The author (geohot) argues that while LLM progress is genuine and exciting, the field suffers from two hype problems: (1) fear-mongering narratives about missed opportunities and societal decline designed to drive talent to expensive hubs, and (2) inflated "superintelligence" claims that leap from "fancy autocomplete" to civilization-ending scenarios. He contends that frontier AI labs capture less value than their valuation suggests because AI progress stems from Moore's law and general computing advances, not proprietary breakthroughs. On a practical note: he reports LLMs measurably improve his productivity as a programming tool—similar to compilers or regex libraries—but warns they increase cognitive fatigue and current generated code remains poor. The core claim is actionable: treat LLMs as productivity multipliers in specific domains (coding assistance, compilation), ignore both doomist and utopian narratives equally, and recognize the commodification risk means competitive moats are weaker than valuations imply. One sharp question: if "they won't capture it" (the value), who does, and does that change the current investment thesis?