pith. sign in

arxiv: 2512.07526 · v2 · pith:6NBCTLCRnew · submitted 2025-12-08 · 💱 q-fin.RM · econ.GN· q-fin.EC· q-fin.GN

The Suicide Region: Option Games and the Race to Artificial General Intelligence

David Tan This is my paper

Pith reviewed 2026-05-21 17:58 UTC · model grok-4.3

classification 💱 q-fin.RM econ.GNq-fin.ECq-fin.GN
keywords AGI racepreemption gamereal optionsexistential risksuicide regionoption gamesrisk cancellationsystemic ruin
0
0 comments X

The pith

In the AGI race, shared existential risks cancel from players' decisions and create a suicide region of forced early deployment despite negative value.

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

The paper models the AGI race between sovereign actors as a continuous-time preemption game that includes endogenous existential risk. A systemic ruin parameter D, correlated with development speed and borne globally, enters both players' payoffs. Because this shared disutility appears symmetrically, the risk term drops out of the equilibrium indifference condition that sets the investment threshold. The resulting suicide region is the portion of investment space where competitive preemption compels rational agents to deploy AGI systems ahead of the point where risk-adjusted net present value turns positive. The model further shows that sub-existential warning shots leave the winner-takes-all structure intact and therefore do not restore delay.

Core claim

When the cost of global ruin is embedded in both players' payoffs, the risk term mathematically cancels from the equilibrium indifference condition of the continuous-time preemption game. This cancellation produces a suicide region in which competitive pressures force early AGI deployment even though the risk-adjusted net present value remains negative.

What carries the argument

The equilibrium indifference condition of the preemption game, from which the shared systemic ruin term cancels.

If this is right

  • Warning shots or sub-existential disasters leave the winner-takes-all structure unchanged and therefore fail to slow acceleration.
  • The race stops only when the cost of ruin is internalized to each player, making safety research economically necessary before deployment.
  • A critical private liability threshold exists that restores the option value of waiting.
  • Targeted liability or insurance mechanisms can shift the equilibrium back toward safer research sequences.

Where Pith is reading between the lines

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

  • If the ruin cost is only partially shared or imperfectly correlated with speed, the cancellation weakens and some value of waiting may reappear.
  • Treaties that assign ex-post liability for global harm could replicate the internalization effect without requiring perfect symmetry in D.
  • The same cancellation logic may apply to other winner-takes-all technological races that carry correlated global downside.

Load-bearing premise

The model assumes a systemic ruin parameter D that is correlated with development velocity and shared globally across players.

What would settle it

Empirical observation that competing actors delay AGI deployment when they acknowledge correlated global catastrophe risks would contradict the cancellation result.

read the original abstract

Standard real options theory predicts delay in exercising the option to invest or deploy when extreme asset volatility or technological uncertainty are present. However, in the current race to develop artificial general intelligence (AGI), sovereign actors are exhibiting behaviors contrary to theoretical predictions: the US and China are accelerating AI investment despite acknowledging the potential for catastrophic failure from AGI misalignment. We resolve this puzzle by formalizing the AGI race as a continuous-time preemption game with endogenous existential risk. In our model, the cost of failure is no longer bounded only by the sunk cost of investment (I), but rather a systemic ruin parameter (D) that is correlated with development velocity and shared globally. As the disutility of catastrophe is embedded in both players' payoffs, the risk term mathematically cancels out of the equilibrium indifference condition. This creates a "suicide region" in the investment space where competitive pressures force rational agents to deploy AGI systems early, despite a negative risk-adjusted net present value. Furthermore, we show that "warning shots" (sub-existential disasters) will fail to deter AGI acceleration, as the winner-takes-all nature of the race remains intact. The race can only be halted if the cost of ruin is internalized, making safety research a prerequisite for economic viability. We derive the critical private liability threshold required to restore the option value of waiting and propose mechanism design interventions that can better ensure safe AGI research and socially responsible deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper models the race to AGI between two sovereign actors as a continuous-time preemption game with endogenous existential risk. A systemic ruin parameter D, correlated with development velocity and shared globally, is embedded in both players' payoffs. The central claim is that this risk term cancels from the equilibrium indifference condition between investing immediately (leader value) and waiting (follower continuation value), producing a 'suicide region' in which rational preemption forces early deployment despite negative risk-adjusted NPV. The manuscript further argues that sub-existential warning shots fail to deter acceleration and derives a private liability threshold plus mechanism-design interventions to restore the option value of waiting.

