@iScienceLuvr The improvement in test-time compute is clear, but did the 100k+ GPUs actually lead to a significant improvement in pre-training?