「上海把 AIE 带到上海」：Swyx 谈大会 franchising、Kill My SaaS 与中美 AI 差分

信源：YouTube 视频 QmDLDy3ZSNE（Latent Space 频道；发布时间未在转写中自述，内容判断为 AI Engineer 上海站（AIE Shanghai）首办前夕录制——片头独白称「明天开幕」，主持人口径为「还有两周」，两处时点不一致，按「开幕前夕」理解）。时长 4344 秒 ≈ 72 分钟；转写共 37 个 [Ns - Ms] 时间段、完整覆盖 0s→4344s，两遍通读无遗漏，本文依据英文 ASR 转写整理（附录含全量中文翻译）。受访者在节目中自述：新加坡人、第五代华人移民后裔、曾在北京留学、金融业（交易/投行/对冲基金，首份工作 Two Sigma）约八年后转科技（Amazon、Airbyte 等）、四年前独立、现为 Cognition 顾问（曾正式雇员六个月）——与公开资料中 Latent Space 主编、AI Engineer 大会创始人 Swyx（Shawn Wang）完全互证（转写中作 “Sean”，即 Shawn 的同音），确认级。主持人自述：Latent Space 老听众、三年前创办服务中国创始人出海的社区、在中国办过千人级会议、AIE 上海本地合作方一员（转写作 Stand Up Partners / Startup Partners 两见，具体姓名未自述，按「主持人」泛称）。归属口径：播客内数字与判断一律标「Swyx 认为/主持人称/节目称」，写作者判断标「写作者推断」；Swyx 是 AI Engineer 大会与 Latent Space 的运营方，本场的多数判断与其自身商业利益直接相关，读者自酌。同域旧文（背景不复述，仅互链）：法律垂直 Agent 的 Tabular Review 与 Legora 判断见《看不见数据也要造 Agent：Legora CTO 谈法律 AI》；消费者自建软件见《Matt Berman：软件成本归零，消费者自建》。本文只写本场增量。

先说结论

这是 Latent Space 为 AIE 上海首办录制的一期「给粉丝的特辑」：Swyx 被一位中国主持人反向采访，话题从造词史一路聊到地缘政治。三条主线值得记住：其一，AIE 进中国走的是 JSConf 式 franchising——「我是麦当劳，你是麦当劳中国」，品牌、频道与行为准则打包授权，因为 Swyx 自认无法在中美限制下自己落地（「我不能给自己找白宫的麻烦」）；其二，他刚做完的 Kill My SaaS 实验给出了对中等 SaaS 最直白的威胁陈述——4 万美元年费的大会管理软件，1 万美元奖金、69 份参赛作、一周交卷就克隆出来了，而「模型会点开你软件的每一个界面然后独立复刻它」；其三，他对中美差分的判断与硅谷主流叙事相反：B2B SaaS 泡沫在旧金山，规模、执行强度与创始人密度在中国，中国缺的不是能力而是分发形态。最后他给了两个自我坦白的担忧：AI 寒冬会让他这种「靠热度吃饭的会议公司」破产，以及安全问题的极端版本（生物武器）让他默认「AI 末日假设」。

一、「上海把 AIE 带到上海」：franchising 作为地缘工程

主持人开场把「把 AIE 带到上海」归功于 Swyx，他立刻纠正：「不是我把 AIE 带到上海，是上海把 AIE 带到上海」——本地团队（转写作 Hanhua、Lei Du、Mengying 等）才是操盘者。这不止是客套：Swyx 自述他不是中国公民、在中国没有人脉、无法设立本地主体，且「我的主业在美国，不能给自己找白宫的麻烦」。因此 AIE 的全球化复制的是 JSConf 模式：他永远到不了每个国家，但每个国家都值得有一场头部 AI 会议，于是打包输出品牌、YouTube 频道和社交媒体，收取分成；合作条件只有三条——不滥用品牌、全部内容公开出版、有行为准则与申诉机制。

这个机制的二阶效应已经显形：巴黎是第一个申请者，「好玩的事实是 AIE 巴黎的主办团队被 Mistral 整体收编了」——如今欧洲头部实验室运营着 AIE 巴黎。新加坡站则另有注脚：Swyx 是新加坡人，却完全没有参与组织，原因是此前新加坡最大的「AI 会议」其实是一场加密货币会议（转写作 SuperAI），他直言「这让我为国家尴尬」。而 AIE 世界嘉年华（World’s Fair）层面的价值，Swyx 给出一个地缘性的稀缺品：实体清单上的中国实验室与美国公司「不能正式会面、甚至不回邮件——不是因为他们是混蛋，是他们不被允许」，但同一会场里的「不期而遇」没有禁令。他把「经我撮合的婚姻和婴儿数量」列为自己的第一 KPI（节目原话，含玩笑成分），今年世界嘉年华目标 6000 人、实际到 7000+，线上观看 150 万人次，30 个分论坛覆盖「成为一名 AI 工程师意味着什么」。

内容资产化做得比多数会议彻底：所有演讲全部转写、编目，以数据集和 MCP 形式放在 ai.engineer/data，「把我们的整个网站喂给 LLM，它就能教你任何东西」，并对每位讲者与组织做深度调研，目标是「AI 的维基百科」。

二、「AI Engineer」是怎么造出来的：身份比技能更值钱

造词史的核心一幕发生在 ChatGPT 发布后的一次黑客松：一位非技术背景的女性问 Swyx「我 all-in AI 了，我能做什么？」他的结论是残酷的——「和工程师相比，她什么都做不了。工程师可以同时控制 100 个 AI 并编程，非技术人员的上限是把 prompt 写得更好」。而当时「 prestige（有声望的）」的 AI 工作是预训练、后训练、PhD、GRPO、DeepSeek 式的 hyper-connections（转写语）；工程侧的 harness 工作不需要高等数学与超算经验，只需要「足够好、且心态开放的软件工程师」。市场上公司在招、开发者想成为，缺的只是一个名字——AI Tinkerers、LLM Engineer、Applied AI Engineer 都有人试，Swyx 选中 “AI Engineer” 的依据是 James Clear《原子习惯》式的洞察：要与人深度连接，不要连接他的工作或爱好，要连接他的身份认同；给他一条成为更好的自己的路径，并配上一个能挣钱的职位头衔，他就会忠诚于你。

他特意把这套逻辑对着亚洲观众讲了一遍：亚洲文化强调 credentialism（资格主义）——「你要有『资格』，要那张纸；不是学位就是证书，不是证书就得有相关工作，而工作又要学位」，死循环；他自己的路径是 learning in public：从金融转科技的顿悟是「金融是零和的，我知道什么告诉你，这个知识就贬值；科技是正和的，分享让价值增加」；在 Two Sigma 遇到坏老板、无聊到把精力全部投入工作外的写作演讲，反而起飞——「learn in public 的终极回报是你可以自己选老板」。他也补了反鸡汤：「努力不等于你有资格赢。很多人很努力但不够好，太糟了；要努力且够好，且做的是实质性的东西。」

主持人贡献了一个中国平行案例：中国最大的 AI 社区（转写作 Way to AI，据其「飞书协作文档、Stable Diffusion 起家、ComfyUI 普及推手、百万级用户」特征，写作者推断为 WaytoAGI/通往AGI之路）与 Latent Space 几乎同时诞生于 Stable Diffusion 时刻——同一个「公开学、一起学」的模式在两个语言区各自长成了头部社区。

角色本身三年间的演化，Swyx 认为名号没变、热点在换：早期是 RAG 与上下文管理，现在是 coding agent、harness 工程、loop 工程、memory；正在分化出 agent engineer、AI architect（公司级「全公司怎么做 AI」的决策人）、AI designer、AI PM。他给出的判据朴素而硬：「凌晨四点爬起来修系统的人才是工程师。」同一个判据被他用来抨击「从不写代码的大空头」：那些宣称「AGI 两年后到来、你们都没工作了，要为下一步做准备—— conveniently，下一步恰好长得像他们的工作」。

三、Kill My SaaS：4 万美元软件的 1 万美元克隆

这是全场信息密度最高的部分。Swyx 的会议公司预算每次大会 300 万–1000 万美元，其中约 30 万美元是软件，一年 SaaS 总支出约 100 万美元——「大约 10% 的预算/利润率，付给又慢又丑、不完全符合需求的别人的软件」。作为老板，员工申请工具他都会批，「加起来就是一大笔钱」；而作为客户提需求，得到的是「排进六个月后的路线图」。于是他做了实验：员工要买的 4 万美元 SaaS，他悬赏 1 万美元克隆成开源（参赛者的 coding agent 订阅费他都报销），69 份有效提交、一周交卷，多数可用、约 5 份「我们愿意付费」。他由此给出对行业的直白判断：假设一家 SaaS 有 50 个客户、每个 4 万美元，客户一旦集体意识到可以克隆，「一年之内从年入 2000 万到归零」；「模型非常擅长点开你软件的每一个界面然后独立复刻——只要我能访问你的软件，我就能克隆它。祝你好运」。

落到 SaaS 业主的出路，他认为没有人知道正确答案，但方向只有两条：要么转成垂直领域的 agent builder（对标 Harvey、OpenEvidence），要么转做专有数据生意，「中间地带的一切，别人都会自己写一个」。Salesforce 级没事，危险的中间层。而他更想建议的其实是买方——中小老板的「黄金时代」：他 10 个员工年薪合计 150 万美元，协作效率提升 10% 就是 15 万美元、实际可能是 5 倍即 750 万美元量级；「过去顾问说下周给你方案，现在我要今天下午」。他自己的软件栈已经是 audience-of-one 的样板：自建 Calendly（两个版本）、自托管 Google Drive、GitHub、Dropbox、Slack、开源版 DocuSign——「签署完成的瞬间文件自动归档进 Drive 正确的文件夹，DocuSign 本尊做不到这个」；对「DocuSign 有网络效应」的提问他的回答是「完全没有：电子签名到处合法、没有专利、所谓网络就是邮箱地址」。他的收束句可以当创业选题库用：「你人生与生意里的任何问题都能找到开源版本，自己托管、定制、集成；每一次上百万次的复制粘贴里都藏着一个绝佳的创业。」

两个限定条件他也主动交代：一是维护不是零，但「你不是 99.999 可用性的关键基础设施，你只是个大家往里输数据的数据库；宕一天、修一天，换取想怎么定制就怎么定制」；二是 2026 年的今天，「随便雇个看起来懂点技术的年轻人来克隆」大概率翻车——技能水位真实存在，slop（低质产出）对员工和自己都不负责，「有时买久经考验的 SaaS 反而更好，只不过我们公司选择 AI-forward」。

大会管理软件的真实痛点案例：下一次会议 1000 份投稿抢 25 个席位，现有 SaaS 只能一个一个点开看，官方视图「排六个月路线图」，而参赛者直接克隆出了单表排序视图。关于「AI 评审与你团队选的赢家不一致」（主持人注意到），他的解释值得所有做评测的人抄下来：「评审只判断提交是否有效合规，不判断全场最佳；我无法对自己都不知道想要什么的东西做 AI 评审——看到 70 个例子之后，我自然知道要哪个。」他还想把这套「悬赏克隆」平台化：当年 Latent Space 的 logo 是在 99designs 上 400 个方案里选出来的，「软件界缺一个 99designs for SaaS」——规格、评审标准、打款流程全都得主办方自己搭，这是他做这场实验的隐性成本（评测花了两个月，五种角色账号来回切换）。

四、Jev/System One：把模型从「祈祷对象」变成 if 语句

主持人问到他与 Diego（转写 Diogo/Diego 两见）的那期对话——产品名转写作 Jev/Jeff（存疑），公司为 TypeSafe。Swyx 的复述构成本场的方法论增量：现在的 AI 工程是「围绕模型搭一个大 harness，做一堆工作然后交接给模型，说『拜托了模型大大，给我正确答案，你可以想十天』」；而 Jev 的立场是「模型是为人服务的，我们不给模型打工」——把模型的作用域收缩到只剩一件事：做选择、估概率，并把它放进关键工作的代码路径里，「像写 if 语句一样随意地写进普通编程语言」。这就是 Diego 说的 smarter software：软件应该在每一点上都聪明，而不是只有一点聪明、其余全笨；速度是次要的，可靠性才是他真正的卖点（公司名 TypeSafe 的由来）。主持人补充了中国侧用例：一位不懂代码的中文播客主播用 Jev 判断播客切片会不会火。

为什么「小模型撒进代码里」这件事几千人喊过、只有 Diego 做成了？Swyx 给了三条：OpenAI 出身的信誉背书（InstructGPT 联合作者，该工作后来变成 ChatGPT）；System One vs System Two 的框架化表达清晰到可以现场跑一个 Doom demo（「你绝无可能用 LLM 做这件事」的直观反证）；以及表现力——粉外套、激情演讲，「表现力+技术能力，我自己是这两者的缩小版，他是双十倍」。顺带一句对机器人创业圈的扫射：中国机器人创始人「全都做洗衣」，Unitree 最好但仍在做 Unitree 会做的事（本周的功夫动作「我以前见过」）——「给我看点我没见过的，我才会真正注意」。

五、给出海创始人的 GTM 与「品味」定义

主持人的社区服务中国创始人出海三年，观察是：技术执行力强、GTM 与本地市场理解是断层（「中国创始人闯美国，技术你很强，营销你很差，去找擅长的人」）。Swyx 的回应用了画龙点睛的比喻（转写 garbled 为 hua long ting ting）：「龙是你自己画的——你必须真的在某一件事上是世界最好；我只帮你点睛。有些人没有龙却以为自己有，那是自我。」工程侧的通病是把营销当最后一公里：「上线前一晚让我看博客?我什么也做不了」——GTM 应该在产品成型之前就进入：选哪个细分市场、怎么讲故事；而现实中 GTM 团队往往是公司里最junior 的（实习生、22岁运营社媒）。他给的自查清单：客户是谁、他们在哪里出没、他们用什么词描述自己的问题、看到你的东西后生活会怎么变好。配套原则是 show, don’t tell：TypeSafe 出圈的 Doom demo 不是官方做的、是 GTM 团队做的——「给我看这个东西正在产生好处，别给我 PDF 和 PPT」。

关于被投资人用滥的「品味」，他的定义自嘲「没什么用但准确」：在巨大的可能性空间里，快速、低成本地收敛到一个回望时「显然就该是这个」的子集；蛮力试 1000 个也行，但有品味的人说不清理由地知道另外 999 个不行。例证：iPhone 之前智能手机都有键盘，乔布斯坚持触屏平板形态、研发十年，之前 General Magic 试过并失败——现在全世界的手机都长这样。AI 产品的品味样本：Cursor 的每个决定（如今做 Grokbot，逼得 OpenAI 和 Claude 来抄——「别人开始抄你刚发布的东西，就说明你有品味」），以及法律领域的 Legora：大规模并行的文档评审，「任何人都能想出这个点子，点子很便宜，但你没做，Legora 做了」（本场再度提及，与旧文互为印证）。

六、上海与旧金山的差分：规模、996 与「错的国家」

被问到期待在 AIE 上海看到什么，Swyx 的答案是「差分」：旧金山有 B2B SaaS 泡沫，而中国用户的工具栈完全不同——他问主持人「你们用 ComfyUI 吗」，得到的回答是「我们这儿不怎么用，我们喜欢我们的 Canva 和 Fair AI（转写存疑）」；OpenClaw 在中国的真实声量（「我听说过，但我还没亲眼见过那种能量」）。中国的尺度感对他造成冲击：「对你们来说一百万人的社区是中等规模——在这里已经是顶流」；网红工厂的规模、竞争烈度都是另一个量级。996 的对照更直白：「旧金山的人说自己 996，其实没有——周六周日还在放松；在中国是动真格的」；主持人体感是「创业公司和大厂都正常」；Swyx 补了一个法律对照：「在这儿（美国）法律上你不能要求员工那样工作，是违法的；我见过的最顶尖的美国创业公司，也没有中国拼。」他最重的一句评价给了 BAT 那一代创始人：「他们是那么好的创始人，只是生活在错误的国家；如果他们在美国，会干掉 Amazon、干掉 Facebook。他们没有更富的唯一原因是他们住在中国。」以及比亚迪：2007 年他在北京留学时 BYD 还说「我们立足中国」，如今「BYD 在接管美国，疯狂」。

