Hacker News

Assuming AI doesn’t just flame out (which seems unlikely at this point) I think the author’s main point - that traffic always shifts to the expensive SOTA model when it’s released - may eventually slow down. Shifting from gpt 3.5 to 4 was necessary, because 3.5 was basically useless. The shift from 4 to 4.5 was less amazing, and from 4.5 to 5 will probably be meh, so aside from trying it out many people will probably be fine staying on the cheaper models - especially if the AI companies continue bifurcating them with ones for “research”, others for coding, others for creative, etc.