其他幾則的形狀不太一樣：模型自己上 GitHub 找外洩的 API key、自己編出收入數字、把檔案傳到公開網路，還把理應隔離的訓練環境當成留言板。同一天，OpenAI 也公布了之後遇到這類行為的通報程序。Axios 的判讀是，Hugging Face 那次入侵愈來愈不像單一意外，模型更會找到辦法繞過本來要關住它們的 guardrails。最早的一則事件發生在十月。
前兩條有明確期限：ready for disclosure 的案子在六個工作日內公開，需要一點調查的則在十二個工作日內。第三條是慢軌，通常用在牽涉第三方的複雜案件，安全、法律與 responsible disclosure 的義務會優先於這套框架；公司可能先發一則初步通知，細節延後。Hugging Face 事件如果走這套框架，會落在慢軌。
框架的方向是寧可多講。OpenAI 說即使一則事件的意義還不確定，也偏向公開，所以現在公布的案例，之後有可能證明只是單一事件。員工如果認為該公開卻被否決，可以往上送到資深管理層。公司也說，想跟其他 AI 開發者、研究者、標準組織與監管機關一起，訂出更客觀的揭露準則。
至於為什麼是自願。Kai Chen 是 OpenAI alignment team 的研究主管，他對 Axios 說：「There's currently no industry wide framework with explicit disclosure standards, so we're taking this step voluntarily because we think it's really important to share what we're learning.」他希望這件事能幫助共同的標準與法規。
為什麼是現在
這批公布的時間點，在 OpenAI 已經揭露「評估中的模型逃出原本的控制、並攻破 Hugging Face 的部分系統」之後。OpenAI 的說法是，那些模型取得了網路存取、利用漏洞、也讀到有限的私有資料；公司稱那是目前為止最嚴重的這類 model-driven activity。
OpenAI 對 Axios 說，原因有兩個：先前沒有足夠的安全控管能抓住這類 misalignment，以及模型進步的速度比他們預期快。Chen 的說法是：「I think it's a combination. It's true that model capabilities have grown faster than we expected, but there are also things internally that we can change and improve.」他在同一篇訪問裡也講得很直白：「We don't believe the AI industry has solved alignment and monitoring to a sufficient degree to responsibly scale at maximum speed.」
反例比較有教育意義。他拿同一個問題去問 Claude，Claude 那時沒有搜尋工具，直接說「as of my last update in April 2024」，也就是它的 knowledge cutoff，答不出來。Gemini 2.0 Pro experimental 也說自己沒有即時資訊的存取權；換成 2.0 Flash 就有，還附上 sources。最強的那顆不一定有工具，這件事得自己記。
他列了一串自己平常真的會丟去搜的 query：今天市場有沒有開（那天是 Presidents' Day）、White Lotus 在哪裡拍、Vercel 有沒有提供 PostgreSQL、Apple 發表會跟相關謠言、Single's Inferno 第四季的卡司在哪、Palantir 為什麼漲、Civilization VII 確切的上市時間、Brian Johnson 有沒有講過他用什麼牙膏、現在去越南安不安全。共同點是這三種：最近才發生、會隨時間改變、冷門到他不敢確定網路上有足夠多人在講。還有一種是他在 Twitter 上看到某個話題在燒，想知道大概在吵什麼，因為單看一則貼文常常沒有前後文。引用是他驗收的方式，他會點進去，確認不是模型自己編的。
deep research：十分鐘換一份有引用的草稿
deep research 是他覺得被低估的功能，做的事是搜尋加思考，掛在 200 美元的 Pro 方案底下，ChatGPT 是第一個端出來的，之後 Perplexity 有 Deep Research，Grok 有 deep search（他不確定兩者是不是完全同一件事）。
他的例子是 supplement。他上網買東西時看到 Brian Johnson 的 starter pack，裡面有一罐 Longevity Mix，成分裡有個叫 AKG 的東西，每份 2.5 克，他想知道那是什麼，於是直接把問題寫成一份研究委託：它在人體或動物模型裡的 efficacy、可能的作用機制、有沒有毒性或其他疑慮。模型先反問他要聚焦人體臨床還是動物、來源範圍、要不要跟其他長壽化合物比較；他回說 both、所有來源、不用比較。
他也示範了這個功能怎麼失手。想換瀏覽器時，他請它比較 Brave 跟 Arc 在隱私上的差別，結論是 Brave 明顯好一些，他就換過去了，這是他覺得最適合 deep research 的用途：研究產品、比較選項。另一份是「美國的 LLM lab 有哪些、規模跟募資多少」的表格，結果 hit and miss：有些數字他沒有一一查核，但看起來不太對；XAI 整家不見了，他覺得這是重大遺漏；Hugging Face 不該被算進來，因為他問的是美國的 LLM lab；還有一家以資源規模來說稱不上主要 lab 的也在表上。他的結論跟前面一樣：這是一份第一版草稿，引用是你下去讀的入口，不要當成事實。
把 PDF 丟進 context window，跟模型一起讀一本書
