Kurzweil Scorecard: The Face Got Its Computer. The Sky Did Not.

A clean line runs through Kurzweil’s 2005 computing batch. On one side: predictions about bolting a computer to a face — glasses that overlay the world, earpieces that skip the eardrum, electrode arrays that tap the motor cortex. On the other: predictions about bolting a computer to a star — Dyson shells, Matrioshka brains, the “ultimate cold laptop” at the thermodynamic ceiling, Anders Sandberg’s Zeus and Uranos designs.

Twenty-one years on, the face side is shipping. The sky side is still a chalkboard. One number makes the split visible: the fastest supercomputer on Earth runs at 1.8 × 10¹⁸ operations per second. Kurzweil’s “ultimate cold laptop” sits at 10⁴² — a factor of 10²⁴ away. Twenty-one years of the steepest compute scaling in history has not closed that gap by a trillionth.

The predictions

Kurzweil’s 2005 arc had three scales. Personal — AR overlays, anywhere-anytime VR, bone-conduction and directional audio, implantable computing. Planetary — Earth intelligence expanding exponentially toward the matter-energy limit. Cosmic — Dyson-shell computers, nested Matrioshka brains, black holes as the ultimate substrate, Sandberg’s Zeus (Earth-size, 10⁶¹ cps) and Uranos (solar-shell, 10⁵¹ cps) concepts.

In The Singularity Is Nearer (2024) he restated the near-term version: “Augmented reality will be projected constantly onto our retinas from our glasses and contact lenses. It will also resonate in our ears and ultimately harness our other senses as well… In the 2030s, medical nanorobots will begin to integrate these brain extensions directly into our nervous system.” The 2005 book placed most of this in the 2010s. Most of it is, in fact, happening — a decade late, through roughly the mechanisms he named.

Where we actually are

The face got its computer. Kurzweil wrote that “augmented reality will overlay the real world with real-time guidance, explanations, reminders about people and places” (ch. “…on the Human Brain”), placing this in the 2010s. US patent grants for AR overlay systems climbed from 16 in 2015 to 91 in 2025 — a six-fold expansion, most of it after 2018. The grants read like applied versions of Kurzweil’s sentence. US 12,576,205 (March 2026) describes an AR headset that watches an injectable medication administration in real time, overlays needle targeting, tracks plunger movement, cross-references the vial against the prescription, and produces timestamped compliance records. US 12,536,922 (January 2026) generalizes the pattern across clinical workflows; US 12,239,381 (March 2025) puts surgeons into an arthroplasty overlay. It all landed ten years late, in a form factor Kurzweil got wrong. The dominant shipping device is not a retinal projector but a lightweight pair of Ray-Bans: Meta and EssilorLuxottica sold seven million smart glasses in 2025 — triple their entire prior-year run, roughly 76 percent of the global market. Apple’s Vision Pro, the closest thing to Kurzweil’s retinal-resolution full-immersion device, collapsed (≈45,000 units expected for Q4 2025 after a Luxshare production halt); Apple has reportedly pivoted toward smart glasses too.

VR “anytime” still isn’t “full-immersion anytime.” High-resolution VR headset grants have plateaued at two to four per year since 2020, and the Vision Pro flop signals the market is not yet pulling on the high-immersion end. The 2010s deadline is missed; the 2020s are closing the gap via smart glasses, which is a different product.

The skull became an audio bus. Kurzweil’s 2005 claim was essentially an observation — bone-conduction MP3 players, army skull-conduction helmets, Hypersonic Sound, Holosonics Audio Spotlight all existed. He was right, and the industry scaled. Shokz (the former AfterShokz) went from one US patent in 2014 to 131 in 2024; the bone-conduction market is forecast to grow from ≈$1.05B in 2024 to over $8B by 2034 at ~22.8% CAGR. Holosonics launched its X-Series Audio Spotlight at InfoComm 2025, claiming ten times better isolation than conventional speakers, shipping into museums, kiosks, and banking counters. The niche is, at this point, plumbing.

Implants moved from thought experiment to outpatient. Neuralink placed its N1 device into 13 human subjects across the US, Canada, and the UK by late October 2025, with 20–30 more targeted by year-end. Two of Neuralink’s recent grants — US 12,369,863 (neural signal compression, July 2025) and US 12,248,629 (dynamic multiplexing for high-density recording, March 2025) — describe the engineering problem “routine” implantation actually requires solving: getting thousands of simultaneous neural channels off a chip via short-range wireless without burning out the battery or the tissue. Synchron’s endovascular alternative (US 12,333,074, US 12,248,631) routes signaling through a blood-vessel stent without craniotomy.

