TLDR:
Any improvements in LLM output in the last 3 years have come from brute force recursion, not from fundamental improvements. This means a couple things; the technology itself has plateaued, and they’re not going to be able to continue spending at this rate to pretend it hasn’t.