[vonneumannstan]

It's pretty clear that the reasoning models are using mass amounts of synthetic data so it's not a bottleneck.