Why nobody can tell you how many years you have

The most quoted prediction, and the sentence that follows it on the tape

In 2016 Geoffrey Hinton said that people should stop training radiologists now, and that within five years deep learning would be better. Almost every retelling stops there. On the conference organiser's own recording, the next thing he says is that it might be ten years, and that there are plenty of radiologists already.

We record this as a forecast: a named person with standing, on a date, in an attributable statement. The stage exists precisely so that the claim stays checkable and sits on the same page as the measurements about that occupation. We publish no verdict on whether it came true — judging that is itself a judgement, and this site does not make those. The reader does the arithmetic.

What the record shows about that occupation since

The technology arrived. The FDA's list of AI-enabled medical devices reaches 1,614 authorisations, of which 1,230 are radiology — by far the largest category, and a regulator's own count rather than a vendor's. Korea's AI Framework Act places AI medical devices inside high-impact AI, with obligations attached.

And the number of people going into the occupation went up. US diagnostic radiology training posts rose from 997 to 1,083 across five Matches, filling every year. Both of those are true at once, and neither is a rebuttal of the other: a thousand authorised devices and a growing training pipeline are exactly what you would see if the technology entered the work rather than replacing the worker.

What that history shows is not that the prediction was wrong about the technology. It is that a prediction about technology was read as a prediction about employment, and those are two different claims with two different kinds of evidence behind them. This site keeps them in separate columns for that reason.

Six times the arrow pointed backwards

A year-estimate assumes adoption only moves one way. It does not. Klarna's chief executive said the company would hire humans for customer service again, saying cost-cutting had gone too far. Zillow wound down its algorithmic home-buying business and cut about a quarter of its staff. Figma disabled its prompt-to-UI feature a week after launch. Booths removed self-checkouts from 26 of its 28 stores. Miso Robotics' own SEC filing shows its installed base of kitchen robots falling from 17 to 14. A New York school district paused an approved humanoid teaching robot.

None of these says the technology failed in general. Each says something narrower and more useful: the deployment conditions did not hold in that setting. That is why this site weights a pilot differently from a deployment, and why a single rollback is recorded rather than filed under noise.

The forecasts being made now, recorded the same way

OpenAI's chief scientist wrote in 2026 that he expects the current pace to be sustained. We record that under the same rule as Hinton's: named, dated, attributable, and unable to move any task's assessment, because a prediction is not an observation no matter who makes it or how many people make it.

Recording a forecast is not endorsing it and not dismissing it. It is making sure that in five years the sentence is still findable in the form it was actually said, next to what happened.

What to watch instead of a year

Every occupation page on this site answers this question with a signal you can check yourself instead of a date — because the signal is available to you and the date is available to nobody. The general shape: watch the thing that has to change before your work can change, and watch it where you are, not nationally.

Concretely, the three that recur across occupations: whether the checking step in your work is visible to the person who sets budgets; whether the part of your job that has an automatic referee is growing as a share of your week; and whether the entry level in your occupation is still hiring, since the payroll evidence shows movement there first and by a wide margin.

Everything cited here

Each one opens the full record: its source and source tier, the dates, the scope it applies to, who verified it and what it does not establish.

1. Geoffrey Hinton said in 2016 that people should stop training radiologists now, that within five years deep learning would do better, and added that it might be ten — Creative Destruction Lab — Geoff Hinton: On Radiology, recorded at the 2016 Machine Learning and Market for Intelligence Conference, Toronto
2. FDA's AI-enabled device list reaches 1,614 authorisations, 1,230 of them radiology — U.S. Food and Drug Administration — AI-Enabled Medical Device List
3. US diagnostic radiology training posts rose from 997 to 1,083 across five Matches, filling at 98% or above throughout — National Resident Matching Program — Results and Data: 2026 Main Residency Match
4. Korea's AI Framework Act places the use of AI medical devices inside high-impact AI, obliging the user to keep it under human management and supervision and to be able to explain its output — 국가법령정보센터 (Korea Ministry of Government Legislation) — 인공지능 발전과 신뢰 기반 조성 등에 관한 기본법, 법률 제21311호
5. Klarna's CEO said the company would again hire humans for customer service, saying cost had been 'too predominant' in its AI-first approach and quality had suffered — CX Dive (reporting a Bloomberg interview)
6. Zillow wound down Zillow Offers, its algorithmic home-buying business, cutting about 25% of staff; its CEO wrote that unpredictability in forecasting home prices far exceeded expectations — Zillow Group (Exhibit 99.1 to Form 8-K, SEC EDGAR)
7. Figma disabled its Make Designs prompt-to-UI feature a week after launch when 'weather app' prompts produced screens closely resembling Apple's; the CEO cited insufficient QA; relaunched Sept 2024 — Figma — blog retrospective
8. UK supermarket chain Booths removed self-checkouts from 26 of its 28 stores and returned to staffed tills, citing machines that were slow, unreliable and impersonal and trouble with loose produce — Fortune (quoting BBC)
9. Miso Robotics' FY2025 SEC annual report discloses $514,798 of Flippy net revenue, lists all three CaliBurger sites as uninstalled, and states substantial doubt about it continuing as a going concern — Miso Robotics, Inc. — Form 1-K annual report for FY2025 (SEC EDGAR)
10. A New York school district paused an approved 60,000 dollar humanoid teaching robot after state officials, the teachers union and parents objected — AP News
11. OpenAI's chief scientist wrote that he expects the current pace to be sustained into recursive self-improvement, and that systems will increasingly drive their own development — OpenAI — An Alien Mind, by Jakub Pachocki, Chief Scientist
12. Employment of 22-25-year-olds in the two most AI-exposed occupation quintiles fell about 11% from November 2022 to June 2026, while the least-exposed quintiles grew about 10% — Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (August 2026)
13. Reading the routine study — task on Radiologist