Is antithetic sampling doing anything for CMA-ES?

To me it looks like we are just re-using the same seed twice which would waste compute and effectively cut our population size in half(seen here in [train.py](https://github.com/hardmaru/WorldModelsExperiments/blob/244f79c2aaddd6ef994d155cd36b34b6d907dcfe/carracing/train.py#L304))?
It seems like we are only making use of antithetic sampling in OpenES and PEPG.

Am I missing something here?