Significance. If the cancellation result is rigorously established, the work supplies a game-theoretic account of why observed AGI acceleration can be consistent with rational behavior under shared catastrophic risk, extending real-options analysis to races with global externalities. The proposed liability threshold and safety-research prerequisites offer concrete policy levers. The model is falsifiable via its predicted dependence of the suicide region on the correlation structure of D and velocity.

major comments (2)
  1. [§3.2, Eq. (15)] §3.2, Eq. (15) (indifference condition): the derivation asserts that D cancels because it enters symmetrically in leader and follower continuation values. However, the leader's stopping time τ_L precedes the follower's τ_F, so the integrated hazard rates over [0,τ] differ unless the post-deployment ruin probability is explicitly independent of role and velocity. The manuscript must display the explicit integral expressions for both continuation values and show that the D terms are identical after substitution; without this step the cancellation is not guaranteed by global sharing alone.
  2. [§4.1, Proposition 2] §4.1, Proposition 2 (suicide region): the existence of the region where NPV < 0 yet investment occurs is load-bearing on the cancellation result. If the D integrals do not cancel, the threshold reverts to the standard real-options form and the suicide region disappears. A direct comparison of the derived threshold with and without the symmetry assumption on D would clarify the scope of the result.
minor comments (2)
  1. [§2] Notation for the hazard rate λ(v) and its dependence on velocity v is introduced in §2 but used without re-statement in the continuation-value integrals of §3; a brief reminder equation would improve readability.
  2. [Figure 2] Figure 2 (investment-space diagram) labels the suicide region but does not indicate the numerical values of D and correlation parameter used to generate the boundaries; adding these parameters would allow readers to reproduce the plotted region.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments, which help clarify the scope of our cancellation result. We address each major comment below, indicating planned revisions to strengthen the exposition without altering the core model.

read point-by-point responses

  1. Referee: [§3.2, Eq. (15)] §3.2, Eq. (15) (indifference condition): the derivation asserts that D cancels because it enters symmetrically in leader and follower continuation values. However, the leader's stopping time τ_L precedes the follower's τ_F, so the integrated hazard rates over [0,τ] differ unless the post-deployment ruin probability is explicitly independent of role and velocity. The manuscript must display the explicit integral expressions for both continuation values and show that the D terms are identical after substitution; without this step the cancellation is not guaranteed by global sharing alone.

    Authors: We agree that explicit verification is needed. In the model, D represents a global existential cost realized upon AGI deployment by either player, with the hazard rate λ(t) driven by cumulative velocity up to the first stopping time. Because the post-deployment ruin is triggered globally and independently of which actor leads (the deployed AGI affects the shared world state), the integrated term -∫ D · λ(s) ds from 0 to τ_L in the leader value equals the corresponding term in the follower continuation value after the leader's deployment (the follower then faces the same global D from τ_L onward). We will insert the full integral expressions for V_L and V_F immediately before Eq. (15), substitute the common D factor, and demonstrate algebraic cancellation under the global-sharing assumption. This addition will also note the modeling choice that post-deployment risk does not depend on role. revision: yes

  2. Referee: [§4.1, Proposition 2] §4.1, Proposition 2 (suicide region): the existence of the region where NPV < 0 yet investment occurs is load-bearing on the cancellation result. If the D integrals do not cancel, the threshold reverts to the standard real-options form and the suicide region disappears. A direct comparison of the derived threshold with and without the symmetry assumption on D would clarify the scope of the result.

    Authors: We accept that the suicide region is conditional on the cancellation. In the revised manuscript we will add a short subsection after Proposition 2 that compares the equilibrium investment threshold under (i) the baseline global D with role-independent post-deployment ruin (yielding the suicide region where investment occurs for NPV < 0) and (ii) a counterfactual where D is either private or role-dependent (in which case the threshold reverts to the standard real-options form with no suicide region). This comparison will be presented both analytically and via a numerical illustration to delineate the precise conditions under which the result holds. revision: yes

Circularity Check

1 steps flagged

Cancellation of systemic ruin parameter D from indifference condition is imposed by symmetric global-sharing assumption rather than derived from differential stopping times

specific steps
  1. self definitional [Abstract]
    "As the disutility of catastrophe is embedded in both players' payoffs, the risk term mathematically cancels out of the equilibrium indifference condition. This creates a 'suicide region' in the investment space where competitive pressures force rational agents to deploy AGI systems early, despite a negative risk-adjusted net present value."

    The paper claims the risk term cancels because D is embedded in both payoffs, directly yielding the suicide region. In a preemption game, however, leader and follower continuation values integrate the hazard over different intervals (earlier deployment for leader). Global sharing alone does not equate these integrals unless the model additionally assumes the ruin probability is role-independent and identical regardless of who moves first. The cancellation is therefore imposed by the symmetric embedding assumption rather than derived, making the suicide region result reduce to that modeling choice.

full rationale

The paper's central result—the suicide region where agents deploy despite negative NPV—rests on the claim that D cancels from the equilibrium indifference condition because it is embedded in both players' payoffs. This cancellation is asserted to follow from D being systemic, correlated with velocity, and shared globally. However, in a continuous-time preemption game the leader's stopping time precedes the follower's, so the continuation values contain distinct integrals over the hazard rate. Cancellation therefore requires an additional modeling restriction that the post-deployment ruin probability is independent of role and identical for both players. The abstract presents this cancellation as a mathematical consequence of embedding, but the skeptic analysis shows it is not automatic from global sharing alone. The result is therefore partially forced by the choice of how D enters the leader and follower values, producing a circularity score of 6.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The model rests on introducing D as a global shared ruin cost correlated with velocity; this parameter drives the cancellation and is not derived from external benchmarks.

