Kurzweil Scorecard: The Panoply of Criticisms, Twenty Years On

In March 2025, a London startup called Vaire Computing fabricated a 22-nanometer test chip named Ice River. The chip does something ordinary silicon has never done: when it finishes a calculation, it sends most of the energy that drove the logic back to where it came from, like a pendulum returning to its swing. An adder running on the prototype reclaimed 41 percent more energy than a conventional square-wave driver. The idea was sketched in the 1970s by Charles Bennett and Ed Fredkin. In 2005, Ray Kurzweil used it to answer a whole shelf of critics who told him computing had nearly hit the wall.

Batch 9 of our Kurzweil scorecard covers that reply: twelve claims assembled under A Panoply of Criticisms in The Singularity Is Near. They are the structural timbers that hold the rest of his forecast up. If any had rotted, the whole 2045 edifice would sag.

The criticisms Kurzweil was answering

Critics argued, roughly: transistors will stop shrinking; software is too fragile to keep up; even perfect computers cannot outrun the physics of heat; and the brain is a million times more complex than any machine. He answered with interlocking claims. The most testable:

• “Reversible computing can reduce energy requirements and heat dissipation by many orders of magnitude” (ch. “A Panoply of Criticisms”).
• “Moore’s Law for shrinking transistors on flat integrated circuits is expected to hit a limit around 2022” (ch. “The Criticism from Malthus”).
• “Computers in 2005 are on the order of one million times simpler than the human brain” (ch. “The Criticism from Ontology”).
• “Complex software is not inherently brittle, as shown by mission-critical systems such as automated airplane landings, critical-care monitoring, intelligent weapons, and hedge-fund automation” (ch. “The Criticism from Software”).
• A cluster of historical claims about algorithmic efficiency in FFTs, modems, partial differential equations, and defense avionics.

In The Singularity Is Nearer (2024), he restated the central wager: “after integrated circuits have reached their limits, new paradigms using nanomaterials or three-dimensional computing will take over” (ch. 2). He cites a 2022 study finding that better algorithms “halved compute requirements for a given level of performance every nine months from 2012 [to] 2021”.

Where we actually are

The flat chip ran out, on roughly the schedule he predicted. In April 2025 Intel moved its 18A process into volume production, ahead of TSMC’s comparable N2 node. Both replace the planar FinFET transistor with vertically stacked nanosheet channels — Intel calls it RibbonFET, TSMC calls it nanosheet FET — paired with backside power delivery. These are not flat integrated circuits anymore. In our patent corpus, US filings on gate-all-around and nanosheet transistors went from eight in 2015 to 109 in 2021 and have stayed above seventy every year since; patents on three-dimensional stacking of NAND, transistors, or chiplets run 450–530 grants per year. Kurzweil’s 2022 deadline was a node or two early — traditional shrinks continued through 3 nanometers — but the structural break to three-dimensional geometry is exactly what took over. On track, with a one-node slip.

Reversible computing stopped being a thought experiment. In 2018 a group at Trinity College Dublin published Single-Atom Demonstration of the Quantum Landauer Principle in Physical Review Letters (doi:10.1103/physrevlett.120.210601), confirming the energy-erasure bound in a single trapped ion — 110 citations and counting. A 2018 Nature Physics paper demonstrated the same effect in a molecular nanomagnet. A 2021 PRL paper derived the “universal bound on energy cost of bit reset in finite time” (doi:10.1103/physrevlett.127.190602), closing the theoretical gap between reversible and finite-speed logic. On the patent side, US 10,720,924 (Adiabatic logic cell, 2020) and US 11,206,021 and US 12,021,522 (Quasi-adiabatic logic circuits, 2021 and 2024) describe pullup/pulldown networks clocked by a periodic resonator that drives voltages gradually between rails — the classical Bennett recipe, now in granted claims. In March 2025, Vaire’s Ice River demonstrated the first commercial CMOS chip in which a running adder recovers energy at a measured 1.41 ratio against a square-wave baseline. Vaire’s stated roadmap target is 4000-fold efficiency gain over fifteen years. Kurzweil said “many orders of magnitude” — one is now in silicon. Behind schedule but validated.

