HackerNews中文版

代理的核心能力完全来自于底层的 LLM。因此，代理的未来完全取决于 LLM 的现状。那么，LLM 目前究竟发展到什么程度了呢？我认为我们目前正处于 AI 的“手工业”（或自给自足）阶段——工业化的黎明。用一个历史类比：我们刚刚发明了第一批蒸汽机。它们笨重、固定，只能用于从煤矿中抽水。我们离拥有蒸汽火车还差得远呢。目前，定制代理的构建正在大规模爆发。但如果你仔细观察，它们几乎完全是“自给自足”和孤立的。每个人都在为自己的特定用例构建自己的代理，但很难将其适配或扩展以供更广泛使用。这就像每个家庭都有自己的织布机，自己织布，从不使用别人的。为什么会这样？这归结于 LLM 目前的局限性。如果我们暂时抛开多模态能力，基于文本的 LLM 基本上有四个核心支柱：自然语言理解 (NLU) 自然语言生成 (NLG) 工具调用推理前三个已经高度成熟和可靠。但第四个——推理——仍然是一个充满幻觉的雷区。然而，代理开发者最痴迷于什么？推理。为什么？因为它在演示中看起来很酷。这种痴迷正是我们还无法真正“工业化”代理的原因。这也是为什么在实际应用中很难找到真正可靠的通用代理的原因（最近关于 Manus 的炒作和现实检验就是一个教科书式的例子）。当然，总有一天 LLM 的推理能力可能会超越 99% 的人类。当这一天到来时，我们将最终看到真正强大、通用的代理。但老实说，没有人确切知道这个时间表何时会到来。我的观点是：如果我今天正在构建一个用于生产的通用代理，我将严格使用 NLU、NLG 和工具调用。我将远离依赖“推理”。最近和一些朋友关于 AI 的对话让我思考。我的观点似乎引起了他们的共鸣，所以我在这里分享出来，听听大家的想法。

查看原文

An Agent's core capability comes entirely from the underlying LLM. Therefore, the future of Agents is strictly dictated by the present state of LLMs.So, where exactly are LLMs right now?I believe we are currently in the "cottage industry" (or subsistence) phase of AI—the very dawn of industrialization. To use a historical analogy: we just invented the first steam engines. They are bulky, stationary, and only good for pumping water out of coal mines. We aren't anywhere close to having steam locomotives yet.Right now, there's a massive explosion of custom Agents being built. But if you look closely, they are almost entirely "self-sufficient" and siloed. Everyone is building their own Agent for their own specific use case, but it's incredibly hard to adapt or scale them for broader use. It’s like every household having its own loom, weaving its own cloth, and never using anyone else's.Why is this happening? It comes down to the current limits of LLMs. If we put aside multimodal capabilities for a moment, text-based LLMs basically have four core pillars:Natural Language Understanding (NLU)Natural Language Generation (NLG)Tool CallingReasoningThe first three are already highly mature and reliable. But the fourth—Reasoning—is still an absolute minefield of hallucinations.Yet, what do Agent developers obsess over the most? Reasoning. Why? Because it looks cool on a demo. This obsession is exactly why we can't truly "industrialize" Agents yet. It’s why it is so damn hard to find a genuinely reliable, general-purpose Agent in the wild (the recent hype and reality check around Manus is a textbook example of this).Sure, one day LLM reasoning capabilities might surpass 99% of humanity. When that day comes, we will finally see truly powerful, general-purpose Agents. But honestly, nobody knows exactly when that timeline will hit.My takeaway: If I am building a general Agent for production today, I am strictly utilizing NLU, NLG, and Tool Calling. I am staying the hell away from relying on "Reasoning."A recent convo with some friends about AI got me thinking. My take seemed to resonate with them, so I’m sharing it here to hear your thoughts

为什么大多数通用 Agent 会失败，以及我为什么避免使用 LLM 的“推理”