Dario recently made it clear that cutting-edge AI could take inspiration from the approach of the U.S. Federal Aviation Administration: Mandatory testing and third-party audits before model deployment, and if the risks are deemed unacceptable, the government should have the authority to block deployment. Sacks directly pinpointed the most vulnerable aspect of Anthropic's business model: the competitive edge of AI models is an asset that depreciates extremely quickly. If an airplane takes an extra year to get certified, its competitiveness might not necessarily disappear; but if Claude is delayed by a few months, it’s a completely different story. Sacks gave the example that new airplane models in the U.S. typically take years to get certified, whereas cutting-edge AI models iterate every few months. According to his judgment, if this kind of "approve first, release later" system were applied to AI, Anthropic’s original technical lead window of just a few months could be completely exhausted while waiting for approval. This touches on Anthropic’s most practical business issue: why would users be willing to pay a premium for Claude? A big part of the reason is its significant lead in scenarios like coding, agents, and long tasks. OpenAI and Anthropic are essentially on a never-ending treadmill. As soon as Gemini, Grok, Chinese models, or open-source models catch up to being "good enough," the premium on APIs and subscriptions will immediately start to shrink.