Diminishing Incremental Gains¶
Core Idea¶
Diminishing incremental gains is the general structural pattern that each successive unit of a contributing input or effort produces a smaller increase in output, benefit, or value than the one preceding it, once some threshold is exceeded, so that the relationship between input and output is concave over the relevant range rather than linear or accelerating. The essential commitment is broader than any particular technical law: it is a domain- general heuristic claim about the structure of many real-world input-output relationships — learning curves, utility from consumption, health benefits of exercise, returns to effort, accuracy of measurement, security from redundancy — that saturate, taper, or asymptote as more of the input is applied. Every diminishing-gains articulation specifies (1) the input variable — effort, time, money, practice hours, feature complexity, redundant copies, study time; (2) the output variable — skill, satisfaction, reliability, security, accuracy, conversion; (3) the functional shape — typically concave, often logarithmic, power-law with exponent < 1, exponential approach to an asymptote, or a sigmoid past its inflection; and (4) the regime of applicability — after an initial linear or accelerating phase (ramp-up), diminishing returns set in; at very high input levels, some systems experience actual decline (not merely diminishing gains) due to overtraining, satiation, or diminishing returns turning negative. The construct pre-dates formal economics as a practical wisdom ("past a point, more is not better") and was formalized in economics (law of diminishing marginal returns, diminishing marginal utility), in learning theory (power law of practice), in engineering (redundancy-reliability trade- offs), and in public health (dose-response curves), among many other domains.
How would you explain it like I'm…
Less Wow for Each Try
Each Extra Helps Less
Concave Input-Output Curves
Structural Signature¶
A diminishing-gains relationship between input x and output y is characterized by a concave function f(x) with f'(x) > 0 (still increasing) and f''(x) < 0 (but at a decreasing rate). Common functional forms include y = A log(1 + Bx) (logarithmic), y = A x^α with 0 < α < 1 (power law, concave), y = A(1 − exp(−Bx)) (exponential approach to saturation), and the upper (concave) portion of a sigmoid. The marginal benefit dy/dx decreases monotonically with x. At saturation, additional units produce negligible gain. In regions where the relationship turns negative (e.g., overtraining reducing performance), the construct extends to a full inverted-U or yields to its own structural tensions.
What It Is Not¶
Common misclassification: Treating diminishing incremental gains as identical to the economic law of diminishing returns. They are closely related but distinct in scope: see diminishing_returns — the economic law is a formal microeconomic proposition about production functions in which one input is varied while others are held fixed, with specific technical content and empirical domain. Diminishing incremental gains is the broader generalized heuristic applicable to effort, learning, utility, and many other settings where the structural pattern appears without the full production-function apparatus.
Not identical to marginal utility or marginal analysis in general: see marginal_utility — marginal utility is a specific concept in consumer theory (the additional utility from one more unit of consumption), often but not always exhibiting diminishing marginal utility. Marginal analysis as a method is broader still. Diminishing incremental gains is the structural pattern; marginal utility is one theoretical framework in which the pattern commonly appears.
Not synonymous with "saturation" or "plateau": saturation means the output ceases to increase at all; diminishing gains means each increment is smaller but still positive (until or unless saturation is reached). A system can exhibit diminishing gains without reaching saturation within the observed range.
Not universal: many relationships are linear, convex, or S-shaped. Early regions of many processes show increasing returns (learning-by-doing, network effects); only past a threshold do they become concave. Assuming diminishing gains universally misleads in these cases.
Not always monotone: in some settings (overtraining, overconsumption, excessive redundancy cost), returns do not merely diminish but become negative. "Diminishing gains" is a concave-increasing shape; "inverted-U" or "too-much-of-a-good-thing" is a distinct pattern where the output actually falls at high input levels.
Not universally intuitive about magnitude: even when the shape is correctly perceived, the severity of diminishing gains (how quickly marginal benefit declines) is often underestimated or overestimated, leading to systematic over- or under-investment.
Cross-references: see diminishing_returns (the formal microeconomic law — tight pair); see marginal_utility (the consumer-theory sibling); see concavity (the mathematical shape); see optimization (where concavity simplifies analysis); see saturation (the limiting case).
