Pith. sign in

REVIEW 3 major objections 3 minor 60 references

A generative-AI provider's optimal freemium menu reduces to two user-value thresholds, and a positive-quality free tier is optimal exactly when data value beats privacy-sensitive compute cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-05 00:12 UTC pith:BV5FNZ7A

load-bearing objection The central closed-form optimality claim is invalidated by an algebraic error in Proposition 10.6, and the dynamic-to-static jump is unsupported. the 3 major comments →

arxiv 2608.00823 v1 pith:BV5FNZ7A submitted 2026-08-01 cs.GT

Freemium Is All You Need

Maxim Bichuch , Aviv Yaish This is my paper
classification cs.GT MSC 91B2434D23
keywords generative AI economicsfreemium pricingquality dynamicstwo-threshold policyprivacy-separated menussteady-state qualitydata-driven quality improvement
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks when a generative-AI provider should offer a free tier that trains on user data, alongside a private paid tier, and how to price both. Its central claim is that this pricing problem reduces to choosing two endogenous user-value thresholds—one that fixes the flow of training data and thus steady-state model quality, and one that sets the price given that quality. For uniformly distributed request values the optimal strategy is given in closed form (Proposition 10.8) and divides into four regimes, with a positive-quality free tier optimal in exactly one of them. The regime boundary is a simple ratio: Cφ/(KM), where Cφ is the expected compute cost of privacy-sensitive users and K = αλ(1−φ)/δ is the market's data-driven quality potential. If correct, a provider's freemium decision becomes a one-line inequality check rather than a dynamic optimization.

Core claim

Under the model, paid requests are training-exempt and private, free requests train the model, and perceived quality decays at rate δ; every fixed menu has a unique steady-state quality Q* = (α/δ)N*ϕ, globally reached (Props. 7.2–7.6). Maximizing steady-state profit reduces to two value thresholds v1 = c/Q and v2 = c/(Q(1−γ)), separating monetization from data quality. For uniform values on [0,M] the optimum is closed-form (Props. 10.5, 10.8): when Cφ/(KM) is small, both thresholds are interior and the free tier has positive quality (γ>0); otherwise the optimum collapses to a paid-only single threshold or to corner regimes. The sufficient condition for a positive-quality free tier is equatio

What carries the argument

The central object is the change of variables from controls (c, γ) to the user-value thresholds v1 = c/Q and v2 = c/(Q(1−γ)), where Q is paid-tier quality and γ the free tier's quality ratio. In steady state Q = K F(v2) with K = αλ(1−φ)/δ (the quality-generation potential), so the profit objective becomes a quadratic function of the normalized thresholds x1 = v1/M, x2 = v2/M. The decisive dimensionless ratio Cφ/(KM)—expected compute cost of privacy-sensitive users divided by the data-driven quality value—classifies the optimum into four closed-form regimes; the regime with γ > 0 (positive-quality free tier) is exactly the interior one where x*1 < x*2 satisfy equations (8)–(9).

Load-bearing premise

The load-bearing premise is that maximizing discounted infinite-horizon profit is equivalent to maximizing steady-state per-period profit; if the optimal policy is not stationary, or if transition dynamics and the discount rate matter, the closed-form thresholds may not be optimal for the original problem.

What would settle it

For fixed parameters (M, φ, α, δ, C, λ, r), solve the original discounted problem by dynamic programming over the state Q with controls (c, γ), using the demand formulas of Prop. 6.2; compare the discounted profit of the optimal non-stationary policy with the profit of the paper's steady-state thresholds from Prop. 10.8. A measurable profit gap for any parameter set—especially in the interior regime—would show that maximizing steady-state profit does not maximize the stated discounted objective. A cheaper version: simulate the system with the paper's policy for the four regimes and check wheth

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • An operator with uniform user values can compute the optimal price and free-tier quality directly from closed-form formulas, with no dynamic programming.
  • A positive-quality free tier is optimal only in the interior regime where the data value from free users outweighs the compute cost of the privacy-sensitive segment; outside it, the model predicts a paid-only menu or a data-only corner.
  • The free tier never starves: steady-state demand on the free tier is always strictly positive, and quality converges to its unique steady state from any starting point.
  • Raising the paid price or making the free tier closer in quality to the paid tier both increase steady-state quality, because both channel more users into training-eligible demand.
  • In the high-cost corner, the model predicts effective market shutdown (all users priced out), so cost control, not pricing, is the binding lever.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the two-threshold decoupling appears distribution-free—Q = K F(v2) and the mark-up equation separate for any CDF—so the same structural result should hold for log-normal or Pareto values, with the closed form replaced by a numerical root-finding step.
  • Beyond the paper: the regime boundary suggests an empirical test: measure Cφ/(KM) for a deployed freemium GenAI service (using cost logs, privacy-sensitivity surveys, and quality-decay estimates) and check whether the presence of a positive-quality free tier matches the predicted regime.
  • Beyond the paper: the model treats quality as a single stock; extending it to multi-dimensional quality (e.g., safety vs. capability) would likely produce a family of thresholds and a higher-dimensional analogue of the Cφ/(KM) condition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper studies a monopolist GenAI provider offering a paid training-exempt tier and a free training-eligible tier, with model quality governed by Nerlove-Arrow dynamics. It derives demand under privacy separation, proves existence/uniqueness/global convergence to a unique steady state for fixed stationary strategies, and then analyzes optimal pricing and quality policies. For uniform valuations it claims a closed-form optimal policy expressed through two thresholds, and a characterization of when a positive-quality free tier is optimal. The steady-state and convergence results are plausible, but the optimality analysis contains a load-bearing algebraic error in the uniform case and rests on an unproved replacement of the dynamic objective by steady-state profit; the privacy-separation constraint is also not enforced.

