Technical Summary: Racing to Ruin

1. Problem Statement

The paper investigates R&D competition between duopolists in the presence of an existential risk: advancing the technology frontier increases the hazard rate of a catastrophic event that permanently drives all firms' flow payoffs to zero. Unlike standard patent race models where the goal is to be the first to capture a prize, here the objective is to determine how far the technology frontier advances before firms voluntarily stop.

The core tension arises from the trade-off between competitive incentives and coordination. Flow profits increase with a firm's own technology but decrease with the rival's. When risk is low, the competitive motive dominates, and firms race regardless of the rival's actions. When risk is high relative to marginal profits, firms prefer to stop if and only if their rival also stops, transforming the problem into a coordination game. The paper asks:

1. How far does the technology frontier advance in equilibrium?
2. How do transparency (the speed of monitoring a rival's stop) and trust (belief in the rival's rationality) shape this frontier?
3. Under what conditions does "racing to ruin" (infinite scaling leading to certain disaster) become an equilibrium outcome?

2. Methodology and Model Setup

The authors develop a continuous-time dynamic game between two firms ($N=\{1,2\}$).

• Technology and Stopping: Firms advance technology at unit speed until they stop. Stopping is irreversible. In the baseline, firms perfectly and immediately observe the rival's stopping decision.
• Payoffs: Firm $i$ receives flow payoff $\pi(t_i, t_j)$, where $t_i$ is its own technology level and $t_j$ is the rival's. The function $\pi$ is strictly increasing in own technology and strictly decreasing in the rival's.
• Disaster Risk: A market collapse occurs at a rate $\lambda(T)$, where $T = \max(t_1, t_2)$ is the frontier technology scale. Upon collapse, all payoffs drop to zero. The hazard $\lambda$ is non-decreasing.
• Representative Agent Benchmark: The authors construct a "representative firm" that holds misspecified beliefs: it assumes the rival's technology is frozen at the current time $t$. This agent solves an optimal stopping problem comparing the marginal gain of scaling against the disaster risk. The solution, $\tau_R$, serves as a critical benchmark.
• Extensions: The baseline model is extended to:
  1. Imperfect Transparency: Stopping is observed with a random delay $E \sim \text{Exp}(\omega)$.
  2. Imperfect Trust: Firms may be "crazy types" (Kreps-Wilson style) that never stop, with probability $1-p$.

3. Key Results

A. Perfect Transparency (Baseline)

Under perfect monitoring and common knowledge of rationality, the equilibrium frontier is bounded within a specific interval $[\tau, \tau_R]$.

• Upper Bound ($\tau_R$): The frontier never exceeds the stopping time of the representative agent ($\tau_R$). If a firm stops at $\tau_R$, the rival observes this immediately and also stops. Backward induction ensures that no firm continues past $\tau_R$.
• Lower Bound ($\tau$): The frontier advances at least to $\tau$, the stopping time of a monopolist (where the rival's technology is zero). Due to log-supermodularity, the marginal return to scaling is higher when the rival is active, ensuring the frontier exceeds the monopoly level.
• Equilibrium Structure: Equilibria are either "bunched" (both stop simultaneously at $\tau_R$) or "spread" (one stops earlier, the other stops later, but the frontier remains within $[\tau, \tau_R]$).
• Comparative Statics: Higher hazard rates pull both bounds inward. Stronger competitive intensity (higher sensitivity of profits to relative position) pushes the upper bound $\tau_R$ outward.

B. Imperfect Transparency (Delayed Monitoring)

When stopping is observed with a lag (rate $\omega$), the coordination mechanism is disrupted.

• Racing to Ruin: If monitoring is sufficiently noisy ($\omega < \bar{\omega}$), there exist equilibria where both firms race forever, leading to disaster with probability one. The delay prevents a unilateral stop from being reciprocated in time to save the other firm.
• Overshooting the Ceiling: When monitoring is precise ($\omega > \bar{\omega}$), every equilibrium stops in finite time. However, the frontier may overshoot the perfect-transparency ceiling $\tau_R$.
• Quantitative Bound: Under a "tail condition" (ensuring competitive losses do not vanish and hazards are bounded), the expected overshoot is $O(1/\omega)$. As monitoring becomes perfect ($\omega \to \infty$), the frontier converges to $\tau_R$.
• Mechanism: Firms engage in a "war of attrition," waiting to see if the rival stops first. The cost of waiting is the continued hazard and erosion of profits, while the benefit is the small gain of being the sole survivor during the rival's detection lag.

C. Imperfect Trust (Crazy Types)

When firms are unsure if their rival is rational (probability $p$ of being rational), the dynamics change significantly. The authors identify three zones of trust ($\ell = p/(1-p)$):

1. Low Trust ($\ell < \underline{\ell}$): Every equilibrium results in racing to ruin. The fear that the rival is a "crazy type" who will never stop makes unilateral stopping too costly.
2. Intermediate Trust ($\underline{\ell} \le \ell \le \bar{\ell}$): Both "immediate joint stopping" and "racing to ruin" are equilibria. Firms can coordinate to stop, but the temptation to wait for verification (free-riding on the signal) sustains the bad equilibrium.
3. High Trust ($\ell > \bar{\ell}$): In every equilibrium, the probability of racing forever (conditional on both being rational) vanishes quadratically in the prior odds ratio $(\bar{\ell}/\ell)^2$.

The Double-Edged Sword of Transparency:
Transparency interacts differently with the two coordination modes:

• Sequential Coordination (Stopping First): Faster monitoring ($\uparrow \omega$) helps by increasing the prize of reciprocation (the rival stops sooner after detection).
• Simultaneous Coordination (Stopping Together): Faster monitoring hurts by reducing the cost of the deviation "wait and verify." If monitoring is fast, a firm can wait to confirm the rival stopped before stopping itself, undermining simultaneous coordination.
  Consequently, increasing transparency can first destroy the "stop together" equilibrium (moving the system from Medium to Low trust) before eventually enabling "stop first" coordination (moving from Low to High trust).

4. Contributions and Significance

The paper contributes to the literature on dynamic R&D, wars of attrition, and the economics of existential risk.

• Strategic Source of Excessive Risk: The authors identify a strategic mechanism for excessive risk-taking that persists even among risk-neutral, patient agents. The inability to coordinate on stopping, driven by transparency and trust frictions, can lead to disaster even when all firms would prefer to stop.
• Bounds on the Frontier: The paper establishes that under perfect information, the frontier is strictly bounded by monopoly and representative-agent thresholds, providing a clear theoretical limit to "racing" behavior.
• Role of Information and Beliefs: The analysis reveals that transparency is not unambiguously beneficial. While it prevents infinite racing in the long run, it can destabilize intermediate coordination regimes. Similarly, trust is a critical determinant of whether the "bad" equilibrium (ruin) is selected.
• Policy Implications: The results suggest that in high-stakes technology races (e.g., AI), the design of monitoring regimes and the cultivation of trust are crucial. Simply making actions more transparent may not suffice to prevent ruin if it simultaneously incentivizes free-riding on verification.

The paper concludes that while perfect transparency recovers the efficient ceiling in the limit, imperfect transparency and trust create a complex landscape where "racing to ruin" can be a self-enforcing equilibrium, highlighting the fragility of coordination in the shadow of disaster.