One could imagine a version of Plan A that addresses this concern by limiting capabilities primarily via limiting compute instead of algorithms. This has some other benefits as well like being more enforceable than our version of Plan A.
I also wanted to point out this problem! At the time of writing, the median estimate of a secret project’s compute was 1.5M H100e, aka 1.5-6M petaFLOP/second. The problem is that a human brain requires one petaFLOP/sec to run. But the secret project should not be allowed to create the ASI by raising a simulated 100-times-as-fast brain for a quarter of a simulated century aka 3 months, then upskilling the brain for simulated decades in 10 thousand directions…
I think this consideration goes the other way, unless I’m misunderstanding. Algorithms diffuse to covert projects, but compute doesn’t, in the Plan A scenario, we deliberately scale compute for this reason while heavily limiting algorithmic progress.
(Also note that the 1.5M H100e is before accounting for detection)
I struggle to understand this. Section 2 on covert AIs contains the following phrase: “We estimate that, under competent US execution of our policies in Plan A, a competently-executed PRC diversion effort beginning one year in advance of the deal would be able to acquire between0.1% and 1.4% (80% CI) of the world’s AI-relevant compute at the start of the deal, without the US getting unambiguous evidence of violation,” and the median estimate is the very 1.5M H100-equivalents. Additionally, what prevents the secret project from recreating the algorithms which the larger humanity tabooed?
I also wanted to point out this problem! At the time of writing, the median estimate of a secret project’s compute was 1.5M H100e, aka 1.5-6M petaFLOP/second. The problem is that a human brain requires one petaFLOP/sec to run. But the secret project should not be allowed to create the ASI by raising a simulated 100-times-as-fast brain for a quarter of a simulated century aka 3 months, then upskilling the brain for simulated decades in 10 thousand directions…
I think this consideration goes the other way, unless I’m misunderstanding. Algorithms diffuse to covert projects, but compute doesn’t, in the Plan A scenario, we deliberately scale compute for this reason while heavily limiting algorithmic progress.
(Also note that the 1.5M H100e is before accounting for detection)
I struggle to understand this. Section 2 on covert AIs contains the following phrase: “We estimate that, under competent US execution of our policies in Plan A, a competently-executed PRC diversion effort beginning one year in advance of the deal would be able to acquire between 0.1% and 1.4% (80% CI) of the world’s AI-relevant compute at the start of the deal, without the US getting unambiguous evidence of violation,” and the median estimate is the very 1.5M H100-equivalents. Additionally, what prevents the secret project from recreating the algorithms which the larger humanity tabooed?