Significance. If correct, the paper would be a valuable contribution to the growing literature on GenAI platform economics: it connects freemium data acquisition to dynamic model quality and offers rare closed-form policy characterizations. The steady-state existence, uniqueness, and global convergence propositions (7.3, 7.4, 7.6) appear sound and are useful in themselves. However, the central abstract claim of a closed-form optimal service strategy is not supported because the uniform-case feasibility conditions contain an algebraic error, the dynamic problem is replaced by a static steady-state problem without proof, and the privacy-separation constraint L≥cγ is never imposed in the optimization. These are load-bearing issues in the main result, so the paper's headline contribution is not established.

major comments (3)
  1. [Proposition 10.6 / Eq. (9)] The feasibility condition x2*≤1 is derived incorrectly. From Proposition 10.5, x2*≤1 is equivalent to φx1*≥2φ−1, i.e. 5−6φ≥√[1−12S(1−φ)], where S=Cφ/(KM). This inequality is impossible for φ>5/6; for φ≤5/6, squaring yields S≥3φ−2, the reverse of the printed S≤3φ−2. The proof squares without requiring 5−6φ≥0 and reverses the inequality direction. Concretely, φ=0.9 and S=0.45 satisfy both Eq. (8) (0.45<0.661) and the printed Eq. (9) (0.45≤0.7), yet x1*≈0.490 and x2*≈2.80>1. Thus Proposition 10.8's Case 1 is applied to infeasible parameters, and the closed-form 'optimal' strategy and the sufficient conditions for offering a positive-quality free tier are invalid as stated.
  2. [Sections 8 and 9] Definition 5.1 maximizes the discounted infinite-horizon profit Π=∫e^{-rt}π(t)dt. The optimization then replaces this with the static per-period profit Eq. (2), asserting that maximizing Definition 5.1 is equivalent to maximizing Eq. (2) at all τ, and later sets Π=(1/r)(Rev−Cost). Neither step is proved: no argument shows that an optimal policy is stationary, that the transition path can be ignored, or that discounting does not matter. Hence the thresholds derived in the paper solve a steady-state profit problem, not necessarily the dynamic problem in Definition 5.1. This gap affects every optimality claim, including the abstract's 'closed-form optimal service strategy'.
  3. [Privacy constraint, Proposition 6.1 and Section 10] Proposition 6.1 shows that privacy-sensitive users stay off the free tier iff L≥cγ. The optimization in Section 10 changes variables to (x1,x2) and maximizes Eq. (6) subject only to 0<x1≤x2≤1. The constraint cγ = v1 KF(v2)(1−v1/v2) = KM x1 x2(1−x1/x2) ≤ L is never imposed or checked. Proposition 10.2 invokes cγ≤L at the outset, but the candidate optima in Proposition 10.8 need not satisfy it. Consequently the proposed policies may violate the privacy-separation assumption on which Proposition 6.2's demand functions are built, so the profit objective itself would not be the one governing the modeled user behavior.
minor comments (3)
  1. [Proof sketches, Propositions 7.6 and 10.8] Typos: 'quaity' in the proof sketch of Proposition 7.6, and 'weather' for 'whether' in the proof of Proposition 10.8.
  2. [Proposition 10.2 proof] The proof begins 'Due to Proposition 6.1, we have cγ≤L'. Since Proposition 6.1 is an iff, the paper is implicitly restricting to the privacy-separated regime; this should be stated explicitly before using it as a premise.
  3. [Sections 5 and 9] The symbol Π is used both for the discounted objective in Definition 5.1 and for the steady-state value (1/r)(Rev−Cost) in Section 9. Please use distinct notation to avoid conflating the dynamic and static problems.

