Kurzweil Scorecard: The Robots Went Civilian. The Battlefield Waited.

In 2005 Ray Kurzweil wrote a chapter titled “on Warfare: The Remote, Robotic, Robust, Size-Reduced, Virtual-Reality Paradigm.” He imagined, roughly in parallel, a near future of guided bullets, dust-sized reconnaissance devices, snake-form military robots, nanotube-reinforced combat vehicles light enough to airlift anywhere in four days, and an artificial olivocerebellar circuit that would one day command both military machines and prosthetic limbs. In the same span of predictions he tucked in a modest footnote: Carnegie Mellon had just finished a 3,000-mile “hands-off-the-wheel” drive from Pittsburgh to San Diego.

Twenty-one years later the footnote has eaten the chapter.

The most striking pattern in this batch of predictions is not a systematic miss on timelines. It is a systematic miss on who would build the autonomous machines. The Pentagon spent two decades failing to deliver a brigade in 96 hours, a dust-sized reconnaissance drone, or an operational guided bullet for infantry. Meanwhile, a single Alphabet subsidiary quietly put 3,000 driverless robotaxis on public roads across ten U.S. metros and logged 200 million fully autonomous miles.¹ Kurzweil got the robots right. He got the uniform wrong.

The predictions

Batch 61 contains nine predictions from The Singularity Is Near (2005), mostly drawn from the warfare chapter and the “Response to Critics” appendix. Eight are classified as robotics, one as nanotech. The unifying logic: Kurzweil believed the 2010s and early 2020s would see rapid convergence of miniaturization, autonomy, and smart-material engineering — and that the U.S. military, backed by DARPA, would be the fastest mover.

Where we actually are

Guided bullets. Kurzweil wrote that “providing bullets with a measure of intelligence, analogous to smart missiles, is a military objective” (ch. “on Warfare”). DARPA’s EXACTO program demonstrated in 2014 and again in 2015 that a .50-caliber round could change course mid-flight to hit a moving target.³ The Pentagon then went quiet. EXACTO has published no public results since 2015. The XM25 “Punisher” airburst grenade launcher — the closest fielded analog — was deployed in small numbers to Afghanistan in 2010, then formally terminated in July 2018.⁴ The M7 rifle, fielded in March 2024 as the Army’s new service weapon, uses conventional 6.8×51mm cartridges.

The research never stopped. US 12,241,721, granted March 2025, claims “at least one sensor, configured to determine a position, motion and/or orientation of said smart bullet while… in-flight; at least one actuator(s) configured to steer said smart bullet in response to communications” over a wireless link.⁵ Claim 20 goes further: the bullet “is configured to penetrate a target, sense condition(s)… within said target, and communicate information… wirelessly to a computer system in a remote location.” That is recognizably the object Kurzweil described. It just doesn’t have a line number in any U.S. procurement catalog.

Autonomous driving. Kurzweil’s 2005 reference point was CMU’s “3,000-mile hands-off-the-wheel trip from Pittsburgh to San Diego” (ch. “Response to Critics”). In The Singularity Is Nearer he updated the number: “Waymo’s self-driving vehicles have traveled well over 20 million fully autonomous miles at the time of this writing.”

That figure has since multiplied by an order of magnitude. Waymo reached 100 million fully autonomous rider-only miles in July 2025, 170 million by December 2025, and 200 million by March 2026 — completing 500,000 paid rides per week across ten metropolitan areas.¹ A peer-reviewed comparison to human benchmarks at 7.1 million rider-only miles found an 85% reduction in injury-causing crashes.² Through 2025 there were zero human fatalities caused by a Waymo vehicle in the rider-only fleet.²

The patent record is the story in compressed form. U.S. filings combining “autonomous vehicle” with “driving” or “steering” went from a handful per year in 2006 to 758 in 2024. Filings pairing LiDAR or radar with autonomous driving climbed from four in 2007 to 126 in 2024. Scientific papers on deep learning for self-driving exploded from 3 in 2005 to 2,447 in 2025. Kurzweil picked the right technology. He aimed it at the wrong decade and the wrong service branch.

Search-and-rescue robots. Kurzweil claimed these “were in practical use by 2004” (ch. “Response to Critics”). The prediction looks unambitious in retrospect. After the February 2023 Turkey-Syria earthquake, thermal-imaging drones, 3D-photogrammetry mapping drones, and ground robots assessing structural stability deployed within hours across the disaster zone. Boston Dynamics’ Spot has become the default industrial inspection platform for hazardous environments, with routine deployment for safety inspection beginning mid-2024 and, as of April 2026, integration of Google DeepMind’s Gemini Robotics model for autonomous reasoning. The field evolved from “in practical use” (a few research platforms at specific sites) to “routine infrastructure” without the prediction ever missing.

Dust-sized reconnaissance. This is the batch’s clearest miss. Kurzweil wrote that “before the end of the 2000s decade, devices the size of dust particles will be able to carry out reconnaissance missions” (ch. 8, “Promise and Peril of GNR”). The goal was compound: dust-scale and autonomous and reconnaissance-capable.

