Kurzweil Scorecard: The Swarm Came for Us, Not the Abrams

The 2005 robotics-and-warfare chapter reads strangely in 2026. Kurzweil opened with a boast: the M1 Abrams tank, only three combat casualties in twenty years of use, a testament to intelligent armor. Two paragraphs later he described DARPA’s then-new plan for a 120-robot swarm built by iRobot, directed by commanders through specialized virtual-reality consoles. The future of ground combat, as he framed it, belonged to us.

By June 2025, 27 of 31 M1A1 Abrams tanks delivered to Ukraine had been destroyed, captured, or abandoned. Most were killed by single-use kamikaze drones, coordinated in small flights, increasingly autonomous, sometimes in self-organizing groups of three to twenty-five. The Pentagon’s 2026 budget allocates $107 million for stealth paint, top-armor kits, and laser-warning receivers — survivability upgrades forced by the same swarm paradigm Kurzweil predicted would be ours.

This batch of ten predictions covers human-level robots, cyborgs, military VR, the Moravec machine-vision program, and Kurzweil’s “human body version 3.0.” Most score “wrong mechanism” or “right destination, different vehicle.” One is overtaken by events.

The predictions

The 2005 claims in this batch share a single thesis: reverse-engineering the brain and retina would give us human-level perception and dexterity, and those capabilities would propagate first into military systems, then into consumer bodies, and finally into a redesigned human form. Human-level robots by 2029. Cyborg foothold by the 2030s. Body 3.0 by the 2040s.

Kurzweil wrote that “the robotics revolution will produce human-level robots whose intelligence is derived from human intelligence but redesigned to far exceed human capabilities” (ch. “GNR: Three Overlapping Revolutions”). In The Singularity Is Nearer (2024), he restated the arc: “In 2000, Honda’s ASIMO humanoid robot wowed experts by gingerly walking across a flat surface without falling over. By 2020, Boston Dynamics’ Atlas robot could run, jump, and tumble across an obstacle course with greater agility than most humans.” He frames the lineage as a straight line from reverse-engineered biology to commercial humanoids. The data does not support that lineage.

Where we actually are

Humanoid robots have arrived, but not through the retina. The top USPTO humanoid-robot assignees since 2021 are UBTECH Robotics (Shenzhen), Sanctuary Cognitive Systems (Vancouver), and Figure AI (Sunnyvale) — not Honda, not Moravec’s laboratories. Sanctuary’s US 12,409,570 (September 2025) describes a transformer platform that pivots between a standing bipedal configuration and a seated wheeled one via a pedestal linkage against an omnidirectional wheel. Figure’s US 12,420,434 claims a hand whose fingers are self-contained single-motor modules, each removable from the frame. These are mechanical-engineering patents. The intelligence layer riding on top is a vision-language model trained on human demonstrations — the opposite of the retina-first program Kurzweil described.

Commercially, 2026 is “mass production year zero.” Unitree’s 2025 output alone exceeded the combined output of Tesla, Figure, and Agility. Unitree has committed to 75,000 humanoids annually; AgiBot hit 10,000 units in March 2026. But only Agility’s Digit earns revenue from productive work — 100,000+ totes moved at GXO warehouses, contracts with Toyota and Mercado Libre. Figure 03 is in a single BMW pilot. Atlas is shipping to Hyundai with no external customers. Optimus, by Musk’s own 2026 admission, is not doing useful work. Against the “far exceed human capabilities by 2029” bar, this is a miss. Against the quieter 2024 restatement — agile humanoid platforms by the early 2020s — it is roughly on schedule.

Machine vision arrived early, along a path that treated the retina as irrelevant. Kurzweil wrote that “Hans Moravec has reverse engineered retinal and early visual processing, and only recently has microprocessor power become sufficient to replicate human-level feature detection for robots navigating unplanned complex environments” (ch. “The Visual System”). Modern robots do navigate unplanned environments with human-comparable visual competence. But the winning stack is convolutional and transformer architectures trained on internet-scale image data. None of it begins from retinal circuitry. 1X CEO Bernt Bornich wrote in 2024 that humanoid progress suggested Moravec’s paradox “might be debunked, and we just didn’t have the data” — a statement about data scale, not about retinal modeling.

