Shek
846 posts
Shek
@ShekOMP
Co-Founder of https://t.co/XHMLSwnVHj
Katılım Ağustos 2020
312 Takip Edilen210 Takipçiler
@joannakurylo @zocomputer You’re too kind! It was great meeting you too. Singapore is even better when it brings good people together 🙌
English
One of my favorite moments of my trip to Singapore has been meeting the GOAT himself, @ShekOMP
Shek@ShekOMP
Grateful to have shared at yesterday's "Kill Your SaaS" event. I have been using @zocomputer for a few months, and it quickly became my most-used daily AI tool. I use it to write and publish social posts, create videos, build my own dashboard, maintain my own LLM-wiki inspired by @karpathy, build slides, schedule daily AI news, create lead-gen landing pages, and much more. Someone asked yesterday: how much building time does it take to replace a SaaS? The honest answer is: it really depends. But in many cases, you can start with a mature open-source project and ask Codex / Claude Code to customize it for your own specific workflow. It is not only about saving money. It is about creating a tool that is hyper-personalized for me, the kind of internal system that would normally require hiring an AI team to build. Because Zo is a cloud computer, you can install almost any open source on it. For example: • HyperFrames + FFmpeg for video rendering, then combine it with a SKILL to create marketing videos for your own brand • Cal. diy for scheduling, then combine it with a SKILL to automatically provide different time-slot selection links for different types of clients • You can even install Hermes Agent and OpenClaw into it I've been paying for software as a “service” that never actually served me. They just gave me a tool to self-serve. • Canva did not design for you. It gave you a drag-and-drop tool so you could design by yourself. • Xero did not do your accounting. It gave you a ledger interface so you could do it yourself. But today, software powered by AI agents — or even a personal computer — can finally help me complete the actual task. That’s what a real service should look like.
English
I originally prepared my intro with @MiniMax_AI using my own cloned voice, but time was tight, so I skipped the intro and focused directly on a few real use cases.
English
Grateful to have shared at yesterday's "Kill Your SaaS" event.
I have been using @zocomputer for a few months, and it quickly became my most-used daily AI tool.
I use it to write and publish social posts, create videos, build my own dashboard, maintain my own LLM-wiki inspired by @karpathy, build slides, schedule daily AI news, create lead-gen landing pages, and much more.
Someone asked yesterday: how much building time does it take to replace a SaaS?
The honest answer is: it really depends.
But in many cases, you can start with a mature open-source project and ask Codex / Claude Code to customize it for your own specific workflow.
It is not only about saving money. It is about creating a tool that is hyper-personalized for me, the kind of internal system that would normally require hiring an AI team to build.
Because Zo is a cloud computer, you can install almost any open source on it.
For example:
• HyperFrames + FFmpeg for video rendering, then combine it with a SKILL to create marketing videos for your own brand
• Cal. diy for scheduling, then combine it with a SKILL to automatically provide different time-slot selection links for different types of clients
• You can even install Hermes Agent and OpenClaw into it
I've been paying for software as a “service” that never actually served me.
They just gave me a tool to self-serve.
• Canva did not design for you. It gave you a drag-and-drop tool so you could design by yourself.
• Xero did not do your accounting. It gave you a ledger interface so you could do it yourself.
But today, software powered by AI agents — or even a personal computer — can finally help me complete the actual task.
That’s what a real service should look like.
English
@wilsonsfh_ @zocomputer @bytheophana @0thernet @joannakurylo It’s great to see students in Singapore so passionate about learning AI.
English
caught the entire @zocomputer team whos here in singapore after a more than full house event!
@bytheophana @0thernet @joannakurylo @ShekOMP
English
First time successfully joining an AI Meetup in Singapore — and wow, there are events almost every night, nearly all packed. Plus free pizza 🤭
One attendee told me that when the Claude Code team came last time, over 1,000 people showed up.
Tonight’s topic was Personal AI Agents. I shared how I built an LLM Wiki in ZO and used Hyperframes to automate video creation.
First time sharing in English — even after 10 years of training experience, speaking outside my home ground still made me a little nervous.
Funny thing is, the ZO team noticed the growth of Hong Kong users and found my YouTube channel. Since they were in town for ClawCon and AIE, they invited me last-minute to share some personal use cases.
Great vibe, lots of people stayed behind to chat. I think I’ll be joining more AI meetups in Singapore soon.
Brian Chew@brianchew
Shek next, a quick sharing on how @zocomputer was used to automate a marketing workflow within it! They made use of hyperframes.heygen.com to automatically curate news, and automate certain marketing video workflows. Super cool! 🔥
English
pretty excited for voice models to get great
its interesting to watch how people are already starting to change the way they interface with AI
English
If you don't enjoy hassle, just use @zocomputer
Riley Brown@rileybrown
OpenClaw starts out very fun, then the headaches start happening after a few weeks. I’ve talked to many AI agent power users who are also struggling with OpenClaw. Gateway issues, cronjobs not firing, skills not being used consistently, memory fall off, integrations need re-authentication, oscillating between overeager and undereager, not placing files in the right place. Etc
English
Do you prefer queuing or steering in Codex? What's your use case for each?
