“Lari Hämäläinen believes that as of early 2026, standard LLM benchmarks are saturated and no longer a meaningful way to differentiate model performance.”