DeepSeek-R1-Zero’s RL Training Used 30× Less Compute Than Estimated
[Crossposted from my Substack.] An update to Ege Erdil’s “What went into training DeepSeek-R1?” TL;DR * I estimate that DeepSeek-R1-Zero's RL training used ~2e22 (1.5e22-3e22) FLOPs, 30x less than Ege Erdil's estimate of 6.1e23 in the Epoch AI newsletter, mainly due to a batch size correction in the estimate. *...
Nov 74