<p>Ask GPT-4 to recommend an underrated sci-fi film. It says <em>Moon</em>. Ask Claude the same question — also <em>Moon</em>. Try Gemini — <em>Moon</em> again. A NeurIPS 2025 Best Paper, <a href="https://arxiv.org/abs/2510.22954">Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)</a>, quantifies this phenomenon at scale: different language models give strikingly similar answers to open-ended questions.</p>

<p>RAG re-retrieves, re-assembles, and re-reasons on every query. Ask something that requires synthesizing five documents and the model has to find all five, stitch them together, and derive the answer from scratch. Ask ten times, retrieve ten times. Nothing accumulates.</p> <p>Karpathy recently posted a gist called LLM Wiki proposing a different approach: instead of retrieving at query time, have the LLM pre-compile knowledge into a structured wiki and query the compiled result.</p>

<p>RAG 的工作方式是每次提问都重新检索、重新拼接、重新推理。问一个需要综合五篇文档才能回答的问题，模型每次都得从头找到这五篇，拼起来，再给你答案。问十次，找十次。什么都没积累下来。</p> <p>Karpathy 前两天发了一个叫 LLM Wiki 的 gist，提了一个不同的思路：别让模型每次现场检索了，让它把知识预先编译成一个结构化的 wiki，查询的时候直接查编译好的结果。</p>

<p>Dense retrieval has become the default first stage in RAG pipelines. Encode documents into vectors, encode queries into vectors, compute cosine similarity, done. But a basic question rarely gets asked: for a d-dimensional embedding, how many distinct top-k retrieval results can it actually represent?</p> <p>An ICLR 2026 paper from Google DeepMind and JHU, <em>"On the Theoretical Limitations of Embedding-Based Retrieval"</em>, gives a mathematical answer: not enough. Not even close.</p>

<p>向量检索（dense retrieval）这几年几乎成了 RAG 的标配。把文档编码成一个向量，查询也编码成一个向量，算个余弦相似度就能检索。但一个基本问题很少被认真讨论过：一个 d 维向量，到底能表示多少种不同的 top-k 检索结果？</p> <p>ICLR 2026 这篇来自 Google DeepMind 和 JHU 的论文 <em>"On the Theoretical Limitations of Embedding-Based Retrieval"</em> 给出了一个数学上的回答：不够。而且远远不够。</p>

<p>There's a module inside Claude Code called autoDream. Its prompt title reads "Dream: Memory Consolidation."</p> <p>This isn't a metaphor. Claude Code actually spins up a background sub-agent that reviews transcripts from past sessions, consolidates scattered memories — merging, deduplicating, correcting — and writes them back to disk. The whole thing is invisible unless you dig into the background task list.</p>

<p>Claude Code 内部有一个叫 autoDream 的模块。它的 Prompt 标题是"Dream: Memory Consolidation"。</p> <p>这不是什么隐喻。Claude Code 确实会在后台启动一个子代理，回顾过去多个会话的记录，把零散的记忆整理、去重、纠错，然后写回磁盘。整个过程你看不到，除非刻意去翻后台任务列表。</p>

<p>Last June, Shopify CEO Tobi Lutke posted that he preferred "context engineering" over "prompt engineering." Karpathy retweeted with a +1. Simon Willison wrote a blog post saying the term might actually stick. Phil Schmid published a full definition. Half the AI community switched terminology within a week.</p> <p>By early 2026, Phil Schmid introduced another term: Agent Harness. It didn't generate the same buzz, but anyone building Coding Agents quietly nodded along.</p>

<p>去年 6 月 Shopify CEO Tobi Lutke 发了条推，说他觉得 context engineering 比 prompt engineering 好。Karpathy 转发 +1。Simon Willison 写了篇博客说这个词可能真能立住。Phil Schmid 做了完整定义。半个 AI 圈在一周之内集体换了术语。</p> <p>到了 2026 年初，Phil Schmid 又抛了一个新词：Agent Harness。这次没有上一轮那么热闹，但写 Coding Agent 的人几乎都默默点了头。</p>

