I think we're nearing the saturation point of how good LLMs can become not because they can't get better, but because we're running out of benchmarks to even understand what better means.