The Law of Accelerating Returns (LOAR) is Ray Kurzweil's claim that information technologies follow exponential, often doubly-exponential, improvement curves, and that progress itself accelerates because each generation of technology enables the next. Kurzweil generalizes Moore's Law (transistor density doubling) into a cross-domain thesis spanning compute, storage, bandwidth, biotechnology (cost per genome base pair), AI, and more. He articulated it at book length in *The Singularity Is Near* (2005) and developed it earlier in The Age of Spiritual Machines (1999). LOAR is an intellectual ancestor of current transformative-AI timeline arguments and is contested by slow-diffusion and "AI as normal technology" counter-frames.
Statement
Kurzweil's formulation holds that information technologies follow exponential, often doubly-exponential, improvement curves, and that progress itself accelerates because each generation of technology enables the next (The Singularity Is Near — Kurzweil (2005)). The graph Kurzweil made widely known plots cross-paradigm computation cost (calculations per $1,000) as an uninterrupted exponential from 1900 to the present.
Core mechanisms
Kurzweil attributes the pattern to three mechanisms:
- Positive feedback — better tools lead to better science, which leads to better tools, shortening the doubling time.
- Paradigm cascading — when one technology's exponential curve plateaus, a new paradigm picks up the curve (vacuum tubes to transistors to integrated circuits to parallel computing to quantum).
- Cross-domain transfer — compute advances enable biotech advances, which enable materials advances, which enable further compute advances.
Timeline predictions
In the 2005 book Kurzweil set out specific dated milestones:
- A desktop computer matches human-brain compute in the 2010s.
- Human-level AI, including a passed Turing test, by 2029.
- The Singularity — the point at which $1,000 buys compute equal to all human brains combined — by 2045.
Kurzweil reaffirmed the 2029 estimate multiple times between 2014 and 2024 and pointed to 2024–2025 large-language-model benchmarks as vindication.
Track record as of 2026
Kurzweil's curves and predictions have held in some respects and not in others.
Among the elements that have broadly tracked the LOAR projection: compute doublings have remained roughly on schedule, though with documented slowdowns (the end of Dennard scaling in 2006, the end of single-thread Moore's Law around 2016, and a recovery in the GPU/TPU era); brain-compute equivalence milestones broadly arrived when Kurzweil predicted; LLM capabilities on benchmarks met or exceeded the 2005-era "2029 Turing test pass" prediction by 2024; and consumer AI adoption outpaced 2005-era expectations in some areas.
Among the elements that have not tracked: nanobots in the 2020s have not materialized; brain uploading by 2045 has no established scientific path; biological immortality remains off Kurzweil's curve despite real biotechnology progress; and the empirical dynamics of an intelligence explosion appear richer than Kurzweil's single smooth-exponential takeoff curve (see Three Types of Intelligence Explosion (Davidson, Hadshar, MacAskill)). Commentary on LOAR also notes that it functions more as an observed pattern than a law, holding conditional on specific scientific and economic circumstances that could change, including energy limits, data walls, and hardware physics.
Debates and positions
LOAR is an intellectual ancestor of much current transformative-AI trajectory writing.
Sources extending LOAR-style projections include Situational Awareness: The Decade Ahead ("The Decade Ahead"), framed as an explicit continuation; The Intelligence Age (Altman) and Abundant Intelligence (Altman), in which Sam Altman's framings echo LOAR logic; Machines of Loving Grace, in which Dario Amodei's "compressed 21st century" is described as a softened LOAR; The Industrial Explosion (Davidson, Hadshar) and Three Types of Intelligence Explosion (Davidson, Hadshar, MacAskill), in which Tom Davidson's intelligence-explosion taxonomy extends Kurzweilian curves; AI 2027, Daniel Kokotajlo and Scott Alexander's 2027 AGI / 2028 ASI scenario; Superintelligence: Paths, Dangers, Strategies — Bostrom (2014), which shares the trajectory while differing on the normative conclusion; and Schmidt's San Francisco Consensus essay, which identifies the pairing of scaling laws, shorter timelines, and abundant benefits as a Silicon Valley consensus.
Sources contesting LOAR include AI as Normal Technology, in which Arvind Narayanan and Sayash Kapoor reject LOAR-style extrapolation and argue that AI diffuses slowly like prior general-purpose technologies; AI Snake Oil — Narayanan and Kapoor (2024), the same authors' book-length critique; Why I Think AI Take-Off Is Relatively Slow (Cowen), Tyler Cowen's slow-takeoff framing; an empirical retrospective grading AI 2027's predictions (Source: blog.aifutures.org); and The Least Understood Driver of AI Progress, which argues that algorithmic progress is narrow and partially saturates.
Empirical tests
Several lines of work bear on whether the LOAR curve can continue: Epoch AI — Can AI Scaling Continue Through 2030? examines whether scaling curves can continue; Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? examines energy constraints on continuation; RAND — AI's Power Requirements Under Exponential Growth (2025) examines infrastructure constraints; AI Software Progress tracks algorithmic-efficiency trends (roughly 10× per year); and AI Benchmarks and Evaluation treats benchmark saturation as a proxy.
Relationships
- depends-on: The Singularity Is Near — Kurzweil (2005) — canonical source.
- supports: Scaling Laws — scaling laws are the LOAR-era formalization for neural networks.
- supports: Compressed 21st Century — Amodei softening.
- supports: AGI Timelines — concept page for timeline estimates broadly.
- supports: Recursive Self-Improvement (RSI) / Three Types of Intelligence Explosion (Davidson, Hadshar, MacAskill) — LOAR extrapolated into takeoff.
- contradicts: AI as Normal Technology / AI Snake Oil — Narayanan and Kapoor (2024) / Why I Think AI Take-Off Is Relatively Slow (Cowen) — rival framings.
- related: Situational Awareness: The Decade Ahead / AI 2027 / The Intelligence Age (Altman) / Abundant Intelligence (Altman) / Machines of Loving Grace — LOAR-descended current writings.
- related: (Source: blog.aifutures.org) / Situational Awareness: One-Year-Later Retrospectives — retrospectives.