Outperform GPT-3 with Karpathy's llm.c using just 1/3 training tokenstwitter.com6 pointstosh2 years ago