[Osmium]

What would regulation look like if it was based on energy usage rather than capabilities?

A guardrail on mass deployment that is not linked to specific model size or aspects of model performance that are difficult to quantify.

[Lerc]

Energy usage of training or inference?

As a guard on capabilities, it would permit a rise in abilities with gains in both hardware and software efficiency. Is this desirable?

[Osmium]

Overall energy usage, from the perspective that energy usage is the primary immediate harm of AI. This is a thesis, not an opinion.