Benchmaxxing for one set of benchmarks can actually degrade performance on other benchmarks. They’ve plateaued for a while, all they can do is scale inference compute up and down (at logarithmically poor rates of exchange) and trade performance in one area for performance in others