The Kurzweil Scorecard: Cross-Provider Synthesis Report

Tracking 147 Predictions Against Real-World Outcomes Through March 2026

Executive Summary

• The "86% accuracy" claim is demonstrably inflated, but the "7% accuracy" critique is equally misleading. The defensible independent accuracy rate for specific, time-bound predictions is approximately 40–50%, rising to 70–75% for directional accuracy (correctly identifying what technology would arrive, regardless of when). The gap between these figures is the central methodological dispute in evaluating Kurzweil's record.

• A domain-stratified accuracy pattern is the most actionable finding. Kurzweil's predictions achieve ~85–95% accuracy in pure information technology (computing hardware, networks, digital media), drop to ~65–70% in AI/software, fall further to ~40–60% in robotics and autonomous systems, and collapse to ~15–20% in biotechnology and nanotechnology — revealing that his Law of Accelerating Returns applies reliably only where progress is purely computational and fails systematically where physical-world deployment, regulatory friction, or biological complexity intervenes.

• The "5–10 year timing optimism" pattern holds but has become domain-differentiated. For pure software/AI predictions, the gap is narrowing (possibly to zero or even negative for AGI). For physical-world deployment predictions (self-driving, VR/AR, nanobots, longevity), the gap has widened to 10–20 years, suggesting Kurzweil's framework is becoming simultaneously more accurate in one domain and less accurate in another.

• The 2029 AGI prediction is Kurzweil's most consequential near-term test. All six providers agree it has shifted from fringe speculation to mainstream plausibility since 2022, but disagree on whether current LLM trajectories constitute genuine progress toward AGI or sophisticated narrow capability. The prediction's outcome within three years will either be his greatest vindication or his most prominent miss.

• Longevity Escape Velocity by 2029–2032 is his most vulnerable near-term prediction. All providers independently assess this as highly unlikely on his stated timeline, with realistic estimates ranging from 2035 to 2045+, representing a consistent 10–15 year optimism bias in the biological domain — the same structural error visible across his entire biotech prediction history.

Cross-Provider Consensus

Finding 1: The True Accuracy Rate Lies Between 40–55% for Specific Predictions

Providers: Perplexity, Anthropic, OpenAI, Gemini, Gemini-Lite, Grok-Premium Confidence: HIGH

All six providers independently converge on rejecting both Kurzweil's self-assessed 86% and the harshest critic figure of ~7%. The independent academic benchmark from the FHI/Armstrong study (~42%) is cited by five of six providers as the most credible anchor. Providers differ on whether to apply generous partial credit (pushing toward 50–55%) or strict binary scoring (holding at 40–45%), but none accept either extreme.

Finding 2: Directional Accuracy (~70–75%) Substantially Exceeds Timing Accuracy (~40–50%)

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

Five providers independently identify this as the central insight for practitioners. Kurzweil reliably identifies which technologies will become important but systematically overestimates the speed of mainstream deployment. This distinction has direct strategic implications: his predictions function better as technology roadmaps than as investment calendars.

Finding 3: Domain-Stratified Accuracy — IT Strongest, Biotech Weakest

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

All five substantive providers independently produce a similar domain hierarchy:

• Computing/networks: ~80–95% accurate
• AI/software: ~65–70% accurate
• Robotics/autonomous systems: ~40–60% accurate
• Biotech/longevity/nanotech: ~15–25% accurate

The consistent explanation across providers: Kurzweil's Law of Accelerating Returns is a valid descriptor of computational progress but fails to account for regulatory, biological, and physical-world friction that doesn't follow exponential curves.

Finding 4: Self-Driving Vehicles Are ~15 Years Behind Kurzweil's Timeline

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

All five providers agree that Kurzweil's 2009 prediction for autonomous highway driving was wrong on both timing (off by ~15–20 years) and mechanism (he predicted smart roads; reality delivered smart cars). All note that as of 2026, Waymo and limited robotaxi services represent genuine progress, but ubiquitous consumer autonomy remains 5–10 years away even from the current date.

