"You're talking to a zip file. If you stream tokens to it, it will respond with tokens back."00:04:52
他还给这个实体写了一段自我介绍,是全场最好用的心智模型:
「Hi, I am ChatGPT. I am a 1 terabyte zip file. My knowledge comes from the internet, which I read in its entirety about six months ago, and I only remember vaguely. My winning personality was programmed by example, by human labelers at OpenAI.」00:05:16(白板原话)
00:01:20 · 画面上的原始白板:LLM ≈ 1TB lossy, probabilistic “zip file of the internet”,标注 pre-training ~$10M / 3 months of training on internet documents,加 post-training(SFT、RLHF、RL on conversations);下方是 “New chat” 的 token 条,箭头指向 context window,左侧 User 头像往 token 里写字
"I usually try out the non-thinking models because their responses are really fast. But when I suspect the response is not as good as it could have been... I will change it to a thinking model."00:22:38
00:19:40 · 左上角模型下拉:GPT-4o(Great for most questions)、o1(Uses advanced reasoning)、o3-mini(Fast at advanced reasoning)、o3-mini-high(Great at coding and logic)、o1 pro mode(Best at reasoning);右侧是粘进去的多层感知机代码与 “My gradient check fails and I'm not sure why.”
"Different models, possibly different apps, have different amount of integration of this capability. So you have to be kind of on the lookout for that."00:30:52
00:28:30 · ChatGPT 未手动点搜索,自己判定要联网:「The White Lotus season 3 premiered on February 16, 2025, and airs new episodes every Sunday at 9 p.m. ET/PT on HBO and Max. The season consists of eight episodes, with the finale s…」
"There are no guarantees that there are no hallucinations here. Treat this as your first draft. Treat this as papers to look at. But don't take this as definitely true."00:40:37
00:38:40 · grok.com 的 DeepSearch 输出:页面上标着「52 web pages」,正文分 Key Points / What is CaAKG and Why Is It in the Longevity Mix? / How Is It Associated with Longevity?,并写着 CaAKG 是 α-ketoglutarate 的钙盐、在 Bryan Johnson 的 Longevity Mix 里每份 2.5g
"Different LLMs have different tools available to them, and you kind of have to keep track of it. And if they don't have the tools available, they'll just do their best, which means that they might hallucinate a result for you."00:56:03
08 · 让它动手改文件
Artifacts / Cursor:从「给你代码」到「直接把 app 跑在你面前」
Claude 的 artifacts 把 React 组件直接加载进浏览器——他拿 Adam Smith 的维基介绍生成 20 张抽认卡,再让 Claude 写一个刷卡片的小 app。但他真正的日常不是这个,是 Cursor(底下调 Claude 3.7 Sonnet),用 Composer 从空目录起一个井字棋,然后加 confetti、加音效、加 git commit,一路 vibe coding:
"Vibe Coding just refers to... giving control to Composer and just telling it what to do and hoping that it works."01:09:44
01:04:20 · Claude 侧栏 artifact「Conceptual Diagram of Adam Smith's Division of Labor」,左边是它自己写的五条论点(Central Relationship / Geographic Factors / Historical Examples / Comparative Analysis / Economic Development Path),右边 Preview 里是画出来的关系图,左上角有 Code / Preview 切换01:11:20 · Cursor 里 Composer 正在改 App.tsx:红绿 diff 显示它在改 calculateWinner 相关的行,聊天区写着「Let's also add some additional styling to make the winning cells stand out」,右侧浏览器预览是井字棋,显示 Next player: X
"I like to do it in two steps whenever possible... first I want to transcribe this into text, because I want to make sure that the model is seeing the values correctly."01:33:15
01:33:10 · ChatGPT 里贴了 Bryan Johnson “Lengevity Mix” 的营养成分表截图,用户消息是「Nutrition label for Bryan Johnson “Lengevity Mix” is as follows in the image. First, please transcribe it.」,下面已经开始按 Supplement Facts 逐项输出表格
00:58:40 · “OpenAI Valuation Over Time With Exponential Fit (Log Scale)” 的代码与图同屏:图右端标注 2027.1B,而它下面的文字写着 “the expected valuation in 2030 is approximately $1.7 trillion”
没有工具它会硬猜,而且猜得像真的
Grok 3 没有 Python 解释器,那道大数乘法它给了个「非常接近」的答案;Gemini 2.0 Pro 同理,加大难度就错。这比干脆拒答危险得多——他专门提醒这种错「remarkably close」。