Several converging observations motivate this analysis. First, consumer technologies smartphones, laptops, and increasingly automobiles exhibit strong convergence in form, capability, and usage patterns. Performance continues to improve, but largely within stable design envelopes, suggesting maturation rather than rapid conceptual diversification. Second, in several foundational scientific domains, progress increasingly appears cumulative and technically demanding, with fewer frequent, field-defining conceptual ruptures relative to earlier historical periods, despite substantial ongoing advances. Third, many prominent recent breakthroughs most visibly in artificial intelligence have been driven primarily by scale, data availability, and engineering optimization built on established theoretical foundations, rather than by the introduction of fundamentally new elementary theories.

 Taken together, these observations raise a policy-relevant question: are we experiencing a genuine plateau in discovery, or is the character of progress changing in ways that require different interpretation and institutional response? This article synthesizes economic, technical, and scientific evidence to address that question and proposes concrete institutional actions. The sections that follow summarize the strongest empirical findings on discovery dynamics and examine their implications for the future of computation and scientific innovation.

Empirical Evidence That Discovery Is Becoming More Resource-Intensive

A growing body of economic research suggests that ideas are, empirically, becoming harder to find. Bloom and colleagues show that across multiple sectors, increasingly large research teams and resource inputs are required to achieve levels of scientific and technological advance comparable to those of previous decades. Their analysis documents a rising input-per-output ratio in innovation, indicating declining productivity of traditional research investments across many domains.

 In computing, a similar pattern is visible in hardware progress. The long-standing pace described by Moore’s Law has slowed, and its physical and economic limits are increasingly apparent. While engineering advances in architecture, specialization, and energy efficiency continue to deliver meaningful gains, straightforward transistor-density scaling is no longer the dominant driver of visible performance improvements. Technical commentary and industry analyses document this transition and its implications for device- and system-level innovation.

 Machine learning provides a complementary illustration. Empirical scaling laws show that recent improvements in large language models and related architectures have followed predictable relationships with model size, data volume, and compute expenditure. At present, capability gains are largely achieved through scale and engineering refinement rather than through simple conceptual breakthroughs, even as these systems demonstrate striking practical utility. The transformer architecture and subsequent scaling work exemplify this dynamic.

 Finally, frontier scientific experimentation increasingly depends on high-cost, capital-intensive infrastructure. Large accelerators, advanced fusion facilities, and similar installations have enabled significant progress but also impose substantial financial and organizational barriers to entry. The National Ignition Facility’s recent results illustrate both the power of such investments and the degree to which modern discovery is shaped by large-scale engineering and long-term institutional commitment.

Mechanisms Shaping the Tempo and Visibility of Breakthroughs

Several interacting mechanisms help explain why contemporary progress may appear slower or less revolutionary.

 Exhaustion of low-hanging fruit. Early phases of scientific and technological development often resolve the most accessible problems first. Remaining questions tend to be conceptually subtler and technically more demanding, increasing the effort required per unit of discovery.

 Rising costs and capital intensity. Many modern research frontiers require substantial capital investment and long planning horizons, reorganizing how research agendas are set and which questions are pursued.

 Incentive and institutional effects. Funding allocation mechanisms, evaluation metrics, and corporate investment horizons often favor lower-risk, incremental projects that reliably yield measurable outputs. Empirical economic studies indicate that such incentives can shift research portfolios away from speculative, high-risk work with potentially transformative payoff.

 Scale-driven engineering gains. In several domains particularly machine learning rapid application-level progress has been achieved through scale and engineering optimization rather than through new core theories, accounting for both impressive practical results and persistent conceptual uncertainty.

 These mechanisms are not mutually exclusive. In practice, they interact to shape both the pace and the form of contemporary discovery, reinforcing the need to distinguish between incremental improvement within mature paradigms and the rarer emergence of genuinely new conceptual frameworks.

Contemporary Technical Landscape: What It Is and Why It Appears Incremental

 Moore’s Law and the changing driver of progress. What Moore’s Law is. Gordon E. Moore observed in 1965 that the number of components (transistors) that could be economically placed on an integrated circuit was increasing at an approximately exponential rate. This empirical observation later termed Moore’s Law became both a descriptive rule and an industry-wide coordination mechanism for performance expectations. The companion concept of Dennard scaling explained why power density and clock frequency could increase alongside transistor density for several decades. When Dennard scaling ended, transistor density alone no longer produced proportional gains in performance or energy efficiency.