但他同样点出中国的结构性短板——分发：中国活动受直播限制、呈现形态也不对西方观众胃口。他举的例子是阿里云把整个阿里巴巴路线图连同 Qwen 一起在一场企业发布会中途讲掉，「美国没人看到，因为那是嵌在别人代讲的 corporate presentation 里——是我自己跑去给美国观众逐条翻译的。跟我们一起做嘛，放到 AIE 上讲，他们就会看到」。MiniMax 今年已参展旧金山站；他的立场声明是「我们不碰政治，只做技术」。

七、他兴奋什么、怕什么

兴奋清单：第一名为 System One 模型（「无聊但明确的答案」）；真正被低估的是 interaction models——Thinking Machine 在做的全双工交互，「你在说的同时我在听、你能打断我、我立刻调整——这才应该是 AI 的样子」；推论是持续学习（continual learning）：不再有预训练/后训练的切分，「用着用着一直在学」；以及无需解释的世界模型。担忧两条，都带着罕见的诚实：一是 AI 寒冬——「这是人类历史上最长的 AI 盛夏，AI 史约 50 年，冬夏交替；如果冬天来，什么时候来？我会不会破产？」一个靠行业热度吃饭的会议公司的真实风险敞口；二是安全——实验室的朋友们「真的非常担忧」，而社会调整速度跟不上；不是假新闻、不是失业，「是更糟的东西，现在以网络安全的形式表达（我每天都被黑），极端版本是生物武器——oops，我发明了 COVID-2」。他的收束立场是节目组的默认口径：「默认假设应该是 AI 末日：人类历史上，更先进的文明遇到更落后的文明，后者灭绝；而这东西的进化速度大约是我们的 1000 倍，某个时刻它会比我们先进，我们必须想好怎么办。」内容推荐环节顺带暴露了他的新实验 aeo.latent.space：拿 200 个「最好的 X」问题去问各个模型（最好的支付公司、银行、数据库……），把模型答案转写成排行榜——「AEO（答案引擎优化）是人们所给答案的代理，也是实验室选数据集的人的品味代理；他们选了 Latent Space，排名第一到第三」。

机制、二阶影响与读者启发

1. 大会 franchising 的地缘工程。变化：单一品牌无法覆盖所有国家 → JSConf 式授权（品牌+频道+分成+行为准则）。机制：Swyx 保留品牌与分发，本地团队承担全部落地与风险——规避了他「无法在中国设主体、不能惹白宫」的约束。二阶：本地主办团队被本地龙头收编（巴黎→Mistral）；实体清单双方的「同场偶遇」成为稀缺的合规交流接口。观察指标：AIE 上海是否续办、中国实验室是否借 AIE 通道向西方分发（阿里云 Qwen 路线图无人看到 vs「放到 AIE 上讲」）。启发：地缘限制越紧，「可合法相遇的场域」越是资产；分发渠道本身就是护城河。
2. 克隆成本坍缩 → 中间层 SaaS 的生存挤压。变化：coding agent 把克隆一个中等复杂度 SaaS 的成本压到「1 万美元奖金+一周」（69 份提交的实证，节目称）。机制：会编程的客户变成竞争者（「我们这些老板才是他们软件真正的竞争对手」），且模型「点开每个界面就能复刻」。二阶：mid-tier SaaS 要么转垂直 agent builder、要么转数据生意；SMB 进入 audience-of-one 自建时代；可能出现「99designs for SaaS」的悬赏克隆平台。反方/限定：技能水位真实、slop 有害，有时买成品更好；维护非零；他自己是「因为要做播客内容才愿意付这个组织成本」。启发：软件价值从功能集合转移到数据、集成与责任边界；评测者无法对自己不知道想要什么的东西做 AI 评审——先看够多例子再定标准。
3. 造词=最便宜的护城河。变化：harness 工程长期「无 prestige 却被雇佣市场需要」。机制：给身份（AI Engineer）而非给课程，借 James Clear 的身份认同理论把散沙开发者变成职业共同体。二阶：岗位分化（AI architect / agent engineer / AI PM）、工程师对工程师的大会生态（AIE 复制 GDC 模式）、反 credentialism 的示范效应。启发：在任何「有需求、无名字」的缝隙里，命名+社区+头衔是三方共赢的杠杆；个人版本是 learning in public——分享让知识增值（正和）且让你「自己选老板」。
4. 中美差分=分发差分。机制：中国的规模、执行强度（真 996）、创始人密度被 Swyx 判为全球顶级，但直播限制+发布形态使进展在西方不可见，需要人工翻译中介。二阶：AIE/播客这类「翻译界面」成为跨文化影响力的瓶颈资产。启发：判断跨国 AI 格局时，把「谁能看到」与「谁做得出来」分开评估；内容分发形态的适配（对西方观众的讲法）本身就是中国团队的出海课题。

未采用条目及原因

• 片头独白感谢名单（Meng Xiao 长期翻译每期节目、Mengying 与团队组织）：人名无法核实（转写存疑），仅入校正表未展开。
• 「被撮合的情侣已基本结婚」等社区轶事：并入第 1 条链条的 KPI 语境，未单独展开。
• 主持人对 Jev 中国用例、出海社区运营的自述细节：归属主持人且未独立核实，正文已按「主持人称」保留主线、未展开。

转写校正表

转写原文（修正版）

[0s - 120s] Oh, this is a special podcast because it’s about AIE Shanghai. AIE Shanghai is the first AIE in China. It’s brought to us by Stand Up Partners, which is one of the local partners that we have. We have a few AIEs worldwide that my team doesn’t run, and this is the first one in Shanghai. I’ve always wanted to bring AIE to China. I think there’s a lot of East meets West, and a lot of progress in China that needs to be shared and distributed among the global community. China is super interested in us. There’s been many Chinese fans of the Latent Space, and I’ll flash a few of them up on screen here on YouTube. And primarily, I want to recognize Meng Xiao, who is our primary fan, always translating basically every episode into Chinese, which I which I am super appreciative of. And in general, I think I want to also congratulate and celebrate the progress that China has made, not just in open source models, but in robotics and a lot of other sort of startups. You know, who can remember Manus that was started this time last year and then moved to Singapore? And obviously, OpenClaw had a big run, and we talk a lot about what that like, what that’s like in terms of the overall mood on the ground and how the scale of AI community is so different in China. And the type of things get they get adopted, as well as the I guess the startup culture. People hear a lot about nine nine six. We talk about the reality of that versus in a place like Cognition. And finally, we also talk a little bit about the origin story of the Latent Space, which some of you may not have heard if you’re newer to the podcast. Talk about when it was started and why it was started, and how we think about running the multiple shows that we’re running, and a little bit about safety and alignment and all that other stuff. So there’s a little bit here for everyone. It’s basically for the fans if you’ve been following the Latent Space for a while, and if you’re interested in AIE Shanghai, it’s not being live streamed, but

[120s - 240s] We’re going to post it up on the YouTube at some point. It starts tomorrow, so anyone here in China who’s watching this, I’m really looking forward to having it. And thanks to Mengying and team for organizing all of this. I also have to shout out Hanhua and Lei Du, who basically got this thing started. You know, it nearly did not happen a few times, but it’s so so exciting to have AI in Shanghai and to bring latent space to China in general. So thanks for watching and look forward to the interview. The models are very good at clicking through every screen of your software and then replicating it independently. Oh yeah. Right. So the moment I have access to your software can clone it. So good luck. There’s a lot of talking heads who have never coded a single day in their lives. They all like my timeline to AGI is like two years away. All of you are going to have not going to have jobs in the future. You need to prepare for what’s next. Conveniently, whatever is next looks like their job. I have my own calendar. I have my own Google Drive. I have my own GitHub. I’m serious. There are such good founders. Yes. They live in the wrong country. If they live in the US, they’ll kill Amazon. They’ll kill like Facebook. The only reason they’re not richer is they live in China. AI doom. In all of human history, when a more advanced civilization comes along and meets a less advanced civilization, the less advanced civilization goes extinct. And this thing is evolving roughly 1,000 times faster than us. Hi Sean. So thank you for joining me for this show. So I’ve been a good like long time listener for the latent space podcast. I know that many founders in AI engineers in China as well have already learned a lot from your work and conversations on the podcast. Well, and also I’m grateful that you know you are bringing the AIE the AI engineer conference to Shanghai the first time this year. So I think also like seven months ago I read your piece actually on July 4th about your essay on you know your relationship with the Silicon Valley. You move back and.

[240s - 360s] Force from Singapore and here, I guess that’s why how you actually coined or invented the the role of AI engineer. The last time you moved to Silicon Valley, and also this just makes me feel like it’s very special. And that this time you bring the AIE to to Shanghai the first time. So tell us a little bit about your background and your relationship with Silicon Valley. I I should correct you on one thing. That’s that’s wrong. I did not bring AIE to Shanghai. Shanghai brought AIE to Shanghai. Okay, right. It’s like Hanhua and like Lei Du and Mengying and those those folks as startup partners. So I’m very grateful for them because we. I always wanted to do something in China, but I’m not a Chinese citizen and I have no contacts or relationships there. But I think that there’s a lot of interest on both sides of the world on U. S. And China. China Chinese AI in general, you know, whatever we can do to to surface like high quality conversations, I think that AIE in China can help there. So my TLDR is I’m from Singapore. You know, fifth generation Singaporean. Oh sorry, fifth fifth generation Chinese, right? Most Singaporean is majority Chinese. I think I am from. People ask me like, “你的家族在在哪里?” Oh, nice Chinese. I I was liuxue (studying abroad) in Beijing. Oh, cool. So and I don’t know, and like that’s that’s like how it feels like to be like Chinese diaspora, right? Like we don’t know what our origins are because it just makes that nobody kept any record. It’s like nobody even knows our birthdays, like sometimes, you know. But fifth generation Singaporean, grew and raised up in Singapore and went to the U. S. For university and did my first career in finance. I was in trading and investment banking and hedge funds for ten years, sorry eight years, and then switched to tech.

[360s - 480s] So I worked in Amazon and a few startups, and then four years ago became independent. Wrote latent space, figured out that there was going to be this AI engineer thing, and then I I started AIE. Cool, and I know that actually you have different roles right now. You are you are you are doing the latent space podcast. You are writing newsletters, and also I find out that you have different roles in several startups like Cognition, etc. Yes, yeah. Just wonder, you know, you know what your current titles or roles in different different roles and. Yeah, so it’s very funny because I recently had to go over my taxes. I figured out like what money do I have? So Cognition is the biggest one. Is I’m an advisor. I’m still an advisor to Cognition. I was an employee for a long times for for for a while. So basically, the idea was I have three companies. I have AIE. I have latent space, and then I have miscellaneous. It used to be small AI, which is my personal sort of startup vehicle. Basically, got eco hired into Cognition, worked there for six months, and then stepped back to be an advisor because AIE is now you know a global conference conference organization, and I had to be CEO. I couldn’t be part time employee part time here. Cognition is very big. Fireworks is very big. Models very big. Supabase. So I think it’s a very common thing if you’re a Silicon Valley influencer or you’ve been in developer relation circles for a long time. So that that is my specialty in which is that I am a technical, but I also write and speak, and that is a good combination of skills to have for people because any startup will want to pay you to like talk about their stuff, and I think you know then then you start working with a bunch of things. So like a lot of this is not a normal career. I actually don’t even recommend it for most people, but it is my role in life that I I can advise a few different companies and work with them in in exchange for stock, and that’s the that’s the rough idea. I just want to deep dive a bit more into the.

[480s - 600s] Early days of latent space, because I know I read your essay that you start to you know learn in public and then tell everyone what you learned in the early days and then and then you felt the need you need to you know transfer the conversation you talk to with different researchers online and then become the early some conversation of latent space probably just you know yeah about the past days okay build up latent space and there’s there’s kind of two different things that you’re mixing together right the first thing is the principle of learning in public which is something I realize when I left finance and I worked in tech why is tech so much better going so much better for me it feels so much more comfortable and the very core reason for people who are interested in both fields both are high paying jobs like prestigious jobs the difference is finance is zero sum right if I know something and I tell you that that truth loses value right but in tech it is most stuff is things are positive sum if I know something I tell you it increases value and that’s a very different paradigm for people who are crossing over from finance to tech and it’s very liberating it means that people will get on stage and they open source their work they share their code they screw up they share that they screwed up in finance no like you want to pretend that you never screwed up and everything is top secret and I think that is something that I really took to heart and I think I realized that oh the the other thing that happened to me was my first job was at Two Sigma which is still in finance but I was on the software engineering side and I realized that I had a bad boss so then I got really bored and I started to do a lot of writing and speaking because at work I wasn’t able to do the things that I wanted to do because they didn’t let me but like obviously because I was so bored at work and then I put all my energy into the the stuff outside of work then obviously that thing took off because no one else is doing that and by comparison.

[600s - 720s] That becomes you being able to choose your own boss, right? Like most people, when they have a job, they join a company, they are assigned a boss, they are assigned a team. Most of most of the time, you don’t even know if like the team is good or not, or like whatever, and then you find out it’s not good. You’re stuck. When you learn in public, you actually can choose your own boss. You choose you like work for yourself, choose your own boss, all those things. So it’s very very good. I think the fast forward to Latent Space that was created when Stable Diffusion happened. This is before ChatGPT. Yeah. When Stable Diffusion happened, I was in a startup that was not going very well. It’s called Airbyte. It’s still doing okay today, but nobody would call it a rocket ship. That’s just that’s objectively true, and I would say that to the founders. But because I was like, okay, I have I did like some machine learning courses, but I’m not an AI expert. I am so interested in this field. There’s there’s only one way for me to get into the field, which is to learn in public. And so that’s what I did. I started a blog and I started analyzing and learning and like writing up my learnings into a blog. And at first, nobody will notice. But then eventually, if you start talking about people’s things, like literally, Emad Mostaque, who’s the CEO of Stability AI, which created Stable Diffusion, which is the the reason I started the blog, read the blog post analyzing Stable Diffusion and said this is good. And I think like that work was basically me learning AI like the past like five years of like just grinding and and and writing about that. And I think that is really the you know like you are not qualified, and I’m still probably not that qualified today, but I’m a bit more qualified than I used to be four years ago. So that’s good. And I think that’s the only way you can learn. Like I think especially in Asian cultures, there’s a lot of emphasis on credentialism. Yeah. Like you need some资格, right? Like you need that piece of paper. You need like some qualification. Like if it’s not if it’s if it’s not a degree, it’s a certification. If it’s not a certification, you must have had a job to do it. And how do you get the job? You must have a degree. Like.

