as work / token (engine efficiency) increases, because agents are getting smarter, my agent (locomotive) count is increasing, because I find more uses for smarter agents, and my overall token use is increasing.

what's odd is we must take this back to literal coal in order to complete the analogy. tokens are ethereal; energy inputs are not.

cost / token is decreasing while model performance stays roughly the same or gradually improves. cost is a rough proxy for tons of coal consumed.

I think smarter models is what's driving greater coal use overall.