Project MPG: towards a generalized performance benchmark for LLM capabilities