Finding 5: The 2029 AGI Prediction Has Shifted from Fringe to Plausible

Providers: Perplexity, Anthropic, OpenAI, Gemini, Gemini-Lite, Grok-Premium Confidence: HIGH

All six providers independently note that the emergence of GPT-4 and successor models has dramatically increased the perceived plausibility of Kurzweil's 2029 AGI prediction. All note that in 1999 this prediction was considered absurd by mainstream AI researchers; by 2026, it represents a mainstream debate rather than a fringe claim. However, providers diverge on whether current systems are on a trajectory to meet the 2029 date specifically.

Finding 6: Longevity Escape Velocity by 2029 Is Highly Unlikely

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

All five substantive providers independently assess the 2029 LEV prediction as the most vulnerable near-term claim. The consensus realistic timeline is 2035–2045, representing a 10–15 year optimism bias consistent with Kurzweil's historical pattern in biological domains. All note that AI-driven drug discovery is genuinely accelerating but that regulatory timelines, clinical trial requirements, and biological complexity create irreducible friction.

Finding 7: VR/AR Adoption Is 10–15 Years Behind Kurzweil's Timeline

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

All five providers agree that Kurzweil's predictions for retinal-projection AR glasses by 2010 and fully immersive VR by the late 2010s failed significantly. The 2025 VR/AR market data (cited by Perplexity and Anthropic: $1.75B market, declining headset shipments in 2025) confirms that consumer adoption remains far below his projections. All providers note the technology exists but adoption has been dramatically slower than predicted.

Finding 8: The Systematic Timing Optimism Pattern Persists and Has Become Domain-Differentiated

Providers: Perplexity, Anthropic, OpenAI, Gemini, Grok-Premium Confidence: HIGH

All five providers confirm that the 5–10 year timing optimism pattern identified in earlier analyses continues to hold, but with an important evolution: the gap is narrowing for pure AI/software predictions (possibly to near-zero for AGI) while widening for physical-world deployment predictions (now 10–20 years for biotech, robotics, and nanotech). This bifurcation is a new and important finding not present in earlier analyses of Kurzweil's record.

Unique Insights by Provider

Perplexity

• The "Kurzweil Constant" quantified by prediction horizon: Perplexity uniquely breaks down timing error by prediction horizon — near-term predictions (5–10 years out) show 3–7 year delays; medium-term (10–20 years) show 5–12 year delays; long-term (20+ years) show 10+ year delays. This suggests the optimism bias compounds with prediction horizon length, which has direct implications for how to weight his 2045 Singularity prediction. No other provider quantified this relationship explicitly.

• Metaculus forecasting data on AGI timelines: Perplexity uniquely cites specific Metaculus prediction market data showing that the median "strong AGI" forecast extended from ~2031 to ~2033 during 2025 alone, even as AI capabilities improved. This counterintuitive finding — forecasters becoming more pessimistic even as capabilities advance — suggests the community is updating on bottlenecks (data quality limits, inference-compute scaling) rather than raw capability demonstrations. This is the most specific quantitative data point on AGI timing in the entire report.

• Specific AI benchmark data for March 2026: Perplexity cites GPT-5.4 achieving 75% on OSWorld-Verified desktop navigation (exceeding human performance at 72.4%) and 92.8% on web navigation tasks. This is the most specific and recent AI capability data point across all providers and directly bears on the AGI 2029 assessment.

Anthropic

• The "timing offset grows with domain complexity" framework: Anthropic uniquely formalizes the domain-differentiated timing error into a four-tier taxonomy with approximate offsets: pure computing/software (0–3 years), software + consumer adoption (5–8 years), hardware + regulatory + physical world (10–20 years), biology + nanotechnology (20+ years). This is the most actionable framework for practitioners adjusting Kurzweil's predictions for real-world use.

• Smart glasses vs. VR headset divergence: Anthropic uniquely notes that while VR headset shipments declined 42.8% in 2025, smart glasses grew 211.2% in the same year. This divergence — the form factor Kurzweil predicted (glasses) succeeding while the form factor the industry bet on (headsets) failing — represents a nuanced partial vindication of his directional prediction that no other provider captures.

