Kurzweil Scorecard: Physics at the Edge and the Law of Accelerating Returns

In 2005 Ray Kurzweil devoted a short, strange chapter of The Singularity Is Near to whether computation would eventually descend past the atom — into picometer, then femtometer scales — and whether the speed of light might turn out to be adjustable. The chapter reads today like a footnote. But it sits inside the same logical arc as his flagship claim: that progress is accelerating at a doubling rate, that the twentieth century’s achievements equaled “about twenty years of progress” at the rate of 2000, and that twenty more such years would arrive by 2014.

The most interesting finding in this batch isn’t that Kurzweil missed on picotechnology. It’s that his core acceleration curve — the mechanism he built his forecasts on — has been independently rediscovered, with numbers that match his to the month, by a research group with no obvious reason to vindicate him.

The predictions

This batch covers twelve predictions from the philosophical scaffolding of The Singularity Is Near: the Law of Accelerating Returns, the Six Epochs framework, the GNR revolutions (genetics, nanotechnology, robotics), three physics benchmarks from 2005 that were supposed to foreshadow sub-atomic computing, and one speculation about the speed of light. Eight are testable. Four are long-term metaphysics — intelligence saturating matter, the universe waking up — that can only be assessed by whether their preconditions are advancing.

Where we actually are

The doubling-time claim. Kurzweil wrote in 2005 that “because the paradigm-shift rate doubles every decade, the twentieth century’s achievements were equivalent to about twenty years of progress at the rate of progress in 2000” (ch. “The Intuitive Linear View Versus the Historical Exponential View”). In The Singularity Is Nearer he restated it in narrower, measurable form: “the compute used to train a state-of-the-art artificial intelligence model has been doubling every 5.7 months since 2010.”

Epoch AI, the nonprofit tracking frontier AI compute, published a 2024 analysis showing training compute for frontier language models has grown 5× per year since 2020 — a 5.2-month doubling time, within rounding of Kurzweil’s figure. The top-5 models’ training compute grew roughly ten-thousand-fold in five years. In the same window, price-performance of compute per dollar rose 11,200× between Near and Nearer, using Kurzweil’s own number. On the specific, falsifiable version of the acceleration claim, the evidence is cleanly “ahead of schedule.”

Twenty years of progress by 2014. The companion claim — that fourteen calendar years from 2000 would deliver twenty “year-2000-equivalent” years of advancement — is harder to audit because the rate itself is the unit. But look at the fields Kurzweil expected to compound fastest. Attosecond-physics papers grew from 26 in 2005 to 104 in 2014, then 259 in 2025. CRISPR patent filings, effectively zero until 2013, reached 135 in 2025. The curve shape — near-flat through 2000, visibly exponential by 2014, still accelerating in 2025 — is what the prediction implied.

Attosecond imaging. Kurzweil wrote that “Cornell scientists had demonstrated X-ray scattering imaging that could record movies of a single electron with four-attosecond frames and one-angstrom spatial resolution” (ch. “Going Beyond the Ultimate: Pico- and Femtotechnology and Bending the Speed of Light”). That 2005 benchmark has been demolished. The 2023 Nobel Prize in Physics went to Pierre Agostini, Ferenc Krausz, and Anne L’Huillier for building the apparatus that made attosecond electron movies routine. In August 2025, a paper on arXiv reported 25-attosecond pulses from a tabletop Yb-based laser at more than 10¹² photons per second — a three-order-of-magnitude brightness gain over any prior soft X-ray source. A December 2025 result produced 19.2-attosecond soft X-ray pulses, below the atomic unit of time (24 attoseconds). Zeptosecond-electron-source proposals are in peer review at Scientific Reports and Physical Review Applied.

Atomic-force microscopy resolution. “German scientists had created an atomic-force microscope capable of resolving atomic features only seventy-seven picometers across” (ch. “Going Beyond the Ultimate”). That figure — the separation between the closest minima in a 2004 IBM Zurich charge-density image — has been matched and extended by the qPlus sensor approach Franz Giessibl introduced in the late 1990s. qPlus-based AFMs now routinely resolve individual chemical bonds and image atomic orbitals in the sub-picometer motion regime; interferometric AFMs published through 2024 operate at a sub-picometer noise floor across a motion range of many microns. The instrument Kurzweil pointed to as the state of the art is now a commodity in surface-science labs.

