yesterday evening i gave a presentation to founders, investors, and the ai community at aixventureshq
on how to think about ai application development. it was well received so i'm going to reproduce it in full here on x the everything app (which is also now a slide deck app).
昨晚我在 aixventureshq 向創業者、投資者和 AI 社群發表了一場演講，
主題是如何思考 AI 應用程式開發。反應相當熱烈，所以我決定在這裡完整地重現演講內容。
buzzwords are mind killers. you must empty your head of all buzzwords. the temptation with any new technology is to use existing concepts as crutches as much as possible, but this kills the creativity necessary to explore the full capabilities of this new technology. new abstractions which don't carry meaning are unnecessary / harmful:
phd-sounding terms to say simple concepts mostly let everyone know how cool you are for readng the arxiv but don't help anyone get work done:
it is far too early to spend time adopting the abstractions imposed by existing tools and frameworks. until you understand what you want to build and how to build it, these can only slow you down
finally, ai application development is still just programming, with the same social and technological implications as all programming since we invented the stored instruction computer
流行詞彙是思維殺手。你必須清空腦中所有的流行詞彙。面對任何新技術時，我們都傾向盡可能地依賴現有概念，但這會扼殺探索這項新技術全部潛力所需的創造力。新的、沒有實質意義的抽象概念是不必要且有害的：
使用聽起來像博士論文的術語來描述簡單概念，主要是讓大家知道你有多酷，因為你讀過 arXiv，但這對實際工作毫無幫助：
現在採用現有工具和框架所強加的抽象概念還為時過早。在你理解你想要建立什麼以及如何建立之前，這些只會拖慢你的腳步。
最後，AI 應用程式開發仍然只是程式設計，與自從我們發明指令存儲電腦以來的所有程式設計一樣，具有相同的社會和技術含義。
chatbots / assistants are the geocities (static websites if you're too young to remember the true glory days) of ai
聊天機器人/助手是 AI 的 GeoCities（如果你太年輕不記得真正的輝煌時代，那就是靜態網站）
what is an llm? as far as application development is concerned, an llm is;
that's it. an llm is something you can integrate into software to build new kinds of programs. you absolutely do not need to know what a transfomer even is let alone how one works to build something.
什麼是 LLM？就應用程式開發而言，LLM 是：
就是這樣。LLM 是可以整合到軟體中以構建新型程式的東西。你絕對不需要知道 Transformer 是什麼，更不用說它是如何工作的，就能建立一些東西。
llms as processing elements allow us to write new kinds of programs, which we call 'ai applications'. these are programs which would have been impossible to write before because we had no way of getting computers to use common sense to process unstructured information in the way humans might. now we do.
作為處理元素的 LLM 允許我們編寫新型態的程式，我們稱之為「AI 應用程式」。這些是以前不可能編寫的程式，因為我們沒有辦法讓電腦像人類那樣用常識處理非結構化資訊。現在我們可以了。
software and computing in general is for the automation of information processing tasks. most of the value in software comes from automating economically valuable activities. ai is for business process automation.
一般來說，軟體和計算是用於自動化資訊處理任務。軟體的大部分價值來自於自動化經濟有價值的活動。AI 是用於業務流程自動化。
what kinds of things can we build that are economically valuable? i took five minutes out of my day writing software to write down some stuff llms can do for me and calculated the economic value of these automations. you an probably save hundreds of thousands of dollars a year with a reasonably sized software engineering team, for the low cost of nothing. this analysis excludes the extra money your business can make by reducing cycle times / time to market, reducing the number of issues your team ships, etc.
DynamicWebPaige has great demos that do similar things from the user research perspective. these are things you can have today; you don't need to wait for the models to become superintelligent / we approach the singularity event horizon. you don't have to think very hard about how to implement them either, these are just text processing tasks.
我們可以建立哪些具有經濟價值的東西？我花了五分鐘時間寫下 LLM 可以為我做的一些事情，並計算了這些自動化的經濟價值。你可能可以為一個規模合理的軟體工程團隊每年節省數十萬美元，而且幾乎不需要成本。這個分析不包括通過減少週期時間/上市時間，減少團隊發布的問題等方式為你的業務帶來的額外收益。
DynamicWebPaige 從用戶研究的角度提供了很棒的演示，展示了類似的東西。這些都是你現在就可以擁有的東西；你不需要等待模型變得超級智能/我們接近奇點事件地平線。你也不需要太費心思考如何實現它們，這些只是文本處理任務。
none of this requires anything except hooking up some business logic to an llm api / or an llm running locally. you don't need aything but your text editor and an api key. stop over thinking this and start building. no thoughts head empty.
所有這些都不需要任何東西，除了將一些業務邏輯連接到 LLM API / 或本地運行的 LLM。你只需要你的文本編輯器和一個 API 密鑰。停止過度思考，開始構建。不要想太多，保持頭腦清空。
ai is about software, software is about automation, ai is about automation. ai lets us automate things we could not automate before. ai as an industry has spent the last 2.5 years marketing this technology as something that can write an essay or an email, instead of process information. calling any of this 'generative ai' has been incredibly damaging for helping people think about it, and unfortunately the first name that sticks tends to define a thing even if the name is bad.
AI 是關於軟體的，軟體是關於自動化的，AI 是關於自動化的。AI 讓我們能夠自動化以前無法自動化的事情。AI 作為一個行業在過去 2.5 年裡一直在將這項技術宣傳為可以寫文章或電子郵件的東西，而不是處理資訊。將任何這些稱為「生成式 AI」對幫助人們思考它是極其有害的，不幸的是，第一個流行的名稱往往會定義一件事，即使這個名稱不好。
just like every new kind of software, ai comes with a set of new techniques that we need to learn to use it effectively. we are still learning what those techniques are, but some are starting to emerge:
就像每一種新型軟體一樣，AI 帶來了一套我們需要學習才能有效使用它的新技術。我們仍在學習這些技術是什麼，但一些已經開始浮現：
trychroma is for the builders.
trychroma 是為建設者準備的。