Agreements Have a Clock: Compute Growth and the Verification of AI Restraint
summary
The gist
Agreements have a clock because compute keeps growing during a compliant pause, so the time a country has to act on a secret advance shrinks even while the capability gap stays fixed The paper…
In short
The paper models how an agreement to pause frontier AI has a finite time limit because computational power continues to grow during the pause, shrinking the window for action. It shows that mutual restraint is supportable if monitoring costs are low and cooperation is sustained by matching safety research and implementing hardware/algorithmic security.
Key concepts
- The Nature of the Time Constraint
- Agreements have a clock because computing power increases during a pause, meaning the time available to act on a secret AI advance shrinks even if the capability gap remains fixed. This shrinking window necessitates higher monitoring to deter advances, which in turn limits what both parties can afford.
- Dynamics of Catch-Up Time
- Followers catch up through two ways: their own research and algorithmic diffusion. Diffusion is fast when the gap is large, causing a halted leader to stay ahead longer than raw lead suggests. This catch-up time decreases as the capability gap closes.
- Value of Mutual Restraint
- Mutual restraint benefits both countries by pausing advancement while continuing safety research, preventing one from gaining an unfair advantage. This cooperation is viable when monitoring costs are less than the gains from cooperating, though each country still has an incentive to advance secretly.
Terminology used across episodes
This episode discusses
- Agreements Have a Clock: Compute Growth and the Verification of AI Restraint · Paper Radio
- Nuclear Arms Control Verification and Lessons for AI Treaties
- Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity · Paper Radio
- Safety Research and the Allocation of Frontier Compute · Paper Radio
The paper
Agreements Have a Clock: Compute Growth and the Verification of AI Restraint · Read on arXiv
Li Gan
Texas A&M University
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Agreements Have a Clock".
Warren: Here's the gist of the paper. Agreements have a clock because compute keeps growing during a compliant pause, so the time a country has to act on a secret advance shrinks even while the capability gap stays fixed The paper investigates what an agreement to pause frontier AI needs by modeling how diffusion fades as the gap closes, leading to a finite response window The Nature of the Time Constraint The core finding is that agreements have a clock because compute keeps growing during a compliant pause, so the time a country has to act on a secret advance shrinks even while the capability gap stays fixed This shrinking window raises the monitoring needed to deter a secret advance, and monitoring costs bound what both countries will pay, so cooperation can stop being supportable before the capability gap closes In an example with a fixed leader window, cooperation is supportable through month 18.9 The Dynamics of Catch-Up Time The follower advances through two distinct mechanisms: own research and algorithmic diffusion Diffusion is fast while the gap is wide and fades as the gap closes, so a halted leader stays ahead longer than the raw lead suggests The catch-up time falls from 12 months at the start to 5.74 months at month 24 The Value of Mutual Restraint Mutual restraint creates gains for both countries, as both pause capability advancement while continuing safety research, so neither concedes progress to an advancing rival These gains can finance verification, but each country still has an incentive to advance secretly Cooperation is supportable when the monitoring needed to deter a secret advance costs no more than either country gains from cooperating The Limits of Supportability Cooperation is supportable only if the follower’s window is at least the implementation delay plus ξi/mmax and the leader’s window also passes During a compliant pause with fixed frontier states, aB > 0, γB > 0, and the maintained δ ≥ 0, Tcatch(d) is strictly decreasing and converges to zero The clock runs faster when the two countries are more alike Policy Levers for Extension Three policy channels do distinct work: domestic capability and safety standards can restrain commercial racing and raise safety effort, but a constant research share must stay at about 0.85% or more to keep the foreign lead Hardware controls and algorithmic security lengthen catch-up time (Table 5), widening both the unilateral pause limit and the response window Verified reciprocal restraint works only if restarting is credible (14), the monitoring intensity lies inside the deterrence-and-participation interval (15), and catch-up time exceeds the window bound (17) The Critical Window The critical window is that of the country that binds at the last supportable date The clock comes from the physics; the date comes from chosen inputs Monitoring matters more between near equals because a follower with the leader’s compute, still six months behind, catches up in 4.4 months at the start of the pause, against 12.0 in the main example, so deterring it takes about 2.7 times as much monitoring The critical window is that of the country that binds at the last supportable date The Role of Matched Safety Research An agreement can lengthen its own clock by committing both countries to the same safety research and verify that commitment along with the pause With matched safety research, a compliant 36-month mutual pause lowers the global catastrophe probability from 2.3% to 1.
Title and authors: Ruth: So we're looking at this paper called "Agreements Have a Clock: Compute Growth and the Verification of AI Restraint," and it's about how agreements to pause frontier AI actually have a time limit built into them.