• Neuralink clinical trial specifics: Anthropic provides the most detailed BCI data, noting 21 Neuralink participants as of early 2026 (up from 12 in September 2025), with a third ALS patient demonstrating typing capability, and Synchron raising $200M Series D backed by Bezos and Gates. This granularity is absent from other providers.

OpenAI

• The "external shock" blind spot in Kurzweil's methodology: OpenAI uniquely emphasizes that Kurzweil's framework, grounded in exponential technological trajectories, has no mechanism for incorporating geopolitical shocks, pandemics, or financial crises. The 2008 financial crisis (which Kurzweil predicted would not occur, forecasting continued economic boom) and COVID-19 (which temporarily reversed life expectancy gains globally) are cited as structural failures of his model that no amount of timing adjustment can correct. This is a methodological critique, not merely a timing critique.

• The "adoption vs. capability" distinction applied to speech recognition: OpenAI provides the most detailed analysis of the speech recognition prediction failure, noting that the technology was capable enough for mainstream use before 2009, yet adoption didn't follow because of competing preferences and ingrained behaviors. This case study is used to argue that Kurzweil's framework systematically conflates technological capability with social adoption — a distinction with broad implications for evaluating all his predictions.

Gemini

• The "Kurzweil Constant" as a domain-specific adjustment tool: Gemini uniquely proposes applying a systematic "physical-world delay" correction factor to Kurzweil's upcoming predictions: AGI (no penalty, purely computational), LEV (add 10–15 years → 2040–2045), nanobots in neocortex (add 20+ years → 2050s+). This is the most operationally useful framework for practitioners who want to use Kurzweil's predictions as inputs to planning.

• The Soviet Union prediction as an early validation: Gemini uniquely highlights Kurzweil's prediction of the Soviet Union's collapse (made in The Age of Intelligent Machines, written 1986–1989) as an early and often-overlooked validation of his framework — specifically his argument that decentralized communication technologies would render authoritarian information control impossible. This prediction, made years before the collapse, demonstrates that his framework has genuine predictive power beyond technology domains.

• The "non-causal model" critique: Gemini uniquely identifies that Kurzweil uses "non-causal models" — extrapolating historical data points without mapping specific underlying physics — which explains both his successes (correctly extrapolating trends) and failures (missing when underlying mechanisms change). This is a more sophisticated methodological critique than the "vague predictions" argument made by other providers.

Gemini-Lite

• The "process vs. event" distinction: Gemini-Lite uniquely articulates the distinction between Kurzweil as a "process forecaster" (predicting when a technology becomes viable) versus a "calendar forecaster" (predicting when it achieves mass adoption). This framing, while less detailed than other providers' analyses, offers a charitable and potentially correct reframing of how his predictions should be evaluated — not as calendar predictions but as capability emergence predictions.

Grok-Premium

• The pre-registration critique: Grok uniquely recommends that future audits of futurist predictions use pre-registered probabilistic forecasts for better calibration, rather than retrospective self-assessment. This methodological recommendation — essentially applying superforecasting standards to futurism — is the most forward-looking methodological contribution in the report.

• The "valuable exercise regardless of accuracy" argument: Grok uniquely argues that the act of making many falsifiable predictions decades ahead is itself rare and valuable, independent of accuracy rate. This meta-point — that Kurzweil's framework has inspired researchers and shaped investment even when specific predictions failed — is underemphasized by other providers who focus primarily on accuracy scoring.

• Drone/unmanned warfare as a strong hit: Grok uniquely highlights unmanned systems in warfare as a domain where Kurzweil's predictions were strongly validated, noting that drones have become central to modern conflict in ways that align with his forecasts. Other providers focus on consumer technology and largely overlook this domain.

Contradictions and Disagreements

Contradiction 1: The Baseline Accuracy Rate

The Dispute: Providers disagree on the appropriate baseline accuracy figure, with estimates ranging from 40% (strict independent scoring) to 55% (generous partial credit) to 86% (Kurzweil's self-assessment).