Circularity Check

0 steps flagged

No significant circularity: the derivation is self-contained; closed-form results follow from model primitives and fixed-point equations, not from fitted inputs or self-citations.

full rationale

The paper's derivation chain is self-contained. Demand (Prop. 6.2), steady-state quality (Prop. 7.2), and the uniform-value objective (Eq. 6) are obtained by substituting the model primitives (utility functions, arrival process, uniform CDF) into the definitions; no parameter is calibrated to data and no advertised 'prediction' is a renamed fitted constant. The closed-form thresholds in Prop. 10.5 are solutions of the first-order conditions of J, and Props. 10.6 and 10.8 are constraint-screening conditions for those candidates; this is ordinary constrained optimization, not circularity. The paper cites standard external tools (Nerlove-Arrow depreciation, Lyapunov stability, Leibniz rule), but these are not self-citations and do not import the target conclusion. The main caveats are non-circular correctness and assumption issues: Section 8 replaces the discounted infinite-horizon objective with per-period steady-state profit under an asserted equivalence ('maximizing Definition 5.1 is equivalent to maximizing Eq. (2) at all τ'), which is an assumption gap rather than a circular step; and the proof of Prop. 10.6 contains a suspicious algebraic step (squaring 5−6φ ≥ sqrt(...) without requiring 5−6φ ≥ 0, and apparently reversing the resulting inequality), which would affect the case classification in Prop. 10.8. But even if those concerns are valid, they are mathematical errors or modeling choices, not cases where the output is equivalent to the input by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The main uncharged inputs are the behavioral assumptions above, especially the steady-state equivalence, the indifferent-user tie-breaking, and the dropped privacy constraint.

free parameters (1)
  • L (privacy loss)
    Introduced in Section 3 to model privacy-sensitive users, used only in Proposition 6.1, and absent from the optimization in Sections 9 and 10. The required constraint L >= c*gamma is never enforced, so L effectively plays no role and any optimum may violate privacy separation.
axioms (5)
  • domain assumption Quality evolves as Qdot = alpha*N_phi - delta*Q (Nerlove-Arrow)
    Section 5, Eq. (1). This is the core dynamic law; no empirical justification is given for this functional form beyond citing prior advertising and LLM deprecation literature.
  • domain assumption The provider restricts to a binary privacy-separated menu: a free training-eligible tier and a paid training-exempt tier.
    Section 5, motivated by Propositions 4.6 and 4.8. The general mechanism design results are not carried into the dynamic optimization.
  • ad hoc to paper Maximizing the discounted infinite-horizon profit (Definition 5.1) is equivalent to maximizing steady-state per-period profit.
    Section 8 says 'under the assumption of no dynamic state equations' and Section 9 uses 'In a steady-state, as Rev and Cost are constant'. No proof is given that the optimal policy is stationary or that transition costs and discounting can be ignored.
  • ad hoc to paper Public users who are indifferent between a zero-quality free tier and no service still choose the free tier and generate training data.
    Proposition 6.2 and the limit in Proposition 7.3 (lim Q->0+ N_phi = (1-phi)lambda) require all non-private users below v2 to use the free tier even when gamma=0 and free utility is 0. This tie-breaking rule is not stated or justified.
  • ad hoc to paper Privacy separation holds at all candidate optima: L >= c*gamma.
    Proposition 6.1 gives this as the condition for private users to avoid the free tier, but Sections 9 and 10 optimize over (v1,v2) with only 0 < v1 <= v2, never including L >= c*gamma. The demand functions used in the optimization are invalid if this is violated.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Freemium Is All You Need." pith.science (2026). https://pith.science/paper/BV5FNZ7A

@misc{pith2026260800823,
  author       = {Pith},
  title        = {Pith review of: Freemium Is All You Need},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BV5FNZ7A}},
  note         = {Machine review of arXiv:2608.00823}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Free access to a generative AI (GenAI) service requires costly compute, yet can also produce data that improve future service quality. We study a service provider whose users differ in request value and privacy preference, under the constraint that paid requests are kept private, while free requests can be used to improve model quality which otherwise reverts toward a baseline. In particular, we characterize service demand and prove that quality converges to a unique steady state for every stationary service strategy. Next, we analyze optimal pricing and quality policies and show that these can be expressed using two endogenous user value thresholds. For uniform values, we provide a closed-form optimal service strategy and characterize the sufficient conditions for offering free services as dependent on inference cost.

Figures

Figures reproduced from arXiv: 2608.00823 by Aviv Yaish, Maxim Bichuch.

