What 18 LLM papers taught me about the real bottleneck