• Perplexity synthesizes to 45–52% strict, 55–65% generous, 70–75% directional
• Anthropic synthesizes to 35–50% specific/time-bound, 70–80% directional
• OpenAI cites 40–50% independent, 86% self-assessed
• Gemini cites 42% (FHI study) as the anchor
• Grok synthesizes to 40–60% strict, 70–85% directional

Why It Matters: The disagreement is not merely academic — it determines whether Kurzweil's framework should be used as a primary planning input (if ~80% directionally accurate) or a supplementary signal requiring heavy adjustment (if ~40% strictly accurate). The resolution depends on whether practitioners care about direction or timing, which varies by use case.

Investigative Note: The FHI/Armstrong study (42%) is the most methodologically rigorous independent assessment and should be treated as the floor. The directional accuracy figure (~70–75%) is the most useful for strategic planning. Neither Kurzweil's 86% nor the harshest critic's 7% survives scrutiny.

Contradiction 2: Whether the AGI 2029 Prediction Is "On Track" or "Still Too Early"

The Dispute: Providers diverge significantly on whether current AI progress supports the 2029 AGI timeline.

• Gemini argues the 2029 prediction may actually be conservative, citing Musk's 2026 prediction and Huang's similar timeline, and suggests current AI scaling laws support the date
• Perplexity notes Metaculus forecasters extended their median "strong AGI" estimate from 2031 to 2033 during 2025, suggesting the community is becoming more pessimistic even as capabilities improve
• Anthropic assigns 40–60% probability to the prediction being "broadly correct," the most explicitly probabilistic assessment
• OpenAI labels it "likely too early" with human-level AI possibly slipping to the 2030s
• Grok calls it "on track / too early to fully score" — the most agnostic position

Why It Matters: This is the highest-stakes near-term prediction in Kurzweil's entire body of work. The disagreement reflects genuine uncertainty about whether LLM capability improvements represent progress toward AGI or sophisticated narrow capability expansion. The Metaculus data (Perplexity) and the Gemini/industry-leader data point in opposite directions.

Investigative Note: The contradiction may be definitional — "AGI" means different things to different evaluators. Kurzweil's own definition (passing a sophisticated Turing test) may already be partially met; a stricter definition (autonomous agency across all human cognitive domains) is clearly not met. Future audits should specify which definition is being evaluated.

Contradiction 3: Whether Kurzweil's Timing Optimism Is "5–10 Years" or "10–20 Years"

The Dispute: Providers disagree on the magnitude of the timing error.

• Perplexity and OpenAI characterize the systematic error as "5–10 years" across most domains
• Anthropic and Gemini argue the error has widened to 10–20 years for physical-world predictions (self-driving: ~15–20 years late; VR: ~10–15 years late; biotech: ~15–20 years late)
• Grok takes a middle position, noting the error varies by domain

Why It Matters: If the timing error is 5–10 years, Kurzweil's 2029 AGI prediction could be off by only a few years. If the error is 10–20 years for complex domains, his longevity and nanotech predictions may not materialize until 2045–2060, fundamentally changing their strategic relevance.

Investigative Note: The Anthropic framework (domain-stratified timing error) is the most empirically grounded resolution. The "5–10 year" characterization appears to be an average that obscures important domain-specific variation. Practitioners should apply domain-specific corrections rather than a uniform offset.

Contradiction 4: Whether the Law of Accelerating Returns Is Slowing

The Dispute: Providers disagree on whether Moore's Law and the LOAR are continuing or encountering limits.

• Perplexity explicitly notes that Moore's Law is slowing (doubling period lengthening back to 2–3 years), that LLM scaling is encountering bottlenecks, and that recent AI improvements came primarily from inference compute rather than fundamental breakthroughs — suggesting the exponential foundation of Kurzweil's framework may be weakening
• Gemini and Grok largely accept the LOAR as continuing, citing ongoing compute improvements and AI capability gains as validation
• Anthropic and OpenAI take intermediate positions, acknowledging both continued progress and emerging limits

Why It Matters: If the LOAR is genuinely slowing, Kurzweil's entire predictive framework loses its foundation, and his 2029–2045 predictions become significantly less reliable. If it continues, his framework remains the best available tool for technology forecasting.

