when LLM calls are the bottleneck, the programming language stops mattering.