Eric Xu (e/Mettā)

Announcing Noah — a tool to address every aspect of your computer issues, running locally on your computer (Mac/Windows). It diagnoses problems, explains what's wrong in plain English, asks before touching anything, and remembers what it's learned about your setup. For now, it uses Claude API key, but soon will support other/local models. Open source: github.com/xuy/noah

20 years ago, my first startup was all about enterprise search. Two decades later, we’re still building search engines. The technology has shifted from NLP to NN and the users from humans to agents. but searching is still the core. opensource the fastest bm25 engine:

And the antidote is motivated reasoning to dismantle your own hypothesis.

@xtbot @JosephJacks_ @Zai_org @MiniMax_AI @deepseek_ai @Kimi_Moonshot If intelligence is proportional to # of parameters perhaps it is a great heuristics. X B model = $ X B evaluation 😂

@JosephJacks_ @Zai_org @MiniMax_AI @deepseek_ai @Kimi_Moonshot Are those # of parameters? 🤣

Chinese AI lab valuations (all open weights) $50 Billion ~ : @Zai_org $50 Billion ~ : @MiniMax_AI $50 Billion ~ : @DeepSeek_ai $20 Billion ~ : @Kimi_Moonshot

Claude design is genuinely impressive. Preparing to ship the next version of Noah in a few days, and I cannot wait to see those beautiful dynamic panels rendered in an intuitive way. --- 争取下一个版本的 Noah 能够做到这么漂亮！

The Anthropic loop you described is missing one primitive: `evaluate` before spend. You're paying Meta to tell you which headline works. I built a synthetic panel that front-loads the basic optimizations — so Meta only has to confirm, not discover. Open source as a SKILL: github.com/xuy/sgo

@TheAhmadOsman 😂 Compute Anonymous welcomes you.

My name is Ahmad and I have a Compute problem

@wangyinrec DM 联系

@xleaps 非常有意思的角度，我非常感兴趣，我想帮你补去那些真实维度的信息，如果好找的话，看看能否结合你这个生成维度，发现点什么。

@tinyfool 然后做市场调研 AB 测试或者说任何在一群人上的函数优化问题，就从一个小样本外推问题变成一个大样本拟合问题 拟合比外推简单很多 解锁的是这种群体系统的梯度计算问题 具体比如站点设计 功能组合 定价策略 广告策划等等

@xleaps 然后呢

一个中国人的名字里藏着他出生那一年的新闻联播。比如叫建国的，我们都能猜出他的年龄。 为了做 2000 万有机虚拟中国人的数据集，我花了几天让模型学会了名字和出生时代的关联。记录如下：👇 x.com/xleaps/status/…

@tinyfool 名字是第一步 最终是一个虚拟的性别 职业生涯 收入 地区 民族 消费偏好 婚姻 子女 等等一切统计上与真实毫无二致的数据集 人类的每个子群体形态都可以用一个虚拟数据集以统计上无二致的形式刻画和模拟

@xleaps 做一个虚拟名字集的目的是啥？

@yaojingang 所以很好奇 愿意跟踪一下下面的进展 看看这家公司和客户之间如何分配各自的效率提升步骤

@yaojingang 我之所以知道这件事情 是因为我之前做的 Noah 表面上看是给普通用户解决计算机问题的 实际上在后台一直是卖给 MSP 解决企业 IT 问题的 然后因此就和从事 IT 外包的企业深度整合过 这些企业的效率能够在 AI 帮助下得到提升 但痛点不再开发 在流程 在其他意想不到的地方

朋友在操盘的一个案例，很有意思 一家年营收10亿左右的传统公司 今年把整个IT部门裁了，四五十人 然后，把开发、维护、升级，全部交给了他的这个10人的技术团队 他们接手后，花了一个月时间，用AI把原来的所有系统，重写了一遍 代码、文档以及系统逻辑，重新进行了重构和梳理 很多程序员过去的一个“护城河”， 其实是对历史系统的理解，比如很多代码逻辑，只存在某个程序员的“脑子”里 但是，在AI的这种重构能力面前，这层的壁垒似乎没有了 系统一旦被AI重构完成，后面的维护、迭代、升级 不再需要那么多人 IT外包这个商业模式，在中国还挺成熟的，诞生了不少规模较大的公司，比如软通动力、中软等 这种模式，在会驾驭AI的技术团队里，会让整个系统开发和运营的效率变得很高，也就是用10人外包的方式去替代原本50人左右的技术团队，难度并不大 真正难度大的，其实是解决信任的问题，以及后续的持续良性协作 一个营收十亿级的企业老板，愿意通过外包的方式来解决IT效率与管理，对这个团队的信任要非常高才行 另外，这种模式，适用条件其实挺有限的，并不适合绝大部分的企业

