Comment by WanderPanda

Why did you stop training shy of the frontier models? From the log plot it seems like you would only need ~50% more compute to reach frontier capability