 Historical context: switching elements through time. The historical sequence of switching technologies illustrates how paradigm shifts reconfigure the computational design space. Early electromechanical relays, used in telegraphy and telephone exchanges, were replaced by vacuum tubes to enable high-speed electronic switching. Vacuum tubes, in turn, gave way to semiconductor transistors in the mid-twentieth century, enabling integrated circuits and dramatic reductions in size, cost, and power consumption. Subsequent integration produced the microprocessor and mass consumer computing. Each transition opened architectural possibilities that were not merely “faster relays,” but qualitatively different substrates for computation. As transistor scaling slows, industry emphasis has shifted toward architectural specialization (GPUs, tensor accelerators), heterogeneous systems, and software–hardware co-design as the principal levers of progress highly productive, but less visually dramatic per generation than earlier substrate changes.

 AI as scale, architecture, and a tool for humans (responsible usage). Recent progress in machine learning, particularly since the transformer era, illustrates how architectural innovation combined with scale (data, compute, and optimization) can yield rapid growth in system capabilities. Foundational algorithmic ideas backpropagation, gradient-based optimization, and probabilistic modeling are decades old. Transformer architectures (introduced in 2017), together with empirically observed scaling laws, enabled the construction of substantially larger and more capable models. In practice, modern AI is best understood as a powerful, general-purpose set of tools that augment human judgment and productivity when used responsibly. Framing AI primarily as an autonomous cultural actor risks obscuring the engineering-tool nature of most real-world deployments. Consequently, AI should be treated as a technology that requires explicit governance structures, ethical frameworks, and responsible operational practices, in the same way as other transformative technologies. International policy efforts and ethics frameworks increasingly aim to make that stewardship systematic.

 Emerging substrates: photonics, quantum, and neuromorphic approaches. Photonics, quantum hardware, and neuromorphic substrates each offer domain-specific advantages: photonics for extremely high bandwidth and low-latency analog or optical signal processing; quantum platforms for particular classes of simulation and optimization problems; and neuromorphic devices for energy-efficient, sparse, event-driven computation. Recent experimental progress in all three areas is substantial, including integrated photonic accelerators and continued improvements in qubit coherence and fidelity. At present, however, these substrates function primarily as specialized computational tools rather than as fully general, unifying paradigms. The critical question is whether, and under what conditions, these substrates catalyze genuinely new theoretical abstractions about computation, or whether they remain primarily hardware-level performance multipliers.

Historical Precedents

 Historical episodes demonstrate a recurring pattern: conceptual reframing that resolves persistent anomalies often precedes broad and durable scientific revolutions. Three canonical examples are particularly instructive.

 Thomas Young and the wave theory of light. In the early nineteenth century, Thomas Young’s interference experiments produced clear fringe patterns that a purely corpuscular model of light could not naturally explain. Young’s work, together with later analytic elaborations by Fresnel and others, shifted the dominant conceptual account toward the wave theory of light. At the time, debates about the nature of light were active; many contemporaries favored the corpuscular tradition established by Newton. Young’s experimental results helped displace entrenched expectations about the completeness of existing optical theory. This episode illustrates how carefully designed experiments that expose anomalies can force reconsideration of foundational assumptions.

 Max Planck and quantization. Around 1900, unresolved discrepancies in black-body radiation spectra resisted explanation within classical electrodynamics and thermodynamics. Planck’s audacious hypothesis that energy exchange occurs in discrete quanta resolved the empirical regularities and initiated the development of quantum theory. Planck’s proposal was initially puzzling to many of his contemporaries, and its implications were not immediately understood. This episode illustrates how confronting persistent empirical anomalies can motivate entirely new conceptual machinery. While some late-nineteenth-century physicists believed their field was nearing completion, historians and philosophers of science have since nuanced this narrative, noting that significant anomalies were already recognized. Nevertheless, the broader lesson holds: anomalies, combined with a novel conceptual postulate, can reconfigure an entire domain.

