The gap between what frontier labs claim to control and what they actually control is widening faster than their business models can absorb. OpenAI lobbies California to strengthen safety rules it once opposed, yet leading labs still have no publicly documented containment plans for rogue models, a contradiction that suggests regulatory theater matters more than operational readiness. Meanwhile, the market is already voting with its wallet: Anthropic's Claude 3.5 Sonnet struggles to convert corporate users despite technical sophistication, while cheaper alternatives and narrow-application tools like Legora in legal tech and Inherent's Faraday in scientific replication capture actual demand. Harvard's AI instructor avatars and the push toward AI teammates signal that the real economic value isn't in foundation models themselves but in task-specific deployment and human-AI collaboration at scale. Automation in air traffic control shows the pattern repeating in infrastructure: the systems get faster and more complex, but the regulatory and operational frameworks lag, creating new failure modes no one planned for. The labs are racing to appear responsible while building containment theater; the market is racing past them toward applications that work, even if they're narrower. When Faraday outperforms Claude and GPT-4 at replicating research, and when companies prefer cheaper models over the best ones, the conversation about safety and control becomes secondary to the conversation about who actually ships products users will pay for.
Sloane Duvall