When does AGI arrive? The predictions compared, 2027 to never
Kurzweil holds to 2029, Amodei extrapolates to 2026 or 2027, Altman expects very capable systems by the end of the decade, and Bostrom, Yampolskiy and chip analysts give their own dates and doubts. Seeing the forecasts and their reasoning together helps readers weigh the confidence behind each.
The dated forecasts for human-level AI in these conversations sit in a narrow band, from 2026 to the early 2030s. The spread looks smaller than it is, because the forecasters are not predicting the same thing. And a capability arriving is different from the world being able to afford and absorb it.
Ray Kurzweil has held 2029 since 1999, for a computer passing an hours-long Turing test with an expert judge. When he first said it, a poll of AI experts averaged 100 years, and 20 percent said never. A recent poll said 2030. In his words, "I haven't changed at all. I stayed with 2029". His basis is a chart of 80 years of exponential growth in computing per dollar.
Dario Amodei, the head of Anthropic, gives the earliest date and the hardest definition: a system smarter than a Nobel Prize winner across most fields, running as millions of fast copies. Eyeballing the curve suggests 2026 or 2027, which he says is not a scientific prediction. Data limits or a disruption to chip production could derail it. He expects the effects to take five to ten years more, slowed by physical limits and institutions.
Sam Altman, the head of OpenAI, calls the question poorly formed and still expects quite capable systems by the end of the decade. From inside, progress feels continuous. He admits reasoning is not solved and long tasks rarely work. Nick Bostrom, the philosopher, takes short timelines seriously: superintelligence could come in a year or a few years, though probably a bit later.
Roman Yampolskiy, a safety researcher, cites prediction markets pointing to 2026 and says any such date is too early, since "we don't have a working safety mechanism in place, or even a prototype for one". More compute buys more capability, and more money does not buy more safety.
The skeptics are the analysts closest to the hardware. Nathan Lambert says probably after 2030 and discounts the forecasts of company heads because of their incentives. Dylan Patel expects the capabilities around 2027 or 2028 but says the cost of running them will slow deployment.
The reasoning behind early dates is nearly always a curve extended: computing power, spending and benchmark scores. The doubts concern definitions, incentives, energy, chip supply, a ceiling on spending and unsolved abilities. Nobody in these conversations now says never.
What to do with it: ask for the definition before the date. Keep two dates in mind, one for capability and one for impact. Weigh the forecaster's position, and ask how prediction markets defined the term. Watch what the experts watch: the cost of training runs, benchmarks of real work, whether long tasks become reliable, and signs that AI is speeding up science. Watch the derailers too. Do not bet your skills on one date, since the change may come years late. Treat the safety schedule as part of the forecast. Write down what you expect and check it later.