The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements | Read Paper on Bytez