Isn't being open source creating incentives for the AI companies to optimize their LLMs for the specific benchmark? I thought all those benchmarks are deliberately closed source primarily for this reason.