Kurzweil Scorecard: The Picometer Fell. The Speed of Light Held.

Most of Ray Kurzweil’s predictions live in the comfortable middle distance — chips, genomes, neural nets, the stuff of quarterly earnings calls. This batch is different. Here he reaches past the silicon and bets on the edges of physics itself: the smallest distance a machine can measure, the fastest a signal can travel, and the final destination of intelligence once it has eaten every available atom. These are his vertigo predictions, and they fail and succeed in instructive ways.

The short version: the measurement bet he won outright, and then some. The cosmic bets are quietly turning against him. And the exponential engine underneath all of it is still running — just on different fuel than he said it would.

The predictions

In 2005, near the end of The Singularity Is Near, Kurzweil noted that scientists at UC Santa Barbara had built a detector able to sense beam flexing as small as one picometer — a trillionth of a meter, about a hundredth the width of a hydrogen atom (ch. “Going Beyond the Ultimate: Pico- and Femtotechnology and Bending the Speed of Light”). He offered it as evidence that engineering was reaching down toward the picometer scale, a rung on the ladder toward manipulating matter at the level of individual subatomic structure.

In the same stretch of the book he made two far larger wagers. One was conditional and cosmic: if subtle deviations from the speed-of-light limit exist, a future civilization will ultimately harness superluminal information transfer (ch. “The Six Epochs”). The other was the thesis of the entire book — that the law of accelerating returns will continue until nonbiological intelligence comes close to saturating nearby matter and energy with computation, after which expansion proceeds outward through the cosmos (ch. “The Singularity Is Near”).

And undergirding all of it was the engine: his claim, from the book’s opening argument, that after 2014 humanity would compress the next twenty years of progress — measured at the year-2000 rate — into only seven years (ch. “The Intuitive Linear View Versus the Historical Exponential View”).

Where we actually are

The picometer barrier didn’t just hold — it shattered, by six orders of magnitude. Kurzweil cited a one-picometer detector as a frontier. Twenty years later, that frontier is a footnote. In 2021, a team publishing in Science demonstrated electron ptychography that reaches “atomic-resolution limits set by lattice vibration” — meaning the imaging itself is no longer the limiting factor; the jiggling of the atoms is (DOI 10.1126/science.abg2533, 415 citations). The picometer is now the easy regime.

Above it sits the productized version. US 12,247,856, granted in 2025, claims a “micro-displacement measurement system having picometer scale resolution” built from fiber Bragg grating sensors, a broadband light source, and a lock-in amplifier packed into a stainless-steel probe — picometer metrology as a shippable instrument, not a lab heroic. US 11,598,628 (2023) covers “high dynamic range picometer metrology systems.” Picometer measurement, in other words, has crossed from physics paper to patent claim.

And below the picometer, the gap to Kurzweil’s frontier becomes almost comic. LIGO’s gravitational-wave interferometers routinely measure mirror displacements on the order of an attometer — 10⁻¹⁸ meters, about a thousandth the diameter of a proton, and a million times finer than the picometer Kurzweil flagged as the cutting edge. Compact tabletop Michelson interferometers now reach sub-picometer sensitivity as a matter of routine engineering. On precision measurement, reality didn’t just meet the prediction — it ran past it and kept going.

The speed-of-light bet is aging badly, and the evidence is the reason. Kurzweil’s superluminal prediction was carefully hedged — if deviations from the light-speed limit exist. The trouble is that twenty years of increasingly brutal experiments keep finding that they don’t. A 2015 rotating-cavity experiment constrained Lorentz symmetry violations for the photon to the 10⁻¹⁸ level. The most punishing test came from GRB 221009A, an exceptionally bright 2022 gamma-ray burst whose photons spanned an enormous energy range across cosmological distance: if light speed depended even faintly on energy, the high-energy photons should have arrived measurably early. They didn’t. A November 2025 measurement reconfirmed the constancy of light speed “with unprecedented accuracy,” shrinking the room for any deviation by roughly another order of magnitude.

