That's Too Much Intelligence To Solve Stupid Problems

The case for a smaller model, made by somebody who wants one. Nico's worry is not price, because he expects cost per token to keep coming down anyway. It is that pointing a frontier-scale model at routine work is the wrong fit no matter what it costs. He puts a date on the discomfort: if by December the cheap fast model is running on the biggest model's intelligence, he is going to feel bad about using it that way. What he wants is something smaller and less overstuffed, aimed at the specific thing he needs checked.