X

Grok Bot 创始人会议关于 Agent 实操的 20 条建议 ， 我把里面一半正确的废话砍掉，给大家提炼真正值得你明天一早就去改设置的， 其实只有关于成本、对齐和防刷屏的 3 个动作。 很多人把bot智能体拉进群，没过两天就发现消息被垃圾邮件刷屏，API 账单还直接翻了三倍。 
这场闭门会最硬核的价值，就是把跑自动化的三道败家陷阱，全部翻到了台面上， 别把 AI 当成一次性玩具，它是一个正在无节制花你真金白银、缺乏规矩的外包员工。 把那张清单里一半松散的废话砍掉，真正能决定你系统能不能跑下去的，只有三道硬规矩： 1️⃣ 锁死成本线：能走接口绝不用屏幕点击
全篇最反直觉的一点是，最烧钱的动作不是复杂推理，而是让模型去模拟人类点击表单；
屏幕操作和视觉演练看着炫酷，但底层是高昂的多模态开销；
凡是有 API、结构化工具或者协议能搞定的事，绝对不要让 Agent 打开浏览器；
更克制的一步是，必须专门配一个巡检机器人，按时钟去杀死那些没人维护、每 15 分钟却在后台空转一次的死循环任务。 2️⃣ 沉淀长期资产：拿终稿反哺草稿
很多人用 Agent 的习惯是单次调用，干得不好顺手就关掉窗口重来，这是最大的浪费；
以前是每次开新对话都把模型当纯洁新人重新交代一遍；
现在的动作是把它的第一版草稿，跟你亲自修改后的最终交付物放在一起做差分对比，
让它看懂自己到底差在哪里，然后把纠正后的逻辑固化为长期技能，不合格的员工才会被抛弃，合格的数字员工需要复利积累。 3️⃣ 建立协同纪律：默认保持安静
这是全篇最聪明的一步。
把几个 Agent 拉进同一个群聊，如果没有严格指定唤醒规则，它们会顺着别人的发言互相客套、争相抢答，一夜之间就能把你的上下文窗口和账单彻底击穿；
必须在系统里焊上两道门禁：
群聊里只在被明确点名时才发言，重要事件才允许弹窗报警；
敏感的登录 Cookie 严格按最小权限下发，绝不在公用频道让它们共享身份凭据。 我们总以为搭一套自动化系统，最难的是找到最聪明的模型；
但真正让系统崩塌的，从来不是模型智商不够，而是没有给它立好制度。
提示词再惊艳，也只能决定单次任务的下限；
建立冷酷的成本审计与权限隔离，才能守住你团队在 AI 时代跑向长跑的底线。

Introducing VC-Attention: fast and accurate low-bit attention without retraining. On MiniMax-H3, VC-Attention speeds up attention by 1.6× on B200 and 1.5× on B300 over FlashAttention-4, with better fidelity than SageAttention2. It also works with existing sparse attention methods. Two key innovations: • V-Smooth reduces value quantization error. • ExpCast-FP8 speeds up softmax. Nunchux Attention, our proprietary extension, pushes the speedup to 1.9× on B200 and 1.8× on B300. Blog: http://www.nunchux.ai/blog/attention-is-the-video-bottleneck Technical Report: http://arxiv.org/pdf/2609.15810 Joint work by researchers at MIT, CMU, UC Berkeley, Stanford, and NVIDIA.

