It might be worth trying to coordinate a slowdown between AI labs, rather than a pause.
I could be wrong about this, so sorry if this has been suggested elsewhere, but I don’t think I’ve really seen this idea anywhere—We coordinate frontier labs to iteratively slow down deployments.
I think most pause advocates were pushing for immediate hard stops, like the Future of Life Institute’s “Pause Giant AI Experiments” open letter explicitly called for an “immediate pause for at least 6 months” on training systems more powerful than GPT‑4. But there’s obvious reasons why that isn’t palatable to labs.
Most public “pause” advocacy has been framed as interventions at the frontier: stop training above a capability threshold now (at least for a period of months). There’s a moral clarity there… but it also raises the salience of the exact objection that labs raise, any lab that slows alone risks losing: first-mover advantage, ecosystem lock-in, and investor confidence.
A phased slowdown for frontier AI releases could be framed like a reciprocal arms-control measure and not unilateral stopping. This would also come with some benefits: a way to lengthen decision time, reduce race pressures, and preserve optionality while still avoiding the commercial and political shocks of a hard stop.
Let’s take the “AI arms race” framing seriously here for a second and recall that historically, major arms-control agreements worked through things like ceilings, timetables, verification, and phased reductions rather than demands to immediately cease all weapons research or deployment.
A couple of frontier-lab leaders have indicated a slower pace would be desirable if it could be coordinate. Demis Hassabis said a slightly slower pace might be better for society and Dario Amodei said he’d prefer such a slowdown… if it were enforceable across competitors. So there’s appetite, it’s just a matter of getting buy-in, and maybe you can make the deal more attractive.
Some historical analogs
The START (Strategic Arms Limitation Talks and subsequent) agreements didn’t require the United States or the Soviet Union/Russia to stop all at the same time. First they just agreed to limits and then, in later treaties, verifiable reductions over time.
START I used phased implementation over years, and New START gave the parties seven years after entry into force to meet central warhead and launcher limits.
Obviously AI isn’t identical to nuclear weapons, but the relevant takeaway is that rivals often accept gradual reciprocal constraint more readily than immediate unilateral restraint…
So there could be a negotiated AI trajectory that slows the competitive cycle while preserving mutual visibility and the ability to respond to defection. A phased slowdown just asks labs to stretch out the interval between frontier releases by a small amount at each step, so that everyone slows together and no signatory gives up much relative position at any given time step.
This preserves option value for a given lab: if an outside actor defects or circumstances change, participants can collectively shorten intervals again instead of remaining frozen with costly startup times.
A simple calendar time illustration
For the sake of simplicity, let’s use calendar time as the deciding variable for this slowdown schema, though you might be able to use other things like model size.
Let’s assume the core commitment is a minimum time gap between frontier releases by the same lab. You might define a “frontier release” as “the first public deployment of a model that exceeds agreed capability or compute thresholds”, minor product updates and patches wouldn’t count for these purposes.
The agreement starts from a baseline interval. Let’s say for example everyone agrees to delay their next model release by 2 weeks, and then we lengthen that minimum interval by a small increment after each qualifying release (Ex. by two weeks at a time) until it reaches a negotiated ceiling such as 12 or 24 week or whatever. That creates the iterative “ease off the gas” effect, without forcing labs to jump overnight from rapid release cycles to a dead stop.
You take the spare time and explicitly have it allocated to deeper evaluations for safety like third-party red-teaming, external review, and publication of model-risk summaries… before the next frontier step.
That makes the delay legible as a governance checkpoint, and a safety measure, instead of cartel-like suppression of model launches.
Why Chinese firms might be interested
There’s some public evidence that leading Chinese models remain several months behind the U.S. frontier on average, even if the gap is smaller than many observers assumed a few years ago.
Demis Hassabis said Chinese models may be only a matter of months behind Western models and Bloomberg summarized his estimate as about six months, let’s assume 3-6 months.
Right now the frontier moves at roughly a 4-6 week release cadence. If Chinese labs like Alibaba, DeepSeek, and MiniMax are on roughly monthly release schedules then a coordinated 1-2 week delay per quarter is a fairly marginal cost to a company like DeepSeek specifically. That’s because they’re already in a position where a few weeks represents a small fraction of their existing gap. The asymmetry might actually work in their favor as a negotiating position. They would give up little, and if the US labs were to slow down too, they would benefit proportionally more because the gap they’re closing narrows faster relative to the total, though still not by a ton in the grand scheme of things.
If that characterization is right, the marginal cost to top Chinese firms of delaying an additional few weeks could be lower than the cost to a leading U.S. lab. especially if the agreement preserves catch-up opportunities for them. And so if I’m a Chinese AI lab maybe that makes me more likely to agree.
Note that is structurally similar to how SALT worked. Soviets researchers were slightly behind the US in delivery systems when early talks began, and so a freeze that locked in a (relatively) little US advantage was still preferable to an unconstrained race they could end up losing badly.