Broad Use¶
Diminishing incremental gains appears in economics (consumer utility, production, marketing spend vs conversions, advertising response), in learning and skill acquisition (power law of practice — Newell & Rosenbloom 1981, though challenged by Heathcote, Brown & Mewhort 2000 as often exponential), in statistics (sample size vs precision — √n behavior), in engineering (reliability vs redundancy, performance vs resource consumption), in public health (dose-response curves, vaccination coverage vs herd immunity), in information retrieval (precision vs number of retrieved items), in cost-benefit analysis (effectiveness of incremental investment), in software engineering (feature complexity vs user value), and by analogy in nearly any domain where incremental effort yields progressively smaller gains. It is a cognitive-level heuristic that shapes decision-making about when to stop adding resources.
Clarity¶
Diminishing incremental gains clarifies a ubiquitous structural fact about resource allocation: most beneficial activities have a "good-enough" point past which additional investment is wasteful, and rational resource use requires identifying that point. It explains why "more is not always better," why diminishing returns set in even for valuable activities, and why balanced portfolios (across several diminishing-gains activities) typically outperform concentrated investment in any one.
Manages Complexity¶
The construct manages complexity by collapsing many specific diminishing-returns phenomena (learning, exercise, advertising, redundancy, study time) into a single heuristic pattern whose implications transfer across domains: allocate resources where marginal return is highest, stop adding when marginal return falls below opportunity cost, balance across activities to equalize marginal returns. This reduces decision-making to a comparison of marginal rates across options.
Abstract Reasoning¶
Diminishing-gains reasoning proceeds by identifying the input-output relationship, estimating or measuring its functional shape (concave, linear, convex, sigmoid), locating the inflection or diminishing-gains regime, and allocating resources to equalize marginal returns across competing uses. It licenses cost-benefit stopping rules, portfolio diversification arguments (against excessive concentration), and the heuristic of "incremental investment should cease when marginal return falls below alternatives." It also supports diagnostic arguments: when observed relationships are surprisingly linear or accelerating, that is itself informative (network effects, increasing returns, tipping points).
Knowledge Transfer¶
| Role | Learning form | Health form | Marketing form | Engineering form |
|---|---|---|---|---|
| Input | Hours of practice | Dose of medication / servings | Marketing spend | Number of redundant components |
| Output | Skill / performance | Health outcome | Conversions / reach | Reliability |
| Shape | Power law or exponential approach to asymptote | Sigmoid, often with plateau | Concave (often log-shaped) | Concave approach to asymptotic reliability |
| Key caveat | Deliberate-practice component matters; ceiling effect | Toxicity / overdose reverses gains | Saturation + wear-out effects | Cost scaling with redundancy |
| Practical rule | Stop when marginal learning < alternative | Dose to target response, avoid toxicity | Spend to equal marginal return | Add redundancy while cost per reliability gain is favorable |
A learning-theorist's diminishing-gains reasoning transfers to health (dose-response), to advertising (saturation curves), to engineering (redundancy trade-offs), and to personal productivity ("after a few pomodoros, the extra is not worth as much"). The structural core is the same concave shape; what varies is the substrate and the parameter values.
Example¶
Formal case — sample size and the precision of a sample mean: For a sample of size n drawn from a distribution with variance σ^2, the standard error of the sample mean is σ/√n. To halve the standard error requires quadrupling n; to halve it again requires sixteen-folding the original n. Each additional observation provides progressively less improvement in precision. This is a formal, quantitative instance of diminishing incremental gains: the precision function f(n) ∝ 1/√n is concave in n (after a very small initial region), and the marginal gain per additional observation drops toward zero. Cost-benefit analysis of data collection balances the cost of additional observations against the diminishing precision gain.
Structurally-faithful non-formal case — hours of studying before an exam: A student studying for an exam typically finds that the first few hours of focused review deliver substantial score improvement, later hours yield smaller improvements, and very late hours (especially at the expense of sleep) may produce near-zero or negative net gain (the inverted-U in extreme). The underlying shape — not formally a production function — is structurally faithful to diminishing incremental gains: concave input-output relationship over the productive range, with practical stopping rules emerging from marginal-gain reasoning. Students (and managers, and investors) who correctly intuit the shape make better allocation decisions than those who don't.