Investigative Note: This is the most fundamental unresolved question in the report. The semiconductor industry's own data on transistor scaling (TSMC, Intel roadmaps) and AI lab data on training efficiency improvements would be the most direct evidence. This warrants dedicated follow-on research.

Detailed Synthesis

The Accuracy Debate: Resolving the 86% vs. 42% Dispute

The most important preliminary finding is that the widely cited "86% accuracy rate" is a self-assessed figure derived from a specific methodological approach that independent researchers have consistently been unable to replicate [Perplexity, Anthropic, OpenAI, Gemini, Grok]. In 2010, Kurzweil published a 147-page retrospective audit of predictions made in The Age of Spiritual Machines (1999) for outcomes expected by 2009, claiming 115 "entirely correct" and 12 "essentially correct" predictions — a combined 86% [OpenAI, Gemini].

The FHI/Armstrong study, conducted by a panel of nine independent volunteers evaluating the same 147 predictions, arrived at approximately 42% [Gemini, Anthropic, OpenAI]. This discrepancy stems from three methodological differences: Kurzweil's generous interpretation of ambiguous language (any portable device counting toward specific form factor predictions), his willingness to count technologies that existed in prototype form as "correct" even without mainstream adoption, and his tendency to score predictions as "essentially correct" when they were several years late [Perplexity, Anthropic].

A middle position emerges from LessWrong-style independent reviews, which found approximately 50% of predictions more or less correct by 2019 — suggesting that many 2009 predictions that were "too early" in 2009 had materialized by 2019 [Perplexity]. This is a crucial finding: the appropriate accuracy figure depends not just on how you score but when you score. Kurzweil's predictions improve in retrospect as technologies eventually arrive, which is consistent with his "too early, not wrong" defense but also consistent with the critique that his timing is systematically optimistic [Anthropic, OpenAI].

The most defensible synthesis is a three-tier accuracy framework [Anthropic, Perplexity, Grok]:

• Strict accuracy (prediction correct within 1–2 years of stated date): ~40–50%
• Generous accuracy (prediction correct within 5–10 years): ~55–65%
• Directional accuracy (technology type correctly identified regardless of timeline): ~70–75%

Domain-Stratified Performance: The Core Finding

The most actionable insight from cross-provider analysis is that Kurzweil's accuracy is not uniform — it follows a clear domain hierarchy that maps directly onto the degree to which progress in that domain is purely computational versus requiring physical-world deployment [Perplexity, Anthropic, Gemini, Grok].

Computing and Information Technology (~85–95% accurate): This is Kurzweil's strongest domain and the empirical foundation of his entire framework [Gemini, Grok]. His Law of Accelerating Returns was first validated here, and it continues to hold. The price/performance of computation improved by a factor of 20 quadrillion between 1939 and roughly 2020 [Perplexity]. Specific hits include: the internet's explosive growth from 2.6 million users to billions [Gemini, Grok], the emergence of cloud computing [Perplexity, OpenAI], the shift to portable computing (laptop sales exceeding desktops in 2008) [Anthropic, OpenAI], and the proliferation of wireless connectivity [OpenAI, Gemini]. Deep Blue defeating Kasparov in 1997 — three years ahead of Kurzweil's 2000 prediction — is the most frequently cited example of him being early in the right direction [Anthropic, OpenAI, Gemini, Grok].

Artificial Intelligence and Machine Learning (~65–70% accurate): Kurzweil's AI predictions have been his second-strongest domain [Anthropic]. Narrow AI achievements — speech recognition, machine translation, image recognition, game-playing AI — arrived roughly on his timeline or slightly late [OpenAI, Grok]. His prediction that AI would create original art and music by the 2020s has been validated by DALL-E, Midjourney, and ChatGPT [Anthropic]. The significant miss in this domain is speech recognition as the primary text input method by 2009 — the technology worked, but human behavioral preferences kept keyboards dominant [Perplexity, Anthropic, OpenAI]. This case study is particularly instructive: it demonstrates that Kurzweil's framework correctly predicts technological capability but systematically underestimates adoption friction.