The political attraction is that we aren’t asking Chinese firms, or any firm, to concede the race entirely. We’re asking all participants to slow the cadence of frontier transitions while retaining the option to speed up again if the other side defects.
Verification and institutional design
A calendar-time protocol is attractive because it’s easy to publicly observe. The dates of frontier releases are always clear even if proprietary technical details are still obscured. That lowers the verification burden compared with a pure compute-cap regime (although some compute and capability thresholds would likely still be needed, just to determine which releases actually count)
A lightweight version could be done with public commitments, model trackers, and independent scorekeeping by outside groups. Meanwhile a stronger version of this plan might combine those public commitments with government reporting requirements for things like: large training runs, safety case disclosures, common red-team standards, etc. Together it creates a hybrid system of industry norms and regulatory backstop.
However, the structural advantages of this scheme would be that it wouldn’t really even require government intervention.
How this differs from a pause
Let’s be clear about this: A phased slowdown scheme isn’t claiming that current systems are already safe enough or that a hard pause is always unjustified. I view it more as adapting to meet our circumstances. Getting a voluntary commitment to a pause seems unlikely, and every day the apparatus moves faster and faster. There’s a different intervention logic here: instead of one immediate cliff, we have a staircase of reciprocal delays, buying society more time to adapt while development goes on. As fast as the game is now moving every second we might be able to buy could end up making a real difference. Will Macaskill has talked about the difference that even a month long pause could make, why not just spread that out over a longer time span?
Hard pauses inevitably bring up arguments about unilateral disarmament, enforceability, and sudden economic disruption, but a phased protocol can be defended as competitive stabilization. We’re just slowing the rate of escalation while keeping the door open to renegotiation, verification, and emergency acceleration if things dramatically change like there’s a real breakthrough in alignment.
In Sum
Frontier AI labs and governments should be open to pursuing a reciprocal, calendar-based (or other phase based) slowdown in frontier model releases.
The goal wouldn’t be to stop all AI progress immediately, but to create progressively longer intervals between frontier releases. We can use that time for evaluations and governance, while still preserving the option to accelerate again if a major actor defects, we come into conditions that make moving faster safer, or an emergency requires it.
This approach fits the historical pattern of serious arms control schemas better than abrupt pauses. In other contexts successful de-escalation between competing states was obtained by limiting tempo and scale first, and then building the trust and verification institutions needed for stronger constraints.
What about AI slowdowns instead of pauses?
It might be worth trying to coordinate a slowdown between AI labs, rather than a pause.
I could be wrong about this, so sorry if this has been suggested elsewhere, but I don’t think I’ve really seen this idea anywhere—We coordinate frontier labs to iteratively slow down deployments.
I think most pause advocates were pushing for immediate hard stops, like the Future of Life Institute’s “Pause Giant AI Experiments” open letter explicitly called for an “immediate pause for at least 6 months” on training systems more powerful than GPT‑4. But there’s obvious reasons why that isn’t palatable to labs.
Most public “pause” advocacy has been framed as interventions at the frontier: stop training above a capability threshold now (at least for a period of months). There’s a moral clarity there… but it also raises the salience of the exact objection that labs raise, any lab that slows alone risks losing: first-mover advantage, ecosystem lock-in, and investor confidence.
A phased slowdown for frontier AI releases could be framed like a reciprocal arms-control measure and not unilateral stopping. This would also come with some benefits: a way to lengthen decision time, reduce race pressures, and preserve optionality while still avoiding the commercial and political shocks of a hard stop.
Let’s take the “AI arms race” framing seriously here for a second and recall that historically, major arms-control agreements worked through things like ceilings, timetables, verification, and phased reductions rather than demands to immediately cease all weapons research or deployment.
A couple of frontier-lab leaders have indicated a slower pace would be desirable if it could be coordinate. Demis Hassabis said a slightly slower pace might be better for society and Dario Amodei said he’d prefer such a slowdown… if it were enforceable across competitors. So there’s appetite, it’s just a matter of getting buy-in, and maybe you can make the deal more attractive.
Some historical analogs
The START (Strategic Arms Limitation Talks and subsequent) agreements didn’t require the United States or the Soviet Union/Russia to stop all at the same time. First they just agreed to limits and then, in later treaties, verifiable reductions over time.
START I used phased implementation over years, and New START gave the parties seven years after entry into force to meet central warhead and launcher limits.
Obviously AI isn’t identical to nuclear weapons, but the relevant takeaway is that rivals often accept gradual reciprocal constraint more readily than immediate unilateral restraint…
So there could be a negotiated AI trajectory that slows the competitive cycle while preserving mutual visibility and the ability to respond to defection. A phased slowdown just asks labs to stretch out the interval between frontier releases by a small amount at each step, so that everyone slows together and no signatory gives up much relative position at any given time step.
