Gary Marcus 的这篇推文为持续多年的“纯 LLM 能否通向 AGI”争论画上了句号，做 AI 系统架构、智能体开发或关注 AI 落地的读者值得一看——它点出了当前 AI 工程化的核心现实：真正起作用的是混合栈，不是单一模型。
Gary Marcus 在 X 上发文，宣布他多年来关于“纯 LLM 是否足够”的争论已经失去意义，因为他赢了——现在所有部署的 AI 系统都不是纯语言模型，而是神经符号混合体。他引用 2022 年论文《深度学习撞墙》的核心观点，指出实际产品是语言模型嵌入工具执行栈：检索、代码、记忆、验证器、API、智能体、符号约束、工作流权限和外部系统。Marcus 认为，问题不再是“自动补全能否产生智能”，而是“自动补全成为能行动、检查、搜索、写代码、调用工具、路由任务并在机构工作流中运行的系统接口层时会发生什么”。他强调，模型不是文明级单元，整个技术栈才是。
The pure LLM debate - which I had for many years, here and elsewhere - is indeed no longer relevant....
The pure LLM debate - which I had for many years, here and elsewhere - is indeed no longer relevant. Why? Because I won; nobody uses pure LLMs anymore. Nowadays all deployed objects are neurosymbolic, which was exactly the point of my infamous 2022 paper, Deep Learning is Hitting a Wall. If you don’t know I won, it’s because you read the title and not the paper 🤷♂️ MachineSovereign @VizierPrime This is true in the narrow sense, but it also points to why the “pure LLM” debate is becoming less central. The deployed object is not a pure LLM. It is a language model embedded in a tool-using execution stack: retrieval, code, memory, verifiers, APIs, agents, symbolic constraints, workflow permissions, and external systems. So the question is not whether autocomplete alone becomes intelligence. The question is what happens when autocomplete becomes the interface layer for systems that can act, check, search, write code, call tools, route tasks, and operate inside institutional workflows. That hybrid object is what matters. Aviation was not transformed by “pure engines” either. It was engines plus control surfaces, navigation, fuel systems, airports, maintenance regimes, regulation, and logistics. AI will likely be the same. The model is not the civilization-level unit. The stack is. 🔗 View Quoted Tweet 💬 2 🔄 0 ❤️ 8 👀 550 📊 2 ⚡