Warren: That sounds like something that could really matter for how we think about international competition right now. What’s the main takeaway when you first look at this?
Ruth: Basically, the paper argues that because computers keep getting faster even when everyone agrees to pause development, the window a country has to act on a secret advance shrinks.
Iris: So if I'm hearing you correctly, it’s not just about the initial gap between countries changing; it's about how compute growth during that pause shortens the response time for both sides.
Ruth: Exactly. The core finding is that agreements have a clock because compute keeps growing during a compliant pause, so the time a country has to act on a secret advance shrinks even while the capability gap stays fixed.
Warren: That shrinking window means you need more monitoring to keep things safe, and those monitoring costs end up setting limits on what both sides are willing to pay for cooperation.
Iris: And they show that if the leader has a set window, cooperation is supportable through month eighteen point nine of the pause time.
Ruth: That eighteen point nine months is just an example based on fixed national values, but it shows how long you can actually keep cooperating under these conditions without breaking down.
Warren: It sounds like if you want to sustain that cooperation, you need to figure out what the actual cost of monitoring is versus what both sides gain by cooperating.
Iris: And they model the catch-up time, showing it drops from twelve months at the start down to five point seven four months by month twenty-four.
Ruth: That catch-up time dropping suggests that diffusion is fast when the gap is wide, but it fades as the gap closes, meaning a paused leader stays ahead longer than you might expect from just looking at their raw lead.
Warren: So if you're trying to keep up with a secret advance, you have to account for that catch-up time dynamic when planning your response.
Iris: And they point out that mutual restraint creates gains for both sides because both countries pause capability advancement while still doing safety research, so neither side gives up progress to a rival.
Ruth: Those gains can actually be used to finance the verification process, but each country still has an incentive to advance secretly, which is why the cooperation has its own limits.
Title and authors: Warren: So what’s the threshold for when mutual restraint stops being sustainable? What's that point where it becomes too risky?
Iris: The paper says cooperation is supportable only if the follower’s window is at least the implementation delay plus xi over m max, and also for the leader’s window to pass.
Ruth: That sounds like a very specific mathematical condition that needs to be met for any kind of agreement to hold up.
Warren: They also show that during a compliant pause with fixed frontier states, the value of B is greater than zero, gamma B is greater than zero, and the maintained delta is greater than or equal to zero.
Iris: And when those conditions are met, catch-up time Tcatch d is strictly decreasing and it converges to zero as the pause gets longer.
Ruth: That convergence to zero suggests that no matter how long you pause, the window eventually closes, which brings us to the critical window concept.
Warren: The paper defines a critical window as that of the country that binds at the last supportable date, which is determined by both physics and your chosen inputs.
Iris: And they emphasize that monitoring matters more between near equals because if a follower has the leader’s compute but is still six months behind, they catch up in four point four months at the start of the pause against twelve point zero in the main example.
Ruth: That difference shows how much more intense monitoring is needed when you are close to parity, because deterring a secret advance takes about two point seven times as much monitoring.
Warren: So for someone just listening, it means that if you're not totally equal, the risk and the required vigilance change dramatically based on how close you are.
Iris: They also discuss policy levers for extension: domestic capability and safety standards can restrain commercial racing and raise safety effort.
Ruth: But they stress that a constant research share must stay at about zero point eight five percent or more to keep the foreign lead in check, otherwise catch-up time doesn't lengthen as much.
Warren: And hardware controls and algorithmic security are another lever to extend catch-up time, which widens both the unilateral pause limit and the response window by extending that catch-up time Tcatch d over F d and FB d.
Iris: Verified reciprocal restraint only works if restarting development is credible, meaning you have to make sure that restarting is a real option.
Title and authors: Ruth: That credibility comes from setting monitoring intensity m such that it satisfies min D plus l, R less than Tstar G zero point zero in order to preserve a positive gap after a secret advance.
Warren: So it seems like the strategy isn't just about pausing; it's also about using specific levers like matched safety research and verified restraint to manage the time constraint.
Iris: The paper also shows that verified matched safety research can lengthen its own clock by committing both countries to the same safety research and verifying that commitment along with the pause.
Ruth: With matched safety research, a compliant thirty-six month mutual pause lowers the global catastrophe probability from two point three percent down to one percent.
Warren: That drop in probability is significant, especially when you consider how much compute growth is happening during that time period.
Iris: So while the clock runs faster when countries are more alike, matching safety research seems like a concrete way to slow down the clock for everyone involved.
Ruth: This paper, "Agreements Have a Clock: Compute Growth and the Verification of AI Restraint," really shows how dynamics change when we try to manage high-stakes technological competition through international agreements.
