What makes compute usage regulation promising for AI governance?
Recent AI progress has relied on the “AI triad”: algorithms, data, and compute
Shorthand for “computing power”. It may refer to, for instance, physical infrastructure such as CPUs or GPUs that perform processing, or the amount of processing power needed to train a model.
-
Detectability: Compute consists of physical hardware that can be observed, unlike data or algorithms. Data centers are up to several football fields in size and require extensive infrastructure for cooling and power. The process to construct them is complex and expensive. All this makes data centers easy to see.1
The authors note that data centers could potentially be concealed, and knowing that there’s a data center somewhere doesn’t tell you what it’s being used for. -
Excludability: Whoever controls a computer can prevent others from using it. This is in contrast to data and algorithms, which can be copied easily as soon as they’re public, in a way that’s much harder for governments to prevent.
-
Quantifiability: Compute can be clearly quantified in terms of how many chips of a given kind are being used, or how many FLOP of computation they’re doing. Other associated infrastructure, for networking and power and cooling, is also quantifiable.
-
Supply chain concentration: The global supply chain for the most advanced AI chips is highly concentrated. These chips are mostly designed by a single company (Nvidia) and manufactured by a single company (TSMC) using lithography machines produced by a single company (ASML), in an extremely complex process that relies on tacit knowledge and is hard for others to replicate. Many of the chips are then bought by a few cloud computing providers.
(Image source: Sastry et al.)
Some of these properties (like excludability) are intrinsic to compute, but others (like supply chain concentration) are a contingent feature
A feature of a region of input space that corresponds to a useful pattern. For example, in an image detector, a set of neurons that detects cars might be a feature.
A risk of human extinction or the destruction of humanity’s long-term potential.
The authors note that data centers could potentially be concealed, and knowing that there’s a data center somewhere doesn’t tell you what it’s being used for. ↩︎