Vol. XII · No. 05 · May 2026
Jake Cuth.

Eighty-two named predictions.
Most of them aged badly.

In 1996, Donald Knuth said he expected to be using a fountain pen in fifty years. He was right about the pen. This page collects eighty-two confident, dated, named technology predictions made between 1940 and 2026. Each one named a year. The year came. Most of the predictions did not. A handful did, but arrived years later than promised.

The pattern across all of them is the same: smart, well-resourced people consistently underestimate the time it takes to ship a hard technology to scale. The point isn't to humiliate Elon Musk or Mary Meeker. The point is that everyone in tech does this. Including the smartest people. The pattern is structural, not individual.

Sources are annotated per prediction. v1 dataset; URL verification pending. Predictions marked low confidence are flagged in the archive.


Each dot is one prediction. The X-axis is the year it was said; the Y-axis is the year it pointed at. The diagonal is "predictions about the present" — said in 2018, pointing at 2018. Predictions above the diagonal point at the future; the further above, the more ambitious. Color is category. Open circles did not happen; filled circles did, with delay. Hover any dot for the quote; click for the full context.

Category View

How to read it

A tight cluster far above the diagonal is a pattern of long-shot predictions — autonomous vehicles bunched around "by 2020" said in 2014–2018. A cluster just above is short-horizon overconfidence — "shipping next year" said every year. The 3D TV cluster of 2010–2011 is a brutal three-year band. The diagonal itself catches up over time: predictions resolve as the calendar moves.


Six small multiples. Each panel uses the same axes as the lede chart but is restricted to its category. Category-specific patterns emerge: 3D TVs are a brief band of confident wrongness; autonomous vehicles are an annual repeat through one decade; AR is long-running and unresolved; flying cars are a comedy that crosses generations.


Every prediction in the dataset, sortable. The archive is the primary data product — the chart above is the visual summary. Click any row for the full quote and outcome note.

Predictor Said Targeted Category Prediction Outcome Source

Four candidate explanations for the pattern. None of them excuse it; together they suggest why no individual predictor escapes it.

1 · The demo-to-deploy gap

Working in a controlled demo is roughly five percent of the problem. The remaining ninety-five percent is corner cases, regulatory approval, manufacturing scale, and a long tail of edge cases that erode reliability targets. The first 90% takes 10% of the time. The last 10% takes 90%.

2 · Selection bias in the predictor pool

The predictors who make headlines are the ones with confident timelines. Pessimistic experts — "we don't know when this will work" — don't get speaking invitations, don't get funded, don't get retweeted. The corpus of public predictions is filtered for confidence at the source.

3 · Career incentives

Optimistic predictions raise capital and attract talent. Pessimistic predictions don't. Selection persists at the lifecycle level: the optimist builds the company, gets on stage, and makes more predictions. Survivorship bias inside the sampling.

4 · Hofstadter's Law

"It always takes longer than you expect, even when you take into account Hofstadter's Law." — Douglas Hofstadter, Gödel, Escher, Bach, 1979. Recursive correction does not help. The error has no fixed prior to anchor against.

A fifth, optional

Proxy targets versus deployment targets. The prediction is usually about a demo benchmark or technical milestone (Level-5 capable / supremacy demonstration / billion users) but what the public hears is "deployed at scale." The predictor and the audience score different things and only notice the divergence when the deadline arrives.



Sourcing

Each prediction includes a predictor, a year said, a year targeted, a quote (verbatim or paraphrased — flagged), a source title, and a verification level. v1 of the dataset does not yet have primary URLs for most entries; URL verification is the v2 pass. Predictions marked verified_level: low are flagged in the archive and should be read with caution.

Outcome scoring

Each entry is scored as one of five outcomes: did-not-happen, did-not-happen-yet, fulfilled-with-delay, fulfilled-on-time, or partially-fulfilled. Outcome notes specify the criteria — "self-driving by 2020" can mean Level 2, Level 4, or robo-taxis at scale, and the reading matters.

What this lab is not

Not a takedown of individuals. The lab does not score predictors against each other. The argument is structural, not personal. Read it as "this is what happens when smart people predict hard timelines", not "these specific people are stupid."

Companion piece

This lab pairs explicitly with AGI Horizon. That lab makes the same point about the most ambitious technology prediction of all — human-level AI. If you smiled reading this one, that one will look slightly less funny.

Caveats
  • Selection bias: this archive collects predictions that came due unfulfilled. Successful technology predictions exist — they just don't make a compelling archive. The "Right but late" filter partially addresses this.
  • Anglosphere bias: the dataset skews toward English-language predictors. Tencent, Alibaba, Baidu, Samsung, Sony predictions are underrepresented.
  • Outcome interpretation: ~20% of entries require a judgment call on what "fulfilled" means. Outcome notes specify the reading; reasonable people will disagree on some.
  • Quote fidelity: entries marked verbatim: false are paraphrases, faithful to the public statement but not necessarily to its exact wording. Pre-2010 quotes are particularly likely to be paraphrased until a v2 verification pass is done.