free parameters (1)
  • systemic ruin parameter D
    Correlated with development velocity and shared globally; specific functional dependence not derived from first principles.
axioms (1)
  • domain assumption AGI race modeled as continuous-time preemption game between sovereign actors
    Assumes standard game-theoretic payoffs and timing apply directly to US-China AI competition.

pith-pipeline@v0.9.0 · 5786 in / 1347 out tokens · 80720 ms · 2026-05-21T17:58:01.372959+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

  1. [1]

    Acemoglu, D., & Robinson, J. A. (2019). The Narrow Corridor: States, Societies, and the Fate of Liberty. New York, NY: Penguin Books

    work page 2019

  2. [2]

    Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv preprint arXiv:1606.06565

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  3. [3]

    Armstrong, S., Bostrom, N., & Shulman, C. (2016). Racing to the precipice: A model of artificial intelligence development. AI & Society, 31(2), 201–206

    work page 2016

  4. [4]

    Aschenbrenner, L. (2024). Situational Awareness: The Decade Ahead. Retrieved from https://situational-awareness.ai/

    work page 2024

  5. [5]

    Bartlett, L. (Host). (2023, July). Dario Amodei, CEO of Anthropic [Audio podcast episode]. In The Logan Bartlett Show

    work page 2023

  6. [6]

    Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford, UK: Oxford University Press

    work page 2014

  7. [7]

    Bostrom, N., Douglas, T., & Sandberg, A. (2016). The unilateralist’s curse and the case for a principle of conformity. Social Epistemology, 30(4), 350–371

    work page 2016

  8. [8]

    Christiano, P. (2019). What failure looks like. AI Alignment Forum. Retrieved from: https://www.alignmentforum.org/posts/HBxe6wdjxK239zajf/what-failure-looks-like

    work page 2019

  9. [9]

    K., & Pindyck, R

    Dixit, A. K., & Pindyck, R. S. (1994). Investment under Uncertainty. Princeton, NJ: Princeton University Press

    work page 1994

  10. [10]

    Now, here, you see, it takes all the running you can do, to keep in the same place

    Good, I. J. (1965). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31–88. 17 The term refers to the 'Red Queen' effect originally proposed in evolutionary biology by V an Valen (1973), derived from Lewis Carroll’s Through the Looking-Glass: “Now, here, you see, it takes all the running you can do, to keep in the same...

    work page 1965

  11. [11]

    Goertzel, B. (2014). Artificial general intelligence: Concept, state of the art, and future prospects. Journal of Artificial General Intelligence, 5(1), 1–46

    work page 2014

  12. [12]

    F., Thomas, S., Weinstein-Raun, B., & Brauner, J

    Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., & Brauner, J. (2024). Thousands of AI authors on the future of AI. arXiv preprint arXiv:2401.02843

    work page arXiv 2024

  13. [13]

    Grenadier, S. R. (2002). Option exercise games: An application to the equilibrium investment strategies of firms. The Review of Financial Studies, 15(3), 691–721

    work page 2002

  14. [14]

    J., & Kort, P

    Huisman, K. J., & Kort, P. M. (2003). Strategic investment in technological innovations. European Journal of Operational Research, 144(1), 209–223

    work page 2003

  15. [15]

    Jones, C. I. (2024). The A.I. dilemma: Growth vs. existential risk. American Economic Review Insights, 6(4), 575-590

    work page 2024

  16. [16]

    (2024, May 17)

    Lubetkin, D. (2024, May 17). OpenAI dissolves team focused on long-term AI risks. The Wall Street Journal

    work page 2024

  17. [17]

    (2023, May 1)

    Metz, C. (2023, May 1). 'The Godfather of A.I.' Leaves Google and Warns of Danger Ahead. The New York Times. Retrieved from https://www.nytimes.com/2023/05/01/technology/ai-google- chatbot-engineer-quits-hinton.html Naudé, W., & Gries, T. (2022). Breakthroughs, Backlashes and Artificial General Intelligence: An Extended Real Options Approach. IZA Institut...

    work page 2023

  18. [18]

    Nordhaus, W. D. (2021). Are we approaching an economic singularity? Information technology and the future of economic growth. American Economic Journal: Macroeconomics, 13(1), 299–332

    work page 2021

  19. [19]

    O'Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J., & Dafoe, A. (2020). The windfall clause: Distributing the benefits of AI for the common good. Centre for the Governance of AI

    work page 2020

  20. [20]

    Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. New

    work page 2019

  21. [21]

    Suleyman, M. (2023). The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma. New York, NY: Crown. The Economist. (2024a, April 25). China and America are locked in a race to become the world’s leading AI power. The Economist. The Economist. (2024b, May 26). OpenAI’s former board members: The company cannot be trusted to govern...

    work page 2023

  22. [22]

    Weeds, H. (2002). Strategic delay in a real options model of R&D competition. The Review of Economic Studies, 69(3), 729–747

    work page 2002