Robotics and Autonomous Systems (~40–60% accurate): This domain reveals the first major gap between computational capability and physical-world deployment [Anthropic, OpenAI, Gemini]. Self-driving vehicles are the canonical example: Kurzweil predicted autonomous highway driving by 2009, but as of 2026, commercial robotaxi services (Waymo operating 450,000+ rides per week across multiple cities) represent genuine progress while ubiquitous consumer autonomy remains years away [Anthropic, OpenAI, Grok]. The mechanism was also wrong — Kurzweil predicted smart roads, but the actual solution was smart cars [Anthropic, OpenAI]. Household robots remain largely limited to Roombas, despite decades of predictions about general-purpose domestic automation [OpenAI, Gemini]. Unmanned military systems (drones) represent a strong hit in this domain that is underemphasized in most analyses [Grok].

Biotechnology, Longevity, and Nanotechnology (~15–25% accurate): This is Kurzweil's weakest domain and the area where his timing optimism is most severe [Perplexity, Anthropic, OpenAI, Gemini]. His prediction that "most diseases would be eliminated by 2019" failed comprehensively [Gemini, Anthropic]. Average life expectancy has not approached his predicted 100+ years [OpenAI]. Nanobots for medical applications remain largely theoretical [Gemini, Grok]. The structural explanation is consistent across providers: biology is not a purely information technology, and the complexity of biological systems introduces irreducible friction that exponential computational progress cannot overcome on Kurzweil's timeline [Perplexity, Anthropic, Gemini].

The Critical 2025–2029 Window

AGI by 2029: This is Kurzweil's signature prediction and the one that has undergone the most dramatic reassessment in recent years [all providers]. When made in 1999, it was considered so extreme that Stanford hosted a conference where hundreds of AI experts dismissed it as fantasy [Gemini]. By 2026, it represents a mainstream debate. The emergence of GPT-4, Claude, Gemini, and their successors has demonstrated broad competence across thousands of domains that seemed impossible even five years ago [Perplexity, Anthropic].

However, providers diverge on whether current trajectories support the specific 2029 date. Gemini notes that prominent figures including Elon Musk and Jensen Huang have predicted AGI or human-level capabilities between 2026 and 2028, suggesting Kurzweil may be conservative. Perplexity counters with Metaculus data showing forecasters extended their median "strong AGI" estimate from 2031 to 2033 during 2025 alone, even as capabilities improved — suggesting the community is updating on bottlenecks rather than raw capability demonstrations. Anthropic assigns 40–60% probability to the prediction being broadly correct. OpenAI labels it "likely too early" with human-level AI possibly slipping to the 2030s.

The definitional problem is central: if AGI means "passes a sophisticated Turing test," current systems may already qualify in many contexts [Perplexity]. If it means "autonomous agency across all human cognitive domains," current systems fall far short [OpenAI, Anthropic]. Kurzweil's own definition — AI that "can do everything that any human can do" — remains unmet as of March 2026 [Perplexity].

Longevity Escape Velocity by 2029–2032: All providers independently assess this as Kurzweil's most vulnerable near-term prediction [Perplexity, Anthropic, OpenAI, Gemini, Grok]. The current rate of life expectancy gain is approximately 4 months per year in developed nations [Perplexity, Grok]. Reaching LEV would require tripling this rate in three years — an acceleration that would require not just scientific breakthroughs but simultaneous regulatory approval and global deployment of multiple complementary therapies [Anthropic]. AI-driven drug discovery (AlphaFold, Retro Biosciences, AI-designed senolytics) is genuinely accelerating biological research [Perplexity, Gemini], but the gap between laboratory discovery and clinical deployment remains a multi-year irreducible constraint. Aubrey de Grey, who coined the LEV concept, places the 50% probability date in the mid-to-late 2030s [Anthropic, Gemini]. George Church suggests 2050 [Gemini]. Applying Kurzweil's historical 15-year optimism bias in biological domains yields a realistic estimate of 2044–2047 [Gemini].