Figure 1
Figure 1. Figure 1: Private requests either pay or exit at v1, while public requests use the free tier below v2 and pay above it. [48] Hal R. Varian. “Chapter 10 Price Discrimination”. In: Handbook of Industrial Orga￾nization. Vol. 1. Elsevier, Jan. 1989, pp. 597–654. doi: 10.1016/S1573- 448X(89) 01013-7. [49] Ander Artola Velasco, Stratis Tsirtsis, Nastaran Okati, and Manuel Gomez-Rodriguez. Is Your LLM Overcharging You? Tok… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 44 canonical work pages · 1 internal anchor

  1. [1]

    Optimal Bidding in Multi-Item Multi- Slot Sponsored Search Auctions

    Vibhanshu Abhishek and Kartik Hosanagar. “Optimal Bidding in Multi-Item Multi- Slot Sponsored Search Auctions”. In:Proceedings of the 13th ACM Conference on Electronic Commerce. EC ’12. New York, NY, USA: Association for Computing Ma- chinery, June 2012, p. 2.isbn: 978-1-4503-1415-2.doi:10.1145/2229012.2229015

    work page arXiv 2012

  2. [2]

    Monetizing Freemium Com- munities: Does Paying for Premium Increase Social Engagement?

    Ravi Bapna, Jui Ramaprasad, and Akhmed Umyarov. “Monetizing Freemium Com- munities: Does Paying for Premium Increase Social Engagement?” In:MIS Quarterly 42.3 (Sept. 2018), pp. 719–735.issn: 0276-7783, 2162-9730.doi:10.25300/MISQ/ 2018/13592

    work page doi:10.25300/misq/ 2018

  3. [3]

    Social Referral Programs for Freemium Platforms

    Rodrigo Belo and Ting Li. “Social Referral Programs for Freemium Platforms”. In: Management Science68.12 (Dec. 2022), pp. 8933–8962.issn: 0025-1909, 1526-5501. doi:10.1287/mnsc.2022.4301

    work page arXiv 2022

  4. [4]

    Dirk Bergemann, Alessandro Bonatti, and Alex Smolin.The Economics of Large Language Models: Token Allocation, Fine-Tuning, and Optimal Pricing. Feb. 2025. doi:10.48550/arXiv.2502.07736. arXiv:2502.07736 [econ]

    work page doi:10.48550/arxiv.2502.07736 2025

  5. [5]

    Deming, Nicola Fuchs-Schündeln, and Jonas Jessen.Mind the Gap: AI Adoption in Europe and the U.S.WorkingPaper.Mar.2026

    Alexander Bick, Adam Blandin, David J. Deming, Nicola Fuchs-Schündeln, and Jonas Jessen.Mind the Gap: AI Adoption in Europe and the U.S.WorkingPaper.Mar.2026. doi:10.3386/w34995. National Bureau of Economic Research:34995

    doi:10.3386/w34995 2026

  6. [6]

    Bernard Bolzano.Rein analytischer Beweis des Lehrsatzes, daß zwischen je zwey Werthen, die ein entgegengesetztes Resultat gewähren, wenigstens eine reelle Wurzel der Gleichung liege.Prague:GottliebHaase,1817.url:https://dml.cz/bitstream/ handle/10338.dmlcz/400016/Bolzano_24-1817-1_1.pdf

    work page

  7. [7]

    How Is ChatGPT’s Behavior Chang- ing Over Time?

    Lingjiao Chen, Matei Zaharia, and James Zou. “How Is ChatGPT’s Behavior Chang- ing Over Time?” In:Harvard Data Science Review6.2 (Apr. 2024).issn: 2644-2353, doi:10.1162/99608f92.5317da47

    doi:10.1162/99608f92.5317da47 2024

  8. [8]

    Personalized Pricing in the Presence of Privacy Concerns

    Zhiqi Chen and Mengyu Zhang. “Personalized Pricing in the Presence of Privacy Concerns”. In:Management Science(May 2026).issn: 0025-1909.doi:10 . 1287 / mnsc.2024.07170

    work page arXiv 2026

  9. [9]

    AdsThatStick:Near-OptimalAdOptimizationthroughPsychologicalBehavior Models

    Kailash Gopal Darmasubramanian, Akash Pareek, Arindam Khan, and Arpit Agar- wal.“AdsThatStick:Near-OptimalAdOptimizationthroughPsychologicalBehavior Models”. In:The Fourteenth International Conference on Learning Representations. Oct. 2025.url:https://openreview.net/forum?id=A8wfXZkoMs

    work page 2025

  10. [10]

    Definite and Semidefinite Quadratic Forms

    Gérard Debreu. “Definite and Semidefinite Quadratic Forms”. In:Econometrica20.2 (1952), pp. 295–300.doi:10.2307/1907852. 15

    work page doi:10.2307/1907852 1952

  11. [11]

    2025.url:https: //andreyfradkin.com/assets/LLM_Demand_12_12_2025.%5Cgls%7BPDF%7D

    Demirer, Mert, Andrey Fradkin, Tadelis, Nadav, and Sida Tadelis.The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs. 2025.url:https: //andreyfradkin.com/assets/LLM_Demand_12_12_2025.%5Cgls%7BPDF%7D

    work page 2025

  12. [12]

