Case in point: AI models could be written to be more efficient in token use
They are being written to be more efficient in inference, but the gains are being offset by trying to wring more capabilities out of the models by ballooning token use.
Costs have been dropping by a factor of 3 per year, but token use increased 40x over the same period. So while the efficiency is contributing a bit to the use, the use is exploding even faster.
They are being written to be more efficient in inference, but the gains are being offset by trying to wring more capabilities out of the models by ballooning token use.
Which is indeed a form of Jevon’s paradox
Costs have been dropping by a factor of 3 per year, but token use increased 40x over the same period. So while the efficiency is contributing a bit to the use, the use is exploding even faster.
I think we’re meaning the same thing.