[720s - 840s] Like it’s just like a infinite loop of things that basically mean that when you’re 16, you need to decide what you want to do in life. But like I think life is a lot more flexible than that, and you can choose to pivot if you want to. You just have to be really persistent and also realize that just because you worked very hard, you’re not entitled to win. You have to work very hard and be good. I think a lot of people try, a lot of people fail, and they are like, “Why did it fail? " Well, you’re not good. Too bad. But some people try and they worked and and they had judgments. They really did work on substantial things rather than insubstantial things, and that’s what that’s generally what happens with learning public. Like, um, so so yeah, that start this started Latent Space, and then very very interestingly, Latent Space was just a blog. It was just starts was just a blog. Yes, just a blog, and I had no no other idea apart from this is Swyx’s writing about AI. That’s it. But we happened to host some events like some gatherings like, oh, like you know, ChatGPT came out like obviously the most exciting thing of all time, and then we wanted to host some hackathons, so we hosted a hackathon, and at the hackathon, that’s where the idea for AI engineer came about, which is the realization. It actually was this woman who came up to me and said, “I’m I’m all in on AI. I’m non-technical. What can I do? " And I realized that the simple answer is based compared to an engineer, she can do nothing. Engineers are about to be so much more powerful than the non-engineer because engineers can control 100 AIs at once and program it. Yeah. Whereas the the maximum that non-technical person can do is prompt better. Then I was like, oh, like I think I have seen this before. So my my career throughout my career, I’ve seen like professionalization of roles that are non prestigious. So what do I mean by non prestigious? It is not prestigious to be a prompt engineer. The prestigious thing is you train large language models. You do pre training, you do post training. You have to go through a PhD. You make fancy algorithms like G.

[840s - 960s] RPO, you know, and and you you do manifold hyper connections, which is what DeepSeek does. That’s prestigious, and that continues to be prestigious. But I think the overcorrection that people have is that looking at everything else that is not that ML research, that is worthless. That actually was untrue. You actually do want to build harnesses, and harnesses do not require you to know advanced advanced math or have a lot of experience in supercomputing. Like actually, you just need to be a very good software engineer that is open minded enough to understand how to put AI into traditional software code, which is a harness. We didn’t have the word for harness at the time, but it was very clear that engineers would be able to do things like this that normal people cannot. Right, and they would be in demand. And there’s not only that companies are trying to hire this kind of people, and developers want to become this kind of people. They want they want a term to call themselves something. And so, all you have to do is figure out the right term, which is the simplest one is AI engineer, and make it a industry, make it a discipline. And so that’s what I did. Actually, I can share with you a side story in China is that at the time when you you just mentioned the the advent of Stable Diffusion, the open source of Stable Diffusion, it was actually at the time that you know right now the largest AI community in China, which is called WaytoAGI. Actually, that was also started from a bunch of people who are ambitious or lovers of those Stable Diffusion technology. I want to transform into different like images generation, and also it become the foundation of the comfy UI. You definitely familiar with that comfy? Yes, comfy. Comfy. We’ve had him on the podcast. Yeah, yeah, yeah. And in China as well. And then so do they use comfy UI? Yes. Oh wow. And also and also then it like gradually invite more people to learn in public and contribute more contributions on online document, which is on the Lark, which is Feishu, but.

[960s - 1080s] Which was on the lock, so everybody contribute their knowledge,s share their use cases about stable diffusion, comfy UI, and then become larger and larger. I think right now the online community has around a million users learning together, and several million users like learning together, so become the largest AI community in China as well. So that’s also I think that happened at exact exactly at the same time. I think you yeah you start to feel like you need a we need a new word for for this type of engineer, and then which is you just publish the essay of um the AI engineer. Yeah, I think people. So here are some of the alternatives. I think you you can best understand it by the alternatives that people were trying at the time. So there was AI tinkerers, which is another significant community. I know Joe and Alex well, and then there people were saying LLM engineers or applied AI engineer or all these all these other terminology. And I think basically you have to focus on the job title that people want to have, and I think that for us that was very key. And I think I picked this up from the James Clear. Is it James Clear that wrote Atomic Habits? I think so. I forget I forget I forget the authorship relationship, but I think it’s James Clear. Basically, if you want to deeply understand or connect with somebody, you have to connect with them not on the level of their job, not on the level of their activities or hobbies or whatever. It is their identity, like who are you? Right. What What do you aspire to be? If you can give them the path towards being the better version of themselves, then you can build a community around that. And so, a lot of non AI people wanted to be an AI, and what is the path for them? While you give them a job title that pays money, like yes, then they will be loyal to you. Yeah, yeah, right. So so yeah yeah now there’s a lot more AI engineers in the world.

[1080s - 1200s] Yeah, and to follow up, just to my interest as well. So, I know that there are just a lot of works or to do, all right, as an AI engineer. Right now, you mentioned harness. We a lot of developments happen in the AI. So, do you think the role of the world AI engineer has been evolved as well, or is different from what you just mentioned, like three years ago, once you started this work? Honestly, it hasn’t really changed that much. I think that what’s hot within AI engineering has changed, but the job title is pretty much correct. I think there are people in the early days. It used to be rag, yeah, and managing context and all that. And and then more recently, it’s been coding agents. It’s been harness engineering. It’s been maybe loop engineering or whatever. There’s also other disciplines on memory and all these things. So for those who are interested in all the fields of AI engineering, in AIE World’s Fair, which is the the flagship, biggest conference, we have thirty different tracks of what it means to be an AI engineer. And evals is part of that. Multimodal is part of it. You know all that stuff. So I think it has shifted over time. I think agent engineer maybe is is one of them, or AI architect. You know, so AI engineer is like the individual contributor version of this. But who is like the company level, company wide boss? And AI architect seems to be the emerging term for for that that person that decides this is how everybody in the company does AI. Yeah, that’s AI architect. So and then like there’s there’s all the all the other like AI designer, AI PM, all these things. So it’s just like who does that work and who writes code and who is responsible for checking that code if if something fails, are you responsible? No, you’re not an AI engineer. Right. You know, if the person who wakes up in like four a.m. in the morning to go fix it, that’s the engineer. Yeah. Right. And so we want to serve those people, and that’s the real core of it. With the other problem with AI is that there’s a lot of bullshit. There’s a lot of talking heads who have never coded a single day.

[1200s - 1320s] In their lives, but they are very big picture. They’re very very senior. They are like, oh, I’m CTO, blah blah blah, and but they have they don’t touch code. And so a lot of AI conferences and communities focus on those guys because they have big fancy titles, but then they all sound the same. They all like, oh, my timeline to AI is like two years away. You know, all of you are going to have not going to have jobs in the future. You need to prepare for what’s next. Conveniently, whatever is next looks like their job. So it’s very self serving, and I think that that’s one of the reasons that AI engineer works is that we try to focus on the lower level engineer that usually doesn’t get the spotlight that actually did the work and is able to talk about all the war stories and share with their fellow engineers about what they’re doing. And so I borrowed this idea from other parts, other conferences I’ve seen in my career, like GDC and KubeCon and JSConf and PyCon and all these other conferences that I’ve spoken at, where basically it is engineer to engineer, and they are the decision makers for the tech stack, and they are the people who create and innovate the new things, and they’re not necessarily the most famous person, but they are definitely correct when it comes to the engineering. There’s no bullshit, right? And I think there’s a lot of bullshit in the other things. Yeah, yeah, you know. And then it comes to the the first AIE summit that you had organized in like three years ago, and then now it’s actually a three day full summit with over seven thousand people joined this year. Just just for the audience who are not that familiar with what this event is and different kinds of contents that or workshops or the topics you actually a AI engineer is caring about, and then how did you actually bring this like small event to this such a big scale scale event this year? AIE World Fair is basically a similar industry.

[1320s - 1440s] Conference similar to the ones that every other industry already has. So, gaming has GDC, thirty thousand people in San Francisco. Security has two. Security has RSA. Security has Black Hat. If you’re in the cybersecurity field, you go to one of those things, and that’s that’s where you learn. I really like the gaming one because gaming is very similar to labs. You have AAA studios. They work on the hardest games in the world, and it is extreme engineering. And they’re all top secret. They don’t do anything else, but one time a year they come to GDC, and they will share like this is how I did game of the year, and that’s the only way. If you’re trying to be a game developer and you want to work at a AAA game studio, it’s the only way you learn. Yeah. So I think that now the model labs are similar to game studios. They have a lot of money. They’re working on very very big major projects. There’s a lot of smaller companies that are you know trying to compete with them, but the labs are the important ones, and you want to learn from the labs. And where are you going to learn from them? You know they they do like very small conferences where they like they launch like Developer Day for OpenAI or Code with Claude for Anthropic, but those are super small and like very selective and mostly for customers, and they’re not for engineers talking to other engineers. And then there’s like fifty other labs that are not OpenAI and Anthropic that also have something interesting, right? So they usually every industry needs one big like everyone comes here and then you just walk around the room and just talk to everybody, and then they can talk to each other, yeah, which is very interesting because especially for Chinese labs, some of whom are on the entity list, is this something that you guys talk about? Yes, they cannot officially take meetings with you. Yes, a lot of American companies cannot officially take meetings with you. They’re not allowed. They don’t even respond to you. It’s not because they don’t want to. It’s not because they’re assholes. They’re not allowed. But if you’re in the same location, you run into each other, right? Right. And I think that’s that’s kind of interesting as for me to create opportunities for for magic to happen, and it’s like people get hired, people find.

[1440s - 1560s] A co-founder. I have one couple that started dating, and now they, you know, they’re they’re basically married. And like, yeah, that that is my number one KPI number number of marriages and babies that I’ve created. And so, so I I really like that. I I think that we can create that kind of high quality community. And you know, we we’re aiming for six thousand people this year. We managed to hit seven thousand plus, one point five million online watched our talks. So there are people who cannot afford to come, don’t live in San Francisco, don’t live in the US. They can still watch our videos and catch up on all the content we publish. Every single talk, every single workshop, and like I said, it’s like thirty different tracks. There’s too many tracks to name, but you can see on the website. But I I basically think that I really try to focus on things that are working, that you can make money. They’re clear role models for you. Like if you’re trying to learn this field or you’re a startup trying to build something, there’s somebody who has probably talked about your thing at AIE. Right. That becomes a very valuable resource for anyone looking up anything. So another thing I’ve actually recently done that I haven’t really been super public about is I transcribe everything and I catalog everything. So you don’t even have to watch the talks. You can just take all of our website and feed it into an LLM, and then it can teach you whatever. Oh, are you actually creating a CLI a base on top of everything? Everything is is MCP. MCP. But yeah, everything has been transcribed. It’s available as a dataset. If you want to train your LLM on it, yeah, go ahead. Oh, have you published yet? Yeah, it’s it’s on AI. Dot engineer slash data. Oh, nice. Which also yeah shows the shows the MCP and the data sets that you can download. It just depends how you want to consume it. You know, we transcribe everything. We also deep research every speaker and every organization. So whoever has ever spoken at AIE will you have a profile and will basically try to make the Wikipedia of AI. Nice. And just one last question to follow up the AIE is that when did you actually bring the AIE?

[1560s - 1680s] Or AI to global like different markets. I know. Yeah. Now this year we have Singapore, but I think a few years ago, yeah, Paris was the first. Paris is the first one. So how did how was their journey end? Mostly that is basically exactly following the JSConf model. JSConf is the number one JS conference. Yeah. And it’s a very simple model, right? Like I will never reach every country in the world, but every non-country in the world will probably want to do some AI. So obviously their country deserves to have a leading conference. They can either make their own or they can borrow our brand and our YouTube channel and our social media. So basically, we just set up a partnership program. Paris was the first to apply. Okay. Very fun fact: the organizers of AIE Paris got acquired into Mistral. Oh. So now now it’s run by you know France’s and Europe’s leading AI lab is runs AIE Paris, which is very cool. Like I want that for every single you know country to like yeah you you have AIE India you have AIE Japan right yeah so I’m working with all those those companies now and it’s very cool like I will never know that region as much as them but like they do they deserve a good conference yeah absolutely yeah and then they can do business and and hire people and all that all that stuff right yeah the world is very big so we we follow that model so basically it’s the idea that you have there’s there’s three things right like one you agree to not abuse the brand name and like you know we have some kind of like revenue share thing two you’re publishing all everything that you’re you’re you’re doing and then three you have a code of conduct where like if you know if someone is being abused or mistreated there’s some way of of resolving that I think that’s mostly an an ethical thing rather than anything else so we take a lot of risk by doing this because we’re sharing our brand with people that we don’t work with and so what if they throw a bad AIE yeah so far it hasn’t happened and so far yeah because I think like mostly we’re we’re careful about who we want to work

[1680s - 1800s] With, but then also the the kind of people that are drawn to us anyway are pretty enthusiastic and organically interested in like yeah especially for Singapore AI Singapore yeah first of all I’m Singaporean but that actually did not matter at all the current top conference in Singapore that is an AI conference is actually a crypto conference right turn out to be a lot of crypto people want to be want to be yeah yeah yes it is super AI and I think they know that I don’t like them and so like they’re like this is embarrassing for my country you’re telling me this is the number one thing in my country we need something better and so yeah now I think AI Singapore is is up there it’s at least competitive and maybe maybe better than super AI and I’m very proud of the team for doing that and I didn’t organize any of it so I think that is the important thing for every country in the world to have it and we can support that funny thing is that because I saw you went to Singapore for the first first AI event and I because I knew Lay Lay very well and I asked him that hey AI came to Singapore for the first time when when do you go to go to Shanghai or China and then he told me that he is working with you yeah so China is the tricky one right because I’m not a I’m Singaporean citizen I’m a Chinese and obviously there’s a lot of restrictions between the US and China so that is the main tricky part that which is why I needed Lay and his team to really own the whole thing yeah because I like I cannot even set up a local entity in China right there’s a lot of controls I don’t want to get in trouble because my main business is US right so I cannot get into trouble with the White House but if you want to share the brand and run it as your own thing and it’s a franchise whatever okay like there’s lots of franchises yeah as far as we’re concerned I’m McDonald’s and you’re McDonald’s China yeah yeah like okay yeah happy we can bring there’s KFC KFC China yes great yeah that that’s great Sherry and I guess next topic is something the.

[1800s - 1920s] Context you just shared like few few few weeks ago about the Q my SaaS okay I really like that show because um yeah just tell us something about behind the Q my SaaS and okay I’m wondering how many of your listeners will care about this okay you will only care if you own a business uh huh because most developers do not own businesses they’re like not that interested but if you aspire to own a business someday yeah you will be better care and also you know um because you know SaaS was a very hot topic like several years ago just oh yeah SaaS is dying and dying and you know being discussed a lot this year yeah yeah yeah and so uh I found out that this so this show is very interesting and okay let me introduce the the the overall context yeah my budget for running a conference is somewhere between three to ten million depending on the size of the conference uh roughly about three hundred k of that per conference is software I’m paying like that’s like ten percent of my budget that’s like ten percent of my margin that I could be yeah getting back that is to just somebody else’s software that is shitty slow uh ugly not exactly what we want and you know so I’m paying three hundred thousand maybe I’m only paying one million a year in just like SaaS like like you want this you want that to do your work okay okay okay okay like as the boss like my employee comes to me and wants to use this thing they say it’s important for their work yeah I’m going to say yes but like when it all adds up like it’s a lot of money yeah and then like and then I use it and they use it they’re like yeah it doesn’t give me what we want like it has it’s slow like they request like as a customer of a SaaS what happens is you are in a queue for support right you want this feature and they’re like okay slow hour six months from now you know we’ll do we’ll put you on the Q three roadmap you are very important to us but like it’s you know we are an AI conference company and this is 2026 and there’s coding agents right like we can do this so uh the idea was I took a forty thousand dollars SaaS that my employee was

[1920s - 2040s] Proposing to pay, and I said I will pay somebody ten thousand dollars to clone this whole thing. It’s open source, and it’s exactly what we want. And from here, we can use whatever coding agent, Codex, Claude Code, Devin, to just customize it, whatever, right? And we did it. We, you know, announced the competition. We had sixty-nine submissions. When you have a competition like that, you offer free money. I’ve offered to cover everybody’s coding agent subscription, right? So you have a lot of bad submissions, and so I just like this is obviously not like I didn’t I didn’t give money for that. But we have sixty-nine good ones, and most of them were like usable. About five, I would say that we would pay for. Like it was just like good. Like it was done in one week. Yeah, yeah. As a owner of a SaaS business, you have maybe like fifty customers, all of them paying you forty k. Imagine if all of them realize this and they just just they just they just turn away in one year. Your Your business is gone. You went from earning twenty million dollars a year to nothing. Nothing. Yeah. Yeah, it’s it’s very scary. I think for for a business owner, and I think that’s the tricky thing of of owning a SaaS, right? Like like I own a physical business, so that’s not going away anytime soon. Robots are cool, but they’re not they’re not organizing conference. But like people who do software B two B, who own this like mid tier SaaS, Salesforce is fine, right? Nobody’s worried about Salesforce, but like the mid tier SaaS, where like you know a lot of small S S M B. Do you guys call S M B S M B S M B s They’re going to really struggle, or they are going to pivot to something that’s better. And I think they they really better figure it out now because all of us business owners who are their customers, we also can code, yeah, and it doesn’t take that much long time. We are the actual competitors to to their own software right now. And by the way, the models are very good at clicking through every screen of your software and then replicating it in independently.

