Researcher Proposes Perplexity on Clean Data to Fix AI Benchmarks