Typical LLM pretraining is very inefficient with data. NanoGPT slowrun shows that data can be used much more efficiently.