Well, they're being trained with reinforcement learning now, so I, for one, am excited to see what kinds of unique extremely performant insanity they'll produce in the future =D
What we discovered that LLMs aren't really the answer wrt performance, the whole system around verifying correct and performance is the key.