Structural Tensions and Failure Modes¶
T1: Initial Regime May Be Linear or Accelerating. Many processes show increasing returns in their early stages (learning-by-doing, network effects, economies of scale) before diminishing returns set in. Assuming diminishing gains universally misses this early phase and produces under-investment when returns would still be rising. Failure mode: a project is abandoned at an early stage because "we've plateaued," when in fact the current regime is still ramping up toward higher productivity, as Wright (1936) first quantified for airframe production where unit cost fell predictably with cumulative output before tapering. [1] The tension arises between the general principle's applicability (diminishing gains do apply widely) and its contingency on regime (early-phase processes often violate it).
T2: Over-Application Turns Gains Negative. Past the saturation point, many systems exhibit actual decline (overtraining, overdose, audience fatigue, excessive redundancy cost). Diminishing gains flattens toward zero; the inverted-U pattern goes below zero. Failure mode: diminishing-gains reasoning is applied but the real relationship is inverted-U, so investment is continued past the optimum into the declining region, a dynamic Banister (1991) modeled in elite athletes whose performance peaks at intermediate training loads and degrades under chronic overload. [2] This creates a critical distinction: recognizing where gains diminish (still positive) versus where they reverse (become negative).
T3: Shape Varies Across Domains and Contexts. The specific functional form (logarithmic vs power law vs exponential) and the threshold where diminishing gains set in vary widely. The "power law of practice" (Newell & Rosenbloom 1981) has been challenged as often being better fit by exponential functions (Heathcote, Brown & Mewhort 2000). Failure mode: a specific functional form (e.g., square-root, log, power) is assumed without empirical validation, producing biased stopping-point estimates or extrapolation errors, a hazard Anderson (1982) flagged in his ACT-based account of skill acquisition where production-rule compilation produces concave but not strictly power-law gains. [3] The mathematical shape is not universal; applying the wrong functional form creates systematic prediction errors.
T4: Marginal Gain Is Often Compared to the Wrong Benchmark. The key decision is whether the marginal gain from continuing exceeds the marginal gain from an alternative use of resources. Even small remaining marginal gains can justify continuation if no better alternative exists; conversely, sizeable marginal gains may not justify continuation if alternatives are better. Failure mode: marginal gains are compared to zero ("still yielding something, so continue") rather than to opportunity cost, producing over-investment in weakly productive activities, the very textbook error Mankiw (2014) anchors his treatment of rational choice on—rational decision-makers compare marginal benefit to marginal opportunity cost, not to zero. [4] The principle requires implicit portfolio reasoning—allocating across competing opportunities simultaneously—which is cognitively demanding.
T5: Observable Gain Lags Behind Actual Gain Due to Measurement. The relationship between measured input and measured output often appears less concave than the underlying causal relationship because of measurement noise, delayed effects, and confounding variables. A learning intervention might produce delayed cognitive effects that are invisible in immediate measurement, making the concave relationship appear nearly linear in short-term observation. Failure mode: the true diminishing-gains structure is obscured by measurement issues, leading to continuation of investments that appear productive in the short term but have underlying diminishing marginal effects masked by noise—a pitfall Ericsson, Krampe, and Tesch-Römer (1993) addressed by demanding decade-long longitudinal records of deliberate practice rather than short-window snapshots. [5]
T6: Systemic Effects Create Thresholds That Violate Local Diminishing-Gains Reasoning. Some systems exhibit threshold effects or phase transitions where incremental investment produces near-zero gains until a critical point is crossed, then produces large jumps. Network effects approaching critical mass, phase changes in physical systems, and tipping points in social systems all exhibit this pattern. Local diminishing-gains reasoning, applied at any point before the threshold, recommends stopping investment exactly when it should be accelerated to reach the threshold. Failure mode: incremental analysis fails at systems with tipping-point dynamics; investment is abandoned in the flat phase before threshold, missing the opportunity for large gains beyond it—the very S-curve dynamic Rogers (2003) documents in Diffusion of Innovations, where adoption appears stagnant until critical mass triggers explosive uptake. [6]
Structural–Framed Character¶
Diminishing Incremental Gains sits at the structural end of the structural–framed spectrum: it is a pure relational pattern, the same in any domain where it appears, and nothing about its meaning depends on a particular field's vocabulary or assumptions. It is the pattern in which each additional unit of input adds less to the output than the unit before it, so the input–output relationship is concave over the relevant range.
The same concave shape applies unchanged wherever it shows up — effort and skill, fertilizer and yield, study time and recall — because it is a claim about the curvature of a relation, not about any particular subject matter. It carries no evaluative weight; concavity is simply a shape, neither good nor bad. Its origin is formal, captured by a function with positive but decreasing slope, it can be defined with no reference to human practices, and to use it is to recognize a curvature already present in a relationship rather than to import an outside perspective. On every diagnostic, it reads structural.