    Spillover Effects and Freemium Strategy in the Mobile App Market

    Yiting Deng, Anja Lambrecht, and Yongdong Liu. “Spillover Effects and Freemium Strategy in the Mobile App Market”. In:Management Science69.9 (Sept. 2023), pp. 5018–5041.issn: 0025-1909, 1526-5501.doi:10.1287/mnsc.2022.4619

    arXiv 2023

  13. [13]

    Mech- anism Design for Large Language Models

    PaulDütting,VahabMirrokni,RenatoPaesLeme,HaifengXu,andSongZuo.“Mech- anism Design for Large Language Models”. In:Proceedings of the ACM Web Confer- ence 2024. WWW ’24. New York, NY, USA: Association for Computing Machinery, May 2024, pp. 144–155.isbn: 979-8-4007-0171-9.doi:10.1145/3589334.3645511

    arXiv 2024

  14. [14]

    DynamicOptimalControl Models in Advertising: Recent Developments

    GustavFeichtinger,RichardF.Hartl,andSureshP.Sethi.“DynamicOptimalControl Models in Advertising: Recent Developments”. In:Management Science40.2 (Feb. 1994), pp. 195–226.issn: 0025-1909, 1526-5501.doi:10.1287/mnsc.40.2.195

    work page doi:10.1287/mnsc.40.2.195 1994

  15. [15]

    Differentiation Under the Integral Sign

    Harley Flanders. “Differentiation Under the Integral Sign”. In:The American Mathe- matical Monthly80.6 (1973), pp. 615–627.doi:10.1080/00029890.1973.11993339

    arXiv 1973

  16. [16]

    From Competition to Collaboration: Designing Sustainable Mechanisms Between LLMs and Online Forums

    Niv Fono, Yftah Ziser, and Omer Ben-Porat. “From Competition to Collaboration: Designing Sustainable Mechanisms Between LLMs and Online Forums”. In:Proceed- ings of the 25th International Conference on Autonomous Agents and Multiagent Sys- tems. AAMAS ’26. Richland, SC: International Foundation for Autonomous Agents and Multiagent Systems, May 2026, pp. 3...

    work page doi:10.65109/stkg8016 2026

  17. [17]

    Andrey Fradkin.Demand for LLMs: Descriptive Evidence on Substitution, Market Expansion, and Multihoming. Apr. 2025.doi:10.48550/arXiv.2504.15440. arXiv: 2504.15440 [cs.CY]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.15440 2025

  18. [18]

    Prediction-Sharing During Training and Inference

    Yotam Gafni, Ronen Gradwohl, and Moshe Tennenholtz. “Prediction-Sharing During Training and Inference”. In:Algorithmic Game Theory. Ed. by Guido Schäfer and Carmine Ventre. Cham: Springer Nature Switzerland, 2024, pp. 425–442.isbn: 978- 3-031-71033-9.doi:10.1007/978-3-031-71033-9_24

    work page doi:10.1007/978-3-031-71033-9_24 2024

  19. [19]

    Long-Term Data Sharing under Exclusivity Attacks

    Yotam Gafni and Moshe Tennenholtz. “Long-Term Data Sharing under Exclusivity Attacks”. In:Proceedings of the 23rd ACM Conference on Economics and Computa- tion. EC ’22. New York, NY, USA: Association for Computing Machinery, July 2022, pp. 739–759.isbn: 978-1-4503-9150-4.doi:10.1145/3490486.3538311

    work page arXiv 2022

  20. [20]

    Google AI for Developers

    Google.Gemini API Additional Terms of Service. Google AI for Developers. 2026. url:https://ai.google.dev/gemini-api/terms

    work page 2026

  21. [21]

    Google AI for Developers

    Google.Gemini Developer API Pricing. Google AI for Developers. 2026.url:https: //ai.google.dev/gemini-api/docs/pricing

    work page 2026

  22. [22]

    StabilityinCompetition

    HaroldHotelling.“StabilityinCompetition”.In:The Economic Journal39.153(1929), pp. 41–57.issn: 00130133, 14680297.doi:10.2307/2224214. JSTOR:2224214

    doi:10.2307/2224214 1929

  23. [23]

    Nonparametric Estimation from Incomplete Observations

    Edward L. Kaplan and Paul Meier. “Nonparametric Estimation from Incomplete Observations”. In:Journal of the American Statistical Association53.282 (1958), pp. 457–481.doi:10.1080/01621459.1958.10501452. 16

    arXiv 1958

  24. [24]

    Strategic Content Creation with GenAI: To Share or Not to Share?