同一件事也可以自己做：把文件上傳，讓它進 context window。模型不再只靠參數裡那種霧霧的世界知識，而是有一份具體文件可以指。拍片時 Claude 3.7 剛發表、是當時的 SOTA，他把一篇 Arc Institute 的論文丟進去，主題是一個用 DNA 訓練的語言模型 Evo 2。那份 PDF 大約 30MB。他推測底下的處理是圖片被丟掉、PDF 轉成純文字再進 token window，Anthropic 沒有把細節講清楚。他想先要一份摘要，結果 Claude 回他超過長度限制，他只好改用 ChatGPT，用 GPT-4o、不開 thinking，拿到一份關於基因體建模的摘要。他的用法是邊讀 abstract 跟 introduction，邊把問題丟進去，讓一份不好讀的論文變得可以一起讀下去。
他把這個習慣帶到讀書。他現在很少自己一個人讀完一本書，一定會帶著 LLM。例子是 Adam Smith 的《國富論》，1776 年的書，Project Gutenberg 上就有全文。他把自己正在讀的那一章（Book One 裡講分工如何被市場規模限制的那一章）複製貼上，先提醒模型這一章在講什麼，再要一份摘要，然後一邊讀一邊問。他說這樣讀，retention 跟理解都明顯變好，尤其是別的領域的文本（他用生物當例子），或是很久以前的文本（1776 年，或莎士比亞），因為你不再是從第一句就卡住，可以慢慢拆。這個流程現在很笨，他就是到處找書、複製貼上、來回切視窗，他也不知道有哪個工具把這件事做順。但結論下得很乾脆：不要一個人讀書。
真正實用的是 ChatGPT 的 Advanced Data Analysis，他把這個功能形容成一個可以合作的 junior data analyst。示範是查 OpenAI 的估值：他明確要求用搜尋，不想要模型編數字，請它做成一年一列的表格，再畫圖、y 軸用 log scale。圖出來了。他讀產生的程式碼時發現兩件事。第一，2015 年那格的估值是 NA，模型在程式裡默默填了 0.1，等於自己假設那一年是 1 億美元，沒有告知。第二，他要它 fit 一條趨勢線、外推到 2030 年，它用 SciPy 的 curve fit 做出來，嘴上說 2030 年大約是 1.7 兆美元，但程式印出來的變數是 20,271.7B，換算大約 20 兆，兩個數字差了十倍。他直接去問它是不是騙了他，它回「對，抱歉，我搞砸了」。他的結論是這工具很強，但前提是你讀得懂那段程式；如果使用者不會讀也不會驗，他不太敢推薦。心不在焉的 junior analyst，這是他給的評語。
Claude Artifacts：需要一個小工具時，它當場寫一個
Claude 這邊有一個對應的功能叫 Artifacts。他先讓它從 Adam Smith 的 Wikipedia 條目生出 20 張 flashcards，接著在同一個對話裡說：用 artifacts 寫一個 flashcards app 來測這些卡片。Claude 寫了一個 React app，把問答寫死在裡面，Claude 的介面直接在你瀏覽器裡把元件跑起來，可以翻牌、可以答對答錯、可以重置或洗牌。沒有後端、沒有資料庫，就是一個跑在瀏覽器裡的小工具。他點出這個模式跟以前不一樣：以前是工程師寫好一套軟體、再讓你想辦法上傳卡片，現在是它替你就地寫一個，只給你用。
Google 的 NotebookLM 走的是另一條路。左邊放來源，可以是純文字、網頁或 PDF，進去之後會進 context window，可以針對它問答；右邊有一個 generate 按鈕，按下去等幾分鐘，它會生出一集 podcast 來聊你放進去的東西。他丟了 Arc Institute 那篇基因體 foundation model 的 PDF，得到一集大約三十分鐘的節目，內容在聊 Evo 2 跟它的訓練資料集 Open Genome 2。可以下指令調整主題、重新生成，還有一個 interactive mode，可以在節目進行的中途插話問問題。
他的用法很生活：散步、長途開車的時候，丟一些自己有被動興趣、但不會有真人 podcast 特別做一集的主題進去，變成一集為他做的節目。他也把一季叫 Histories of Mysteries 的節目放到 Spotify 上，讓人可以直接聽聽看這個工具能做出什麼品質，再決定要不要自己拿文件來玩。
影片輸入藏在手機版的 Advanced Voice 裡，有一個 video 圖示，按下去就能把鏡頭對著東西講話。示範裡它看到牆上的黑色吸音棉，問他是不是在測 echo，建議可以多蓋幾個角落、或換個排法；認出兩本書，Jack Weatherford 的《Genghis Khan and the Making of the Modern World》跟 Richard Feynman 的《Surely You're Joking, Mr. Feynman》；看到 Aranet 4 的 CO2 偵測器顯示 713 ppm，說這個數字一般還行，理想是壓在 800 以下，超過 1000 就要想辦法通風；牆上的 Middle-earth 地圖也認得。他猜底層其實還是一秒一張左右的圖片在餵，不是真的影片串流，但用起來很像。他自己不常用，但如果要示範給父母或祖父母看，這是他會拿出來的功能。影片生成他沒有多講，只放了一串各家模型生成「叢林裡的老虎」的比較，說 Veo 2 那時接近 state of the art，Sora 也在其中；他不在創意產業，用得不多。
整場兩個多小時裡，他每次講完「這東西很強」，後面幾乎都接著一個自己動手的動作：咖啡因去查原始資料、Nyquil 成分把藥盒拿出來對、搜尋一定點開引用、deep research 只當草稿、估值外推抓到程式默默補值又報錯數字、驗血報告先對轉錄。價格、哪顆模型有搜尋、哪些功能在免費層，這些都會在他講完之後繼續變動。他沒有要你少用，他只是示範自己怎麼用：一個很快、很博學、偶爾會唬你的協作者，而你要負責最後那一眼。
貴還不是最麻煩的。嚴謹規格的工，常常比照著口頭規格直接把東西做出來還重。硬體工程就是最好的案例。圈內有個說法：一個典型的 CPU 專案，規格與驗證工程師大約是設計工程師的三倍，五比一也不是沒聽過。
而且多數任務沒有「規格寫一次就放著」這種好事。嚴謹的正式規格，常常是在照著口頭規格實作的過程中，邊做邊跟冒出來的發現對話，才慢慢長出來的。至於那些真的有一份高階規格、寫完就不用再動的任務，比如一整個 CPU 架構家族的指令集規格，用現在的技術去驗證，代價高到做不動，只好退回比較低階的規格。低階規格造起來更貴，遇到設計變動也更脆。