The cyborg foothold is real, but two orders of magnitude short. Kurzweil predicted “humans will become cyborgs as nonbiological components gain a foothold in the brain” by the 2030s (ch. “Vexing Question of Consciousness”). Neuralink’s PRIME study now counts 21 participants globally (up from 12 in September 2025), plus seven GB-PRIME patients implanted at UCLH in late 2025. Musk has committed to thousands of implants by end of 2026. The USPTO has issued 32 Neuralink-assigned patents, including US 12,369,863, which claims four neural-signal compression schemes — lossless difference prediction, lossy wavelet transforms, binned-spike, and spike-band power — for a 1,024-channel recording array. Timeframe intact; scale — “humans become cyborgs” meaning mass merger, not 28 paralyzed subjects moving cursors — not.

DARPA’s 120-robot iRobot swarm never happened the way Kurzweil described. The 2003 announcement existed; the program faded. The swarm revolution arrived via drones, not iRobot ground platforms, and via commercial autonomy stacks, not DoD programs. Ukrainian units have deployed coordinated drone swarms in 100+ operations as of late 2025, typically three to eight drones each but tested at 25 and preparing for 100-drone trials. Drones share jamming-region data and redistribute roles when one is lost. Russian Shaheds recovered in 2025 contained Nvidia AI chipsets and thermal-vision modules capable of nighttime target lock. None of it runs on the command-and-control model Kurzweil sketched, where commanders would don specialized VR rigs to visualize the swarm. The actual model is operator dashboards, on-device autonomy, and kill-chain automation. The VR helmet for swarm command is not a thing.

The Abrams prediction is the most savagely overtaken. Kurzweil cited the tank’s three-casualty record as evidence intelligent defensive systems produced battlefield superiority. In 2005, that was true. By 2025, 87 percent of Ukraine’s Abrams fleet had been destroyed or captured, mostly by $500 FPV drones. The Pentagon’s 2026 budget allocates $107M for stealth paint, top armor, belly protection, and laser-warning receivers — an admission that the intelligent-armor era ended when swarm economics flipped the defender-attacker ratio.

The DoD virtual-human training prediction is the cleanest hit. The Institute for Creative Technologies, which Kurzweil named explicitly, is still doing what he described. At I/ITSEC 2024, ICT presented the Watercraft and Ship Simulator of the Future. The Leaders Enhanced & Applied Doctrine System is in evaluation at the Command and General Staff College. ICT’s Virtual Human Therapeutics Lab runs AI-driven interpersonal-skills training for second lieutenants at Fort Benning with a virtual squad leader, “SSG Jacob Garza.” A $4.5M congressional earmark funds the Army Writing Enhancement tool. Substrate upgraded from scripted virtual humans to LLM agents; prediction held.

“Human body version 3.0” is not yet scorable. Kurzweil placed the redesign in the 2030s–2040s, so the clock has not run. The prerequisites he named — nanotechnology, mature tissue engineering, widespread BCIs — are behind schedule. Without that substrate, body 3.0 is a 2045 conversation at earliest.

The scorecard

What Kurzweil got right, and what he missed

He nailed the destinations. Agile humanoid platforms exist. Virtual humans are standard in military training. BCIs are moving from single-subject research to multi-site trials. Drone swarms are operational. Robots navigate unplanned environments. Direction: correct. Arc: bent.

He missed the mechanism, and mechanism determines who wins. His 2005 model assumed the biological substrate was the bottleneck — perception through retinal reverse-engineering, dexterity through biomechanical fidelity. The actual path ran through data. The retinal model is a footnote. The biomechanical humanoid is a platform for a transformer to ride. The lab that wins is not the one that understands visual cortex; it is the one that collects the most video of human hands manipulating objects.

He also missed the economics. His 2005 framing had DARPA and the U.S. military leading the swarm transition. Instead, a Ukrainian infantry company with $500 drones has fielded more operational autonomy than any DARPA program. Two Chinese firms account for roughly 80 percent of 2026 global humanoid shipments. The United States leads in foundation models; it does not lead in robots that move. The kindest read of the batch is that Kurzweil was right about what would become possible by 2026 and wrong about who would build it. That is also the harshest read.

Method note

Evidence for the ten scored predictions came from the USPTO patent corpus (9.3M documents, full-text searched for humanoid, swarm, BCI, and military VR filings 2005–2026), the OpenAlex scientific-literature index (357M papers, filtered to recent high-citation robotics, BCI, and drone-warfare work), and open-web reporting on Neuralink’s PRIME trial, the Ukraine drone war, humanoid production tracking, and USC ICT. Quoted passages are from The Singularity Is Nearer (2024). Abrams loss figures were cross-checked across four sources including Pentagon 2026 procurement documents.