Warren: It’s clear that simply agreeing to pause isn't enough; you have to build in mechanisms for verification and monitoring that account for the fact that compute keeps growing during those pauses.
Iris: The complexity here lies in balancing the incentive to advance secretly against the cost of monitoring, which is what sets the ultimate limit on cooperation.
Ruth: This brings us to where we are with this paper. It suggests that sustained restraint isn't just about setting a pause button; it’s about managing a dynamic clock that speeds up under certain conditions and requires continuous calibration.
Warren: And for the future, it seems like the focus needs to be on those levers—hardware security and matched safety research—to actually extend the window of supportability beyond what static estimates suggest.
Iris: I think the most important part is understanding that the critical window isn't just a fixed number; it’s defined by which country binds at the last possible supportable date based on their specific parameters.
Ruth: Exactly, and we need to keep thinking about those dynamics as compute keeps growing and we try to manage these agreements.
Warren: That’s all for this paper on "Agreements Have a Clock: Compute Growth and the Verification of AI Restraint." We’ll be back next time to look at some other interesting stuff from arXiv.
The paper's summary: Ruth: So, to recap, this paper looks at how agreements to pause frontier AI actually have an expiration date because computer power keeps growing while everyone is waiting for those pauses to finish.
Warren: Right. The core idea is that if you agree to a pause today, the time you actually have left before the country getting ahead of the curve gets too big shrinks, even though that gap between your technology and theirs stays exactly where it is.
Iris: I see what they're doing here with modeling how diffusion fades as the capability gap closes. It shows that while raw compute growth is constant, the speed at which a leader can actually pull ahead slows down as everyone starts catching up.
Ruth: Exactly. And that’s why the time you have to act on a secret advance gets tighter even though the technology gap itself isn't changing much during the pause. It raises what you need to monitor.
Warren: Monitoring costs bound how much both countries will spend, which sets a limit on how long this whole cooperation can last before it becomes too expensive for either side. They find that with a fixed leader window, cooperation is supportable through eighteen point nine months of that pause time.
Iris: That number is based on a specific scenario with a fixed leader window, but it points to the fact that there’s an actual finite response window built into these agreements.
Ruth: And they look at how the follower actually catches up—it happens through two ways: their own research and algorithmic diffusion. The diffusion part is fast when the gap is wide, but it slows down as the gap closes.
Warren: So a leader who stays ahead longer than their raw lead suggests can still be caught up, but only over time. They show that catch-up time falls from twelve months at the start to five point seven four months by month twenty-four of the pause.
Iris: That drop in catch-up time is interesting because it shows that the dynamics of how fast you close a gap change depending on whether you're far behind or closer.
Ruth: Mutual restraint itself creates gains for both countries because both sides can pause their capability advancement while they still keep doing safety research. Neither side completely gives up progress to the other.
Warren: Those gains can actually pay for the verification process, but each country still has an incentive to sneak ahead secretly, which is why that cooperation isn't permanent on its own.
Iris: They lay out the conditions for supportability pretty clearly—cooperation only works if the follower’s window is long enough to cover the implementation delay plus some factor of xi over m max, and the leader’s window also passes.
Ruth: And they show that if you maintain certain values for B, gamma B, and delta—all positive or zero—then catch-up time strictly decreases and eventually goes to zero as the pause gets longer.
The paper's summary: Warren: That convergence to zero is a big thing. It means no matter how long you pause, the window eventually closes because compute keeps growing during that compliant period.
Iris: The paper defines a critical window as being held by the country that binds at the last supportable date, which depends on physics and your chosen inputs.
Ruth: They show that monitoring matters more when you’re near equals because if you have the same compute but are still six months behind, you catch up in four point four months at the start of a pause against twelve point zero in their main example.
Warren: So deterring a secret advance takes about two point seven times as much monitoring when you’re close to parity, which means being near equals is riskier than being far apart.
Iris: They also give policy levers for extension: raising domestic safety standards can restrain commercial racing, but they stress that a constant research share must stay at about zero point eight five percent or more to keep the foreign lead in check.
Ruth: Hardware controls and algorithmic security are another lever to extend catch-up time, widening both the unilateral pause limit and the response window by extending that catch-up time.
Warren: And verified reciprocal restraint only works if you can prove that restarting development is credible, which they do by setting monitoring intensity m such that it keeps a positive gap after a secret advance.
Iris: They also show that matching safety research—committing both countries to the same safety research and verifying it—can actually lengthen its own clock by committing both sides to the same effort.
Ruth: With matched safety research, they find that a compliant thirty-six month mutual pause lowers the global catastrophe probability from two point three percent down to one percent.
Warren: That drop in probability is significant, especially when you think about how much compute growth is happening while those countries are paused.