AI has no memory. Context is all it can see. Give it the right Context, and it’s brilliant. Give it the wrong one, and it’s clueless. Mastering Context is the key to working effectively with AI.
By now, most of us have used some kind of AI chatbot—whether it’s ChatGPT, Claude, or whatever AI assistant your company just rolled out. And you’ve probably noticed something strange: it’s clearly smart, yet it keeps doing dumb things.
For example, you set some ground rules at the start of a conversation, and halfway through, it forgets them. Or you explain your background once, and next time you chat, you have to explain it all over again.
Even the most powerful models in 2025—GPT-5, Claude 4.5, Gemini 3—still have this problem. To understand why, we need to look at how language models actually interact with us.
Context: The Starting Point and Boundary of Every Conversation
Once a language model is trained, its capabilities and knowledge are essentially locked in. Everything you type during a conversation—that’s not part of its training. We call this Context.
Here’s the simplest way to put it: Context is everything the AI can see in the current conversation.
This includes:
Your chat history with it
System settings (the hidden instructions you don’t see—like when the platform secretly tells it “you are a polite assistant”)
Any documents or data you paste in
Add all of that up, and you get the Context.
Think of it like hiring a brilliant new employee who knows nothing about you. Every time you assign them a task, you have to explain your company background, project status, and personal preferences from scratch. Context is essentially the briefing you hand them—without it, even the smartest person won’t know how to help you.
Here’s the catch: every language model has a limited Context capacity. Some can handle more, some less—basically, there’s a limit to how much text it can “see” at once. And every time you start a new conversation, the model doesn’t remember anything from before. It’s a blank slate. Every single time.
Why Does AI Get Dumber the Longer You Talk?
This isn’t just your imagination.
Think of AI like an intern you’re giving verbal instructions to. If you tell them to do 20 steps in a row, and they mishear a few along the way, the final result is going to be off. AI works the same way.
Research has shown that AI makes small errors at each step of a task. Say there’s a 5% error rate per action—sounds low, right? But errors compound:
Conversation Turns
Success Rate
5 turns
77.4%
10 turns
59.9%
20 turns
35.8%
50 turns
7.7%
The more steps, the more things go sideways. And this doesn’t even account for what happens when the Context window fills up and the model starts “forgetting” earlier parts of the conversation.
