Comment by fooker

Yeah, these benchmarks are bogus.

Every new model overfits to the latest overhyped benchmark.

Someone should take this to a logical extreme and train a tiny model that scores better on a specific benchmark.