but AI" 1. Kelly built it on a party conversation from around 2002, in which Google cofounder Larry Page told him the company was "really making an AI" 1, and on a fall 2015 earnings call where Google CEO Sundar Pichai called AI a "core transformative way" to rethink the company's products 1. The prediction contradicted the number printed in its own paragraph: Kelly noted that search was 80 percent of Google's revenue while he was writing 1. A decade later the revenue mix had not flipped: in the quarter ended December 2025, advertising was $82.28 billion of $113.83 billion in revenue, about 72 percent 3. "Google's main product was and continues to be ads," one commenter wrote 2, and another graded the call a flat "Nah" while scoring the set 3 out of 5 2. AI delivered on schedule; an advertising business did not stop being one, because identity is not a cost curve.

The contested call fails on its own first clause

Between the clean hits and the clean miss sits Kelly's bet that "the bulk of AI will be special purpose software brains" 1. The scorecard's verdict is that Kelly failed to foresee the bitter lesson, the argument from 70 years of AI research that general methods leveraging computation beat built-in human knowledge 14. The thread pushed back on that verdict: language models do not drive cars, and "Stockfish still beats the newest, largest LLM at chess" 2. Our adjudication: Kelly's own sentence contains both the failure and the rescue. He wrote that 99 percent of the AI you would interact with, directly or indirectly, would be narrow specialists 1. The directly clause failed; the system people now put their questions to is a general model 2. The indirectly clause holds, on the thread's own evidence about specialists still running production systems 2. A call that survives only on its second clause is a miss with an asterisk, and scoring it a hit without saying which clause you are reading is charity.

Survivorship bias aims at the excerpt list, not the book

Two commenters attacked the sample itself: one asked what the error rate looks like if the excerpts were cherry-picked, and another demanded to see the predictions that failed, because "we don't want survivorship bias" 2. The accusation hits the wrong target. The Inevitable is a fixed, timestamped text; nobody selected Kelly's winners after the fact. What can be cherry-picked is the excerpt list, and the cure already exists in the thread: enumerate the excerpt list's calls into a full public set, which one reader counted at sixteen, and grade all of them 2. That full-set method scores 8 of 16 as hits on a generous read 2. Half right from a generous grader over a public denominator is a strong decade, and it is the only kind of score worth quoting.

Grade a prophet by what the claim would cost

Kelly's most durable line is his most priceable one: he put the manufacturing cost of a single answer at "nanocents" 1, and answers are now a metered API product 2. The test falls out of the scorecard. Before grading any AI prediction, ask what printed price would settle it. If a number on a public sheet decides the question, the forecast is gradeable and probably worth making. If the forecast needs a company to adopt a new identity, as the Google call did, no invoice will ever settle it. A decade of Kelly says the quiet part of futurology out loud: predictions age exactly as well as they can be priced.