[2040s - 2160s] Yeah. Right. So the moment I have access to your software, I can clone it. So good luck. And so I think that is the the reality. I would say that if you’re a SaaS owner, nobody knows that the right answer. I think that you basically end up having to pivot to being an agent builder. So you have to become like the US equivalent is like Harvey or Open Evidence or yeah, yeah. You own a vertical and you’re the top agent in that vertical, or you own some specialized data and you’re basically a data vendor. Anything in between, someone else is going to just code their own thing. Yeah. Probably can follow a question because you already answer my first question. Yes. You already give some advice to those those. I’m more interested. Like so, I’m not I’m not so sympathetic to the those the SaaS owners. I SaaS owners they they’re big big boys. They can figure it out. Actually, I’m more I’m more keen on advising the business owners. A business owner. Yeah. I I run a small business. Yeah. I’m SMB. Yeah. There’s a million people like me. They’re all trying to just I I you know I make an honest living. I pay employees. I pay my taxes, and it’s a good life. It can be better if you vibe code your own. Yes. Yes. Yes. Right. And like my customers are happy. My employees are happy. I’m happy. It’s a cool story to tell. Yeah. You know, SMB is I think is a golden age for us. If you do anything in the real world, that is, you just try and do your job, and there’s a lot of shitty software between you and your your your intended outcome. You can suddenly shorten that timeline by your cloud and yeah yeah yeah yeah yeah. I mean, like you know, forty thousand dollars worth of cloud is a lot of cloud. Yeah, it’s a lot of cloud. I can spend ten thousand and like have a lot of money to like do whatever I want. Yes. So I think that’s why I I call this is a business owner problem, not a developer problem. Developer problems say like, oh, two hundred dollars is a lot. You know, like like I’m a business owner. I understand the productivity benefits of having ten of my employees work together well. Like ten of my

[2160s - 2280s] I’m paying 1.5 million for them, right? Like if if they work together 10% better, there’s 150k. The potentially it’s not it’s not 10% better. It’s like five times better, right? That’s like 7.5 million. Yeah, great. And that is all the agent work that is being done on on my behalf. That people used to say like, well, okay, I’ll you know I’ll come back next week with a proposal. No, I want it this afternoon. Yeah. My thought is that once you created so many software yourself, would that the maintainance, the work would be something you know difficult? Yeah, it’s it’s real. Right. It’s not zero. You have to use it like it’s not zero, but we have coding agents. Oh yeah, maintaining like self self maintaining. Yeah, and a lot of it is. This is why I say like mid tier SaaS. You are not Salesforce. You are not Google. You are not AWS. You’re not this like mission critical 99.999 uptime thing, right? You are just like a database where people type in their thing and they get something out. Yes. So, yeah, okay, it goes down for a day, like whatever, and you can fix it in a day. Right. Right. And there’s a lot of coding agents where like I think I would argue I would take that downtime for in exchange for like I can customize. Yes. I can customize anything. It’s amazing, you know. And and so my team is so much more productive. You know, I actually put this in the episode and kill my SaaS. One of the favorite features, let’s say I have so for example for AIE code my my next conference next month, I have 1,000 submissions and I have 25 seats. Okay, I have to go through 1,000 people to get 25. And in the current SaaS that we use for that piece of software, we have to click one by one by one by one by one every screen. Click click click click click click. And so in the kill my SaaS competition, one of our contestants just changed everything so that you can sort of reorder in one table. This view doesn’t exist in that SaaS. Okay. To in order to get this view, we have to request for six months. Well, but and they just clone it. But you just you just request whatever.

[2280s - 2400s] Where you want, yeah, and like, yeah. Now I can do it, and I can see all everything in one go. Yeah, okay. That’s so quick. So I, I am for listener. You know, if you are thinking that this is a stupid example, it is. I’m simplifying for audio. I mean, I found out that the top one winner is someone who can manage all the split pairs lineups, and that that’s a yeah, and it’s good. It’s good UX. They are testing it from the point of point of view of humans, and I think that there is a lot of difference between the average submission and the top submission. Yeah. So that is true that today in 2026, if you’re a business owner and you decide to hire the youngest kid with like you know that looks vaguely technical to hey, if I quote this SaaS for me, please, maybe they’re not going to do so well because they don’t have any training. But just give it one year, watch a lot of AI videos, and then you’ll pick on it. There is a skill level, right? Like a a lot of it is slop, and you don’t want to use slop, right? Your your employees deserve better. You deserve better. And actually, sometimes buying from tried and tested SaaS is going to be better. Yeah, it is going to be better. Yeah, of course. But for us, we want to be AI forward as a company. We will be more cutting edge. Yeah, and one last thing to follow up because I, you just said it in the podcast that you have AI judge, and actually judges, the judge of the AI was quite different from the the winners that you picked by your team. Yes, wow, you really listen to the podcast. Oh yeah, because I’m interested about the whole process. So yeah, I just wonder why. Oh, so that was partially on purpose and partially not the goal of the judge. The judge was purely: did you submit a valid submission? Right? Did you qualify? Uh huh. But the judge was not aimed for: is this the best among the whole core? Oh, okay. Right. Then obviously there’s a difference. Okay. So. So it’s for your your purpose or your team’s like conference purpose to pick which is more useful rather than just because we didn’t know what we wanted. So.

[2400s - 2520s] So I cannot make an AI judge for for things that I don’t know what I want. Okay. But once I see seventy different examples, and then you will know. Obviously, I know the one. Okay. Yeah, yeah. That is all. And by the way, fun fact: the latent space logo. If you you see the logo for this podcast in this room outside here, okay, that was through a a website called 99designs, which is very interesting. Are you familiar with this? No. Okay. When you want a new logo, when you want any design, whatever you pick from, like you have maybe you ask for your friend for like, hey, who’s who’s good at design, and they give you three names, and you pick one of the three names, and then they give you a logo, and that’s it. Yeah. 99designs. You put up a contest, and I got four hundred different ideas for a logo, and then I award the prize to the the logo that we got, which is a it’s a very good logo. Yeah, yeah. There’s so much better, right? You want diversity and ideas, so actually this idea of like do contests and all hackers, yeah, yeah, which I can do because I’m an influencer. Yeah. Most people cannot do because they don’t have the audience. Okay. But I think that there is such a thing. But 99designs is a very good business because there’s many designers. It actually doesn’t take that much to create a design, right? You just need. It’s more about the quality of the idea than the amount of work. So you just need high quality ideas, and it and like you just need to be able to spray over to a lot of different contests. And so 99designs is like that, and I think there could be a lot of 99designs for SaaS, which is something I’m thinking about. You know what I mean? Like basically, I had to create a 99designs for myself for saying here’s the forty thousand dollars SaaS, here’s the specs, here’s the judging criteria, here’s the blah blah blah, even right down to like here’s how we pay you money when you win. Everything had to be worked out by me. But would I pay some other person to do all this for me so that I just tell them what I want and I get seventy different versions and I pick one? Oh yeah, it’s a lot of admin for me, right? Yeah, yeah. So there should be a 99designs for uh for SaaS and for cloning all these things because as a

[2520s - 2640s] Again, as a business owner, I honestly I’m only doing this because it’s good content. It’s a podcast. You you’re interested and all these things. But like if I didn’t have a podcast, I probably wouldn’t do it. It’s a lot of work. We took like two months to eval everything, partially because I’m busy, but also it’s just it’s a lot of eval. Like just there’s five roles for this particular SaaS because it’s about like organizer, speaker, attendee, sponsor, vendor. They all have to agree on what the conference organization looks like. Yeah. So you have to create an account for each of the in order to test. You have to create an account for each of things, and you have to do some data entry here, and you change accounts, and do some, and you read and data entry here, and you change, change, change, change, change. And so that’s why it’s like complicated. That’s why that’s a but that’s what the mid tier SaaS is. That’s why I pay forty thousand dollars for this because I don’t want to do it. But now that you can, it’s very fun job and problem to do. Yeah, sure. Yeah, I mean it’s a good experiment, and because we’re also running conference in in China as well, like also over thousand attendees before, so I can echo you so much. Yeah. About. But then like think about the rest, right? So I used to pay Calendly. Do you guys have Calendly? I don’t, but but you know what I mean. Like like a booking link. Like that’s how we that’s how we got to this meeting. Yeah, Calendly is like a thirteen thirteen billion dollar business. Yeah. And yes, you can buy Calendly. Yes. I know, of course. I have I have my own. I have I have two versions. You you already own your like. May your own. Yes. Yes. Yes. Yeah. I I have my own Calendly. I have my own Google Drive. I have my own GitHub. I’m serious. I have my own Dropbox. My God. I have my own Slack. All of this is because I have money and I can I have people and I can I can assign them to do things. But I do think that you can build software for yourself for audience of one or audience of one company, and that’s okay. And like like the fact that you can do that now is something that has never existed before for any business owner. Yeah. Which is amazing.

[2640s - 2760s] Yeah, yeah, yeah. Yes, great, great, great. Yeah, yeah. Okay, you do some investing, right? And you’re you’ve you’ve raised money. Uh-huh. So you know DocuSign. Yes. Ten billion dollar company. Yeah. Or maybe seven now. Yeah. Uh, what is DocuSign? You you send a PDF. They say like, oh, yeah, sign. I give you a certificate. Ooh, that’s it. Yeah. Right. And they they send they charge you something like ten dollars per per PDF, or first signature or something for envelope. Per per piece. They call it you know, and like yeah, I I there’s an open source DocuSign and I hosted it for myself. And you know what’s the difference? Like because I have my own drive, and I have my own DocuSign. The moment that the DocuSign is signed, it goes straight into the correct folder on the drive. Oh yeah, because you integrate together. Yes. And that does not exist in DocuSign. Right. Right. Yes. Right. So like my data room has directly the DocuSign or everything that I didn’t have to touch because it already has. Right. Everything. My one thing special about DocuSign is created like because it was built long time ago and it creates a network effect that everyone joins, then the world everyone using it. Are you still thinking it’s a mode for this kind of software? Fuck no, no, no. Electronic signatures are legally enforceable everywhere. Yeah. And is DocuSign has no patent on it. I can do I can do my electronic signature. What the the network is the email address. That’s it. Yeah. It is so stupid. And like I can keep going about why I made each of each of those things. It doesn’t matter. It’s not like I’m encouraging everyone to do this. I think that whatever problem you face in your business or your life, you really can find open source version of it. Yes. Host it for yourself. Customize it. Integrate. Okay, I want this plus this. Boom, they just work because I fucking use them all the time. I copy here, paste there, copy here, paste there. Inside of every one million copy paste, I think there is a fantastic startup and product to be built. Great. And I guess next question or next show, I’m actually quite interested about your conversation with Diego from Jeff because.

[2760s - 2880s] Yeah, it was actually a really also a very hotly discussed topic and product. Yeah, what what does the Chinese developers thinking about it? I think it just you know, um, just you know, creates a new way of thinking about using like LMs from a totally different angle. Yes. And then I even heard that you know one podcast podcaster from China, she didn’t know anything about coding, but she she used Jeff to judge you know the clips from their podcast, whether it would go viral, and then Jeff gives you know gives the answer, and and then he will post it on yeah on the network. So I found so many great use cases based on Jeff. So I just wonder what’s your take with with there and what probably. Yeah, so for those who didn’t listen to the podcast, we were the first podcast. You were the first one because I I forced him because I’m I’m friends with him for four years and say if you don’t if you don’t give me the first one I’m going to kill you. No, he’s he’s he’s obviously like very very accomplished. He was the co-author on on the instruct GPT, which became ChatGPT. And I think the interesting thing about Jeff is that it is kind of the original vision of AI engineering for me, which is that instead of right now AI engineering is is what we talked about is a big harness around a model, and it’s like I’ll do a bunch of work and then I’ll hand off to a model and say please God model, give me the right answer and and you can take ten days to think about it. Jeff is like, well, no models are for us. We don’t work for the model. Right. And so I will reduce the scope of the model to only one thing, which is to make choices or to estimate probabilities, and I will I will put it in the line of path of my critical work. So if you think about what is the failure mode of of what the LM current LM paradigm is, is that you have the big traditional SaaS, and then your LM is this window that’s a chatbot, right?

[2880s - 3000s] That is basically what everything is, or you are you look like Codex or ChatGPT or Claude agent, and it’s like chat window here, yeah, menu here. That’s it. For what Jeff represents and what System One models represent, this is why he doesn’t call it decision model; he calls it System One models. It is that they are in the path of code. You write it as casually as you write an if statement inside of a normal programming language, and I think that is really what he means by smarter software, which I really, really endorse. Which is like your software should just be at every point smarter instead of at one point smart and then everything else is dumb. That’s really dumb. Yeah, I think the the speed stuff is almost secondary to the reliability, which is what he really is trying to say. That’s why the company is called TypeSafe. But yeah, the speed is nice. I just found out that Diego himself also a very you know social character and and he’s really passionate about what he’s been doing. So I just wonder, you know, you you’ve talked to a lot of like researchers and developers or you know AI engineers. What do you think? Why was Diego that made the Jeff? Was this something special or different from other kinds of researchers or people that you’ve talked with? Because I find it quite special about about him talking to the what he’s building and what he’s believing. He has a non-traditional background himself. He didn’t go to like Stanford PhD and you know the kind of things. So he has that sympathy with a regular developer, but he also worked on ChatGPT, and I think that a lot of people who maybe they went through the traditional ML researcher path—you go to a top university and then you work on some interesting paper and then you get hired into a lab and you work on a lab—they’re in the big models cult. Everything will be solved by making our model bigger by having more more flops in our training run and more parameters and more more more more more right right.

