isn’t training “powerful AIs” remaining a highly resource intensive, observable and disruptable process for likely ~decades?
I expect there’s lots of room to disguise distributed AIs so that they’re hard to detect.
Maybe there’s some level of AI capability where the good AIs can do an adequate job of policing a slowdown. But I don’t expect a slowdown that starts today to be stable for more than 5 to 10 years.
It is a bit ambiguous from your reply whether you mean distributed AI deployment, or distributed training. Agree that distributed deployment seems very hard to police, once training took place, implying also that there is some large amount of compute available somewhere.
About training, I guess the hope for enforcement would be ability to constrain (or at least monitor) total compute available and hardware manufacturing. Even if you do training in a distributed fashion, you would need the same amount of chips. (Probably more by some multiplier, to pay for increased latency? And if you can’t distribute it to an arbitrary extent, you still need large datacenters that are hard to hide.) Disguising hardware production seems much harder than disguising training runs or deployment.
Perhaps a counter is “algorithmic improvement”, which is estimated by Epoch to be providing 3x/year effective compute gain. This is important, but: - Compute scaling is estimated (again, by Epoch) at 4-5x/year, so if we assume both trends to continue, and if your timeline for dangerous AI was say, 5 years, and we freeze compute scaling today such that only the largest training run today is available in the future, IIUC you would gain ~7 years (which is something!) But, importantly, the longer timelines you have, if I did the math correctly, you have linearly ~1.5x extra time. (So, if your timeline for dangerous AI was 2036, it would be pushed out to ~2050.) - I’m sceptical that “algorithmic improvement” can be extrapolated indefinitely—it would be surprising if you could train GPT-3 in ~8 years on a single V100 GPU in a few months? (You need to get a certain amount of bits into the AI, there is now way around it.) (At least this should be more and more true as labs reap more and more of the low-hanging fruit of hardware optimisation?)
Also, contingent on point 2. of my original comment, all of the above could be much much easier if we are not assuming a 100% adversarial scenario, where the adversary has willingness to cooperate in the effective implementation of a treaty.
I agree with all your points except this:
I expect there’s lots of room to disguise distributed AIs so that they’re hard to detect.
Maybe there’s some level of AI capability where the good AIs can do an adequate job of policing a slowdown. But I don’t expect a slowdown that starts today to be stable for more than 5 to 10 years.
It is a bit ambiguous from your reply whether you mean distributed AI deployment, or distributed training. Agree that distributed deployment seems very hard to police, once training took place, implying also that there is some large amount of compute available somewhere.
About training, I guess the hope for enforcement would be ability to constrain (or at least monitor) total compute available and hardware manufacturing.
Even if you do training in a distributed fashion, you would need the same amount of chips. (Probably more by some multiplier, to pay for increased latency? And if you can’t distribute it to an arbitrary extent, you still need large datacenters that are hard to hide.)
Disguising hardware production seems much harder than disguising training runs or deployment.
Perhaps a counter is “algorithmic improvement”, which is estimated by Epoch to be providing 3x/year effective compute gain.
This is important, but:
- Compute scaling is estimated (again, by Epoch) at 4-5x/year, so if we assume both trends to continue, and if your timeline for dangerous AI was say, 5 years, and we freeze compute scaling today such that only the largest training run today is available in the future, IIUC you would gain ~7 years (which is something!)
But, importantly, the longer timelines you have, if I did the math correctly, you have linearly ~1.5x extra time.
(So, if your timeline for dangerous AI was 2036, it would be pushed out to ~2050.)
- I’m sceptical that “algorithmic improvement” can be extrapolated indefinitely—it would be surprising if you could train GPT-3 in ~8 years on a single V100 GPU in a few months? (You need to get a certain amount of bits into the AI, there is now way around it.)
(At least this should be more and more true as labs reap more and more of the low-hanging fruit of hardware optimisation?)
Also, contingent on point 2. of my original comment, all of the above could be much much easier if we are not assuming a 100% adversarial scenario, where the adversary has willingness to cooperate in the effective implementation of a treaty.