Why I'm still bearish on LLMs after Navier-Stokes

The author argues that despite headline-grabbing achievements like solving Navier-Stokes, large language models remain fundamentally limited for autonomous knowledge work because they require extensive oversight, guardrails, and rigorous formal specifications to prevent failure and reward hacking. Most real-world tasks lack the clear specifications available in pure mathematics, and the cost of creating such specifications—often exceeding implementation costs—combined with the poor scalability of human review, makes the premise that LLMs will replace knowledge workers economically unfeasible. The author contends that frontier labs are overvalued based on a narrative of imminent full autonomy that their actual performance does not support.