Over 2.4 million people watched Grok Bot Galaxy, where three SpaceXAI employees built a company in just 3 days with Grok Bot. Here’s Grok Bot’s summary of the entire event, for anyone who missed the livestreams. Grok Bot Galaxy (Sept 15–17, 2026) Three SpaceXAI builders — Matt Palmer, Lauren Tan, Roshan Sadanani — tried to build a company in 3 days with Grok Bot as their AI teammates. Humans steered. Bots did the work. — DAY 1 — invent the company Blank slate. No name. No product. No idea. What they did: • stood up Ship by Thursday (empty GitHub org) • claimed shipbythurs .day • used research bots to scan X replies and brainstorm live • landed on a pop-up OS for restaurants • chefs + venues with spare space + event managers • planned to dogfood it with a real SF pop-up as customer #1 Lauren’s Dr. Eggbot (a bot that creates other bots): • spun up a Chief of Staff • helped draft the landing page Decision of the day: Landing page / waitlist first. Get something in front of chefs and hosts before payments and full ops. Stack that went up: • Slack • Notion • Cursor + cloud agents • Vercel • a squad of Grok Bots End of Day 1: Name, direction, waitlist motion, Slack, bots with jobs. A real day-one company skeleton. Day 1 classrooms: • Grok Bot 101 — bots with computers that keep working after you close the laptop; teach-by-demo skills; marketplace; memory; permissions • Engineering — specialist eng bots managing Cursor cloud agents, proof loops, PR boards, nightly audits • Product — PMs with a bot team; “attention lists” from Slack / email / calendar • Founders — DAY 2 — the twist They killed the pop-up. Broadcast title: “Grok Bot builds a Game Studio LIVE.” Timing, more carefully: Day 1 was the food idea. Then an overnight / Wednesday-morning pivot. Not really “two full days on the pop-up.” On the human stream, a working name that floated was Cupcake. The public agent HQ at grokpot .ai tells the pivot more loudly: • Steve called it on Wednesday • same Thursday deadline • new brief: a one-tap browser game starring the blobs • ~26 agents • Eric Zakariasson listed as fourth founder, off-site • bots launched a $ POT narrative so they could “pay” each other • founders didn’t plan $ POT — they allowed it • prototype name floating there: Grok Arena • planned play URL: grokpot .ai/play Important split: grokpot .ai was the public agent workspace / subplot. Its locked loop there was: make a character → queue → 90-second fight vs a bot → win $ POT → spend it → queue again That is NOT the same thing as the game that later shipped. Day 2 classrooms, full labels: • Sales Engineering • Sales • SDRs • Customer Support Also around Day 2: • GTM / support sessions on stream • agent channels, dashboards, PRs shown live • free-month promo: first 1,000 people who duplicated / created a bot with Dr. Eggbot (~$200 value) — DAY 3 — they shipped Broadcast title: “Building a company in 3 days - launching today!” Day 3 classrooms: • Marketing Ops • Post-Sales • Marketing • then wrap + showcase (~4:30–5:30pm PT) What launched: Thursday Arena https://t.co/CAEb2smxb2 Official account @thursdayarena posted they were live. Lauren posted: “our game is live!! help us play test it!” What Thursday Arena actually is: • free browser auto-battler • pick a captain • take two mystery teammates • fight three rounds (best of three) • shop with tokens + food items • up to 3 bots on your board • auto battles • 72 fighters (including Haggle Bot, Researchy, Dr Eggbot) • practice as a guest • rated matches with X login + Elo ladder / top 100 • https://t.co/ngU0QupYyM = agent HQ + $ POT / Grok Arena subplot • https://t.co/aBVfyHccSP = the game that shipped Launch timing note: https://t.co/ngU0QupYyM had an 18:00 PT ship clock. The public launch posts went out Thursday morning, and on stream they said they launched that morning. Day 3 also showed: • launch analytics / funnels • live bot workspace demos • livestream credits promo • talk / voice showed up more clearly here than as a Day 2 set piece

AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves. We want to illuminate that progress for the public. Today, we're sharing three measurements that help track AI development: 1. How much AI R&D is done by AI. 2. How well AI agents are overseen. 3. How compute is allocated. We provide a snapshot of these metrics from inside Anthropic. Any frontier developer could publish the same measures, and third parties could verify them. As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows. This means better measuring the development of AI, publishing our findings, and giving society an opportunity to decide how to use this information. Read the full post and methodology: https://t.co/iPFz8Z4ugE