If the parabolic graph in the article is roughly accurate, and performance accelerates as the number of parameters in a training model increases, then it will b

If the parabolic graph in the article is roughly accurate, and performance accelerates as the number of parameters in a training model increases, then it will be very economically valuable to generate huge ML models. Depending on the shape of that curve model training could become a primary energy sink for the whole species, dwarfing the ergs spent on bitcoin mines. I could almost wish for diminishing returns.