    Gur Keinan and Omer Ben-Porat. “Strategic Content Creation with GenAI: To Share or Not to Share?” In:Proceedings of the ACM Web Conference 2026. WWW ’26. New York, NY, USA: Association for Computing Machinery, Apr. 2026, pp. 169–180.isbn: 979-8-4007-2307-0.doi:10.1145/3774904.3792308

    work page arXiv 2026

  25. [25]

    Nonlinear Programming

    Harold W. Kuhn and Albert W. Tucker. “Nonlinear Programming”. In:Proceed- ings of the Second Berkeley Symposium on Mathematical Statistics and Probabil- ity. Ed. by Jerzy Neyman. Berkeley: University of California Press, 1951, pp. 481– 492.url:https : / / projecteuclid . org / ebooks / berkeley - symposium - on - mathematical - statistics - and - probabil...

    arXiv 1951

  26. [26]

    SomeExtensions of Liapunov’sSecond Method

    Joseph P. LaSalle. “SomeExtensions of Liapunov’sSecond Method”. In:IRE Transac- tions on Circuit Theory7.4 (1960), pp. 520–527.doi:10.1109/TCT.1960.1086720

    work page arXiv 1960

  27. [27]

    Fine-Tuning Games: Bargaining and Adaptation for General-Purpose Models

    Benjamin Laufer, Jon Kleinberg, and Hoda Heidari. “Fine-Tuning Games: Bargaining and Adaptation for General-Purpose Models”. In:Proceedings of the ACM Web Con- ference 2024.WWW’24.NewYork,NY,USA:AssociationforComputingMachinery, May 2024, pp. 66–76.isbn: 979-8-4007-0171-9.doi:10.1145/3589334.3645366

    arXiv 2024

  28. [28]

    Two-Sided Price Discrimination by Media Platforms

    Song Lin. “Two-Sided Price Discrimination by Media Platforms”. In:Marketing Sci- ence39.2 (Mar. 2020), pp. 317–338.issn: 0732-2399, 1526-548X.doi:10 . 1287 / mksc.2019.1211

    work page arXiv 2020

  29. [29]

    Sur l’application de la méthode des approximations succes- sives aux équations différentielles ordinaires du premier ordre

    Ernst Leonard Lindelöf. “Sur l’application de la méthode des approximations succes- sives aux équations différentielles ordinaires du premier ordre”. In:Comptes rendus hebdomadaires des séances de l’Académie des sciences118 (1894), pp. 454–457.url: https://gallica.bnf.fr/ark:/12148/bpt6k3074r

    work page

  30. [30]

    Pay to (Not) Play: Monetizing Impatience in Mobile Games

    Taylor Lundy, Narun Raman, Hu Fu, and Kevin Leyton-Brown. “Pay to (Not) Play: Monetizing Impatience in Mobile Games”. In:Proceedings of the AAAI Conference on Artificial Intelligence38.9 (Mar. 2024), pp. 9856–9864.issn: 2374-3468.doi:10. 1609/aaai.v38i9.28846

    work page 2024

  31. [31]

    The General Problem of the Stability of Motion

    Aleksandr M. Lyapunov. “The General Problem of the Stability of Motion”. In:Inter- national Journal of Control55.3 (1992). English translation by A. T. Fuller, pp. 531– 773.doi:10.1080/00207179208934253

    doi:10.1080/00207179208934253 1992

  32. [32]

    Strategic Generative AI for Machine Learn- ing in Economic Environments

    Omer Madmon and Moshe Tennenholtz. “Strategic Generative AI for Machine Learn- ing in Economic Environments”. In:ACM AI Lett.1.1 (Mar. 2026), 7:1–7:8.doi: 10.1145/3799994

    work page doi:10.1145/3799994 2026

  33. [33]

    Optimizing Free-to-Play Multiplayer Games with Premium Subscription

    Yunke Mai and Bin Hu. “Optimizing Free-to-Play Multiplayer Games with Premium Subscription”. In:Management Science69.6 (June 2023), pp. 3437–3456.issn: 0025- 1909, 1526-5501.doi:10.1287/mnsc.2022.4510

    work page arXiv 2023

  34. [34]

    Monopoly with Incomplete Information

    Eric Maskin and John Riley. “Monopoly with Incomplete Information”. In:Rand Journal of Economics15.2 (1984), p. 171.issn: 07416261.doi:10.2307/2555674. 17

    work page doi:10.2307/2555674 1984

  35. [35]

    The Value of Per- sonal Information in Online Markets with Endogenous Privacy

    Rodrigo Montes, Wilfried Sand-Zantman, and Tommaso Valletti. “The Value of Per- sonal Information in Online Markets with Endogenous Privacy”. In:Management Science65.3 (Mar. 2019), pp. 1342–1362.issn: 0025-1909, 1526-5501.doi:10.1287/ mnsc.2017.2989

    work page arXiv 2019

  36. [36]

    Incentives in Federated Learning: Equilibria, Dynamics, and Mechanisms for Welfare Maximization

