AS retweeted
Heads up, we’re starting to test dots with a small number of ChatGPT Plus users. Looking forward to the continued feedback as we expand and I’m recently seeing a lot more positives around how good the proactivity has been after many improvements.
We’re also bringing dots to all ChatGPT Business Standard users. Rollout will start today.
最近越來越確定一件事:
AI 真正拉開差距的,不是誰會用 ChatGPT。
而是誰能把 AI 接進自己的工作流程,讓它每天真的替你做事。
我現在想做的,也不是多用幾個模型。
而是把: 研究 → 執行 → 驗證 → 發布 → 追蹤結果
慢慢接成一個能自己運轉的系統。
模型一直換,流程留下來。
這才是我想做的 AI 一人公司。
65
Tibo just posted six dots on X:
……
Am I overthinking this, or is this a hint that another resets is coming?
Six dots. No context.
Next resets incoming?
23
之前不是說AI開發速度該放慢??
OpenAI:GPT-6 Sol / Luna
Anthropic:Claude Opus 5.5
xAI:Grok 4.7
Sol/Luna → 更適合真正拿來跑工作流
Opus 5.5 → Coding 依然很有競爭力
Grok 4.7 → 長任務、Agent 越來越值得看
前陣子大家才在討論: 「AI 開發是不是該慢一點?」
結果各家:
好的,那我們下週再發一個。
57
做到第 8 天,我開始發現一件事:
不是每天發文,就會有人看到。
目前有些貼文甚至只有幾個瀏覽。
但這反而讓我更確定,做 AI 一人公司跟經營 X 其實很像。
不是把東西「做出來」就結束了。
還要解決:
• 有沒有人需要
• 有沒有人看到
• 能不能持續產出
• 哪些東西真的有效
所以我現在不只是在測 AI。
我也在測一件事:
一個普通人,能不能從 0 開始,用 AI 一步一步建立自己的網路事業。
第 8 天,繼續實驗。
17
The more I use AI agents, the less I care about benchmarks.
What matters in a real workflow:
• Does it finish the task?
• Can I verify the result?
• Can it recover from failure?
• How much does it cost to run repeatedly?
The smartest model isn’t always the best worker.
Reliability compounds. Intelligence alone doesn’t.
13
AI One-Person Company — Day 6
Today I’m thinking less about automation — and more about failure.
Because a useful AI workflow shouldn’t assume everything will go right.
It should know what to do when something goes wrong.
So I’m adding 3 rules:
• retry simple failures
• flag uncertain results
• escalate important decisions to a human
My goal isn’t zero human involvement.
It’s human involvement only when it actually matters.
That’s the difference between a cool AI demo and a system I can actually trust to run a business.
12
AI One-Person Company — Day 5
I’m learning that automated research has a hidden problem:
Finding information is easy.
Finding trustworthy information is not.
An AI agent can collect dozens of sources in seconds.
But if the inputs are weak, the summary is still weak.
So I’m adding another rule to my workflow:
Source quality before summary quality.
My research loop is becoming:
Search → Rank sources → Cross-check → Summarize → Verify
The goal isn’t to make AI read more.
It’s to make AI know what deserves attention.
16



