[neuronic]

Would this not mean we need much much more training data to fully utilize the now "free" capacities?

[vonneumannstan]

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