    Aniket Murhekar, Zhuowen Yuan, Bhaskar Ray Chaudhury, Bo Li, and Ruta Mehta. “Incentives in Federated Learning: Equilibria, Dynamics, and Mechanisms for Welfare Maximization”. In:Advances in Neural Information Processing Systems. Vol. 36. Cur- ran Associates, Inc., 2023, pp. 17811–17831.url:https://proceedings.neurips. cc / paper % 5C _ files / paper / 20...

    work page 2023

  37. [37]

    Monopoly and Product Quality

    Michael Mussa and Sherwin Rosen. “Monopoly and Product Quality”. In:Journal of Economic Theory18.2 (Aug. 1978), pp. 301–317.issn: 0022-0531.doi:10.1016/ 0022-0531(78)90085-6

    work page 1978

  38. [38]

    Optimal Auction Design

    Roger B. Myerson. “Optimal Auction Design”. In:Mathematics of Operations Re- search6.1 (1981), pp. 58–73.doi:10.1287/moor.6.1.58

    doi:10.1287/moor.6.1.58 1981

  39. [39]

    Optimal Advertising Policy under Dynamic Conditions

    Marc Nerlove and Kenneth J. Arrow. “Optimal Advertising Policy under Dynamic Conditions”. In:Economica29.114 (May 1962), p. 129.issn: 00130427.doi:10 . 2307/2551549. JSTOR:10.2307/2551549

    work page doi:10.2307/2551549 1962

  40. [40]

    Economics of Free under Perpetual Licensing: Implications for the Software Industry

    Marius F. Niculescu and D. J. Wu. “Economics of Free under Perpetual Licensing: Implications for the Software Industry”. In:Information Systems Research25.1 (Mar. 2014), pp. 173–199.issn: 1047-7047, 1526-5536.doi:10.1287/isre.2013.0508

    work page arXiv 2014

  41. [41]

    Hamidah Oderinwale and Anna Kazlauskas.The Economics of AI Training Data: A Research Agenda. Apr. 2026.doi:10.48550/arXiv.2510.24990. arXiv:2510.24990 [cs.CY]

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.24990 2026

  42. [42]

    ApproximateMechanismDesignwithout Money

    ArielD.ProcacciaandMosheTennenholtz.“ApproximateMechanismDesignwithout Money”. In:ACM Transactions on Economics and Computation1.4 (Dec. 2013), 18:1–18:26.issn: 2167-8375.doi:10.1145/2542174.2542175

    work page arXiv 2013

  43. [43]

    Haoran Sun, Yusen Wu, Peng Wang, Wei Chen, Yukun Cheng, Xiaotie Deng, and Xu Chu.Game Theory Meets Large Language Models: A Systematic Survey with Taxonomy and New Frontiers. Feb. 2025.doi:10.48550/arXiv.2502.09053

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.09053 2025

  44. [44]

    Economics of Content Creation with Generative AI

    Boaz Taitler. “Economics of Content Creation with Generative AI”. In:Proceedings of the AAAI Conference on Artificial Intelligence40.48 (Mar. 2026), pp. 41086–41087. issn: 2374-3468.doi:10.1609/aaai.v40i48.42169

    work page doi:10.1609/aaai.v40i48.42169 2026

  45. [45]

    Braess’s Paradox of Generative AI

    Boaz Taitler and Omer Ben-Porat. “Braess’s Paradox of Generative AI”. In:Proceed- ings of the AAAI Conference on Artificial Intelligence39.13 (Apr. 2025), pp. 14139– 14147.issn: 2374-3468.doi:10.1609/aaai.v39i13.33548

    work page doi:10.1609/aaai.v39i13.33548 2025

  46. [46]

    May 2025.doi:10.48550/arXiv.2505.12386

    Boaz Taitler, Omer Madmon, Moshe Tennenholtz, and Omer Ben-Porat.Data Shar- ing with a Generative AI Competitor. May 2025.doi:10.48550/arXiv.2505.12386. arXiv:2505.12386 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.12386 2025

  47. [47]

    The Theory of Industrial Organization

    Jean Tirole. “The Theory of Industrial Organization”. In:MIT Press Books1 (1988). url:https://ideas.repec.org//b/mtp/titles/0262200716.html. 18 v v1 Privateρ= 1 No Use Paid v v2 Publicρ= 0 Free Paid Figure 1: Private requests either pay or exit atv1, while public requests use the free tier belowv 2 and pay above it

    work page arXiv 1988

  48. [48]

    Chapter 10 Price Discrimination

    Hal R. Varian. “Chapter 10 Price Discrimination”. In:Handbook of Industrial Orga- nization. Vol. 1. Elsevier, Jan. 1989, pp. 597–654.doi:10.1016/S1573- 448X(89) 01013-7

    work page doi:10.1016/s1573- 1989

  49. [49]

