In 2005, Ray Kurzweil sketched a 21st century in which the social fabric — work, retirement, real estate, even the boundary between body and software — would tear loose from its pre-digital moorings. Twenty-one years later, six of those society-scale predictions are old enough to grade. Some of them showed up on schedule. Others arrived early, in the wrong outfit. One is conspicuously late.

The pattern is more interesting than any single verdict. Kurzweil was usually right about the direction of change and consistently wrong about the mechanism. The future he described did happen. It just took a path he didn’t see.

Six predictions, twenty-one years later

The book lays out a worldview in which information technology eats every domain it touches, and the things that used to be expensive — manufactured goods, office buildings, a long retirement — get reclassified as software problems. The six claims in this batch all hang on that frame. They concern firewalls for medical implants, the actuarial assumptions behind Social Security, the speed at which new technologies get adopted, the migration of value from atoms to bits, the fate of physical real estate, and the shape of the long-term growth curve.

In The Singularity Is Nearer (2024), Kurzweil revisits several of these. He observes that virtual reality and augmented reality “will merge into a compelling new layer to our reality” by the late 2020s, in which “many products won’t even need a physical form at all.” He concedes that “current virtual reality systems that incorporate smells or tactile sensations are still clunky and inconvenient.” Both of those caveats matter for the scorecard.

Where we actually are

Firewalls for the body. Kurzweil wrote that “secure private-network-style firewalls and evolving Internet security technologies will be used to protect software running in bodies and brains, though the conflict with malicious code will remain a nervous standoff” (The Singularity Is Near, “Molly 2004 / Ray dialogue”). Nanobots aren’t yet circulating in anyone’s bloodstream. But the antecedent technology — cybersecurity for things implanted in human flesh — is here and patentable.

US 12,217,860, granted in February 2025, describes a method for remotely bonding a “companion device” (typically a phone) to an implanted medical device through an encrypted, authenticated channel “resilient to companion device failure or compromise” — explicitly designed to avoid forcing surgical explantation if a key gets stolen. US 12,225,141 covers the relay: an implant talks low-power short-range to a phone, which then carries the encrypted traffic over the public internet. US 12,375,925 takes a different approach, having the external device deliberately inject noise onto the transcutaneous channel “inhibiting malicious interception” — a kind of in-body jamming. US 12,243,640 covers over-the-air firmware updates for medical devices, a feature that did not commonly exist a decade ago.

The FDA finalized post-market cybersecurity guidance that becomes mandatory in February 2026, requiring continuous monitoring, coordinated vulnerability disclosure, and patch management across all connected medical devices. The “nervous standoff” Kurzweil described is now a regulatory regime.

Social Security’s linear models. Kurzweil claimed in 2005 that government projections out to 2042 rested on “unrealistic linear models of longevity and economic growth.” The 2025 Trustees Report now projects the combined trust funds will deplete in 2034 — eight years earlier than the 2042 figure Kurzweil cited. The Old-Age and Survivors Insurance fund alone runs out in 2033.

He was right that the projections were wrong. He was wrong about why. The bleed-through happened mostly because economic growth disappointed, not because Americans suddenly lived to 100. US life expectancy at birth was 79.0 years in 2024 — a recovery from the 76.1 pandemic floor, but still below the 79.1 pre-COVID peak. The Trustees assume life expectancy of 82.0 years by 2055; the Census Bureau projects 83.7. So the actuaries are still slightly more conservative than mainstream demography, the opposite of the optimism Kurzweil predicted would break the model.

Adoption time halving each decade. This one is uncomfortable to score because it depends entirely on what you measure. If you measure consumer adoption of a single product, ChatGPT crossed 100 million users in two months — a pace UBS analysts called the fastest consumer internet ramp in twenty years. Radio took roughly 38 years to reach 50 million users; television about 13; the World Wide Web about 4. Halving each decade from 1920s radio would have predicted somewhere around two years for a 2020s technology. ChatGPT did it in eight weeks. Threads did 100 million in two days.

If instead you measure how long it takes a powerful new technology to actually change productivity statistics — the way electrification took fifty years to show up in factory output — there is no evidence the pattern has changed. Global GDP growth in the 2020s is on track to average 2.5%, the slowest decade since the 1960s. Adoption is faster; productivity translation is not.