[3000s - 3120s] And so it takes someone who doesn’t come from that mold to really say, “Well, actually, there’s another way to do this, which is smaller stuff that is sprinkled everywhere instead of one big one.” And I think that is, by the way, Diogo was not the first to do that. There’s been thousands of people who’ve tried to say smaller models good, faster, cheaper, better, whatever. But somehow Diogo worked. One because he has the credibility of OpenAI. Two, he made it into a system one versus system two thing, which I think makes the communication very clear to the point where you can play Doom or whatever was. So there’s some game demo that he that he showed in the launch post that was very clear. Like there’s no way you will ever use LM to do this. And I think that showmanship that he wears the pink jacket, he is so passionate. He’s talking that showmanship plus the technical capability is again like similar to the learning public. Like I’m a small version of that, right? Like I’m technically capable and I can talk. But for him, he’s like 10x on both dimensions. And I think that really comes through. And when you talk to him, and also I guess my next question is about your your roles on advising other startups. That just you mentioned, I’m just really curious. You know, for for us as a advising role, like exactly what kind of advices that you are working with those yeah startups. I can I can do. I’m also curious. What can you for your listeners and for me? Yeah, what do you do for other companies? Oh yeah, so we are actually a community that started three years ago that we found out a lot of Chinese based founders want to go global. Yeah, and then I think they were pretty good at building or shipping products quickly, but they found a huge gap between you know go to market marketing and understanding the market. Okay, if a Chinese founder trying to break in the US, you suck at this. You are.

[3120s - 3240s] Very good at your tech, yeah, and you suck at marketing, yeah, and you better talk to someone who’s good, uh, because you cannot do it yourself, yeah. That’s it, yeah, yeah, yeah, right, yeah. So that’s that’s how actually we found out that there’s a huge gap between you know understanding the local market and you know um and then ship the product correctly from your PMF, yeah, yeah. So that’s basically we hold out a lot of meetups and then working with some advisors here and then try to provide different kinds workshops and adversaries to those founders really start to find a little bit they’re you know on the right track of BMF but still lacks some talents and uh you know help from from the go to market side. Yeah, it’s interesting. I we also host meetups. So so Diogo actually before he launched, Jev, he was actually here in this in this room. Oh, we hold this like sort of writers meet up thing where people can show their work and I will give commentary on it on it, um, and so the the sort of Type Safe Manifesto came out of that. But I think it’s very similar to to that, except that I would for some pe some of them I’ll do it for equity, most of it I do it for free, and then some of them they want to have a like a recurring call, um, or they or something they then then they offer some equity. It’s just easier to do that, and it turns out it’s worth a lot of money if it if the company does very well, right? So I think that is interesting in terms of like a job because it’s you cannot eat equity, and like maybe they never sell, like DataBricks is like whatever 15 years old and worth 200 billion dollars, and is like I have my equity is like what is it? But I think it’s it’s very useful to be able to test your theories on a number of different companies. So when you work as an employee in one company, you only see one example, and your advice only works in one company. But now that I have done 30 something, 35, 36 different companies.

[3240s - 3360s] I can really see the universal truths among all of them, and I’m better as an advisor to all of them because I’ve tried everything. Right. Yeah. Talk to everyone. Like, and I I understand the different scenarios of like, okay, my advice is not going to work for you here because I tried it at this other company, and and and you know your situation is like clearly very different, and like I I feel like actually almost all domain subject matter experts like I happen to own developer marketing right that’s my thing, but if you work in I don’t know your strategic finance or you’re working in like GPU compute acquisition or whatever, and you and you are trying to be a global number one in that, you kind of have to go through a tour of duty of just advising everyone, so you can just take fucking test everything. Yeah. Some of it doesn’t work, but so but sometimes I mean some of them is clearly going to work, and then you will learn, and that’s the only like your sample size has to be more than one, and most people’s sample size is like five, but like you know if you can get it up to like fifty, then like you probably are one of the top in the field anyway. And I can share with you some you know some something that I find it quite similar in in those you know go global Chinese founders because right now more and more came from engineering background even research background yeah and yeah they they think that they’re building the best products because they have yes like best hardner skill or you know technical skills and knowledge etc. But actually there’s still sometimes a huge gap between you know what the customer actually wants that the skills and the products you offer and sometimes like different local markets have even have different you know demands and have you seen a similar thing yeah yeah all those technical founders research founders because yes there’s also a lot of founders from the same background. There’s a bunch of ways to to survey this or to tackle this. Uh, the the most important.

[3360s - 3480s] important thing to recognize is that if you are you you have something smart and like you think it’s valuable and global number one yeah really like yeah your world best at one thing like that’s the most important thing so everything else is I always call it hua long dian jing (画龙点睛) right like you have the long okay I will help you with the ting but like you have the long that’s you know and most people don’t have the long right so like that’s that’s good um I think that there are some percentage of people who don’t even have the long and they think they have the long and like then it’s like okay it’s like an arrogance ego thing right but let’s assume that you are genuinely have a very good product and I think the the important thing actually here is that you should not treat marketing as the last mile of like hey I did everything it is the last night before the blog post can you review my blog post and change it there’s nothing I can do like you already did everything but if you involve marketing earlier in the process right down down to like before it’s really built or specked out like what what is the way we’re going to pitch what part the segment of the market we’re going to go then you have more GTM built into the product in the first place and I think that a lot of engineer or researcher defined audiences don’t have that GTM brain that early they usually put it all the way at the end yeah they’re the most junior part of the company is this GTM organization it’s like some interns like some 22 year old person who is like running social media no like actually try to be serious about your the who what customer is supposed to use this thing and how where do they hang out what problems do they face what words do they use to describe their problems and how is their life going to be better after they see your your thing right and so very simple things most and researcher engineer founders have not thought about this but if they have a GTM partner I think that that is often very useful I think the other thing that I and then it comes to more to develop developer tools I find a lot of Chinese founders are robotics right.

[3480s - 3600s] And robotics is also a very interesting space because it is much longer time horizon and still very early days, right? Like, how come all these Chinese robotics founders—they all do laundry, right? Like, Unitree is the best, and at least like, but they’re still Unitree is still doing Unitree things, and I would like them to see things that you never expected Unitree to ever do. But right now, you know, the the the thing they they just launched this week—they’re doing kung fu moves, yeah. But like, okay, I’ve seen that before. You know, like, show me something that I haven’t seen, and then then I’m really going to pay attention. You know, I think that’s that’s something that is hard to show in robotics, but I think with models you can do that still because it’s still so early days, and with innovative like products ideas for like for consumer experiences you can do that. So I think to show the demo, I think is very important, which is why again, obviously, TypeSafe is like very recent, but also very good. It is not the the OGO did not come out with the Doom demo, but the Doom demo is what everyone talks about. Doom demo is done by Devrou like GTM, and that that is that is something that just is a core principle in marketing, which is show, don’t tell, right? Like, show me this thing having the benefits, but don’t tell me in bullet points what this is. Don’t give me a PDF, a slide presentation, PowerPoint that tells me all these things. Just show me. And maybe just a follow up is about the taste because you know everyone is talking about the taste, and then I think go to market and building the product also needs the taste. Yeah, yeah, yeah. That’s probably very true. Actually, I just have becoming a little bit hard for me to define what a taste is because investors talk about. So I just wonder in your view, like, how do you define the taste? Seeing from those engineers or even founders. Okay, I can define taste in a way that is not helpful. Taste is something where you are able to out of

[3600s - 3720s] A large set of possibilities, you’re able to narrow it down quickly, cheaply, to the set of possibilities that, when you look back, is immediately obvious is good. So when you’re at the start, you look out. There’s a thousand different things I can possibly do. Which one do I choose? Right. Yeah. The brute force. You can brute force. You can try one thousand of them, and then you find a good one. But a tasteful person can say, “Yeah, obviously these are bad. I don’t. I cannot explain it. I cannot explain. But obviously these are bad. This one looks. This this tool looks interesting. And then at the end of the journey, they’ll pick one, and you look back and you’re like, “Well, that is. Yeah. What do you mean we’re considering the other 999? Obviously this was the correct one. There’s many versions, applications of this, but the one that’s coming to mind right now is the iPhone, where for a long time the the smartphone had a keyboard, and it took Steve Jobs to say, “No, we will do whatever it takes to have touch screen and virtualized screen because it look needs to look like a screen that’s a flat slab, and that took ten years to R and D because there was a company called General Magic that tried to do this and failed. And but like obviously now every phone in the world looks like this. That is that is taste. In 1990, when people were starting to have smartphones, you could do a number, a huge number of things. But it takes the the person with taste to say like, “Yeah, this is what we want, " and I’m going to say no to everything else. And my first phone I ship was going to be this. If looking back at all those AI products, could you share an example that you think you find it has taste? I actually think that everything that Cursor does has a lot of taste because they were first to the VS Code thing before, because and now they’re they’re doing Grokbot, which is like a good taste thing. Because like now now like OpenAI has to like copy them and what have you, and Claude has to copy them, which like you know you have taste when other people copy the thing that you just shipped. Yeah.

[3720s - 3840s] The interesting way, you know, to go below the labs, I think is interesting to think about like the more boring kind of problems that people solve. Legora, you know, in the legal field, they they have this sort of like document review process that’s like massively parallel. Yeah. And like once you see that, you’re like, I don’t want to do my work anywhere else. I I have to do it in Legora. And like again, like anyone could have come up with it. Anyone, these ideas are cheap. Yeah. But you didn’t do it. Yeah. Legora did it. And so someone at Legora has that taste of like, we will define this billion possible things we can do. We will just define this product that that becomes the way that people do their work. Yeah, I think coming to an end, and I think probably we’ve talked about AI is arriving arrives in Shanghai like two two weeks later. Uh, I just wonder, you know, what do you hope to to see or to learn from the builders? So they develop their what kind of topics actually are you most interested at? It’s tough because it’s more of a differential for me. I want to see how very different Shanghai is from San Francisco, because I think San Francisco has a B two B SaaS bubble, but Shanghai, you you guys use Comfy UI? We don’t use it here. Okay. We like our Canva. We like our Fair AI. Yeah. Don’t use Don’t use Comfy. Like Comfy is more for like a open source tool, and it’s same same for OpenClaw. OpenClaw had more traction here than Comfy, but OpenClaw was really traction in China. Yeah. I bet you heard that. Yeah, yeah. Of course, I know, but like I haven’t seen that kind of energy myself. Yeah. And I think it’s very different. Like the you know for you a medium sized community is one million people. That’s huge in here. Yeah. A million is already the topest the top in China as well. No, look, I’ve seen those kind of like you know Chinese.

[3840s - 3960s] Like influencer farms, like it’s like all these Chinese girls like doing this on the phone. It’s like the scale in China is just so different and is so competitive, right? Like people say they do nine nine six in in San Francisco, they don’t really. Like Saturday Sunday they’re still kind of chilling, but in China they are really nine nine six. Well, it’s kind of normal in startups and also the big tech. Yeah, but I think that I think that’s really intense. People don’t like Toutiao and like Meituan and right. Yeah. Intense. Yes. Right and like people here don’t work that hard. I heard that engineers or startup founders work that hard. I work that hard because I have no equity, but like employees. Employees. Legally, you cannot ask people to work that hard here. It’s illegal. Okay. Yeah, I’ve seen some of the most competitive top tier US startups, and I’m sure they’re not working as hard as China. There’s BAT. There’s like the first generation, then there’s like second. Second generation, whatever the thing is. Yeah, I mean, like those guys. Oh my god! Like. How did you learn? Did you? They’re such good founders. Yes. They’re such good. Like if they, they live in the wrong country. If they live in the US, they’ll kill. They’ll kill Amazon. They’ll kill like Facebook, right? The only reason they’re not richer is they live in China. Also the SaaS, I mean the software as well. Yeah, I mean even even like okay, so when I did my study abroad in Beijing in 2007, BYD was still like oh we’re we’re up and coming like you know we’re we’re Chinese based we’ll always be Chinese based. People love BYD here. Oh really? BYD is taking over the US. It’s crazy, and like, I would never have thought, never, that BYD would ever come here. It’s insane, and I don’t know what I don’t know what conclusion I say this. I think that the Chinese energy and spirit and creativity and grit is something that people don’t have exposure to, and then don’t. Let’s not even talking.

[3960s - 4080s] About the model labs, which obviously like DeepSeek and all those all those labs. So I think there’s many reasons to be excited about exposing that more, especially you know I think Chinese events they’re very there’s regulations against live streaming, but even they’re presented differently than the way that Western audiences like to consume them, so they don’t have the same reach. Right. Right. Ali did like Aliyun did like this thing where they presented the whole Alibaba roadmap with Qwen and everything. No. Nobody saw it here. Why? Because it was in the middle of some corporate presentation presented by other people. So I had to go and translate everything for for the US audience. But like just do it with me, right? Just do it on the AIE. They will see it. Yeah. Yeah. Yeah. Yeah. And also I saw that already there’s some open source like model companies attended AIE this year in San Francisco MiniMax. Yeah. Yeah. So I mean we we hope to have representation from everybody, and I I think that there’s a lot of people interested in in that cross cultural exchange that are not interested in the politics. Right. Like we will not do politics. We’ll only do tech. Yeah. Okay. That’s great and really really lookful for, and so I guess this is my question, but I still have like three quick final questions. Anything just first pop up in your mind? But first is you know what AI development or are you most excited about right now, and actually what actually worries you most? AI development that most excites me, I think System One models is like a very clear number one answer. So it’s boring, so I’m not going to say it. So I think the thing that people are under the radar about is interaction models. Interaction model. Uh. So this is something that Thinking Machine is doing for people who have who missed the announcement. Look it up. It is where you are full duplex back and forth communication at all times. You don’t start and then you take a turn. I take a turn. You take a turn. No, at at like at all times, like you are listening while you’re speaking. Yeah. And

[4080s - 4200s] I’m listening to you while I’m speaking, and I can interrupt you, and you will immediately respond and adjust. Yeah, I saw that demo. Bye. That should be the way of AI, right? There’s a corollary of that, which is continual learning AI, where you don’t have pre train post train done. You are learn, learn, learn, learn, learn, learn, learn as I work with you, and that’s just very that’s a very interesting and important field. And then the last part is world models, which obviously doing very well. I don’t have to explain that. But yeah, what I’m worried about, I think personally, if we have an AI winter, it isn’t AI summer. This is probably the longest AI summer we’ve ever had in human history. But the history of AI has been is about 50 years old. There’s always summer. Everything’s going to work. Yeah. Winter, summer, winter. If we ever have a winter, when is it going to happen? And am I going to go bankrupt? Which yeah, it’s a it’s a real concern for a conference company that relies on hype. You know, more more broadly, I think that there is a real concern about safety that all my friends at all the labs are really, really very worried about, and that society cannot handle and adjust quickly enough. And it is not because of fake news. It is not because of losing your jobs. It is something worse. You know, right now the the way that we express it is cybersecurity. It’s like oh, I get hacked all the time every every day. I so hack, hack, hack, hack, hack. I feel like there’s something even worse on the horizon that we don’t know about. The most extreme version of this is bio weapons, right? Like, oops, like I have invented COVID too. Yeah, what if it’s worse than that? So so the way that we phrase it in on the podcasts that we do is that the default assumption for most people should be AI doom, that in all of human history, when a more advanced civilization comes along and meets a less advanced civilization, the less advanced civilization goes extinct. And this thing is evolving roughly 1,000 times faster than us, and so at some point it’s going to be more advanced.