None of this falsifies Kurzweil — his claim was conditional, and a conditional whose premise keeps failing simply goes dormant rather than wrong. But the direction of travel is unmistakable. Every experiment that tightens the bound on Lorentz violation is an experiment that makes his “if” less likely to ever cash out. The premise is eroding in real time.

The cosmic endpoint remains exactly where he left it: untestable. The prediction that intelligence will saturate local matter with computation and then expand outward is, by construction, not something a 2026 scorecard can grade. There is no instrument, patent, or paper that bears on whether the observable universe eventually “wakes up.” It is a statement of faith about the very long run, and it should be scored as what it is — too early to call, and likely to stay that way for centuries.

The engine is still running — but the fuel changed. The seven-years-for-twenty claim is the one that actually matters, because it is the mechanism the whole edifice rests on. Here the picture is genuinely mixed. Raw hardware price-performance has slowed: GPU FLOP/s per dollar has been doubling roughly every 2.5 years, a touch slower than classic Moore’s law, and former Intel chief Pat Gelsinger has pegged transistor doubling closer to three years (Epoch AI, “Trends in GPU price-performance”). On the narrow metric Kurzweil leaned on hardest in 2005 — cheaper transistors — the curve is bending the wrong way.

Yet by every measure that tracks deployed capability, the acceleration is faster than he predicted. The training compute behind frontier AI models has been growing 4–5x per year — doubling every five to six months, not every two years (Epoch AI). Kurzweil himself, in The Singularity Is Nearer (2024), updates the headline figure: “one dollar buys about 11,200 times as much computing power, adjusting for inflation, as it did when The Singularity Is Near hit shelves.” That is roughly a doubling every fourteen months over nineteen years — faster, not slower, than his original curve, because he now counts specialized AI silicon and not just general-purpose chips.

So the engine is real and arguably ahead of pace. But the fuel is not the one he named. The post-2014 acceleration has come from staggering capital deployment and algorithmic efficiency — deep learning arriving around 2010, transformers in 2017, the scaling era after — far more than from the steady cheapening of transistors. Kurzweil predicted the destination and got there early. He just took a road he didn’t draw on the map.

The scorecard

What Kurzweil missed (and what he nailed)

A pattern runs through this batch, and it is the most consistent finding across the whole scorecard project: Kurzweil is a superb forecaster of direction and a shaky forecaster of mechanism. On the picometer, he didn’t even forecast aggressively enough — he treated as a frontier something that would become a commodity. On the exponential engine, he called the acceleration correctly and even early, but attributed it to a transistor curve that has since slowed, when the real driver turned out to be money and math. On the speed of light, he hedged into a conditional that is slowly being starved of its premise.

The lesson for anyone forecasting technology is uncomfortable. The cheap intuition — “things slow down, the skeptics are usually right” — lost badly on measurement precision and on AI compute. But the maximalist intuition — “the curve is the curve, the mechanism doesn’t matter” — is also wrong, because the mechanism is exactly what determines whether the curve continues. An exponential running on transistor physics and an exponential running on capital expenditure have very different failure modes. Kurzweil got the line right and the engine wrong, and right now that’s working in his favor. The interesting question is what happens the first time the capital stops doubling.

Method note

This scorecard pairs Kurzweil’s original text with current evidence drawn from a corpus of roughly 9 million patents, hundreds of millions of scientific papers, and live web research. Patent and literature counts come from full-text searches of those collections; the picometer and interferometry findings were read from the actual patent claims and paper abstracts, not inferred from counts. Compute-trend figures are from Epoch AI’s published analyses, and the physics results from peer-reviewed Lorentz-invariance and gravitational-wave measurements. Kurzweil’s updated figures are quoted directly from The Singularity Is Nearer (2024). Verdicts are the author’s judgment against that evidence; where a prediction is untestable or its premise unresolved, it is scored as such rather than forced.