Iris: So while the clock runs faster when countries are more alike, matching safety research seems like a concrete way to slow down that clock for everyone involved.
Ruth: This paper really shows how managing high-stakes technological competition through international agreements is a dynamic process that requires constant calibration based on those time constraints.
Warren: It suggests simply agreeing to pause isn't enough; you have to build in mechanisms for verification and monitoring that account for the fact that compute keeps growing during those pauses.
Iris: The complexity here is balancing the incentive to sneak ahead secretly against the cost of monitoring, which ultimately sets the limit on how long cooperation can last.
Ruth: It points toward needing those specific levers—like hardware security and matched safety research—to actually extend the window of supportability beyond what simple static estimates suggest.
Warren: And for the future, it seems like we need to focus on those mechanisms to manage that dynamic clock better.
The paper's improvements: Tom: So, we're looking at how they suggest ways to actually make these agreements work better, and what those fixes mean for real policy.
Ruth: The authors propose three specific levers to extend the time these agreements last. First, they say you need to increase domestic safety research to match foreign levels so it lowers that global catastrophe probability.
Warren: That makes sense from a risk perspective; if you both invest in safety equally, the incentive to rush ahead secretly goes down significantly. They find this commitment can extend cooperation supportable through thirty-one point eight months in the baseline scenario.
Iris: They also suggest implementing hardware controls and algorithmic security because that lengthens catch-up time itself, which widens both the unilateral pause limit and the response window by extending that catch-up time Tcatch d over F d and FB d.
Ruth: That’s an engineering solution to a timing problem; you build in more friction into the system so diffusion doesn't happen as fast. It widens both limits, which is good for both sides.
Warren: If you can extend the window, it buys time for verification and monitoring to catch up with the technology before things get out of control. But they also stress that verified reciprocal restraint only works if restarting development is credible.
Iris: Credibility comes from setting that monitoring intensity m to satisfy a specific condition, min D plus l, R less than Tstar G zero point zero in order to preserve a positive gap after a secret advance. That’s the identification strategy they use to ensure the gap stays open.
Ruth: So it’s not just about pausing; it’s about using safety standards and tech security to actively manage that clock. It changes how we think about what policy levers are actually effective in this kind of race.
Warren: It moves us from just a theoretical limit to a practical set of conditions you can try to meet with actual governance. But they also point out the trade-off: when countries are very similar, the clock runs faster, so matching research is key.
Iris: That’s where it gets interesting—the convergence behavior means that even if you're close to parity, things change quickly. So matching safety research seems like a concrete way to slow down that clock for everyone involved.
Ruth: If you want to keep this cooperation going long-term, the focus has to shift toward those specific commitments and verification methods rather than just relying on a simple pause button.
Warren: It’s about building in resilience so that when the technology inevitably advances, there's still a window left for everyone to react without it immediately becoming an unstoppable force.
Conclusion: Ruth: So, to wrap up, this paper on "Agreements Have a Clock: Compute Growth and the Verification of AI Restraint" shows that pausing frontier AI isn't a static decision; it’s a dynamic constraint that shrinks as computers get faster during the pause.
Warren: Basically, the main implication is that you can’t just set a date on paper; you have to build in mechanisms to manage that time window because compute keeps growing while everyone is waiting.
Iris: I think what this changes for us is realizing that simply agreeing to a pause isn't enough if you don't account for how fast the technology diffuses or how quickly the leaders can just pull ahead.
Ruth: Exactly, and they show that things like hardware controls and matched safety research are the levers we need to actually extend those supportable dates. They give us concrete policy ideas instead of just abstract warnings about a closing window.
Warren: The strategy isn't just about setting a pause button; it’s about using specific verification methods to keep that window open longer, which is what really matters for keeping the competition in check.
Iris: And they do point out that the critical window isn't just a fixed number; it’s defined by which country binds at the last possible supportable date based on their specific parameters. That makes it very personalized to each scenario.
Ruth: So, as we look forward, this paper suggests that we need to focus less on just the initial agreement and more on those continuous calibration mechanisms—the monitoring and research commitments—to keep the restraint viable.
Warren: It’s about moving from a simple pause strategy to a much more complex system of dynamic management that accounts for growth during the wait time.
Iris: I think understanding these dynamics is important because it forces us to look at the costs and benefits of different verification strategies, which ties into other work we’re doing on how we measure policy uncertainty.
Ruth: Well, this was a really thought-provoking look at the incentives behind AI restraint in international competition.
Warren: It definitely shows that if you’re trying to run a long-term strategy like this, you can’t ignore the underlying physics of compute growth.
Iris: It's a strong reminder that econometric models need to be grounded in how the underlying technology actually scales over time, not just static assumptions.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19