 Non-Euclidean geometry and Riemannian reconceptualization. The nineteenth century saw the development of geometries not constrained by Euclid’s parallel postulate, producing self-consistent alternative systems developed by Lobachevsky and Bolyai. These advances culminated in Riemann’s formulation of differential geometry, which later supplied the mathematical language used by Einstein to formulate general relativity. This case demonstrates how mathematical reconceptualization can enable new physical theories and, eventually, new technologies. Collectively, these examples show that conceptual revolutions often arise from rethinking axioms, principles, or explanatory primitives, rather than merely building better instruments.

From Tools to Ideas: Why Instruments Alone Rarely Suffice

 Substrate innovations tend to produce three broad categories of outcomes:

1. Tool amplification. The new substrate accelerates tasks already meaningful under existing theory (for example, faster dense linear algebra).
2. Niche expansion. The substrate enables new but domain-limited capabilities (for example, optical interconnects or specific quantum simulations).
3. Conceptual rupture. More rarely, a substrate motivates or enables new abstractions that reframe computation itself.

 At present, most technological progress aligns with the first two outcomes. To trigger conceptual rupture, research efforts must deliberately pursue theoretical constructs that cannot be recovered by simply scaling existing ideas.

Pathways to Create New Computational Revolutions

Drawing on empirical work on discovery difficulty and the historical precedents discussed above, the following practical strategies are intended to increase the probability of conceptual breakthroughs in computing.

 Prioritize concept-driven, long-horizon funding. Funding agencies should support multi-year programs evaluated primarily on conceptual ambition and designed to tolerate null or negative results, thereby reducing short-term optimization pressures.

 Invest in hybrid theory–experiment centers. The co-location of theoretical computer scientists, physicists, mathematicians, and systems engineers can accelerate cross-disciplinary abstraction and idea transfer.

 Build shared open infrastructure for emergent substrates. Shared photonic foundries, public quantum testbeds, and neuromorphic prototyping platforms can lower capital barriers and broaden the community exploring new abstractions.

 Support conceptual prizes and carefully framed challenge problems. Prize mechanisms should be designed to reward novel abstractions and explanatory frameworks rather than incremental performance improvements alone.

 Reform evaluation metrics. Academic and industrial evaluation systems should explicitly recognize theoretical novelty and enabling artifacts such as frameworks, formalisms, and proof concepts alongside benchmark-driven performance gains.

 Nurture interdisciplinary training that privileges abstraction. Graduate curricula that integrate mathematics, theoretical physics, and systems design can cultivate researchers predisposed to invent new computational formalisms.

 Cultivate ethical, integrative research cultures. Historical precedents from medieval Islamic philosophers like Nasir al-Din Tusi and Avicenna (Ibn Sīnā), whose ethical writings were embedded within scholarly norms, to modern institutions like Bell Labs demonstrate that enduring contributions arise from cultures that value integrity, stewardship, and intellectual risk-taking over short-term metrics. Institutional policies must actively protect exploratory scholarship and mentor researchers in responsible, open scientific practice. A revolution's impact is measured not only by its power but by the wisdom of its application.

Conclusion: Cautious, Constructive Optimism

 Computing today is technically rich: engineering improvements and substrate innovation continue to yield high-value outcomes. Yet many advances appear incremental because prevailing paradigms have matured. Historical precedents suggest that paradigm shifts follow conceptual reframing that resolves anomalies and opens new classes of inquiry. While it is impossible to guarantee when the next conceptual rupture will occur, policy choices and institutional design can increase its likelihood. Longer-horizon funding, hybrid theory–experiment centers, shared infrastructure, prizes for conceptual novelty, reformed evaluation metrics, interdisciplinary training, and a stronger ethical research culture all make conceptual breakthroughs more probable.

 This is pragmatic optimism. The next revolution in computing need not be miraculous; it can emerge from deliberate structural changes that reward abstraction, protect intellectual risk-taking, and connect new substrates to novel theoretical questions. Under these conditions, the steady and rigorous work being conducted today can plausibly seed the conceptual revolutions of the coming decades.

References

 Kuhn, Thomas S. The Structure of Scientific Revolutions. University of Chicago Press, 1962.

 Moore, G. E. “Cramming More Components onto Integrated Circuits.” Electronics (reprinted), 1965.

 Bloom, Nicholas; Jones, Charles I.; Van Reenen, John; Webb, Michael. Are Ideas Getting Harder to Find? NBER Working Paper No. w23782, 2017.