    Is Your LLM Overcharging You? Tokenization, Transparency, and Incentives

    AnderArtolaVelasco,StratisTsirtsis,NastaranOkati,andManuelGomez-Rodriguez. Is Your LLM Overcharging You? Tokenization, Transparency, and Incentives. Sept. 2025.doi:10.48550/arXiv.2505.21627. arXiv:2505.21627 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.21627 2025

  50. [50]

    The Non-Local Existence Problem of Ordinary Differential Equa- tions

    Aurel Wintner. “The Non-Local Existence Problem of Ordinary Differential Equa- tions”. In:American Journal of Mathematics67.2 (1945), pp. 277–284.doi:10 . 2307/2371729

    work page 1945

  51. [51]

    Heterogeneous Data Game: Characterizing the Model Competition Across Multiple Data Sources

    Renzhe Xu, Kang Wang, and Bo Li. “Heterogeneous Data Game: Characterizing the Model Competition Across Multiple Data Sources”. In:Forty-Second International Conference on Machine Learning. June 2025, p. 24.url:https : / / openreview . net/forum?id=o36quGYoQX

    work page 2025

  52. [52]

    Weijie Zhong.Token Is All You Price. Oct. 2025.doi:10.48550/arXiv.2510.09859. arXiv:2510.09859 [econ]. A Proofs Proposition 4.6(FreeTier Necessity).Every mechanism simultaneously satisfying IR, IC, IT, and NS has exactly one positive-quality training-eligible free tier. If more allocations are offered, none charge negative prices, and quality is weakly mo...

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.09859 2025

  53. [53]

    (16), giving Proposition 10.5

    This leads us to choose the negative root in Eq. (16), giving Proposition 10.5. Having now chosen the negative root, we are guaranteedx∗ 1 >0. Proposition 10.6.The unconstrained maximizers Proposition 10.5 are optimal if both x∗ 1 < x∗ 2 andx ∗ 2 ≤1hold. These constraints respectively correspond to: Cφ KM < −1 + 2φ (2−φ) 2 ,(8) Cφ KM ≤(3φ−2).(9) Proof.Whe...

    work page

  54. [54]

    If Eq.(8)and Eq.(9)hold thenx ∗ 1, x∗ 2 follow Proposition 10.5

    work page

  55. [55]

    If Eq.(8)is violated while Eq.(9)holds, thenx ∗ 1 =x ∗ 2 = 1+ √ 1+3 Cφ KM 3 ,andγ= 0

    work page

  56. [56]

    If Eq.(9)is violated and Cφ KM < 3φ−1 4 , thenx ∗ 2 = 1,and Proposition 10.5 holds

    work page

  57. [57]

    If Eq.(9)is violated and Cφ KM ≥ 3φ−1 4 , thenx ∗ 1 =x ∗ 2 = 1, andγ= 0. Proof.1. If both Eq. (8) and Eq. (9) hold, then the optimalx ∗ 1, x∗ 2 are not constrained and are given by Proposition 10.5, Proposition 10.5. Note that this is possible, as (3φ−2)< −1+2φ (2−φ)2, holds forφ∈ h 0, 11− √ 37 6

    work page

  58. [58]

    (8) is violated while Eq

    If Eq. (8) is violated while Eq. (9) holds, then we have that Cφ KM ≥ −1+2φ (2−φ)2. This will result in the constraintx∗ 1 ≤x ∗ 2 being violated. This means that we must set x∗ 1 =x ∗ 2,and both will satisfyx ∗ 1, x∗ 2 ≤1also resulting inγ= 0. In this case the objective Proposition 10.5 collapses toJ(z) =KM(z 2 −z 3)−C(1−φz). Solving its optimizationJ ′(z...

    work page

  59. [59]

    (9) is violated, that isCφ KM <(3φ−2), means thatx ∗ 2 ≥1, so we must setx ∗ 2 = 1

    Similarly, in case Eq. (9) is violated, that isCφ KM <(3φ−2), means thatx ∗ 2 ≥1, so we must setx ∗ 2 = 1. We then need to check weatherx∗ 1 <1, i.e., 2− √ 1−12 Cφ KM (1−φ) 3φ <1. This in turn is equivalent to q 1−12 Cφ KM (1−φ)>3φ−2. This holds if12 Cφ KM (1− φ)<1−(3φ−2) 2 =−9φ 2 + 12φ−3 = 3(−3φ 2 + 4φ−1) = 3(1−φ)(3φ−1),which in turn is equivalent to Cφ ...

    work page

  60. [60]

    exclusivity

    Otherwise, if both Eq. (9) and Eq. (10) are violated thenx∗ 1 =x ∗ 2 = 1, again resulting inγ= 0. The last scenario, namelyx ∗ 1 =x ∗ 2 = 1, withγ= 0also happens when both Eq. (8) and Eq. (9) are violated. 35 B Additional Related Work Privacy and Platform Design.Previous work explores the value of private data. Montes, Sand-Zantman, and Valletti [35] stud...

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.