English
Compaction 是否一個過渡期的技術？
Viviennn@0xViviennn
Hermes在周末突然犯蠢，今天修了两个多小时，最后发现是compaction的问题😭 Hermes 有自动compaction的机制：历史太长时，会调用 auxiliary.compression.model 生成 summary，然后删 掉一部分旧对话，把 summary 塞回上下文里。 问题是：如果这个 summary 模型配置错了、不可用、或者 provider 调用失败，压缩器会报错，但流程不会 停。它还是会删掉选中的历史 turn，然后插入一个静态 fallback placeholder。 也就是说：还没有生成真正的 summary，历史已经被删了。 更麻烦的是，这个失败只写进 agent.log。如果你是通过 Telegram / Discord / Slack 这些 gateway 在用 Hermes，用户侧完全没有提示。你会以为 compaction 正常完成了，但其实上下文已经被截断，而且没有 summary 可以补回来。 这类问题对 agent 很致命，因为它不是“报错失败”，而是“静默成功”。用户继续聊天的时候，只会感觉 agent 突然断片，但不知道原因。 修法大概是：当 summary generation 失败并 fallback 时，把这个状态显式传出来，让 gateway 用户能看 到 warning。 包括两种路径： 1. 自动触发的 auto-compaction 2. 用户手动调用的 /compress warning 里会说明这次丢了多少 turn，并提示检查 auxiliary.compression.model 配置。 已提 PR，救救天才： github.com/NousResearch/h… @Teknium @NousResearch
中文
卧槽，今天 Product Hunt 榜单前列那个叫 Kollab 的 AI 工具，你们注意到了吗？现在排到第一了，票数还在涨。
这个平台已经接入了 GPT Image 2.0 和 Seedance 2.0
点进去看了一下，有点刷新我对「AI 协作工具」的认知。
过去一年我们用 AI 挺傻的：开一个对话，问完就关，产出散落在几十个 tab 里。下次做类似的事，又从零解释一遍背景。团队里人人在用 AI，但没人知道别人在用 AI 做什么，更别说复用。
Kollab 干的事，是把这套流程重新做了一遍。它是一个工作空间，不是一个聊天框。
几个让我觉得方向不错的点：
1、Skill 机制。一次跑通的 workflow 存成 Skill，团队里任何人都能复用。等于把最会用 AI 那个人的方法，沉淀给所有人。
2、Memory 是真的 Memory。项目背景、团队信息、历史产出都记得住，不用每次重新喂。
3、Bot 进 Slack / Telegram。群里 @ 它就能查资料、触发任务，产出自动同步回空间。不用在五个工具之间反复横跳。
4、定时任务 + MCP 连接器。竞品监控、周报生成能自动跑，原生接 Notion、GitHub。
5、写作、做图、搭网页一站式。风格定义一次，后面都能复用。
和谁比也清楚：ChatGPT、Claude 解决问问题，产出用完就没；NotebookLM、Genspark 偏个人向；OpenClaw 要搭环境。Kollab 注册就能用，空间持续在，团队能一起协作。
适合同时推进多个项目的人，以及已经在用 AI 但没统一方法的小团队。
官网：kollab.im，免费版直接用。
@Kollab_AI 是他们官方账号，方向挺值得支持的。
中文
@ajambrosino May i know the difference between a
project , GPT and agent ?
English
If I use Claude Code is there any reason to ever use Cowork?
it seems Cowork is just Claude Code without the dev tooling and with UI for the stuff you're doing..
Every time I use Cowork it seems like I should have just been in Code but Idk if I'm missing something or what
English
当你觉得自己 vibe coding 的东西是坨屎的时候，你不妨看看谷歌。
这可能是他们一个团队干了一年的成果。
歸藏(guizang.ai)@op7418
谷歌在产品上真是太慢了，终于推出了 Gemini 的 Mac 客户端。 全部用 Swift 编写的原生应用. 看了一下，功能相当简陋，很多能力都没有。 比如 Artifact 复杂点，网页都没办法渲染。整个 UI 非常糙，谷歌正常发挥水平。
中文
For a week, @Google has artificially pinned Anthropic’s website to the top of their news section — above every other article about the company or product.
The same article has been refreshed every few minutes for days.
Google is rigging its news feed to boost Anthropic’s PR push.
English
Yesterday, I was using Claude Opus 4.6 via Zo Computer to draft and publish some content, but it completely missed the mark.
The issue was with the character limit on Threads. My SKILL (workflow) specifically dictates that if a post exceeds the limit, it should be split into a thread, with the remainder posted as comments.
Instead, Opus made a baseless assumption and—without any verification—proceeded to blast out four identical posts. It should have been one main post followed by three comments.
It’s honestly a bit surprising. I know a lot of OpenClaw were using Opus before Claude’s restrictions kicked in, but in my experience, it gets pretty shaky when running a SKILL with more than 10 steps.
By comparison, GPT-5.4, mimo-v2, and GLM-5.1 have all been much more reliable. GPT-5.4, in particular, is excellent at following instructions to the letter, though its Chinese writing style still leaves much to be desired.
For now, I’ve had to split this SKILL between two models: using a domestic (Chinese) model for the creative writing and letting GPT-5.4 handle the execution.
English