 Vaswani, A.; Shazeer, N.; Parmar, N.; et al. “Attention Is All You Need.” Advances in Neural Information Processing Systems 30, 2017.

 “Young’s experiment (double-slit).” Encyclopædia Britannica.

 “Planck’s radiation law.” Encyclopædia Britannica.

 “Non-Euclidean geometry.” Encyclopædia Britannica.

 Integrated photonics and computational accelerators. Nature reviews and technical articles.

 Quantum Index Report 2025. MIT.

 National Academies of Sciences, Engineering, and Medicine. On Being a Scientist: A Guide to Responsible Conduct in Research. 4th ed., National Academies Press, 2024.

 Tusi, Nasir al-Din. Akhlaq-i Nasiri (The Nasirean Ethics). Routledge / classical editions.

Chinese Translation / 中文翻译：

我们是否在等待下一个计算范式？

——一次聚焦技术范式的成熟度、历史先例与催生概念革命机制的探讨

多项趋同的观察促使了本项分析的展开。首先，消费类技术——智能手机、笔记本电脑乃至日益增多的汽车——在形式、功能和使用模式上呈现出显著的趋同态势。性能持续提升，但主要是在稳定的设计框架内进行，这表明技术正趋于成熟，而非快速的概念分化。其次，在若干基础科学领域，尽管持续取得重大进展，但相对于早期历史阶段，进步日益显现为累积性和技术密集型特征，能够重新定义学科范畴的概念性突破出现频率有所降低。第三，近年来许多引人注目的突破（尤其在人工智能领域）主要驱动力来自规模、数据可得性以及建立在现有理论基础上的工程优化，而非源自根本性的全新基础理论。

综合来看，这些观察引出了一个具有政策相关性的问题：我们是否正经历真正的发现停滞期，抑或进步的性质正在发生改变，从而需要不同的解读和制度性应对？本文综合经济、技术和科学证据来探讨此问题，并提出具体的制度性行动建议。后续章节将总结关于发现动态的最有力实证发现，并审视其对计算和科学创新未来的影响。

发现正变得愈加依赖资源的实证证据

越来越多的经济学研究表明，从实证角度看，想法正变得越来越难被发现。布鲁姆及其同事表明，在多个领域，要取得与过去几十年相当水平的科技进步，所需的研究团队规模和资源投入正变得越来越大。他们的分析记录了创新活动中投入产出比的上升，表明在许多领域，传统研发投资的生产率正在下降。

在计算领域，硬件进步中可见类似模式。摩尔定律所描述的长期发展步伐已经放缓，其物理和经济极限日益显现。尽管在架构、专业化和能效方面的工程进步持续带来显著收益，但单纯的晶体管密度缩放已不再是可见性能提升的主要驱动力。技术评论和行业分析记录了这一转变及其对设备和系统层面创新的影响。

机器学习提供了一个互补的例证。实证缩放定律表明，近期大型语言模型及相关架构的改进，遵循着与模型规模、数据量和计算投入可预测的关系。目前，能力提升主要是通过规模和工程优化实现，而非通过简单的概念突破，即使这些系统展现出惊人的实际效用。Transformer架构及后续的缩放工作正是这种动态的体现。

最后，前沿科学实验日益依赖于高成本、资本密集的基础设施。大型加速器、先进聚变装置及类似设施推动了重大进展，但也设置了巨大的财务和组织准入门槛。美国国家点火装置近期的成果既彰显了此类投资的威力，也揭示了现代发现如何受到大规模工程和长期制度承诺的深刻塑造。

影响突破速度与可见性的机制

几种相互作用的机制有助于解释为何当代进步可能显得更为缓慢或缺乏革命性。

• 低垂果实的耗尽：科学技术发展的早期阶段通常会先解决最容易触及的问题。遗留问题往往在概念上更精微、技术上要求更高，从而增加了每单位发现所需的努力。

• 成本与资本密集度的上升：许多现代研究前沿需要大量的资本投入和长期的规划视野，这改变了研究议程的设置方式及所探索的问题。

• 激励与制度效应：经费分配机制、评估指标和企业投资视野往往倾向于低风险、渐进式的项目，这些项目能可靠地产生可衡量的产出。实证经济研究表明，此类激励可能使研究组合偏离具有潜在变革性回报的、投机性的高风险工作。