This preserves option value for a given lab: if an outside actor defects or circumstances change, participants can collectively shorten intervals again instead of remaining frozen with costly startup times.
A simple calendar time illustration
For the sake of simplicity, let’s use calendar time as the deciding variable for this slowdown schema, though you might be able to use other things like model size.
Let’s assume the core commitment is a minimum time gap between frontier releases by the same lab. You might define a “frontier release” as “the first public deployment of a model that exceeds agreed capability or compute thresholds”, minor product updates and patches wouldn’t count for these purposes.
The agreement starts from a baseline interval. Let’s say for example everyone agrees to delay their next model release by 2 weeks, and then we lengthen that minimum interval by a small increment after each qualifying release (Ex. by two weeks at a time) until it reaches a negotiated ceiling such as 12 or 24 week or whatever. That creates the iterative “ease off the gas” effect, without forcing labs to jump overnight from rapid release cycles to a dead stop.
You take the spare time and explicitly have it allocated to deeper evaluations for safety like third-party red-teaming, external review, and publication of model-risk summaries… before the next frontier step.
That makes the delay legible as a governance checkpoint, and a safety measure, instead of cartel-like suppression of model launches.
Why Chinese firms might be interested
There’s some public evidence that leading Chinese models remain several months behind the U.S. frontier on average, even if the gap is smaller than many observers assumed a few years ago.
Demis Hassabis said Chinese models may be only a matter of months behind Western models and Bloomberg summarized his estimate as about six months, let’s assume 3-6 months.
Right now the frontier moves at roughly a 4-6 week release cadence. If Chinese labs like Alibaba, DeepSeek, and MiniMax are on roughly monthly release schedules then a coordinated 1-2 week delay per quarter is a fairly marginal cost to a company like DeepSeek specifically. That’s because they’re already in a position where a few weeks represents a small fraction of their existing gap. The asymmetry might actually work in their favor as a negotiating position. They would give up little, and if the US labs were to slow down too, they would benefit proportionally more because the gap they’re closing narrows faster relative to the total, though still not by a ton in the grand scheme of things.
If that characterization is right, the marginal cost to top Chinese firms of delaying an additional few weeks could be lower than the cost to a leading U.S. lab. especially if the agreement preserves catch-up opportunities for them. And so if I’m a Chinese AI lab maybe that makes me more likely to agree.
Note that is structurally similar to how SALT worked. Soviets researchers were slightly behind the US in delivery systems when early talks began, and so a freeze that locked in a (relatively) little US advantage was still preferable to an unconstrained race they could end up losing badly.
The political attraction is that we aren’t asking Chinese firms, or any firm, to concede the race entirely. We’re asking all participants to slow the cadence of frontier transitions while retaining the option to speed up again if the other side defects.
Verification and institutional design
A calendar-time protocol is attractive because it’s easy to publicly observe. The dates of frontier releases are always clear even if proprietary technical details are still obscured. That lowers the verification burden compared with a pure compute-cap regime (although some compute and capability thresholds would likely still be needed, just to determine which releases actually count)
A lightweight version could be done with public commitments, model trackers, and independent scorekeeping by outside groups. Meanwhile a stronger version of this plan might combine those public commitments with government reporting requirements for things like: large training runs, safety case disclosures, common red-team standards, etc. Together it creates a hybrid system of industry norms and regulatory backstop.
However, the structural advantages of this scheme would be that it wouldn’t really even require government intervention.
How this differs from a pause
Let’s be clear about this: A phased slowdown scheme isn’t claiming that current systems are already safe enough or that a hard pause is always unjustified. I view it more as adapting to meet our circumstances. Getting a voluntary commitment to a pause seems unlikely, and every day the apparatus moves faster and faster. There’s a different intervention logic here: instead of one immediate cliff, we have a staircase of reciprocal delays, buying society more time to adapt while development goes on. As fast as the game is now moving every second we might be able to buy could end up making a real difference. Will Macaskill has talked about the difference that even a month long pause could make, why not just spread that out over a longer time span?
Hard pauses inevitably bring up arguments about unilateral disarmament, enforceability, and sudden economic disruption, but a phased protocol can be defended as competitive stabilization. We’re just slowing the rate of escalation while keeping the door open to renegotiation, verification, and emergency acceleration if things dramatically change like there’s a real breakthrough in alignment.
In Sum
Frontier AI labs and governments should be open to pursuing a reciprocal, calendar-based (or other phase based) slowdown in frontier model releases.
The goal wouldn’t be to stop all AI progress immediately, but to create progressively longer intervals between frontier releases. We can use that time for evaluations and governance, while still preserving the option to accelerate again if a major actor defects, we come into conditions that make moving faster safer, or an emergency requires it.
This approach fits the historical pattern of serious arms control schemas better than abrupt pauses. In other contexts successful de-escalation between competing states was obtained by limiting tempo and scale first, and then building the trust and verification institutions needed for stronger constraints.