Substrate Independence¶
Diminishing Incremental Gains is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. The pattern is mathematically universal — a concave input-output relationship where f''(x) < 0 — and the logarithmic and power-law shapes it describes show up in economics, learning curves, medicine, ecology, and cognition alike, with no home-domain dependence. What holds it just below the top is the evidence rather than the structure: the entry asserts this sweep without enough worked examples to fully demonstrate the transfer. It is anchor-adjacent, genuinely cross-substrate in form, simply under-illustrated in practice.
- Composite substrate independence — 4 / 5
- Domain breadth — 5 / 5
- Structural abstraction — 5 / 5
- Transfer evidence — 2 / 5
Relationships to Other Primes¶
Parents (2) — more general patterns this builds on
-
Diminishing Incremental Gains is a kind of Trade-offs
Diminishing incremental gains says successive units of an input produce smaller increments of output once a threshold is passed, so the marginal cost of each next unit of benefit rises along the concave curve. Continuing to improve on the output dimension therefore requires giving up more on the input side — time, money, effort — than for any prior unit. That is the structure of a Trade-off, here driven by concavity in the input-output relationship rather than fixed-frontier opposition between two outputs.
-
Diminishing Incremental Gains presupposes Nonlinearity
Diminishing incremental gains presupposes nonlinearity because its structural claim, that each successive unit of input produces less output than the one before, is precisely a statement of concavity: a nonlinear input-output relation in which scaling does not preserve proportionality. Nonlinearity supplies the general space of relationships in which superposition fails and qualitative features like thresholds and saturation become possible; diminishing gains picks out the concave-saturating subfamily of that space as a domain-general heuristic across learning, utility, exercise, and effort.
Children (1) — more specific cases that build on this
-
Diminishing Returns (Law of) is a kind of Diminishing Incremental Gains
The law of diminishing returns is a specialization of diminishing incremental gains restricted to the technological setting of a production process with at least one fixed factor, where the marginal product of a variable input eventually declines. It inherits the general concave input-output pattern that each successive unit yields a smaller increment, and specializes by fixing the domain to production economics, the cause to the fixed-factor constraint, and the variables to marginal, average, and total products. The broader heuristic of diminishing gains covers learning curves, utility, and effort; the law isolates the production-function instance with formal microeconomic predictions.
Path to root: Diminishing Incremental Gains → Nonlinearity
Neighborhood in Abstraction Space¶
Diminishing Incremental Gains sits in a sparse region of abstraction space (73rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely rather than landing on a neighbor.
Family — Scaling Laws & Nonlinearity (5 primes)
Nearest neighbors
- Diminishing Returns (Law of) — 0.81
- Linearity — 0.78
- Receptor Saturation — 0.76
- Increasing Returns — 0.76
- Complexity (Time/Space) — 0.76
Computed from structural-signature embeddings · 2026-05-29
Not to Be Confused With¶
Diminishing Incremental Gains must be distinguished from Diminishing Returns (similarity 0.772), its closest neighbor, despite the surface similarity in names. The distinction is subtle but load-bearing. Diminishing Incremental Gains is a structural pattern: the observation that successive units of input produce progressively smaller absolute increases in output. If effort hours 1–10 increase productivity by 100 units total, and hours 11–20 increase it by 60 more units, then hours 21–30 might increase it by 40 more units — a diminishing-gains pattern. Diminishing Returns, by contrast, is a formal microeconomic law: it states that when one input to production (labor, for example) is increased while other inputs (capital, land) are held constant, the marginal product of that input decreases. Diminishing Returns is embedded in production-function theory, applies specifically to producers making allocation decisions, and has specific technical content about how to measure "output per unit of input." Diminishing Gains is the broader heuristic pattern; Diminishing Returns is a specialized economic formalization. Many diminishing-gains phenomena (learning, exercise benefits, advertising effectiveness) are not production functions at all and do not fit neatly into the Diminishing Returns framework. Conversely, all Diminishing Returns relationships exhibit diminishing gains, but not all diminishing-gains patterns are instances of Diminishing Returns. The distinction matters because Diminishing Returns has specific policy and optimization implications in production contexts, while Diminishing Gains is applicable much more broadly to any input-output relationship with a concave shape.