• 规模驱动的工程收益：在若干领域（尤其是机器学习），应用层面的快速进步是通过规模和工程优化实现的，而非通过新的核心理论，这既解释了令人印象深刻的实际成果，也解释了持续存在的概念不确定性。

这些机制并非互斥。在实践中，它们相互作用，共同塑造了当代发现的速度与形式，这强化了一种需求：需要区分成熟范式内的渐进式改进与真正全新概念框架的罕见出现。

当代技术格局：现状及其何以呈现渐进性

• 摩尔定律与进步驱动力的转变：戈登·E·摩尔在1965年观察到，能够经济地放置在集成电路上的组件（晶体管）数量正以近似指数速率增长。这一后来被称为摩尔定律的实证观察，既成为描述性规则，也成为行业协调性能预期的机制。伴随的登纳德缩放定律解释了为何在数十年间，功率密度和时钟频率能够随晶体管密度一同提升。当登纳德缩放结束时，仅靠晶体管密度已无法带来性能和能效的成比例增益。

• 历史背景：随时间演进的开关元件：开关技术的历史序列说明了范式转移如何重构计算设计空间。早期用于电报和电话交换的机电继电器，被真空管取代以实现高速电子开关。真空管又在二十世纪中叶让位于半导体晶体管，从而实现了集成电路以及尺寸、成本和功耗的急剧降低。后续的集成催生了微处理器和大众消费计算。每一次转变都开辟了新的架构可能性，这些可能性不仅仅是“更快的继电器”，而是性质不同的计算载体。随着晶体管缩放速度放缓，行业重点已转向架构专业化（GPU、张量加速器）、异构系统以及软硬件协同设计，将其作为进步的主要杠杆——这些手段虽然高效，但每代产品带来的视觉冲击力不如早期的载体变革。

• 人工智能：作为规模、架构与人类工具（负责任使用）：机器学习（尤其是Transformer时代以来）的近期进展，展示了架构创新与规模（数据、计算、优化）相结合如何能推动系统能力的快速增长。基础算法思想——反向传播、基于梯度的优化、概率建模——已有数十年历史。Transformer架构（2017年提出）与实证观察到的缩放定律相结合，使得构建更大、能力更强的模型成为可能。实际上，现代人工智能最好被理解为一套强大、通用的工具集，在负责任使用的前提下能够增强人类的判断力和生产力。若主要将AI框定为自主的文化行动者，则可能掩盖大多数现实世界部署所具有的工程工具本质。因此，应像对待其他变革性技术一样，将AI视为需要明确治理结构、伦理框架和负责任操作实践的技术。国际政策努力和伦理框架日益旨在使这种管理走向系统化。

• 新兴载体：光子学、量子与神经形态方法：光子学、量子硬件和神经形态载体各自提供特定领域的优势：光子学用于极高带宽和低延迟的模拟或光信号处理；量子平台用于特定类别的模拟和优化问题；神经形态设备用于高能效、稀疏、事件驱动的计算。这三个领域近期的实验进展均很显著，包括集成光子加速器以及量子比特相干性和保真度的持续提升。然而，目前这些载体主要作为专门的计算工具，而非完全通用、统一的范式。关键问题在于，这些载体是否以及在何种条件下能够催生关于计算的真正新的理论抽象，还是它们主要作为硬件层面的性能倍增器存在。

历史先例

历史案例展示了一种反复出现的模式：解决持续性反常现象的概念重构，常常发生在广泛而持久的科学革命之前。三个经典例子尤其具有启发性。

• 托马斯·杨与光的波动说：十九世纪初，托马斯·杨的干涉实验产生了清晰的光条纹图样，这是纯粹的光微粒模型无法自然解释的。杨的工作，以及后来菲涅耳等人的分析阐述，将主导性的概念解释转向了光的波动说。当时，关于光本质的争论十分活跃；许多同时代人偏爱牛顿建立的微粒说传统。杨的实验结果帮助动摇了关于现有光学理论完备性的固有观念。这一案例说明，精心设计、能够揭示反常的实验如何能够迫使人们重新思考基础假设。