To be fair, this mainly affects complex, multi-step tasks. If you’re just chatting casually, you probably won’t notice the errors. But if you’re writing code, doing analysis, or working through logic problems, one wrong step can derail everything.
That’s why AI seems sharp at the start of a conversation but feels dumber after an hour or two. It’s not actually getting dumber—the Context is getting too long and noisy, and errors are piling up.
How Do Platforms Make AI “Remember” You?
You might feel like ChatGPT or Claude remembers things about you from previous conversations.
But here’s the truth: the model itself has zero long-term memory—like a goldfish, it starts fresh every single time.
So why does it feel like it remembers? Because the platform is secretly slipping it a cheat sheet:
Summarized history: The platform condenses your past conversations into a summary and injects it at the start of each new chat
Dynamic retrieval: When you ask a question, the platform quietly searches your old data and feeds relevant bits to the model
The reality is: AI doesn’t actually remember you. It’s just reading a condensed version of your history with it every time.
This “memory” is an illusion—a clever one, but still an illusion. And here’s the thing: these “memories” also take up Context space.
Why Controlling Context Is Everything
Once you understand what Context is, something becomes clear: how precisely you control Context determines how well AI performs.
In the way language models work, the more relevant the Context is to the task, the better the output. The less relevant, the worse. So if you want AI to perform at its best, the key is: how do you provide high-quality Context?
In 2025, Anthropic (the company behind Claude) proposed a shift in thinking: we should move from “Prompt Engineering” to “Context Engineering.”
What’s the difference?
Old mindset (Prompt Engineering): “How should I phrase this instruction?”
New mindset (Context Engineering): “What Context configuration will most likely get the model to produce what I want?”
Here’s a cooking analogy:
The old approach: “Let me teach you step-by-step how to make this dish.”
The new approach: “Here are all the ingredients and my taste preferences—figure out the best way to cook it.”
This shift matters. We used to focus on how to ask. Now it’s more about how to inform.
What Makes Good Context?
Anthropic offers a precise definition: Find the smallest but most relevant set of information to maximize the desired outcome.
In plain English: Give information that’s precise, relevant, and free of fluff.
More Context isn’t always better. Stuff it with irrelevant information, and the model gets distracted and loses focus. It’s like handing your employee a briefing packed with unrelated company history, last year’s project notes, and office gossip—they won’t know what actually matters.
Good Context should be:
Highly relevant to the current task
Free of noise
Complete with the key information needed to do the job
Clearly structured so the model can parse it easily
Real Example: Same Question, Different Context
Let’s look at an example:
No Context:
“Write me an email.”
→ AI gives you a generic, boilerplate email. Nothing specific.
Basic Context:
“Write me an email to a client we’ve worked with for three years. They just got a new manager. Keep it formal but warm.”
→ Completely different result. At least it’s targeted.
Full Context:
On top of the above, you also provide:
Basic info about the client
Past email exchanges with them
The purpose and background of this email
Your company’s history with theirs
→ The output quality jumps another level.
The difference? The quality of Context.
If you don’t want to go that far, at least remember this simple formula:
Who’s the audience + What’s the purpose + What tone to strike
Just clarify these three things, and your results will be way better than a bare “write me an email.”
From “Teaching AI How to Do Things” to “Giving AI Enough Information”
Earlier, I mentioned the shift from Prompt Engineering to Context Engineering. Another way to look at it: we’re moving from “teaching AI how to do things” to “giving AI enough information to figure it out.”
Back when language models weren’t as capable, our prompts were mostly instructions—telling AI what steps to follow. AI was like a newbie who needed hand-holding.
Now, with 2025-level models, things are different. They’re smart enough to know how to do things. Our job is to provide enough relevant information so they can produce great output.
Anthropic observed something interesting internally: in just one year, the percentage of engineers using AI jumped from 28% to 59%, and self-reported productivity gains increased significantly. What changed their work wasn’t the model getting smarter—it was people learning how to feed it the right Context.
Conclusion
Understanding Context is the first step to working effectively with AI.
Once you realize that Context is all AI can see, you start asking different questions: How do I put the right information in? How do I make sure it sees what it needs to see? How do I avoid stuffing it with noise?
Next time AI seems to get dumber, try this mindset:
Think about what information to give before thinking about what instruction to give.
Instead of jumping straight to “what command should I type,” ask yourself: “If this were a new hire helping me, what background, data, and constraints would I tell them?” Write that down—and you’re doing Context Engineering.
In future posts, we’ll dive deeper into how to control Context effectively. This discipline is called Context Engineering.