卡帕西在红杉资本AI Ascent 2026炉边谈话中分享了三大主题：首先，LLM远不止是提升已有工具的速度，他通过三个例子说明LLM打开了全新可能性——无需经典代码的menugen应用、用.md技能替代.sh脚本安装软件、以及处理非结构化知识的LLM知识库。其次，他探讨了LLM能力"参差不齐"的根源，认为这与领域的可验证性以及经济因素有关——训练数据分布决定了模型在哪些领域表现出色。最后，他讨论了"智能体原生经济"，即产品与服务向传感器、执行器和逻辑的分解，以及如何让信息对LLM最大可读。
A Hacker News user asks why everyday apps (banks, insurance, delivery) haven't improved despite billions spent on LLM code generation, noting degraded quality in some major apps and no killer features outside AI products themselves.
A software engineer asks on Hacker News which industries are actually replacing SWEs with AI, noting that LLMs are not yet capable of independent production-grade work despite widespread claims that engineers will soon be obsolete.
A new Hacker News user proposes that the fluctuating quality of large language model (LLM) responses, particularly with Claude, may be caused by operators reducing processing depth during peak server times. The author asks the community for their thoughts on this theory.