Noam Brown explains that current evaluation methods are outdated because they don't consider how long a model is allowed to think. Models like o1 can improve performance over weeks of reasoning, but benchmarks only measure static snapshots. This creates a safety gap, as dangerous capabilities may only emerge at higher compute budgets. Brown warns against benchmark-maxxing, where models game evaluations by using more test-time compute. He highlights that while models can optimize existing algorithms, they struggle to create novel ones, making a rapid intelligence explosion unlikely due to time bottlenecks. He also discusses the potential of multi-agent coordination and recommends using current models for high-stakes decisions, trusting their outputs more than human experts. Finally, he advocates for evaluating benchmarks with a cost axis to fairly compare models against those thinking longer.