[4200s - 4320s] And then we have to figure out what to do there. No, cool. Yeah, and and the second question will be well, just a bit light one is that what kind of other newsletters and podcasts or the contents they recommend? I have a list. It’s called the Good AI newsletters and podcast list. Okay, is it on the? It’s on the column. Okay, it’s on my GitHub. But actually, for the more recent one, you can check aeo dot latent dot space. So basically, what I did was I asked every model what are the best AI podcasts for engineers, and it gives me a list, and I transcribe it and put it on a ranking. Okay, I asked two hundred of these questions. What is the best payment startup? What is the best bank? What is the best financial advisor? What is the best database? What is everything? Every every every question that. That’s why you call it aeo. Yes, because agent optimizing for what the models answer, but it’s probably a pretty decent proxy for what the people who choose the data sets in the in the labs choose, and they choose latent space. Okay, the the top, the first one. It does very well. It’s either number one or number between number one number three, and and yeah, so that’s that’s the answer. Cool, cool, cool. I guess I’ll just leave the last question for you to ask me if you have any questions. And what are your hopes for AI Shanghai? Yeah, of course. I I hope to listen to more on topics about like you said some more infrastructure stuffs. For example, like like the safety, the payments, or the things that come into unlock more applications in the world that coming from different like infrastructure, Asian infrastructure. I I’m just really curious about those topic because uh it’s kind of a bit less discussed about in China. But I found out that especially spending few few weeks here, I find more people talking about and also probably RSI or those some frontline researchers. So and I hope that it could be a great conversations between the East and the and the Western like the researchers here and then the the the people in China as well. And I hope it can be a successful one, and then we can come out for continue to host it on next year.

[4320s - 4344s] I hope so too. You know, I wish I wish the the team a lot of luck. Doing the first one of anything, the first one is the hardest, right? Agree. Yeah, I’m I’m so impressed that they’re doing it. Thank you so much, Sean. Thank you. Thank you.

中文翻译（全量）

[0s - 120s] 这是一期很特别的播客，因为它关于 AIE 上海站。AIE 上海是中国第一场 AIE，由我们的本地合作方 Startup Partners（转写亦作 Stand Up Partners）带来。全世界有几场 AIE 不是我的团队直接运营的，这是上海的第一场。我一直想把 AIE 带到中国。我觉得这里有很多「东方遇见西方」的时刻，中国有很多进展需要被分享、传播到全球社区。中国对我们超级感兴趣，Latent Space 一直有很多中国粉丝，我会在 YouTube 上闪现几张他们的画面。首先我想感谢 Meng Xiao（转写存疑），我们的头号粉丝，基本上每一期都翻译成中文，我超级感激。总的来说，我也想祝贺和庆祝中国取得的进步——不只是开源模型，还有机器人和很多其他创业公司。谁还记得 Manus，去年这个时候创立、然后搬去了新加坡？显然 OpenClaw 也大火了一轮。我们会聊很多「当地的真实氛围是什么样的」、中国的 AI 社区规模有多大程度的不同、什么东西会被采纳，以及创业文化。大家听到很多关于 996 的说法，我们会聊它相对于 Cognition 这种地方的现实。最后我们也会聊一点 Latent Space 的起源故事，如果你是新听众可能没听过：它是什么时候、为什么开始，我们怎么运营手上的多档节目，还有一点关于安全与对齐的内容。这里多少有点每个人都能带走的东西。基本上这是给粉丝的一期——如果你一直follow Latent Space，以及如果你对 AIE 上海感兴趣：它不做直播。

[120s - 240s] 我们之后会把内容传到 YouTube 上。大会明天开始，所以在中国看到这期的任何人，我真的很期待你们的到来。也感谢 Mengying（转写存疑）和团队的 organizing。我还要喊话 Hanhua 和 Lei Du（转写存疑），是他们把这件事推动起来的——有好几次差点办不成，但 AIE 能落地上海、Latent Space 能来到中国，实在太令人兴奋了。感谢观看，期待接下来的访谈。（片头金句剪辑）「模型非常擅长点开你软件的每一个界面、然后独立复刻它。」「对，所以只要我能访问你的软件，我就能克隆它。祝你好运。」「有很多从没写过一天代码的空头评论家……我的 AGI 时间线还有两年、你们以后都没工作了、要为下一步做准备——巧了，『下一步』长得正好就像他们的工作。」「我有自己的日历、自己的 Google Drive、自己的 GitHub。我是认真的。」「他们是非常好的创始人，只是生活在错误的国家。如果他们在美国，会干掉 Amazon、干掉 Facebook。他们没有更富的唯一原因，是他们住在中国。」「AI 末日：人类全部历史上，更先进的文明遇到更落后的文明，落后的一方灭绝。而这东西的进化速度大约比我们快 1000 倍。」——你好 Sean，谢谢你参加这期节目。我是 Latent Space 播客的老听众，中国很多创始人、AI 工程师也从你在播客上的工作和对话里学到很多。也很感激你今年第一次把 AIE（AI Engineer 大会）带到上海。七个月前我还读过你 7 月 4 日那篇关于你与硅谷关系的文章，你从新加坡搬来搬去。

[240s - 360s] 我想这也是你真正提出/发明「AI Engineer」这个角色的过程。你上次搬去硅谷，而这次你把 AIE 第一次带到上海，所以感觉很特别。能讲讲你的背景和你与硅谷的关系吗？——我先纠正你一点，你说错了：不是我「把 AIE 带到上海」，是上海把 AIE 带到上海。是 Hanhua、Lei Du、Mengying 这些 Startup Partners 的人。我非常感激他们，因为我一直想在中国做点什么，但我不是中国公民，在那里没有任何人脉和关系。而我觉得美中两边对这个都有很大兴趣——对中国 AI 整体而言，任何我们能做的、把高质量对话浮出水面的事，AIE 在中国都能帮上忙。长话短说：我是新加坡人，第五代新加坡华人。人们问我「你的家族在哪里」，我会一点中文，我在北京留过学。这就是华人散居者的感觉——我们不知道自己的源头，因为没人留下任何记录，有时候连祖辈的生日都没人知道。第五代新加坡人，在新加坡长大，去美国读大学，第一段职业在金融：交易、投行、对冲基金，做了十年——抱歉，八年，然后转科技。

[360s - 480s] 我在 Amazon 和几家创业公司工作过，四年前成为独立工作者，写 Latent Space，意识到会出现「AI 工程师」这个东西，然后创办了 AIE。——我知道你现在有多重身份：做 Latent Space 播客、写 newsletter，还在好几家创业公司（比如 Cognition）担任职务？——对，很有意思，我最近刚为了报税梳理了一遍「我到底有哪些钱和身份」。最大的一块是 Cognition：我现在仍是 Cognition 的顾问，之前当了很长时间雇员。基本格局是我有三摊事：AIE、Latent Space、以及「杂项」——过去是我的个人创业载体 Small AI，后来被整体收编进 Cognition，我在那工作了六个月，然后退回顾问身份，因为 AIE 现在是全球化会议组织，我必须当 CEO，没法兼职。Cognition 很大，Fireworks 很大，模型公司很大，Supabase……如果你是硅谷影响者、或者在开发者关系圈待了很久，这是很常见的模式——我的专长正是「懂技术，同时会写会说」，这个组合很有价值，因为任何创业公司都愿意付钱让你讲他们的东西，然后你就开始同时做多件事。这不是一条正常职业路径，我其实不建议大多数人走，但这就是我的角色：以顾问身份换股权地帮几家公司。

[480s - 600s] 我想再深挖一下 Latent Space 早期——我读过你那篇文章，你早期「公开学习」，把学到的东西讲给所有人，然后把线上和不同研究者的对话转成 Latent Space 最早的对话内容。——这里其实混了两件事。第一件是「公开学习」（learning in public）这个原则。我离开金融进入科技时就意识到：为什么科技对我来说顺这么多、舒服这么多？两个领域都是高薪、体面的工作，核心区别是：金融是零和的——我知道什么、告诉你，这个真相就贬值；而科技里大多数事情是正和的——我知道什么、告诉你，价值反而增加。对从金融跨界到科技的人来说，这是范式级的解放：人们会上台、开源自己的工作、分享代码、搞砸了也公开分享自己搞砸了。金融里不行——你要假装从没搞砸过、一切绝密。这一点我记在了心里。另一件事：我第一份工作在 Two Sigma，还在金融但做软件工程，我遇到了一个坏老板，无聊透了，于是开始大量写作和演讲——工作里不让做想做的事，就把全部精力投入工作之外。那件事起飞了，因为没人做，对比之下我就突出了。

[600s - 720s] 「公开学习」最终变成你可以自己选老板。大多数人有工作时，加入公司、被分配老板和团队，很多时候你根本不知道团队好不好，发现不好时已经被困住了。而公开学习的人实际上可以选自己的老板、为自己工作。快进到 Latent Space 的诞生：那是 Stable Diffusion 发生的时刻，在 ChatGPT 之前。当时我在一家进展不佳的创业公司 Airbyte——它今天活得还行，但没人会说它是一枚火箭，这是客观事实，我当着创始人也这么说。我做过一些机器学习课程，但不是 AI 专家，而我对这个领域太感兴趣了——唯一入场方式就是公开学习。所以我开了博客，开始分析、学习、把学习笔记写成文章。一开始没人注意，但当你持续谈论别人的东西，Stability AI 的 CEO、Stable Diffusion 背后的 Emad Mostaque 亲自读了我分析 Stable Diffusion 的文章并说「这个好」。那基本上就是我过去五年学 AI 的方式：死磕、写出来。我当年「不合格」，今天可能还是不太合格，但比四年前合格多了。我认为这是唯一的学习路径。尤其亚洲文化特别强调资格主义：你需要那张纸、那个证——不是学位就是证书，不是证书就必须有相关工作经验，而工作又要学位，死循环。

[720s - 840s] 这等于让你 16 岁就决定一辈子干什么。但人生比这灵活得多，你想转轨随时可以，只需要真正坚持，并且意识到：努力不等于你有资格赢。你得又努力又够好。很多人尝试、很多人失败，然后问「为什么我失败了」——因为你不够好，太糟了。但也有些人尝试、努力、有判断力，做的是实质性的东西而非虚的——公开学习成功的人通常是这样。Latent Space 起初只是一个博客，就是「Swyx 在写 AI」，没别的想法。后来我们碰巧办了一些活动，比如 ChatGPT 出来之后——显然是有史以来最让人兴奋的东西——我们想办黑客松，就在黑客松上，「AI 工程师」的想法诞生了。起因是一位女士走过来问我：「我 all-in AI 了，我是非技术背景，我能做什么？」我意识到，和工程师相比，简单答案是：她什么也做不了。工程师即将比非工程师强大得多，因为工程师能同时控制一百个 AI 并给它编程；而非技术人员的上限只是把 prompt 写得更好。我随即想到：这个模式我职业生涯里见过——非显赫角色的职业化。什么叫「非显赫」？prompt 工程师不体面；体面的是训练大模型——预训练、后训练、读博、做 GRPO 这种精巧算法、做 DeepSeek 那种 hyper-connections（转写存疑）。那很显赫，也会一直显赫。

[840s - 960s] 但人们矫枉过过了：以为不属于 ML 研究的一切都没价值。其实不然——你确实需要搭 harness，而 harness 不需要你懂高等数学或有超算经验，只需要你是一个足够好、心态足够开放的软件工程师，懂得把 AI 放进传统软件代码里。当年还没有 harness 这个词，但工程师能做普通人做不了的事、并且会被市场需要，这一点很清楚。不仅公司在招这种人，开发者也想成为这种人——他们想要一个名字来称呼自己。你要做的只是选对词，最简单的就是 AI Engineer，然后把它做成一个行业、一个学科。我就这么干了。——我可以分享一个中国的旁注：就在你说的 Stable Diffusion 开源那段时间，中国现在最大的 AI 社区 WaytoAGI（转写作 Way to AI，推断）也是由一群雄心勃勃、热爱 Stable Diffusion 技术的人创办的，做图像生成各种变换，它后来也成了 ComfyUI 普及的基础。Comfy 你熟吧？——熟，我们请过他上节目。——在中国也一样。然后它逐渐邀请更多人公开学习、在飞书协作文档上贡献内容。

[960s - 1080s] 大家在飞书上贡献关于 Stable Diffusion、ComfyUI 的知识和用例，越滚越大。现在这个在线社区大约有百万级用户在一起学习（主持人随后补充「数百万」），是中国最大的 AI 社区。这和你说的发生在完全同一个时刻。——是的，你也感觉到需要一个新词来称呼这类工程师，然后你发布了那篇《AI Engineer》的文章。人们当时在试哪些备选词？可以由备选词来理解这件事：有 AI Tinkerers——也是一个重要的社区，我和 Joe、Alex 都熟；还有人说 LLM Engineer、Applied AI Engineer 等等。关键在于聚焦「人们想要的工作头衔」。这一点我是从 James Clear（《原子习惯》作者，转写注）那里学来的：想与一个人深度连接，不要连接他的工作、活动或爱好，要连接他的身份认同——你是谁？你渴望成为谁？如果你能给他一条通往更好版本自我的路径，你就能围绕它建立社区。很多非 AI 的人想成为 AI 的人，路径是什么？你给他一个能挣钱的工作头衔——好，他们就会忠诚于你。所以，现在世界上多了很多 AI 工程师。

[1080s - 1200s] 追问一下：AI 工程师要干的活很多，你提到了 harness，而这几年 AI 发展很快。你觉得这个角色的内涵有变化吗，和你三年前提出时不同？——说实话没怎么变。变的是 AI 工程学里面的「热点」，头衔本身基本是对的。早期是 RAG、上下文管理这些；最近是 coding agent、harness 工程、也许 loop 工程；还有 memory 等其他方向。对所有关心 AI 工程各个领域的人，在 AIE 世界嘉年华——我们最大旗舰会——有 30 个分论坛来展示「AI 工程师意味着什么」，评测是其中之一，多模态也是。热点在换。「agent 工程师」也许是一个新头衔，还有「AI 架构师」：AI 工程师是个人贡献者版本，那公司级的、决定「全公司怎么做 AI」的那个人叫什么？AI 架构师似乎是正在浮现的词。此外还有 AI 设计师、AI PM 等等。核心判据是：谁写代码、代码出问题时谁负责——出了故障你不负责，你就不是工程师；凌晨四点爬起来修的那个人才是工程师。我们服务的就是这些人。AI 领域另一个问题是废话太多：很多「名嘴」一生没写过一天代码，但个个高瞻远瞩、头衔吓人，「我是 CTO 呢」——不碰代码。很多 AI 大会和社区聚焦这些人，因为他们头衔漂亮，但讲出来全都一个腔调。

[1200s - 1320s] 「我的 AGI 时间线还有两年、你们以后都没工作了、要为下一步做准备」——巧了，「下一步」长得正好就像他们自己的工作。非常自利。AI Engineer 能立住的原因之一，就是我们聚焦那些通常不被聚光灯照到的、真正干活的一线工程师，他们能讲实战故事、和同行分享在做什么。这是我从业内其他大会借来的模式：GDC、KubeCon、JSConf、PyCon——工程师对工程师，他们是技术栈的决策者，是创造和革新的人，未必最有名，但在工程上绝对正确，没有废话。而其他场合废话太多了。——回到第一届 AIE 峰会是你三年前办的，如今已是三天、今年七千多人参加的完整峰会。给不熟悉的听众介绍一下这个会的内容、工作坊、AI 工程师真正关心的话题，以及你怎么把小活动做到这个规模的？——AIE 世界嘉年华本质上就是每个成熟行业都有的那种行业大展。游戏业有 GDC，三万人在旧金山；安全业有两场：RSA 和 Black Hat。你做网络安全，去其中一场，那就是你学习的地方。我特别喜欢游戏那场，因为游戏业和实验室非常像：有 AAA 工作室，做全世界最难的游戏，极致工程，全年保密，什么都不外传——但一年一次他们来 GDC，分享「年度游戏是这么做出来的」。想进 AAA 游戏工作室的开发者，这是唯一的学习途径。我认为现在的模型实验室就像游戏工作室：钱多、项目巨大、有一堆小公司在竞争，但实验室是最重要的，你想向实验室学习。去哪学？他们自己办的小会非常小、非常挑人、主要面向客户——OpenAI 的 DevDay、Anthropic 的 Code with Claude——不是工程师对工程师。

