AI 导读
别误会，编程/工程远未被解决。 天花板依然很高。但如今的 LLM 会在复杂系统上卡住，比如构建一个可扩展的推理系统，能在 neoclouds 之间安全地服务 1B tokens/sec。 这不是一个提示词就能搞定的。
正文
Don’t get me wrong, coding/engineering is far from solved.
The ceiling is still very high. But LLMs today get stuck on complex systems, like building a scalable inference system that can securely serve 1B tokens/sec across neoclouds.
You can’t build that with one prompt.
来源：@Yuchenj_UW · x.com