Nor is Diminishing Incremental Gains the same as Marginal Analysis, though they are deeply related. Marginal Analysis is a mathematical technique: the practice of studying how output changes when input changes by a small amount, computing derivatives (dy/dx), and using those derivatives to optimize or understand system behavior. Marginal Analysis is a tool in the mathematician's and economist's toolkit; it does not commit to any particular functional form or relationship. Diminishing Incremental Gains, by contrast, is a specific claim about relationship shape: it states that the second derivative is negative (f''(x) < 0), that the first derivative decreases with x (f'(x) is decreasing), and that output remains increasing but at a decelerating rate. Marginal Analysis can be applied to any differentiable function — linear, convex, concave, sigmoid — and will reveal the shape of the relationship; Diminishing Gains claims a specific shape (concave-increasing). Marginal Analysis is the analytical method; Diminishing Gains is one of several patterns that Marginal Analysis can reveal. A practitioner using Marginal Analysis might discover that a relationship is linear (constant marginal returns), convex (increasing marginal returns), or concave (diminishing marginal returns); Diminishing Gains is the name for the concave case. They are complementary: Marginal Analysis is the tool that measures diminishing gains when they are present.
Diminishing Incremental Gains is also distinct from Saturation and Plateau, though these are often confused. Saturation means the output has stopped increasing at all — the function has flattened to a horizontal asymptote, and marginal returns are zero. Plateau describes a region where output is roughly constant (flat slope). Diminishing Gains means the output continues to increase but at a decelerating rate: each additional unit still produces a positive benefit, just less than the previous unit did. A learning curve in the diminishing-gains phase still shows performance improvement with more practice, but slower improvement than earlier. Once learning plateaus (no more improvement regardless of practice), the learner has reached saturation or plateau. These are different stages on the same input-output journey: ramp-up (accelerating gains), then diminishing gains (decelerating but still positive), then plateau or saturation (no further gains). Confusing them leads to premature stopping (you quit during diminishing gains when you could still improve, albeit slowly) or wasteful persistence (you keep adding input long past the point where further gains are negligible). The terms are often used loosely in colloquial speech — "diminishing returns" and "plateau" are sometimes treated as synonyms — but Diminishing Incremental Gains specifies a mathematical shape that is distinct from saturation.
Finally, Diminishing Incremental Gains should be distinguished from Increasing Returns or Network Effects, which represent the opposite structural pattern. Increasing Returns means that successive units of input produce larger absolute increases in output: hour 1 of study increases knowledge by 5 units, hour 2 increases it by 6, hour 3 by 7. This concave-upward or convex shape (f''(x) > 0) appears in learning-by-doing contexts, network effects (each new user adds more value than the last as the network grows), and positive feedback loops. Increasing Returns can create tipping points and path dependence; they suggest the opposite optimization strategy from diminishing gains (concentrate investment early, reap accelerating benefits). The distinction is mathematically crisp but practically important: misdiagnosing whether a system exhibits increasing or diminishing returns can lead to reversed resource-allocation strategies.
Solution Archetypes¶
Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.
Built directly on this prime (1)
Also a related prime in 6 archetypes
- Approximation-Target Divergence Mapping
- Iterative Refinement Loop
- Marginal Stop Rule
- Plateau Detection and Switching
- Satiation-Aware Allocation
- Tapering Strategy
Notes¶
Held at High confidence. Generalized heuristic pattern pervasive across domains; tight-pair-flagged with diminishing_returns (#137, the formal microeconomic law) to highlight that the two constructs are structurally related but differ in scope — this v2 entry is the broader generalized heuristic, while #137 is the formal production- function-based microeconomic law. Marginal utility (#136) is the adjacent consumer- theory concept sibling. Opens the diminishing/marginal triad of batch 7.