Speed of light and the fine-structure constant. “Analysis by Steve Lamoreaux and Justin Torgerson of the Oklo natural reactor suggested the fine-structure constant alpha may have changed over two billion years, implying the speed of light may not be immutable” (ch. “Going Beyond the Ultimate”). The 2004 paper claimed a 4.5×10⁻⁸ decrease in alpha at 6σ. That hasn’t held up cleanly. A 2006 re-analysis with a more realistic neutron spectrum found the sign and magnitude depend sensitively on reactor model. A 2015 analysis concluded the Oklo data constrain alpha-variation only weakly. Atomic-clock bounds now pin any present-day drift at 10⁻¹⁸ per year. Kurzweil’s pointer is the weakest single claim in this batch, and subsequent work has not rescued it.

Picotechnology and femtotechnology. Kurzweil wrote in 2005 that “computation at pico- and femtometer scales using subatomic particles may become possible, though it remains speculative” (same chapter). Twenty-one years later the speculation has sharpened one direction only: attosecond probing of electron dynamics is real science, but nothing resembling computation at subatomic length scales exists. The 2024 update drops “picotechnology” and “femtotechnology” entirely. Kurzweil appears to have quietly retired the claim.

The GNR revolutions. The forecast that “the first half of the twenty-first century will be characterized by three overlapping revolutions in genetics, nanotechnology, and robotics” (ch. “GNR Three Overlapping Revolutions”) scores unevenly. Genetics is the clear win: Casgevy, the first CRISPR-Cas9 gene therapy, received FDA approval for sickle cell disease on December 8, 2023; by end-2024, 50-plus authorized treatment centers were active and Phase 3 enrollment had completed for children aged 5–11. CRISPR patents grew from one in 2013 to 135 in 2025. Robotics looks like motion, not arrival: Tesla has begun Gen 3 Optimus production at Fremont, but Elon Musk confirmed on the Q4 2025 earnings call that “no robots are doing useful work yet.” Figure’s humanoids have a narrow BMW factory role. The consumer-deployed general-purpose humanoids Kurzweil’s timeline would predict remain absent. Nanotechnology — specifically the Drexler-style molecular assembler — has stalled; no serious lab is attempting to build one. The push arrived in software compute and in molecular biology, not in atomic-scale mechanical engineering.

The scorecard

What Kurzweil missed (and what he nailed)

The split in this batch is sharper than in any previous scorecard. On the mechanism — that information-processing capacity compounds exponentially and each doubling enables the next — Kurzweil was right to the decimal place. Compute per dollar advanced at essentially the rate he projected. Frontier AI training compute is doubling faster than Moore’s law ever did, and Epoch AI’s 5.2-month figure is within rounding of Kurzweil’s 5.7-month claim. Attosecond physics is now a Nobel-decorated subfield with tabletop instruments. AFM resolution beat its 2005 benchmark within years, not decades.

Where he went wrong, he went wrong in one direction: by assuming that all information-rich domains would advance at roughly the same exponential rate. Molecular assemblers did not. Humanoid robotics did not. The speed of light did not turn out to be soft. What accelerated hardest was the layer closest to pure information — compute, neural networks, genomics — and the physical manifestations lagged. The picometer microscope is on track. The picometer assembler is not.

The cleanest takeaway: Kurzweil’s framework is scoring much better than his inventory. Read the 2005 book as a list of gadgets and the hit rate is uneven. Read it as a claim about which layer of reality compounds fastest, and the 2026 evidence says he identified the right layer.

Method note

Predictions were checked against the U.S. patent corpus (9.3M documents through April 2026), OpenAlex scholarly literature (357M papers), Epoch AI’s compute tracker, Nobel Prize materials for the 2023 Physics award, peer-reviewed attosecond and AFM publications, and targeted web queries on molecular nanotechnology, Optimus, Figure’s factory programs, and Oklo fine-structure-constant analyses. Kurzweil’s 2005 phrasings cite The Singularity Is Near; his 2024 updates cite The Singularity Is Nearer.