[1320s - 1440s] 而且除了 OpenAI 和 Anthropic，还有五十来家实验室都有有意思的东西。每个行业都需要一个「大家都来」的大场子，你在场子里走一圈、和每个人聊，他们之间也能互相聊。——这很有意思，尤其对一些在实体清单上的中国实验室来说，你们会聊这个吗？——他们不能和你正式开会。很多美国公司不能和中国同行正式开会，不被允许，甚至不回你邮件。不是因为他们不想，也不是因为他们是混蛋——是不被允许。但如果你们在同一个地点，就会撞上彼此。我觉得这很有意思：我在为「魔法发生」创造机会——有人被雇用、有人找到联合创始人，有一对在我会上认识、约会、现在基本结婚了。我的第一 KPI 就是我撮合出的婚姻和婴儿数量（笑）。我们想做这种高质量社区。今年目标六千人，最后做到了七千多人，150 万人线上看了我们的演讲。来不了、不住在旧金山或美国的人，仍然可以看我们发布的所有视频、追上所有内容——每一场演讲、每一个工作坊，30 个分论坛，网站上都能看到。我基本只聚焦「行得通、能挣钱、有清晰对标」的东西。你想学这个领域，或者你是创业公司要做某个东西——大概率有人在 AIE 讲过你要做的东西。这对任何查资料的人来说都是很有价值的资源。我最近做了一件没怎么公开的事：我把所有内容转写、编目。

[1440s - 1560s] 你甚至不用看演讲——把我们的整个网站喂给一个 LLM，它就能教会你任何东西。——你是在上面做了一个 CLI？——一切都是 MCP。对，所有内容都转写了，以数据集形式提供。想拿去训练你的 LLM？请便。——发布了吗？——在 ai.engineer/data，也提供 MCP 和可下载数据集，取决于你想怎么消费。我们转写一切，还对每位讲者和组织做深度调研——任何在 AIE 讲过话的人都有档案，我们基本上在做「AI 的维基百科」。最后一个关于 AIE 的问题：你是什么时候把它带到全球市场的？今年有新加坡，但几年前巴黎是第一站？

[1560s - 1680s] ——基本完全照 JSConf 模式。JSConf 是头号 JS 会议，模式很简单：我永远到不了世界上每个国家，但世界上每个国家都大概率想做点 AI，每个国家都配得上一场头部会议——他们可以自己做，也可以借我们的品牌、YouTube 频道和社交媒体。所以我们设了合作计划，巴黎第一个申请。一个好玩的事实：AIE 巴黎的主办方被 Mistral 收编了。所以现在运营 AIE 巴黎的，是法国和欧洲的头部 AI 实验室。我希望每个国家都这样：AIE 印度、AIE 日本……我现在在和这些团队谈。我永远不会像他们那样了解那个地区，但他们配得上一场好会议，他们可以做业务、招人。世界很大，我们照这个模式来。合作条款有三条：一，同意不滥用品牌，另外有某种收入分成；二，你做的所有内容全部公开出版；三，有行为准则——如果有人被骚扰或虐待，要有解决渠道。这更多是伦理层面的要求。这么做我们承担很大风险：把品牌共享给不共事的人，万一他们办砸一场 AIE 呢？目前为止没发生过，一方面我们谨慎挑选合作方。

[1680s - 1800s] 另一方面，被我们吸引来的人本来就很热忱、发自内心想做这件事。新加坡站：首先我是新加坡人，但这完全不重要——新加坡当前最大的 AI 会议，其实是一场加密货币会议（SuperAI，转写作 super AI），结果发现很多币圈的人想蹭 AI。他们知道我不喜欢他们。这是我的国家的尴尬：你告诉我这是我国家的第一名？我们需要更好的。所以现在 AI Singapore 至少有竞争力了，也许比 SuperAI 更好，我为团队骄傲，而我什么都没参与组织。我觉得世界上每个国家都该有这样的事，我们可以支持。——有意思，我看到你第一次去新加坡办活动，我跟 Lei（转写 Lay，存疑）很熟，我问过他「AI 都来新加坡了，你什么时候去上海/中国」，他说正在跟你合作。——中国是最棘手的一个。我是新加坡公民、华人，而美中之间显然有很多限制，这是主要的麻烦，所以我需要 Lei 和他的团队完整拥有这件事。因为我甚至没法在中国设立本地主体，管制很多，我不想惹麻烦——我的主业在美国，我不能给自己找白宫的麻烦。但如果你们愿意共享品牌、当自己的事情来办，那就是特许经营（franchise）——就像麦当劳：我是麦当劳，你是麦当劳中国。或者肯德基，肯德基中国，对吧。很高兴我们能把它带过去。

[1800s - 1920s] 下一个话题，是你几周前发的那个「Kill My SaaS」——我特别喜欢那一期。——我在想你的听众里有多少人在乎这个话题。只有你自己拥有生意才会在乎，因为大多数开发者并不拥有生意；但如果你有志于某天拥有自己的生意，你最好在乎。而且「SaaS 要死了」几年前就是热门话题，今年又被反复讨论。让我介绍背景：我办一场会议的预算大约在 300 万到 1000 万美元之间，取决于规模，其中每个会议大约 30 万美元是软件支出。我付的钱大约是预算的 10%、利润率的 10%——本可以收回来的钱，就为了别人那套又烂又慢又丑、不完全是我们要的软件。一年下来，我可能光 SaaS 就付了 100 万美元：你要这个、你要那个才能干活。作为老板，员工来说「这个对我们工作很重要」，我会说买；但全部加起来就是一大笔钱。然后我们用它——员工的说法是「它给不了我们想要的」，它慢；作为 SaaS 的客户，你想要个功能，你在支持队列里排队，「好的，排进六个月后的路线图，你对 我们很重要」。可我们是一家 AI 会议公司，现在是 2026 年，coding agent 已经存在——我们可以自己干。所以那个实验是：员工提议购买的 4 万美元 SaaS，我说我悬赏 1 万美元把整个东西克隆出来、开源，完全符合我们的需求。

[1920s - 2040s] 然后从这一刻起，随便用什么 coding agent——Codex、Claude Code、Devin——定制成任何我们想要的样子。我们做到了：发布了比赛，收到 69 份提交。办这种比赛等于发免费的钱——我还提出报销所有参赛者的 coding agent 订阅费——所以会有很多烂提交，那些我没给钱。但 69 份都挺好，多数可用，大概有 5 份我们会愿意付费。一周之内做完。作为一个 SaaS 生意的主人：想象你有大约 50 个客户，每个付你 4 万美元；再想象他们全部意识到这件事、在一年之内集体掉头——你的生意就没了，从年入 2000 万美元到归零。非常可怕。我认为对 SaaS 业主来说，这就是这门生意棘手的地方。我自己拥有的是实体生意，短期内不会被取代——机器人很酷，但机器人还组织不了会议。但那些做软件 B2B、拥有中腰部 SaaS 的人——Salesforce 没事，没人担心 Salesforce——中腰部 SaaS、大量服务中小企业的 SMB SaaS，它们会非常挣扎，要么转型到更好的东西上去。他们最好现在就开始想，因为我们这些当客户的老板们，自己也会写代码，而且花不了多长时间。我们现在才是他们软件真正的竞争对手。

[2040s - 2160s] 顺便说一句，模型非常擅长点开你软件的每一个界面、然后独立复刻它。所以只要我能访问你的软件，我就能克隆它——祝你好运。这就是现实。如果你是 SaaS 业主，没人知道正确答案，但你最终大概不得不转型做 agent builder：像美国这边的 Harvey 或 OpenEvidence 那样，拥有一个垂直领域、做那个垂直里最强的 agent；或者你拥有某种专有数据、基本做数据供应商。中间地带的一切，都会有别人自己写一个。——追问一下，因为你已经回答了我的第一个问题（给 SaaS 业主的建议）。我更感兴趣的是反过来：我不太同情 SaaS 业主，他们是成年人，自己能想明白。我更想建议的是生意业主。——对，我经营小生意，我是 SMB。世界上有一百万个像我这样的人：诚实谋生、给员工发工资、交税，日子过得不错。它可以更好——如果你给自己 vibe code 一套的话。我的客户开心、员工开心、我也开心，还是个好故事。对 SMB 来说这是黄金时代：只要你做的事和真实世界有关——你就是想把活干了，而在你和目标之间隔着一大堆烂软件——你现在可以突然把这个链条缩短。

[2160s - 2280s] 四万美元的「云预算」那可是很多算力（笑）；我花一万美元克隆出来，还剩一大笔钱想干嘛干嘛。这就是为什么我说这是「生意业主的问题」而非「开发者的问题」：开发者的问题是「200 美元好多啊」；而我是老板，我懂 10 个员工协作效率提升的价值——我每年付他们 150 万美元，协作好 10% 就是 15 万美元，而且实际上不是好 10%，是好 5 倍，那就是 750 万美元量级，全是替我干的 agent 工作。过去的人说「好，我下周带方案回来」——不，我要今天下午。一个自然的疑问：你自建了这么多软件，维护不会成为负担吗？——维护是真实的，不是零，你得真的去用它；但我们有 coding agent，可以自维护。这也是为什么我说「中腰部 SaaS」：你不是 Salesforce、不是 Google、不是 AWS，不是那种 99.999% 可用性的关键基础设施；你就是一个大家往里输入东西、然后取出东西的数据库。宕一天机？无所谓，一天就能修好。而且有这么多 coding agent，我会说：我愿意拿宕机时间换定制能力——什么都能定制，太爽了，我的团队效率高得多。我把这个写进了那期《Kill My SaaS》。举个最喜欢的功能例子：下个月我的下一个会议，有 1000 份投稿、25 个席位，我得从 1000 人里筛出 25 个。

[2280s - 2400s] 现有那套 SaaS 里，我们只能一个一个点开、一个界面一个界面地点。而比赛的参赛者直接改成了单表可排序视图——原 SaaS 里根本不存在这个视图，想拿到它要走六个月路线图。

[2400s - 2520s] 现在我一秒排完、一眼看全。跟听众说一句：如果你觉得这是个愚蠢的例子——对，它是。我在为音频简化。还有个事实：头名作品做的是把讲者/日程编排全部管理起来（转写存疑），UX 很好，是从人类使用视角去测试的。平均提交和顶级提交之间的差距非常大。——所以说 2026 年的今天，如果你是老板，随手雇个「看起来懂点技术的最年轻的小孩」来克隆 SaaS，可能效果不好，因为他没受过训练。但给一年时间、多看 AI 视频，就能上手。技能水位是真实存在的，很多产出是 slop（低质填充物），你不要用 slop——你的员工配得上更好的，你也配得上更好的。有时买久经考验的成品 SaaS 反而更好。当然对我们来说，公司要 AI-forward，我们会更前沿。——最后一个追问：你在播客里用了 AI 评审，而它选出的和你团队挑的赢家差别挺大。为什么？——部分是有意为之。评审的目标很纯粹：你是否提交了一份有效合规的提交？你合格吗？但评审并不负责判断「这是不是全场最佳」。所以当然有差别——出于我们办会的目的，我们要挑对会议更有用的，因为我们（起初）并不知道自己想要什么。我无法对自己都不知道想要什么的东西做 AI 评审；但当我看过 70 个不同的例子之后，我自然知道——显然就是那个。

[2520s - 2640s] 顺带一个有趣的事实：Latent Space 的 logo——如果你在这个房间外面看到本播客的 logo——来自一个叫 99designs 的网站。熟悉吗？你要新 logo、要任何设计时，常规做法是问朋友「谁设计好」，朋友给你三个名字，你三选一，对方给你一个 logo，就这样。99designs 是你发一个比赛，我拿到 400 个不同的 logo 创意，然后把奖金给我选中的那个——是个非常好的 logo。多样性远胜三选一。办比赛这种事我能做，因为我是网红，有受众；大多数人做不到。但 99designs 本身是门好生意：设计师很多，做一个设计其实花不了多少工时，关键在创意质量而不是工作量——你只需要高质量的想法，然后把它们喷洒到大量比赛上。我认为可能存在很多领域的「99designs」。软件界就该有一个 99designs for SaaS——这正是我在想的。这次我等于给自己手工搭了一个：这是 4 万美元的 SaaS、这是规格、这是评审标准、这是付款方式……全部得我自己设计。如果有人替我把这套流程做了——我只告诉他我要什么，拿回 70 个版本挑一个——我愿意付钱。这对我是一大堆行政工作。再说回动机：老实讲我做这件事主要是因为它是好内容、是播客素材。如果我没有播客，我大概不会做。评测花了两个月——一半因为我忙，一半因为它真的很重：这个 SaaS 涉及五种角色——主办方、讲者、参会者、赞助商、供应商——五方都得就「会议组织流程长什么样」达成一致。你得为每一方建账号、切换账号做数据录入，来回切换。这就是我为什么付 4 万美元买它——我不想干这些。而现在你能干了，它就成了一道很好玩的题。

[2640s - 2760s] （主持人：我们也在中国办过千人级的会，太有共鸣了。）再想想其余的部分：我以前付钱给 Calendly——你们有 Calendly 吗？就是预约链接，我们这次见面就是这么约的。Calendly 是个 130 亿美元级的生意（节目称，存疑）。你当然可以买 Calendly，而我有自己的——两个版本。我有自己的 Calendly、自己的 Google Drive、自己的 GitHub，我是认真的；有自己的 Dropbox、自己的 Slack。这一切是因为我有钱、有人，可以派人去做。但我确实认为：你可以为自己、为「一个人的受众」或「一家公司的受众」造软件，这没问题——这件事现在能成立，是历史上任何生意老板都不曾拥有过的，这很惊人。还有投资视角：你做过投资、也融过资。DocuSign，百亿美元公司——现在也许七十亿吧。DocuSign 是什么？你发一个 PDF，对方「好的，签了」，给你一张证书，就这。他们大概按每个 PDF/信封收 10 美元。有一个开源版 DocuSign，我自己托管了。区别在哪？因为我有自己的 Drive、自己的 DocuSign——签署完成的那一刻，文件直接归档进 Drive 里正确的文件夹。DocuSign 本尊做不到这个。我的资料室自动就是齐的，我一下都不用碰。那 DocuSign 当年建起来的网络效应呢——所有人都用它——你认为这还是护城河吗？完全不是。电子签名在所有司法辖区都有法律效力，DocuSign 在上面没有专利，我自己就能做电子签名。所谓网络，就是邮箱地址，就这么简单。我可以一直讲我为什么自建这每一样东西。倒也不是鼓励所有人都这么做。我的观点是：无论你的生意或人生里遇到什么问题，你真的都能找到它的开源版本——自己托管、定制、集成。「我要这个加这个」，一接就通，因为我天天在用它们：这里复制、那里粘贴。在每一百万次复制粘贴里，都藏着一个绝佳的创业和产品。（主持人随即把话题转向 Swyx 与 Diego 的那期对话。）

[2760s - 2880s] 主持人：那期也是被热烈讨论的话题和产品。中国开发者怎么看？我觉得它开创了一种从完全不同的角度使用 LLM 的新思维方式。我还听说中国有一位播客主播，完全不懂编程，但用 Jef（转写 Jev/Jeff 两见，存疑）来判断自己播客的切片会不会火，Jef 给出答案后她再发到网上。基于 Jef 的好用例太多了。你怎么看？——给没听过那期的听众补背景：我们是第一个播客它的人——是我逼他的，我们认识四年了，我说「你不给我首发我就杀了你」（笑）。他显然非常有成就：他是 InstructGPT 的联合作者，而 InstructGPT 变成了 ChatGPT。我认为 Jef 有意思的地方在于，它其实是我心里「AI 工程」的最初愿景。现在的 AI 工程是什么？围绕模型搭一个大 harness：我做一堆工作，然后交接给模型，说「拜托了模型大大，请给我正确答案，你可以花十天去想」。而 Jef 的立场是：不，模型是为我们服务的，我们不给模型打工。我会把模型的作用域收缩到只剩一件事——做选择、估概率——并把它放进我关键工作的路径上。想想当前 LLM 范式的失败模式是什么：一边是传统大 SaaS，LLM 是旁边一个聊天机器人窗口，基本一切皆如此。

