Comment by BoorishBears

If they really managed this from pre-training a 1.6 T parameter model through to post-training without NVIDIA, Dwarkesh Patel got what he wanted.