People often say US-China deals to slow AI progress and develop AI more safely would be hard to enforce/verify.
However, there are easy to enforce deals: each destroys a fraction of their chips at some level of AI capability. This still seems like it could be helpful and it’s pretty easy to verify.
This is likely worse than a well-executed comprehensive deal which would allow for productive non-capabilities uses of the compute (e.g., safety or even just economic activity). But it’s harder to verify that chips aren’t used to advance capabilities while easy to check if they are destroyed.
This does assume both countries have a good sense of the total number of chips controlled by each country’s developers, which seems doable but not trivial and worth working towards now.
It’s easiest to negotiate and most useful if both sides have a reasonably good estimate of how many chips the other side has. I think it will probably be easy for intelligence agencies on either side to get an estimate within +/- 5% which is sufficent for a minimal version of this. This is made better by active efforts and by government trying to verifiably demonstrate how many chips they have as part of negotiation.
(This is in part because I don’t expect active obfuscation until pretty late, so normal records etc will exist.)
I think it will probably be easy for intelligence agencies on either side to get an estimate within +/- 5% which is sufficent for a minimal version of this.
That seems plausible or even likely but I think it’s hard to be confident. Public estimates of chip production are high variance. For example, SemiAnalysis and Bloomberg’s estimates of Huawei’s 2025 chip production differ by 50% [1]. I doubt either US or Chinese intelligence will be better at estimating these numbers than independent researchers within the next few years. Maybe the true numbers would become clear if governments want them to be, but I could also imagine China rationally not trusting documentation of US chip production created and shared by US chip manufacturers, and vice versa. This is especially hard under time pressure.
The more chips you want to destroy, the more confidence you need. If you want to destroy 90% of your adversary’s chip stock, but you fail to count 5% of their stock, you’ll leave them with 10% + 5% of their original stock, or 50% more than intended.
The main upshot imo is that it’s important for both independent researchers and intelligence agencies to precisely track the chip supply.
Bad analogy, there’s a very significant difference: excessive warheads above the genuine deterrence requirements can’t contribute positively to economy in any way (they can only do so if they contain HEU and it is converted to LEU for the power plants), while even ancient GPUs can be used for some good civilian use (and destroying them irreversibly harms the economy)
I don’t think that makes the analogy bad? I agree that’s a helpful distinction to track, but for people who don’t know about START etc, I think it’s more helpful for them to know than not.
People often say US-China deals to slow AI progress and develop AI more safely would be hard to enforce/verify.
However, there are easy to enforce deals: each destroys a fraction of their chips at some level of AI capability. This still seems like it could be helpful and it’s pretty easy to verify.
This is likely worse than a well-executed comprehensive deal which would allow for productive non-capabilities uses of the compute (e.g., safety or even just economic activity). But it’s harder to verify that chips aren’t used to advance capabilities while easy to check if they are destroyed.
(Cross post from X/twitter)
See also: Daniel’s post on “Cull the GPUs”.
This does assume both countries have a good sense of the total number of chips controlled by each country’s developers, which seems doable but not trivial and worth working towards now.
It’s easiest to negotiate and most useful if both sides have a reasonably good estimate of how many chips the other side has. I think it will probably be easy for intelligence agencies on either side to get an estimate within +/- 5% which is sufficent for a minimal version of this. This is made better by active efforts and by government trying to verifiably demonstrate how many chips they have as part of negotiation.
(This is in part because I don’t expect active obfuscation until pretty late, so normal records etc will exist.)
That seems plausible or even likely but I think it’s hard to be confident. Public estimates of chip production are high variance. For example, SemiAnalysis and Bloomberg’s estimates of Huawei’s 2025 chip production differ by 50% [1]. I doubt either US or Chinese intelligence will be better at estimating these numbers than independent researchers within the next few years. Maybe the true numbers would become clear if governments want them to be, but I could also imagine China rationally not trusting documentation of US chip production created and shared by US chip manufacturers, and vice versa. This is especially hard under time pressure.
The more chips you want to destroy, the more confidence you need. If you want to destroy 90% of your adversary’s chip stock, but you fail to count 5% of their stock, you’ll leave them with 10% + 5% of their original stock, or 50% more than intended.
The main upshot imo is that it’s important for both independent researchers and intelligence agencies to precisely track the chip supply.
[1] Figure 5: https://ifp.org/should-the-us-sell-hopper-chips-to-china/
People looking into this may also want to look at START and New START for possible analogies in nuclear arms reduction
Bad analogy, there’s a very significant difference: excessive warheads above the genuine deterrence requirements can’t contribute positively to economy in any way (they can only do so if they contain HEU and it is converted to LEU for the power plants), while even ancient GPUs can be used for some good civilian use (and destroying them irreversibly harms the economy)
I don’t think that makes the analogy bad? I agree that’s a helpful distinction to track, but for people who don’t know about START etc, I think it’s more helpful for them to know than not.