[2880s - 3000s] 要么你长得像 Codex、ChatGPT 或 Claude agent——聊天窗口在这、菜单在那，仅此而已。而 Jef 所代表的、以及「System One 模型」所代表的——他不叫它决策模型，他叫 System One 模型——在于它们在代码的路径上。你可以像在普通编程语言里写 if 语句一样随意地写它。我认为这才是他说的 smarter software 的真正含义，我非常认同：你的软件应该在每一个点上都变聪明，而不是某一个点聪明、其余全是笨的——那才叫蠢。速度这些几乎是次要的，他真正想说的是可靠性——公司名字 TypeSafe 的由来。当然，速度也不错。——我发现 Diego 本人也是非常外向、对自己做的事充满热情的人。你聊过那么多研究者、开发者、AI 工程师：你觉得是什么让他做出了 Jef？他和别的研究者有什么特别之处？我听他讲自己在做的东西、在信的东西时感觉挺特别的。他自己背景就不传统：没走斯坦福博士那条路。所以他对普通开发者有共情，但他又做过 ChatGPT。而很多走传统 ML 研究员路径的人——顶级大学、有意思的论文、进实验室——他们在「大模型教」里：一切问题都要靠把模型做大来解决，更多 flops、更多参数、更多更多更多。

[3000s - 3120s] 所以需要从一个不属于那个模具里出来的人，才能真正说出「其实有另一条路：把小东西撒得到处都是，而不是一个大东西」。顺便说，Diogo（转写两见）不是第一个这么干的，几千人试过说「小模型好、快、便宜、棒」。但不知怎么 Diogo 成了。第一，他有 OpenAI 的信誉背书。第二，他把它做成了 System One 对 System Two 的问题，这让沟通变得非常清晰，清晰到可以现场跑 Doom——发布帖里有个游戏 demo，意思很明白：你绝无可能用 LLM 干这个。第三是表现力：他穿粉外套、满腔热情地讲。表现力加技术能力——这又回到公开学习：我自己就是这两者的一个缩小版，懂技术、能表达；而他在两个维度上都是十倍。这一点和他一聊就能感受到。——下一个问题关于你给其他创业公司当顾问。具体给哪类建议？也好奇你为我们听众和我这样的角色能做什么。——我们是一个三年前起步的社区，发现很多中国背景的创始人想出海：他们产品做得快、交付强，但在 go-to-market、营销、理解市场之间有巨大断层。

[3120s - 3240s] 中国创始人想打进美国？你技术很强，营销很差，去找擅长的人，因为你自己干不了——就这样。所以我们办了很多 meetup，和一些顾问合作，给已经接近 PMF 但缺人才、缺 GTM 帮助的创始人提供工作坊和建议。——有意思，我们也办 meetup。Diogo 发布 Jef 之前，其实就来过这个房间。我们办一种「写作者 meetup」，大家展示自己的作品，我做点评——TypeSafe 宣言就是从那儿出来的。我觉得和我们做的很像，只是其中一部分我收股权，大部分免费；有的想要固定例会或给股权——那样更省事，而公司做得好的话，股权值很多钱。作为一份职业这很有意思：股权不能当饭吃，也可能永远不卖——Databricks 成立十五年、估值两千亿美元，我的股权呢？但能同时在多家公司身上检验你的理论，这是真正值钱的：当雇员你只看得到一个样本，你的建议只对一家公司有效；而现在我做过三十五六家，我真的能看到它们之间的普适规律，我对每一家都是更好的顾问，因为我什么都试过。跟所有人都聊过。我能分清「这条建议在你这儿不成立，因为我在另一家公司试过、你们的处境明显不同」。几乎所有领域的顶级专家都这样：我恰好拥有「开发者营销」这个领域；但如果你做战略财务、或 GPU 算力获取，想当全球第一，你差不多得先经历一段「给所有人当顾问的服役期」，这样才能把所有东西都测一遍。

[3240s - 3360s] 有些没用，有些明显有用，你会学到东西。样本量必须大于一，多数人的样本量是五；如果你能到五十，你基本就是这个领域的顶尖了。——分享一个我在出海中国创始人身上的相似观察：越来越多是工程甚至研究背景，他们以为自己有最好的硬技能、技术、知识就能做出最好的产品，但客户真正想要的和你的产品之间、以及不同本地市场之间，仍有巨大落差。你见过类似的吗？

[3360s - 3480s] ——有很多种切入方式，最重要的第一件事是：你得真的有个「龙」——你真的在某一件事上全球第一，这最重要，其余都是「画龙点睛」（Swyx 中文原话）：龙是你自己画的，我可以帮你点睛；而多数人连龙都没有。也有人没有龙却以为自己有，那就是自我和傲慢的问题了。假设你真有个非常好的产品，这里真正的关键是：不要把营销当「最后一公里」——「我什么都做完了，博客发出去前一晚，你帮我看看博客？」我什么也做不了。要让营销更早进入流程，早到产品还没建、还没写规格的时候：我们打算怎么讲故事、切哪个细分市场——这样 GTM 从一开始就长在产品里。很多工程师/研究员创始人早期没有这根 GTM 神经，通常把它全堆到最后。

而现实是，公司里最 junior 的往往是 GTM 团队——几个实习生、一个 22 岁的年轻人运营社交媒体。不，请认真对待：谁应该用这个东西？他们在哪里出没？他们面临什么问题？他们用什么词来描述自己的问题？看到你的东西之后，他们的生活会怎么变好？都是非常简单的问题，但多数研究员/工程师创始人从没想过；而如果他们有一个 GTM 搭档，往往非常有用。另一件事：说到开发者工具，我发现很多中国创始人在做机器人。

[3480s - 3600s] 机器人也很有意思：时间维度长得多，而且还在非常早期的阶段。怎么中国的机器人创始人全都在做洗衣（笑）？Unitree 是最强的，但 Unitree 还在做 Unitree 会做的事——他们这周刚发布的，是功夫动作。这个我以前见过。给我看点我从没见过的东西，我才会真正开始注意。这在机器人领域很难展示，但模型领域仍然可以，因为还太早期；消费品体验上的创新产品想法也可以。所以「做 demo」非常重要——这也是为什么 TypeSafe 虽然很新但做得很好：Doom demo 不是 Diogo 本人做的，是他们的 GTM 做的（转写此处人名存疑），而那是营销的核心原则：show, don’t tell。给我看这个东西正在产生好处，别用 bullet points 告诉我它是什么；别给我 PDF、别给我 PPT。给我看。主持人：再追问一下「品味」（taste）——现在人人都在谈 taste，你觉得 GTM 和做产品都需要品味。对，可能真的需要。我现在反而有点难定义品味是什么，投资人天天挂在嘴边。你怎么定义工程师甚至创始人的品味？——我可以给出一个「没什么用」的定义。品味是：你能够从一个巨大的——

[3600s - 3720s] ——可能性集合里，快速地、低成本地收敛到一个子集，而这个子集当你事后回望时，是「一望即知地好」的。站在起点，摆着一千种可能的做法，选哪个？可以蛮力：一千个都试一遍，总能找到一个好的。但有品味的人会说：「这些显然不行。我解释不了，就是显然不行；这个看起来有意思。」走到旅程终点、选定一个之后，你回头看：「对，就是它。你是在告诉我当初还在考虑另外九百九十九个吗？显然它才是正确答案。」这个定义有很多应用场景，现在我脑子里的是 iPhone：智能手机很长一段时间都有键盘，而乔布斯说「不，我们不惜一切代价要触控屏、要虚拟化的屏，因为它必须长得像一块平板」。这花了十年研发——之前有一家叫 General Magic 的公司试过这件事然后失败了。但显然，现在全世界的手机都长这样。这就是品味。1990 年代人们刚开始做智能手机时，可做的事无穷多，但要有品味的那个人说「对，我们要的是这个」，对其余一切说不，而我 shipping 的第一部手机就得长这样。主持人：回望这些 AI 产品，有哪个你觉得是有品味的例子的？——我其实认为 Cursor 做的每一件事都很有品味：他们第一个做了 VS Code 那条路（fork），现在又在做 Grokbot，也是个有品味的决定。因为现在 OpenAI 得抄他们，Claude 也得抄他们——当别人开始抄你刚 shipping 的东西时，就说明你有品味。

[3720s - 3840s] 再往实验室之下去看，我觉得有意思的是那些「更无聊的问题」。法律领域的 Legora：他们有一个大规模并行的文档评审流程，一旦你见过它，你会说「我再也不想在别的地方干活了，我必须用 Legora 干活」。而任何人本都可以想出这个点子——任何人都行，点子很便宜——但你没做，Legora 做了。所以 Legora 里的某个人有那种品味：「我们将从能做的十亿件事里，定义出这个成为大家工作方式的产品。」主持人：接近尾声了。AIE 两周后就到上海了，你最希望在 builder 们身上看到、学到什么？哪些话题你最感兴趣？——对我来说更重要的是「差分」：我想看上海和旧金山有多不同。旧金山有 B2B SaaS 泡沫，而上海——你们用 ComfyUI 吗？主持人：我们这儿不怎么用，我们喜欢我们的 Canva，还有 Fair AI（转写存疑）。别用 Comfy，Comfy 更像开源圈的玩具。OpenClaw 也一样——OpenClaw 在这儿比 Comfy 更有声量。主持人：OpenClaw 在中国是真的火了，你应该听说过。我当然知道，但我还没亲眼见过那种能量。这非常不同。你想想，「一百万人的社区」对你们来说是中等规模——那在这儿已经是中国顶流了。

[3840s - 3960s] 不，我见过那种中国的网红工厂，一堆中国女孩对着手机做内容，中国的尺度完全是另一个量级，竞争也激烈得多。人们说旧金山也 996——并没有，周六周日他们其实还在放松；在中国，996 是动真格的。主持人：在创业公司和大厂确实挺正常的。我觉得这真的很狠。主持人：这边的人不喜欢字节（今日头条）和美团那种。对，太狠了。这儿的人可没那么拼。我听说这边的工程师、创始人有那么拼。我这么拼是因为我没有股权（笑），但普通员工呢？在法律上，你不能要求员工那样工作，那是违法的。我见过一些最顶尖、最竞争激烈的美国创业公司，我可以肯定他们没有中国拼。BAT、第一代、第二代那些创始人——天哪，他们是那么好的创始人。是的，如果他们生活在美国，会干掉 Amazon、干掉 Facebook。他们没有更富的唯一原因，是他们住在中国。说到软件也一样：我 2007 年在北京留学时，比亚迪还在说「我们正在起步、我们立足中国、永远立足中国」——现在这里人人都爱比亚迪。真的吗？BYD 正在接管美国，太疯狂了，我以前绝不会想到 BYD 会走到那一步。我都不知道该怎么下结论。中国人的能量、精神、创造力和狠劲，是外界完全没有接触到的。就更别提——

[3960s - 4080s] ——模型实验室了，显然 DeepSeek 和所有那些实验室。所以有很多理由让人兴奋、去把这些更多地展示出来。尤其我认为中国的活动有个问题：对直播有限制；而且就算办了，呈现方式也和西方观众喜欢消费的方式不同，所以触达不到那边。阿里云做过一件事：把整个阿里巴巴的路线图连同 Qwen 一起发布——美国没人看到。为什么？因为它嵌在一场由别人代讲的 corporate presentation 中间。于是只好由我自己去为美国观众把所有内容翻译一遍。跟我们一起做嘛——放到 AIE 上讲，他们就会看到的。还有我看到今年已经有一些开源模型公司参加了旧金山的 AIE，比如 MiniMax。所以我们希望每一方都有代表出席。我认为有很多人对这种跨文化交流感兴趣、而对政治不感兴趣。我们不做政治，只做技术。好的。——这是我的问题问完了，但我还有三个快问快答，想到什么说什么。第一个：你现在最兴奋的 AI 进展是什么？最担心的又是什么？

[4080s - 4200s] 最兴奋的：System One 模型是非常明确的第一名答案，但太无趣了，所以我不说它。我认为被低估的是 interaction models（交互模型）。Thinking Machine 正在做这件事，错过发布的人可以去查：全程全双工的双向交流——不再是你一句我一句轮着说，而是任何时刻，你说话的同时我在听、我说话的同时你在听，我可以打断你，你立刻响应并调整。我看到那个 demo 了。对，这才应该是 AI 的样子。它有个推论：continual learning AI（持续学习 AI）——不再有「预训练完了、后训练完了」这种切分，而是我与你工作的全程它一直在学，这是非常有趣且重要的领域。最后一块是世界模型，显然发展得很好，不用我解释。至于担心的：从个人角度，如果 AI 寒冬来临——现在肯定不是寒冬，这大概是人类历史上最长的 AI 盛夏，但 AI 的历史约五十年，一直是冬夏交替。如果冬天来，什么时候来？我会不会破产？对一家靠行业热度吃饭的会议公司来说，这是真实的风险。更宏观地：我所有实验室的朋友都真的非常担忧的一件事，是社会无法足够快地适应和调整。不是因为假新闻，也不是因为失业——是更糟的东西。它现在以网络安全的形式表达出来：我每天都被黑，黑、黑、黑。我感觉地平线上还有更糟的、我们尚不知道的东西。最极端的版本是生物武器——「oops，我发明了 COVID-2」。如果比那还糟呢？所以我们在播客里的表述是：大多数人的默认假设应该是 AI 末日——人类全部历史上，更先进的文明遇到更落后的文明，落后的一方灭绝；而这东西的进化速度大约是我们的 1000 倍，所以某个时刻它会比我们更先进。

[4200s - 4320s] 到那时我们必须想好怎么办。——第二个轻松点的问题：你推荐哪些 newsletter、播客或内容？我有一个清单，叫「好用的 AI newsletter 与播客列表」，在我的 GitHub 上。最近的可以看 aeo.latent.space：我把「对工程师来说最好的 AI 播客是什么」这类问题拿去问每一个模型，把答案转写出来做成排行榜。我问了两百个这类问题：最好的支付创业公司、最好的银行、最好的理财顾问、最好的数据库、最好的「一切」。所以叫 AEO（答案引擎优化）？对，优化的是模型给出的答案——而这大概是实验室里选数据集的人的品味的相当不错的代理指标，他们选了 Latent Space：要么第一、要么前三。挺好。——最后我把提问权交给你，或者：你对 AIE 上海有什么期待？当然。我希望能听到更多基础设施类的话题，比如安全、支付，或者那些真正解锁更多落地应用的东西，来自不同亚洲基础设施的进展。我特别好奇这些话题，因为它们在中国讨论得比较少。而且这几周待在这里，我发现谈这些的人变多了，可能还有 RSI（转写存疑）或者一线研究者。我希望它能成为东方与西方之间一场很好的对话——这里的研究者和中国的研究者。也希望它圆满成功，然后我们可以继续把它办下去，明年再见。

[4320s - 4344s] 我也希望如此。祝团队好运。做任何事情的「第一次」都是最难的，对吧？同意。我真的很佩服他们在做这件事。非常感谢你，Sean。谢谢，谢谢。