I am only half joking 😄

Programmers worried about AI agents should take a page from OnlyFans. OnlyFans was revolutionary. It made creation trivial and removed gatekeepers from idea to income. OnlyFans has been the Claude Code for the people who vibe create their way into the adult industry. So the great flood of supply came, ie those who had never been in the industry suddenly found themselves in it. What happened to those who were already in the industry? The reality is that they remained successful. The audience only increases by # of people turn 18 each year, so all things considered attention simply spread thinner and became harder to hold. So what is the underlying dynamics? Well the oversupplied market treats average work the way the internet treats a slow page. The world moves on without ceremony on "average". Skill issue is a harsh synonym for not good enough. So who made money on onlyfans? Those who are oddly specific. Software is walking into similar equilibrium: It helps to be known for something oddly specific. And of course if you already have an audience, try to keep it close and give them a reason to stay. When creation is easy, skill stops being scarce. Being wanted and being in others desire does.

看了好几篇介绍 Claude Managed Agents 的中文长文，越看越不懂这是用来干什么的，最终还是要靠跟 Claude 对话来问明白，然后结合我参与 coding agent 项目的经验来理解。我觉得介绍任何新概念，必须把下面三件事情说明白： 1. 它尝试解决的问题是什么？ 2. 它如何解决了这个问题？ 3. 你如何用它来解决这类问题？ 如果你决定自己写一个 Lovable，你需要有一个云端的 agent 来响应用户在浏览器里面输入的 prompt，然后编辑存储在云端的那个网站，这是 Managed Agent 尝试解决的问题。如果你用 Claude Code 写代码，这事情跟你没半点关系。如果你写客户端的 agent loop，例如说 iOS 应用，这也跟你没关系。 Managed Agent 解决自建云端 agent 以下常见问题： 1. 需要自建 context 存储。用户今天创建一个 Lovable 项目搞两下，关掉浏览器后明天再打开，这个 context 怎么继续下去？你要找个地方存。用 Managed Agent 的话就它帮你存。 2. 需要不停地根据新模型的行为调整 harness。老的模型需要这样 prompt，需要在这时候 compact memory，新模型出来你可能要摸索着改。用 Managed Agent 的话就由 Anthropic 帮你维护 harness 更新，你根本接触不到 agent loop。 3. 需要智能启停云端 agent 实例。用户打开一个 Lovable 项目，启动对应 agent；用户关闭浏览器，过一段时间后停止实例；实例崩溃了但前端还连着，赶紧重启新实例。Managed Agents 可以帮你管理这些实例，你从前端发 prompt 过去就好。（实例仅在 agent loop 跑着的时候计费，停下来等下一个 prompt 就不再计费。） 你自己写的这个 Lovable 可以通过 API 来调用 Managed Agents 创建 context、harness 和 sandbox（临时存储和编辑用户网站的环境），你不再需要手工解决上述问题。 在创建 harness 时你可以提供 system prompt 和 tools，这样云端 agent 知道自己该干什么，懂得调用 tools 来跟 sandbox 打交道或者跟其它云端服务打交道。例如说你写的 Lovable 要支持生成图片，你可以提供一个调用 Nano Banana 的 tool 给 harness。 最后我还是要吐槽那些贩卖信息赚取流量的中文内容，对于一个新热点如果你连顶上三个问题都说不清楚你还是别写了吧。

@Fermi_ko Long Costco!

我有时会想，几百年后Voyager 1会被围绕建成一座博物馆，跟着它一起在宇宙里漂泊，博物馆记着人类是怎么一步步迈进深空的。但对那时候的人来说，这就是个高速路边的服务区，路过看着这展品简介感慨一下，然后买个$1.5的热狗和可乐，就回超空间航道接着赶路了

@tunguz Deal. Let's sign a formal treaty: I won't use my Mythos to one-shot a nuclear warhead your way, and you won't use yours on me. Mutual Assured Destruction achieved. Peace in our time... or at least until the next AGI model drop.

I got access to Mythos this morning. It one shotted an ICBM for me. It's over.

The challenge is that the uncertainty and non-deterministic nature of the world has to be managed by someone. For example in an email exchange with a client, they might first propose a budget and a date/time to meet, and in the same thread later they may say "let me go back to my supervisor for a different number and let's meet next week instead". Your previous graph is now outdated (deal, budget, number) but unfortunately you cannot replace it with a similar RDF triple. Then you'd need a semantic edge to represent that uncertainty. It ends up just writing a Bayesian logic into a graph connected by maybe. The easiest way in the LLM era is to leave the whole thing raw and LLM can assemble the information incrementally, as opposed to grounding it each and every step of the way and be awkward.

@xleaps @svpino @xleaps what if every subgraph was interconnected with deterministically anchored nodes and edges (entities & relationships)?

Knowledge graphs win every single time. Before embeddings and similarity search, knowledge graphs were a game-changer. They are now going to win again. Similarity is not relevance. It never was. If you want relevant search results, you can't rely on similarity alone.