LLM benchmarks aren't useful, but at least this one has drawings.