Software brittleness was not the straw man Kurzweil made it out to be. His Panoply defense leaned on automated airplane landings and hedge-fund algorithms. On July 19, 2024, a single misconfigured content file from CrowdStrike — Channel File 291, 40 kilobytes — bricked 8.5 million Windows machines inside two hours. Hospitals canceled surgeries. Airlines grounded fleets. The U.S. House Homeland Security Committee later put Fortune-500 direct losses at $5.4 billion. Two years earlier, the Boeing 737 MAX’s MCAS system, wired to a single angle-of-attack sensor, killed 346 people across two crashes. The F-35, which Kurzweil cited as proof of manageable complexity — “tens of millions of lines of software code” — has grown to roughly 24 million lines across flight and logistics, and the 2024 GAO report flags software as the program’s largest risk. Wrong — and the evidence has gotten louder since he wrote it.

The “million times simpler than the brain” line was wrong even in 2005. Mainstream estimates put whole-brain computation at 10^17 to 10^18 operations per second. IBM’s BlueGene/L, 2005’s top supercomputer, ran at 2.8 × 10^14 FLOPS — roughly 350 to 3,500 times simpler than the brain, not a million. By 2025, frontier AI training runs have reached 10^25 total FLOPs, and US exascale systems clear 10^18 FLOP/s of sustained compute in a single building. On the raw arithmetic dimension, machines hit brain-scale around 2022. Ahead of schedule by a decade — on Kurzweil’s own terms.

Algorithmic efficiency is the claim that held up best. Kurzweil’s 2005 examples were historical: the Cooley–Tukey FFT’s 200-fold operation count reduction on 1,024-point transforms; modem data rates from 300 bps to 56,000 bps in twelve years; an elliptic PDE problem becoming 300,000 times more efficient between 1945 and 1985. Those figures check out. More important, the pattern has continued: Erdil and Besiroglu’s 2022 study, which Kurzweil cites in The Singularity Is Nearer, showed the compute needed to reach a given image-classification accuracy halved every nine months from 2012 to 2021 — faster than hardware. Current MLPerf training figures improve nearly five times faster than transistor density alone can explain. Ahead of schedule.

The exotic outer claims are still unfalsifiable. The 2.2-pound “ultimate cold computer” at 10^42 operations per second is a Seth Lloyd physics-limit calculation; nothing remotely close has been built. “Exponential trends through the next century” is a conjecture we will not resolve this decade. Too early to call on both.

The scorecard

What Kurzweil missed, and what he nailed

He was right about the shape of progress and wrong about one subsystem: human organizations. Every hardware and algorithmic claim in this batch either landed or is en route. Intel 18A, TSMC nanosheets, the Vaire reversible adder, the nine-month halving of ML compute — variations on a trend he described correctly two decades ago. His million-fold gap between 2005 machines and the brain is off by about three orders of magnitude in the wrong direction, which means reality ran ahead of his forecast, not behind.

The miss is human. Kurzweil argued that the mission-critical systems we already trust — flight control, medical monitoring, financial trading — prove complex software need not be brittle. Twenty years of evidence says the opposite. The failures that matter are not in the instruction-level logic. They live at the organizational seams, where a kernel-mode agent can push a config file to eight million machines with no staged rollout, or where a single sensor input can feed a control loop with authority to push a jet into a dive. The silicon got better. The interfaces between humans, vendors, and regulators did not. Any honest scoreboard for the Singularity has to keep that column too.

Method note

Counts and specific claim text came from the US patent corpus (grants and pre-grants through 2025) and from publication-year trends in the scientific literature. Web searches filled in recent product announcements, Government Accountability Office reports, and peer-reviewed experimental papers cited by DOI. Every figure in this post traces to a specific source we read today: a granted US patent number, a DOI, or a named publication.