Sixteen years past deadline, the state of the art is a 0.3mm × 0.3mm temperature-sensing node from the University of Michigan — no propulsion, no reconnaissance mission planning, no autonomy. Working “smart dust” motes remain at cubic-millimeter scale, typically relying on RF backscatter and fixed positioning. US 12,518,602, granted September 2025, describes a “smart-dust” MEMS system for ATM physical security — intrusion detection, not reconnaissance. The patent describes nodes “less than 1 mm in size” that sense vibration and RF inside a bank machine. The size is right. The mission is domestic.

Snake-form robots. Kurzweil noted that “NASA was developing a snake-form robot design with envisioned military applications” (ch. “on Warfare”). The robot arrived, on the cover of Science Robotics in March 2024 — EELS (Exobiology Extant Life Surveyor), NASA-JPL’s autonomous snake robot designed to slither through the ice vents of Enceladus in search of extraterrestrial biosignatures.⁶ Dozens of patents describe snake-arm robots for engine servicing (US 12,569,947; US 12,186,848), anatomical interventions (US 12,004,825), and industrial inspection with convertibility to aerial drones (US 12,459,680). The envisioned military application is nearly absent from the current patent record.

Artificial olivocerebellar circuit. Kurzweil wrote that “an artificial olivocerebellar circuit being modeled by an international group will be applied to military robots and robots assisting disabled people” (ch. “Other Works in Progress”). The underlying science has matured. A 2024 paper in Frontiers in Neuroscience implemented a 9,504-neuron cerebellar spiking network with 240,484 synapses on FPGA, consuming 0.6 watts, controlling a DC motor under sudden load changes in real time — demonstrably biological in its adaptation profile.⁷ The authors note applicability “to various neuro-devices to augment and extend human motor control capabilities.”

No operational military robot uses such a controller. No commercial prosthetic uses one either. The science is on track. The deployment Kurzweil envisioned is absent.

Brigade Combat Team in 96 hours. “The U.S. Army’s goal was to deploy a Brigade Combat Team in 96 hours and a full division in 120 hours” (ch. “on Warfare”). GAO testing of Stryker brigades, conducted after the unit was fielded, found air deployment actually required 5 to 14 days — not 4. Twenty years on, the Army still describes 96 hours as a “goal” and “design metric,” not a demonstrated capability. Current doctrine assumes a mix of airlift and sealift. This one is behind schedule, arguably permanently.

Nanotube-reinforced combat systems. Kurzweil wrote that “new ultrastrong nanomaterials such as nanotube-reinforced plastics… will dramatically lower the weight of ground combat systems” by the 2010s. Carbon-nanotube composite patents stayed at roughly 100-170 filings per year from 2010 onward — a steady drumbeat, not the breakout he described. In 2025 researchers at Peking University combined heterocyclic aramid with treated long carbon nanotubes to produce a fiber absorbing energy at more than double the prior record; Israeli firm TorTech now manufactures nanotube yarn at commercial quantities. The materials exist. They have not dramatically lowered the weight of any fielded U.S. ground combat system.

Primo Posthuman. Kurzweil cited Natasha Vita-More’s “Primo Posthuman” — the whole-body prosthetic design featuring a metabrain, prosthetic neocortex interwoven with nanobots, smart skin, and high-acuity senses. Vita-More produced the design in 1997, updated it as Primo 3M+ in 2001, and in 2024 transitioned from Executive Director to President Emeritus of Humanity+. The conceptual artifact Kurzweil pointed to exists. Whether any part of it has been physically built belongs to a different batch.

The scorecard

What Kurzweil missed (and what he nailed)

The category hit-rate is mediocre. The pattern underneath it is what matters.

Every military-framed prediction in this batch is behind schedule or arrived via a civilian path. Every civilian or dual-use prediction is on track or ahead. The guided bullet became a patent thicket owned by startups and university labs, not a fielded round. The snake robot is crawling toward an ice moon, not through a bunker. The artificial cerebellum is running on a 0.6-watt FPGA in a neuroscience lab, not inside a prosthetic arm. Meanwhile, the machine Kurzweil treated as a near-offhand mention — an autonomous car — has accumulated more real-world driving miles than the entire U.S. Army has flown operational cargo hours in Stryker airlifts over two decades.

This is a forecasting pattern worth naming. Kurzweil, writing in 2005 with the Iraq War as backdrop and DARPA as the most visible funder of speculative autonomy, assumed the military path would cash its bets first. The budgets were there. The political will was there. But DoD procurement cycles, export controls, and human-in-the-loop doctrine created something closer to a clock running at half speed. Consumer technology ran at double. The gap compounds.

A second pattern: Kurzweil tends to get the object right and the deployment channel wrong. Smart bullets: the object exists in patents, not in rifles. Dust sensors: the silicon exists at sub-millimeter scale, but not as autonomous reconnaissance packages. Snake robots: they exist, but as space probes and surgical tools. If you read his predictions as forecasts about whether a technology will be invented, the hit rate is higher. If you read them as forecasts about who will own and use it, the miss rate jumps.

Method note

This post checked each prediction against the U.S. patent corpus (9.3 million documents through April 2026), the OpenAlex scholarly literature (357 million papers), and targeted web queries for funding, deployment, and expert commentary. Patent numbers quoted here were read claim-by-claim, not just counted. Research paper findings were pulled from full text where accessible. Kurzweil’s 2005 phrasings are cited to The Singularity Is Near; his 2024 updates are cited to The Singularity Is Nearer.