• 马克斯·普朗克与量子化：大约在1900年，黑体辐射光谱中未解决的差异无法在经典电动力学和热力学框架内得到解释。普朗克提出的能量交换以分立量子形式发生的激进假说，吻合了实证规律并开启了量子理论的发展。普朗克的提议最初令许多同时代人感到困惑，其含义也未被立即理解。这一案例说明，面对持续存在的实证反常如何能够激励全新的概念工具。虽然十九世纪末一些物理学家认为他们的领域接近完成，但科学史家和科学哲学家后来细化了这一叙事，指出当时已经认识到显著的反常。尽管如此，更广泛的启示依然成立：反常与新颖的概念假设相结合，能够重构整个领域。

• 非欧几何与黎曼重构：十九世纪见证了不受欧几里得平行公设约束的几何学的发展，产生了由罗巴切夫斯基和波尔约建立的、自洽的替代系统。这些进展在黎曼微分几何的表述中达到顶峰，后者后来为爱因斯坦表述广义相对论提供了数学语言。此案例表明，数学重构如何能够催生新的物理理论，并最终催生新技术。总而言之，这些例子表明，概念革命往往源自对公理、原则或解释基元的重新思考，而不仅仅是制造更好的仪器。

从工具到思想：为何仅有工具鲜少足够

载体创新往往产生三类广泛的结果：

• 工具放大：新载体加速了现有理论下已有意义的任务（例如，更快的密集线性代数运算）。

• 领域拓展：新载体实现了新的、但受领域限制的能力（例如，光互连或特定的量子模拟）。

• 概念断裂：更为罕见的是，新载体激发或促成了新的抽象，从而重构计算本身。

目前，大多数技术进步属于前两类结果。要引发概念断裂，研究工作必须有意识地追求那些无法通过简单扩展现有想法而获得的的理论建构。

催生新计算革命的路径

基于关于发现难度的实证研究和上述历史先例，以下实践策略旨在提高计算领域概念突破的可能性。

• 优先支持概念驱动、长周期的资助：资助机构应支持以概念抱负为主要评估标准、旨在容忍零结果或负面结果的多年期项目，从而减少短期优化压力。

• 投资理论与实验结合的混合中心：将理论计算机科学家、物理学家、数学家、系统工程师聚集一处，可加速跨学科抽象和思想转移。

• 为新兴载体构建共享开放基础设施：共享光子学代工厂、公共量子测试平台、神经形态原型设计平台可以降低资本门槛，拓宽探索新抽象概念的社群。

• 支持概念性奖项和精心设计的挑战性问题：奖励机制应设计为奖励新颖的抽象和解释框架，而不仅仅是渐进式的性能改进。

• 改革评估指标：学术和工业评估体系应明确认可理论新颖性，以及框架、形式体系、证明概念等赋能性成果，与基于基准测试的性能提升同等对待。

• 培育重视抽象的跨学科训练：融合数学、理论物理学和系统设计的研究生课程，可以培养出倾向于创造新计算形式体系的研究者。

• 培养注重伦理、融合的研究文化：从中世纪伊斯兰哲学家如纳西尔丁·图西和阿维森纳（其伦理著作嵌入学术规范之中），到现代机构如贝尔实验室，历史先例表明，持久的贡献源于那些珍视诚信、责任和智力冒险精神、而非短期指标的文化。制度政策必须积极保护探索性学术，并指导研究者进行负责任、开放的科学实践。一场革命的影响力不仅取决于其力量，更取决于其应用的智慧。

结论：审慎而建设性的乐观主义

当今的计算领域技术丰富：工程改进和载体创新持续产生高价值成果。然而，许多进步显得渐进，因为主流范式已然成熟。历史先例表明，范式转移遵循着能够解决反常并开辟新研究类别的概念重构。虽然无法保证下一次概念断裂何时发生，但政策选择和制度设计可以增加其可能性。长周期资助、理论与实验结合的混合中心、共享基础设施、概念新颖性奖项、改革后的评估指标、跨学科训练以及更强的伦理研究文化，都能使概念突破更有可能发生。

这是一种务实的乐观主义。计算领域的下一次革命无需奇迹；它可以源于有意识的结构性变革，这些变革奖励抽象思维，保护智力冒险，并将新载体与新颖的理论问题联系起来。在此条件下，当今正在进行的稳健而严谨的工作，很可能为未来几十年的概念革命播下种子。

参考文献

（此处略去英文参考文献列表，内容与原文一致）