<p>After OpenClaw crossed 350K stars, a narrative started forming in the community: since both run on Opus 4.6 under the hood, the open-source option should be on par with Claude Code. Anyone who has actually used both probably shares the same observation — in long sessions, OpenClaw starts losing context, forgetting files it already read, redoing work it already did. Claude Code does too, but noticeably later, and it recovers much better.</p> <p>Same model, different experience. Why?</p>

<p>OpenClaw 拿下 35 万 Star 之后，社区开始出现一种论调：底层都是 Opus 4.6，开源方案应该能对标 Claude Code。实际用过两边的人大概都有同一个感受，长会话跑到后半段，OpenClaw 开始丢上下文，忘记之前读过的文件，重复做已经做过的事。Claude Code 也会，但明显晚得多，而且恢复能力强很多。</p> <p>模型一样，体验不一样。差在哪？</p>

<p>Cursor's parent company Anysphere has about 150 employees. In November 2025, its <a href="https://e.vnexpress.net/news/tech/personalities/4-mit-graduates-who-built-the-popular-ai-coding-tool-cursor-become-billionaires-4965462.html">ARR crossed $1 billion</a>. OpenAI, as of early 2026, has <a href="https://www.ft.com/content/7ffea5b4-e8bc-47cd-adb4-257f84c8028b">4,500 employees</a>. Its 2025 revenue was $13.1 billion, but according to Fortune, it <a href="https://fortune.com/2025/11/12/openai-cash-burn-rate-annual-losses-2028-profitable-2030-financial-documents/">lost roughly $9 billion</a> and doesn't expect to turn profitable until 2028.</p> <p>An application company that trains zero models is outproducing, per capita, the company that trains them. This is the most telling signal in AI for 2025.</p>

<p>Cursor 的母公司 Anysphere 大概 150 人，2025 年 11 月<a href="https://e.vnexpress.net/news/tech/personalities/4-mit-graduates-who-built-the-popular-ai-coding-tool-cursor-become-billionaires-4965462.html">年收入突破 10 亿美元</a>。OpenAI 到 2026 年初有 <a href="https://www.ft.com/content/7ffea5b4-e8bc-47cd-adb4-257f84c8028b">4500 名员工</a>，2025 年收入 131 亿美元，但据 Fortune 报道，<a href="https://fortune.com/2025/11/12/openai-cash-burn-rate-annual-losses-2028-profitable-2030-financial-documents/">亏损约 90 亿美元</a>，而且预计一直亏到 2028 年。</p> <p>一个不训练任何模型的应用公司，人均产出碾压了训练模型的公司。这组数字是 2025 年 AI 行业最值得琢磨的信号。</p>

<p>In late November 2025, an open-source project called OpenClaw went live on GitHub. Four and a half months later, it had 350K stars, 70K forks, 81 releases, and sponsorships from OpenAI, NVIDIA, and Vercel. For comparison: Open WebUI took two and a half years to reach 130K stars; NextChat took three years to hit 88K. Growth like OpenClaw's is rare in GitHub's history.</p> <p>It isn't a new model, a training framework, or even a "technical breakthrough" in the traditional sense. It's a personal AI assistant that runs on your own machine and talks to you through the chat apps you already use: WhatsApp, Telegram, Slack, Discord, WeChat, Feishu, iMessage, Matrix, and over 25 platforms in total, all connected to a single backend.</p> <p>Why did it break out of the developer bubble?</p>

<p>2025 年 11 月底，一个叫 OpenClaw 的开源项目在 GitHub 上线。4 个半月后，它有 35 万 Star，7 万 Fork，81 个发布版本，OpenAI、NVIDIA、Vercel 做它的赞助商。同期对比：Open WebUI 用了两年半才到 13 万，NextChat 三年到 8.8 万。OpenClaw 的增速在 GitHub 历史上都不多见。</p> <p>它不是一个新模型，不是一个训练框架，甚至不是一个传统意义上的"技术突破"。它是一个个人 AI 助手，跑在你自己的机器上，通过你已经在用的聊天工具跟你对话。WhatsApp、Telegram、Slack、Discord、微信、飞书、iMessage、Matrix，25 个以上的平台，同时接入，一个后台。</p> <p>这篇聊聊它为什么能出圈。</p>