References¶
[1] Wright, T. P. (1936). Factors affecting the cost of airplanes. Journal of the Aeronautical Sciences, 3(4), 122–128. Original empirical derivation of the learning curve: unit cost in airplane manufacturing falls as a log-linear function of cumulative output, the canonical reference for learning_curve_effects as a sibling specialization of increasing returns under the experience channel. ↩
[2] Banister, E. W. (1991). Modeling elite athletic performance. In H. J. Green, J. D. McDougal, & H. Wenger (Eds.), Physiological Testing of Elite Athletes (pp. 403–424). Human Kinetics. Fitness–fatigue impulse-response model in which performance is the difference of two exponentially decaying training responses, predicting an inverted-U where additional training load past an optimum degrades performance rather than producing further gains. ↩
[3] Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89(4), 369–406. Skill-acquisition theory grounded in the ACT production-system architecture: practice yields concave learning gains via knowledge compilation, with the specific functional form (power, exponential) varying by task and stage rather than being universally power-law. ↩
[4] Mankiw, N. G. (2014). Principles of Economics (7th ed.). Cengage Learning. Standard introductory text whose Ten Principles open with "people face trade-offs" and "the cost of something is what you give up to get it," anchoring the equimarginal rule that rational agents compare marginal benefit to marginal opportunity cost rather than to zero. ↩
[5] Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406. Deliberate-practice framework: concave skill-vs-practice relationships reveal themselves only over decade-long longitudinal records, since short-window measurement noise and delayed consolidation effects mask the true functional shape. ↩
[6] Rogers, E. M. (2003). Diffusion of Innovations (5th ed.). Free Press. Canonical synthesis of how novelty spreads through a social network's structure, with adoption and reach governed by non-redundant interpersonal channels across community boundaries; supports the information-theoretic redundancy argument, the organizational knowledge-flow example, and the epidemic/cross-community diffusion-via-bridge example. ↩
[7] Marshall, A. (1890). Principles of Economics (Book IV, Ch. IX–XIII). Macmillan. Foundational treatment distinguishing internal and external economies of scale and the favorable below-optimum regime (fixed-cost spreading, deepening specialization), establishing the lineage in which the long-run average-cost curve and its eventual upturn become explicit objects of analysis. ↩
[8] Varian, H. R. (2014). Intermediate Microeconomics: A Modern Approach (9th ed.). W. W. Norton. Standard intermediate microeconomics text: develops the equimarginal principle (at an interior optimum the ratio of marginal benefits to prices is equalized across uses) as the formal stopping rule for any concave-objective allocation problem. ↩
[9] Pareto, V. (1896). Cours d'économie politique. F. Rouge. Originates the empirical regularity that a small fraction of inputs accounts for the majority of output in income distributions and many other systems—the long-tail signature that motivates portfolio-style reallocation away from low-marginal-return activities. ↩
[10] Solow, R. M. (1956). A contribution to the theory of economic growth. Quarterly Journal of Economics, 70(1), 65–94. Neoclassical growth model with diminishing marginal returns to capital: long-run per-capita growth cannot come from capital accumulation along a fixed concave production function but requires technical progress that shifts the function upward. ↩
[11] Cobb, C. W., & Douglas, P. H. (1928). A theory of production. American Economic Review, 18(1, Suppl.), 139–165. First empirical fit of a multiplicative power-law production function Y = A L^α K^β to U.S. manufacturing data; demonstrates how diagnostic capacity for identifying the true functional form constrains subsequent input-allocation reasoning. ↩
[12] Reinertsen, D. G. (2009). The Principles of Product Development Flow: Second Generation Lean Product Development. Celeritas Publishing. Builds an explicit economic framework around queue size, cycle time, and cost of delay so that the marginal economics of work-in-progress and batch sizes—normally hidden by aggregate metrics—become continuously visible to product-development decision-makers. ↩
[13] Bloom, N., Jones, C. I., Van Reenen, J., & Webb, M. (2020). Are ideas getting harder to find? American Economic Review, 110(4), 1104–1144. Empirical evidence across multiple R&D fields that research productivity per researcher has fallen sharply over time; sustaining a constant rate of innovation requires ever-larger researcher inputs, a paradigm case of diminishing returns to R&D effort. ↩
[14] Howard, R. W. (2014). Why are Hooper et al. wrong? Intelligence, 45, 14–17. Argues that apparent plateaus in chess Elo ratings reflect ceiling and threshold effects rather than absolute limits: extended phases of seemingly diminishing return on practice can give way to further gains once the right threshold of accumulated deliberate practice is crossed. ↩
[15] Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., et al. (2022). Training compute-optimal large language models. In Advances in Neural Information Processing Systems 35 (NeurIPS 2022). Empirical scaling-laws study (the "Chinchilla" paper): a sweep of compute-matched training runs fits concave loss-versus-compute curves and overturns earlier extrapolations about optimal model-size versus dataset-size allocation. ↩