On the whole, I think PauseAI’s strategy is close to the best thing to be doing at current margins. The main reason why I haven’t donated to PauseAI is this:
Frontier AI is being developed in the US and, to a lesser extent, China. PauseAI does not operate in the US or China. Even if PauseAI-style advocacy produces interest in a pause in the UK, France, etc., I expect there’s a large penalty in terms of how this converts to frontier-AI-relevant policy, because the US and China are by far the most important countries.
I want to hear how you would respond to that. What’s your theory of change for how PauseAI increases the odds of a treaty that the US and China participate in?
Every country has some influence with other countries. If governments around the world push for a treaty between the US and China on AI, that makes it more likely that such a treaty will occur. Australia, for example, has relatively good relations with the US and is a major trading partner of China. We have some capacity to influence those nations on this issue.
Is this guaranteed? No. My gut feeling from the start was that there was about a 15% chance of a pause treaty happening. Recent events have increased that in my thinking, but it is still nowhere near 50%.
Naturally countries have influence on each other, but there’s a penalty factor, and the question is, how big is that penalty?
Route 1: US advocacy convinces US policy-makers to establish a treaty
Route 2: Australian advocacy convinces Australian policy-makers to establish a treaty (where the most important parties are the US and China)
My intuition is that route 2 has maybe a 3:1 penalty, i.e. a treaty is 3x more likely to go through if it originates from US policy-makers rather than Australian ones. I have no evidence on what the real number is, and I’m not sure where I’d get evidence from.
I think the binding constraint is getting the US and China to sign, and there, non-aligned actors can present options that don’t get pre-framed as capitulation or unilateral constraint. Australian advocacy wouldn’t sway the large actors unilaterally, to my patriotic regret, but advancing its interests under a coalition might. This is true for others, too; Indonesia, Vietnam and Singapore are all non-aligned states without active security disputes with either major power.
This feels like a useful alternate path off-ramp rather than relying on the two superpowers most tempted by arms race dynamics to de-escalate.
I agree that approaching US policy makers directly is best. But, as an Australian, that option is not open to me, at least! ;) I work alongside people from ControlAI also, and they are directly working on US policy makers. So it is a push by allies across many countries. Of course, if you are deciding where to fund, that does make things tricky. It is a good question for PauseAI’s leadership.
On the whole, I think PauseAI’s strategy is close to the best thing to be doing at current margins. The main reason why I haven’t donated to PauseAI is this:
Frontier AI is being developed in the US and, to a lesser extent, China. PauseAI does not operate in the US or China. Even if PauseAI-style advocacy produces interest in a pause in the UK, France, etc., I expect there’s a large penalty in terms of how this converts to frontier-AI-relevant policy, because the US and China are by far the most important countries.
I want to hear how you would respond to that. What’s your theory of change for how PauseAI increases the odds of a treaty that the US and China participate in?
Every country has some influence with other countries. If governments around the world push for a treaty between the US and China on AI, that makes it more likely that such a treaty will occur. Australia, for example, has relatively good relations with the US and is a major trading partner of China. We have some capacity to influence those nations on this issue.
Is this guaranteed? No. My gut feeling from the start was that there was about a 15% chance of a pause treaty happening. Recent events have increased that in my thinking, but it is still nowhere near 50%.
Naturally countries have influence on each other, but there’s a penalty factor, and the question is, how big is that penalty?
Route 1: US advocacy convinces US policy-makers to establish a treaty
Route 2: Australian advocacy convinces Australian policy-makers to establish a treaty (where the most important parties are the US and China)
My intuition is that route 2 has maybe a 3:1 penalty, i.e. a treaty is 3x more likely to go through if it originates from US policy-makers rather than Australian ones. I have no evidence on what the real number is, and I’m not sure where I’d get evidence from.
I think the binding constraint is getting the US and China to sign, and there, non-aligned actors can present options that don’t get pre-framed as capitulation or unilateral constraint. Australian advocacy wouldn’t sway the large actors unilaterally, to my patriotic regret, but advancing its interests under a coalition might. This is true for others, too; Indonesia, Vietnam and Singapore are all non-aligned states without active security disputes with either major power.
This feels like a useful alternate path off-ramp rather than relying on the two superpowers most tempted by arms race dynamics to de-escalate.
I agree that approaching US policy makers directly is best. But, as an Australian, that option is not open to me, at least! ;) I work alongside people from ControlAI also, and they are directly working on US policy makers. So it is a push by allies across many countries. Of course, if you are deciding where to fund, that does make things tricky. It is a good question for PauseAI’s leadership.