Information-dominated value. The prediction here was that as molecular manufacturing matured, “the value of physical products will be based primarily on the information describing them rather than on manufacturing cost.” Atomically precise nanofabrication is not here. But the principle has shown up in adjacent industries that didn’t exist when Kurzweil was writing.

US 12,311,606, granted in 2025, describes a system in which a product designer sends “discrete packets of operational parameters” to a remote additive manufacturing device. Each packet is encrypted; the next packet is only transmitted after the previous one completes. The physical machine is reduced to a rented executor of someone else’s information. US 12,319,005 is a sibling patent covering pre-manufacture verification of the receiving facility. The information is the product; the printer is a peripheral.

Even the silicon industry has caught the bug. US 12,554,547, “silicon usage metering to provide silicon-as-a-service,” describes a chip that counts how often a particular feature is used and bills for it — software-defined silicon licensed pay-as-you-go. The physical wafer is a sunk cost; the value is gated by a counter and a license check.

Real estate becomes virtual. Kurzweil was right that office space would lose its grip on the economy. He was wrong about how. US office vacancy hit a record 20.7% in 2025; San Francisco crossed 27%, up from 8.6% pre-pandemic. The displacement happened. But it happened through Zoom, Slack, and a laptop on the kitchen table, not through immersive virtual environments. Roughly 28% of American paid workdays are now performed at home — a level Stanford’s Nicholas Bloom and coauthors documented in “Why Working from Home Will Stick” (cited 743 times since 2021).

The immersive replacement Kurzweil described is in retreat. Meta’s Reality Labs has lost between $70 billion and $80 billion since 2020; Horizon Worlds reportedly has fewer than 200,000 monthly active users. The technology industry’s investment narrative pivoted from metaverse to AI between 2022 and 2024. Patent filings tell the more nuanced story: the US patent system granted only a handful of metaverse-tagged patents per year through 2022, then 14 in 2023, 45 in 2024, and 84 in 2025. The vocabulary endures even as the consumer product fades.

Growth rates appear vertical. “Before the middle of this century, technological growth rates will become so steep as to appear essentially vertical to unenhanced humans.” By 2045, in other words. We are halfway there. Global GDP growth is decelerating, not accelerating: 5.9% average for developing economies in the 2000s, 5.1% in the 2010s, 3.7% in the 2020s. World trade growth has fallen from 5.1% in the 2000s to 2.6% in the 2020s. The aggregate curve, viewed at the resolution of national accounts, is bending the wrong way for this prediction’s timeline.

The compute curve, on the other hand, is doing exactly what Kurzweil said it would. It just hasn’t broken into the economy-wide statistics — yet.

The scorecard

What the pattern tells us

Three of the six predictions came true in form but not in substance. Offices emptied, but through a videoconferencing app, not a headset. The federal retirement system buckled, but on its growth assumptions, not its mortality assumptions. Implants got cybersecurity, but the implants are pacemakers, not nanobots.

Kurzweil’s blind spot is consistent. He thinks in technology stacks: each prediction assumes the field that delivers the change is the one closest to the problem in the lab. Virtual offices come from VR. Social Security strain comes from radical life extension. Implant security comes from molecular machines. The actual transitions came from adjacent, less glamorous technologies that were already shippable. Zoom shipped before the metaverse. CMS reimbursement codes shipped before mind uploading.

The compute curve is real and continues to track his projections. The mechanism by which compute transmutes into economic and social change is messier, slower, and more contingent on regulation, real estate leases, and middle managers than the 2005 frame admitted.

Method note

Patent counts and abstracts are drawn from a corpus of 9.3 million US patent grants and applications, queried through standard full-text search. Literature counts come from a 357-million-paper corpus of scientific publications. Verdicts are graded against passages from The Singularity Is Near (2005) and The Singularity Is Nearer (2024). Macroeconomic, demographic, and adoption figures are sourced from the 2025 Social Security Trustees Report, CDC NCHS mortality data brief 548, World Bank GDP series, Moody’s Analytics office vacancy data, FDA cybersecurity guidance, and the Stanford WFH Project. Six predictions in this batch; six verdicts above.