Brain-Computer Interfaces: BCI represents perhaps the most concrete recent progress toward Kurzweil's longer-term vision [Anthropic, Grok]. Neuralink has enrolled 21 participants as of early 2026, with demonstrated cursor control, gaming, internet browsing, and typing via thought [Anthropic]. Synchron's endovascular approach achieved safety endpoints with zero serious adverse events [Anthropic]. However, all current implementations remain in early clinical stages focused on restoring lost function to severely disabled patients — not the consumer cognitive augmentation Kurzweil envisions [Anthropic, Gemini]. His specific mechanism (nanobots navigating capillaries to the brain) faces material science and biocompatibility challenges that make the 2030s timeline implausible [Gemini, Grok]. The electrode-based approach currently being pursued is directionally aligned with his vision but mechanistically different and likely 10–15 years behind his timeline for consumer applications [Anthropic].

Self-Driving Vehicles: The current state — Waymo operating 450,000+ rides per week across Phoenix, San Francisco, Los Angeles, Atlanta, Austin, and Miami, with commercial launches planned in additional cities in 2026 [Anthropic] — represents genuine progress that validates Kurzweil's directional prediction while confirming his timing was approximately 15–17 years optimistic [Anthropic, OpenAI, Grok]. The mechanism error (smart roads vs. smart cars) is a notable miss that reveals a limitation in his framework: he correctly predicted the outcome (autonomous vehicles) but incorrectly predicted the technological pathway [Anthropic, OpenAI]. McKinsey survey data from 2025 indicates that expert adoption timelines have actually slipped by 1–2 years relative to 2023 forecasts, suggesting continued headwinds [Perplexity].

VR/AR Adoption: The 2025 market data tells a nuanced story [Anthropic]. VR headset shipments declined 42.8% in 2025, with Apple and Meta dropping advertising for their headsets — suggesting consumer appetite remains fundamentally limited [Anthropic]. However, smart glasses grew 211.2% in 2025 [Anthropic], suggesting that the form factor Kurzweil predicted (glasses) may be succeeding even as the form factor the industry bet on (headsets) struggles. The global AR/VR market at $1.75 billion in 2025 with a 10.3% CAGR [Perplexity] implies mainstream adoption remains a decade or more away. Kurzweil's retinal-projection prediction for 2010 was off by at least 15–20 years [Perplexity, Anthropic, OpenAI].

The Structural Explanation: Why the Pattern Persists

The consistent explanation across providers for Kurzweil's systematic timing optimism is that his Law of Accelerating Returns correctly describes computational capability but fails to account for the friction between capability and deployment [Perplexity, Anthropic, OpenAI, Gemini]. Anthropic formalizes this as a four-tier taxonomy with domain-specific timing offsets. Gemini describes it as the gap between "computationally feasible" and "societally deployed." OpenAI frames it as the failure to distinguish technological capability from social adoption. Perplexity identifies regulatory barriers, liability frameworks, insurance models, and behavioral adoption patterns as the specific mechanisms that create this friction.

Gemini's "non-causal model" critique adds a deeper methodological dimension: Kurzweil extrapolates historical data points without mapping specific underlying physics, which works when the underlying mechanisms are stable but fails when they change [Gemini]. This explains why his computing predictions (where Moore's Law provided a stable underlying mechanism for decades) succeeded while his biological predictions (where the underlying mechanisms are vastly more complex and less well-understood) failed.

The question of whether the Law of Accelerating Returns itself is slowing remains genuinely contested [Perplexity vs. Gemini/Grok]. Perplexity's observation that Moore's Law doubling periods are lengthening and that recent LLM improvements came primarily from inference compute rather than fundamental breakthroughs is the most specific evidence for a slowdown. If confirmed, this would undermine the foundation of Kurzweil's entire predictive framework for the 2029–2045 period.