Maggie Appleton 写设计、人类学和智能体界面时，习惯先画清楚隐喻，再谈工具。把「The Expanding Dark Forest and Generative AI」整理成可阅读的中文笔记：问题在哪、界面默认会把人带去哪、落地时该改什么。原站导航、评论回响和广告已去掉。
Causal Islands Recording – April 2023
Slides, transcript, and video from my talk on how language models and generative AI could lead us to an epistemic meltdown on the web.
This is an expansion of my essay The Dark Forest and Generative AI published in January 2023 3ya .
Beyond Tellerrand Recording – May 2024
First presented at Causal Islands in Toronto, April 2023. Updated for Web Directions in Sydney, Australia in October 2023, FFConf in Brighton in November 2023, Beyond Tellerrand in Dusseldorf in May 2024, and UX Brighton in November 2024.
Yancey Strikler The Extended Internet Universe Venkat Rao Generative Agents: Interactive Simulacra of Human Behavior Joon Sung Park, et al. Al is being used to generate whole spam sites James Vincent for The Verge A New Frontier for Travel Scammers: A.I.-Generated Guidebooks Seth Kugels and Stephen Hiltner for The New York Times How Spammers and Scammers Leverage AI-Generated Images on Facebook for Audience Growth Renee DiResta & Josh A. Goldstein 'As an A
References in order of appearance
A small footnote: this talk up to date as of about 2 weeks ago which means everything I say about the current state of AI is probably completely irrelevant by now.
I have essentially gathered some thoughts that relate to a moment in time that has just whooshed past us.
Slides and Transcript
First, some context. I’m Maggie and I want to lay out all my biases up front.
I’m a designer at an AI research lab called Elicit . My job involves creating tools that use language models to augment and expand human reasoning.
值得单独记下的观察
- Its outputs are generally indistinguishable from human-made text, roughly at a high school essay level. We should acknowledge that these outputs are astonishingly good. I’ll get to the flaws and weaknesses of language mo
- They’re trained on a huge volume of text scraped primarily from the English-speaking web.
- All language models do is predict the next word in a sequence, which may sound simple but leads to all kinds of complex and potentially useful behaviour.
- Group by different variables
- Surface causal reasoning chains
- Generate research questions
落地时建议先做的 5 件事
- 先写清这个工具在强化思考还是在替人思考。
- 默认交互不要只会谄媚：该追问、该给反例、该要求证据。
- 智能体界面要露出推理步骤，而不是只给一个光滑答案。
- 知识发布用可生长的笔记，而不是一次性营销长文。
- 改产品前先画隐喻：用户以为自己在做什么，系统实际在做什么。
和智能体产品怎么接
龙虾PRO做 OpenClaw 落地时，最该从这类笔记里拿走的是「别把聊天框当唯一界面」。数字员工要能追问、能验证、能把过程摊开，而不是只负责说好话。
本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源
常见问题 FAQ
什么是AI智能系统？
「AI智能系统」可概括为：Slides, transcript, and video from my talk on how language models and generative AI could lead us to an epistemic meltdown on the web. 本文从定义、方法与实践要点展开说明。
为什么要关注AI智能系统？
关注AI智能系统，是因为它直接影响效率、风险与可复制性。文中指出：Slides, transcript, and video from my talk on how language models and generative AI could lead us to an epistemic meltdown on the web.
如何落地AI智能系统？有哪些关键步骤？
建议按以下路径推进AI智能系统：1) They’re trained on a huge volume of text scraped primarily from the English-spe…；2) All language models do is predict the next word in a sequence, which may sound …；3) Group by different variables；4) Surface causal reasoning chains；5) Generate research questions。细节见正文对应章节。
AI智能系统适合哪些人或团队？
AI智能系统更适合：产品/技术负责人、运营与增长团队、需要落地智能体或自动化的中小团队、关注「AI智能系统」方向的读者。若你只需要单次聊天式问答，可先读概念；若要上生产，请重点看步骤、权限与风控相关段落。
关于「Causal Islands Recording – April 2023」，本文给出了什么结论？
在「Causal Islands Recording – April 2023」部分，要点是：The Dark Forest and Generative AI published in January 2023 3ya . Beyond Tellerrand Recording – May 2024 First presented at Causal Islands in Toronto, April 2023. Updated for Web Directions in Sydney, Australia in Octob
关于「Beyond Tellerrand Recording – May 2024」，本文给出了什么结论？
在「Beyond Tellerrand Recording – May 2024」部分，要点是：llerrand in Dusseldorf in May 2024, and UX Brighton in November 2024. Yancey Strikler The Extended Internet Universe Venkat Rao Generative Agents: Interactive Simulacra of Human Behavior Joon Sung Park, et al. Al is bein