I think the issue is that both the alignment stuff and performance in terms of reasoning is largely the result of the reinforcement learning that happened after

I think the issue is that both the alignment stuff and performance in terms of reasoning is largely the result of the reinforcement learning that happened after that model was completed.