The clinical literature tips this toward “ahead on capability, on track on scale.” BrainGate’s An Accurate and Rapidly Calibrating Speech Neuroprosthesis (Card et al., NEJM 2024, DOI 10.1056/NEJMoa2314132) reported 99.6% word accuracy on a 50-word vocabulary within 30 minutes of training, 90.2% on a 125,000-word vocabulary after 1.4 hours, and 97.5% sustained across 248 cumulative hours over 8.4 months — with four 64-channel Blackrock microelectrode arrays in the left precentral gyrus of an ALS patient. Thirteen Neuralink patients and a few hundred industry-wide trial slots is not “routine.” But the capability is ahead of what a 2005 reader would have extrapolated.

The sky did not get its computer. The cosmic-scale predictions are where Kurzweil left them in 2005, with footnotes. Sandberg’s 1999 Jupiter Brains paper remains the reference design for Zeus and Uranos; nothing between Earth-size and solar-shell computing has moved from theory toward construction because the intermediate steps — megastructure-scale space engineering — have not begun. Matrioshka brain literature still cites Bradbury (1997) and Sandberg (1999). Lloyd’s 2000 Nature black-hole ultimate-computer result (5 × 10⁵⁰ ops in the ≈10⁻¹⁹ seconds before Hawking evaporation) was refined in Reilly & Lloyd’s June 2025 preprint Physical complexity and black hole quantum computers (arXiv 2506.16527) — another refinement of the same theoretical framework, not a construction step.

The “ultimate cold laptop” is the sharpest data point. Kurzweil’s 10⁴² cps per liter (rising to 10⁵⁰ cps per kilogram with smarter heat management) is a physics bound, not a roadmap. The fastest supercomputer on the November 2025 TOP500 is El Capitan at Livermore: 1.809 exaFLOPS — 1.809 × 10¹⁸ ops per second — across 11.34 million cores. That is twenty-four orders of magnitude short of the ultimate-laptop bound and roughly twenty-three short of the 10⁴¹-class performance Sandberg’s Uranos assumes. The gap has narrowed by perhaps five or six orders since 2005. The other eighteen are untouched.

Earth-scale intelligence is expanding — but not toward a matter-energy limit. Kurzweil’s “by 2040s” prediction has a measurable near-term proxy: total global data-center electricity use. The IEA’s November 2025 numbers put that at ≈415 TWh in 2024, with 2025 growth of 17 percent year-over-year and a base-case projection of 945 TWh by 2030. That is exponential, roughly doubling on a six-year cadence. But “matter-energy limits” for the Earth’s crust and incoming solar budget sit fifteen-plus orders of magnitude further out. Direction right. Clock to the ceiling: very long.

The scorecard

What Kurzweil nailed and what he missed

The pattern in this batch is unusually crisp. The near-body predictions are landing. The end-of-century ones are not advancing — they can’t, because the engineering path to any of them runs through intermediate steps nobody has started. Dyson shells require solar-shell-scale space construction. Matrioshka brains require Dyson shells. Black-hole computers require compressing a kilogram into a Schwarzschild radius of ≈1.5 × 10⁻²⁷ meters and reading it out in 10⁻¹⁹ seconds. The 2025 literature on these is the 2005 literature plus refinements. If Kurzweil’s core forecasting engine — exponential scaling in cost-performance — is doing the work, the cosmic predictions are not on any visible curve.

What he consistently got right, even with a decade slippage, was direction. The next computer moved to the face (Ray-Bans, AR procedure-guidance headsets, surgical loupe overlays). It moved into the skull (BrainGate, Neuralink, Synchron). The skull became an audio bus (Shokz at drugstore prices). On that axis the 2005 book reads more like an engineering roadmap than a futurist tract.

What he missed was the form-factor rotation. Retinas and contact lenses were the wrong end of the stack to optimize — cheap, light, camera-equipped sunglasses turned out to be enough. Most users do not need 4K-per-eye resolution; they need a camera, a mic, a voice assistant, and open ears.

One final pattern. When Kurzweil gave bounds in calculations per second — 10¹⁴ for a brain, 10⁴² for a cold laptop, 10⁶¹ for Zeus — the modest-end bounds were validated by practice and the extravagant ones have neither been approached nor refuted. They are physics claims, not technology claims, and physics claims age differently from technology claims. Kurzweil occasionally treated them as the same.

Method note

Patent numbers, counts, and claim summaries came from the US patents corpus (9.3M grants and pre-grants indexed through April 2026) via keyword and assignee queries. Literature figures came from a 357M-paper OpenAlex mirror; the BrainGate NEJM 2024 paper was selected for its citation weight and clinical endpoint. Smart-glasses, bone-conduction, and supercomputer figures came from IDC, Counterpoint Research, TOP500 (Nov 2025), IEA (Nov 2025), and Gartner. Kurzweil quotations are from The Singularity Is Near (2005) and The Singularity Is Nearer (2024). Verdicts are the author’s.