“DeepSeek-R1 proved that reasoning-style post-training for language models can be reproduced and scaled by organizations other than the original innovators.”