Haven't looked into how accurate the page is, but the list of regressions on the bottom looks terrifying, at first glance?

Yes, it looks like regressions are frequent, but sometimes performance goes back to baseline quite fast.

how the hell do we even track that? and before someone says...

—"Benchmarks!"

...I'll tell that they can be gamed so easily, and they are on a consistent basis.

Sure they are, but do you think they are continuing training to improve a model after release without bumping the version number, presumably only to game the benchmarks?

If you're willing to cheat, isn't it just a matter of grepping for the benchmark's question and pasting the solution in the chain of thought?

like windows 9