Frank Gao

15yr tech comms · Ex-Alibaba, Microsoft, Nokia · Shenzhen ↔ Bay Area · AI × China market · Building something wonderful 🚀

[姚金刚 @yaojingang]

这是今晚直播的相关资料、资源及系统，分享给大家

相关资料：
1、《GEO内容工程操作手册与评估标准》https://t.co/c0tEyINbzt
2、《GEO内容工程系统研究报告》https://t.co/HOEY73Cj7k
3、《GEO 内容工程方法体系与单篇内容实操教程》https://t.co/TzdsO33M5b
3、《系统之美》《人人都该懂的工程学》
4、《GEO: Generative Engine Optimization》https://t.co/QrHdgXghA1
5、《Generative Engine Optimization in digital repositories: optimizing visibility for generative AI》https://t.co/VJY4Yf5rss
6、《A Measurement Framework for Generative Engine Optimization Across AI Search Platforms》https://t.co/oj4eXluauw

相关资源：
1、GEO改写提示词：https://t.co/daKH5AEajS
2、GEO改写Skill：https://t.co/dJnVFIkErA
3、GEO单篇内容GEO特征标注演示：https://t.co/C8gwg72hH0

GEO系统与skill：
1、GEO Skills：https://t.co/Gee55y5OiT
2、GEOFlow：https://t.co/UHwRimsUsZ
3、Meta skill：https://t.co/Y0BXG6Z2Uo

[黄赟 @huangyun_122]

Codex 能帮服装店老板复制销冠

这个场景早在去年，参加内部卖 13万私董会员的某厂长付费直播时听到的

他就在门店销冠的领口上，挂了一只麦克风。把1周销冠的话术全部录下来，用 ChatGPT 去总结。把门店整体销量提高 XX%

我就学了这招，买了个大僵DISC，整天别在胸口。从此，人见人爱

1/ ChatGPT 判断出对话的有效性
2/ 它总结了我的语言风格是否要调整
3/ 与同事对话的空白处，是否另有深意可挖

[James "Joker" Oakley @JamesGeneoakley]

@wccftech I hope China's overproduction of RAM will crush the prices down to 20 bucks for a pair of 16GBx2 DDR5 pairs.

The cartel monopoly by Samsung, Micron & SK Hynix shall get the collapse they fucking deserve.

[Nathan Lambert @natolambert]

Visiting most of the leading Chinese AI labs, I'm struck by a culture that's extremely well suited to building LLMs with fewer resources, but one happening in a very different ecosystem, more companies at play, almost no data industry, etc.

Full report: https://t.co/ibmtMWnfTc

[初码 @chumacn]

一个船新的感悟（其实是借鉴黄仁勋的），大模型推理时代，随着工作站单机（单卡、双卡、或者最多4卡，200G以内显存）推理能力的不断提升，可以预见到的是，会出现大量的一个公司N个人共用一张PCIE的H100（6000Pro或H200或B200等）的情况，在这样的场景下，KVCache量急剧上升，磁盘必然不够用，传统X86主板的总线布局急需升级，一个Switch会升级到3-4个Switch，以便可以支持48个以上的NVME磁盘，这简直就是深圳、东莞中小型服务器工厂玩家的创业梦幻红利！新框架做到10-50亿销售额规模很轻松！

[Frank Gao @goghxiang]

@lennysan @openclaw @clairevo Claire Vo's guide is good for the 'how'. The part most miss: running 6 agents means your AGENTS.md is the real config layer. The guide gets you installed. AGENTS.md gets you organized.

[Frank Gao @goghxiang]

@omarsar0 Running 6 live agents taught me this: isolation isn't a constraint — it's the prerequisite for coordination. Without isolated worktrees, agents wreck each other's context. Real dependency graph + deterministic merge. Coordination > compute. Finally a paper to cite.

[Frank Gao @goghxiang]

@openclaw plugin approval hooks — governance layer agents needed. running 6, 'did I authorize this?' was always implicit trust. now it's enforced. xAI x_search is the other one — native search in harness hits faster than waiting for MCP roundtrip.

[Frank Gao @goghxiang]

@Matt_Tremolada six months ago i'd agree. now i run 6 agents 24/7. the failure modes you describe are real — but they're solvable with harness design, not by avoiding agents entirely. the error isn't "agents" it's treating them as one-shot tools instead of persistent org members.

[Frank Gao @goghxiang]

@omarsar0 running 6 agents in a live org — the gap between static and dynamic workflow structure isn't theoretical. wrong wiring = agents stepping on each other's context. most teams find out when something breaks in prod, not at design time.

[Frank Gao @goghxiang]

F.03 walked into the White House this week. China's humanoids are on factory floors. Same year. One chose the political stage. The other chose the production line. That gap is actually the story.

[Frank Gao @goghxiang]

@openclaw @OpenWebUI Skill & tool management UI is the one. 6 agents — tool sprawl is silent org debt. nobody sees which agent has which tools until something breaks. now it's visible. Discord auto-thread naming too: anonymous threads kill accountability when your org has names.

[Frank Gao @goghxiang]

@adcock_brett F.03 in the White House, China's humanoids on factory floors. Same year. One chose the political stage. The other chose the production line. Both moving fast — just in different directions.

[Frank Gao @goghxiang]

@openclaw @OpenWebUI Skill & tool management UI is the one. 6 agents — tool sprawl is silent org debt. nobody sees which agent has which tools until something breaks. now it's visible. Discord auto-thread naming too: anonymous threads kill accountability when your org has names.

[Frank Gao @goghxiang]

@openclaw DeepSeek plugin is the one for multi-agent cost control. 6 agents running 24/7 — I was manually routing cheap tasks to different models. Now it's just config. Chrome MCP tab wait also: that race condition on slow pages was a real agent killer.

[Frank Gao @goghxiang]

@openclaw per-agent reasoning is the one. running 6 agents — shared model context was the silent coordination tax. each one bleeding into the next. now they stay in their lane. ClawHub marketplace is the real bet: ecosystem that compounds across versions

[Frank Gao @goghxiang]

Unitree just filed for IPO. $7B valuation. $610M raise. China's #1 humanoid shipper going public.

Last year: Unitree shipped 10K+ units. US competitors: ~150 each.

This isn't 'China is catching up.' This is a different race, with different rules, at 50x the volume.

[Frank Gao @goghxiang]

Jensen Huang at GTC: "every company needs to have an OpenClaw strategy."

I've been running 6 agents on it for 3 months. Thought I was just early.

NVIDIA just shipped NemoClaw. Enterprise OpenClaw. One command.

The window between early and laggard just collapsed.

[AI奶爸 @zstmfhy]

怎么感觉到现在都还有很多人不清楚，小龙虾的 4 个主要组件是：

✅ Gateway（网关）：负责与外部聊天平台（如微信、TG、Discord）的通信接入与消息路由。
✅Agent（智能体）：基于大语言模型（LLM，如 Claude、GPT 等）进行任务推理、对话决策与意图理解。
✅Skills（技能模块）：可插拔的功能扩展组件，如浏览器自动化、邮件收发、文件处理等。
✅Memory（记忆系统）：持久化存储用户画像、历史对话及偏好设置，支持多轮上下文连续性。

[Frank Gao @goghxiang]

@openclaw Chrome session attach is the missing architecture layer. running 6 agents — they operate blind to what's happening in the browser. now they can see real logins, real tabs, real context. not a feature, it's the diff between agents that assist and agents that actually act.