2026-07-27

Lenny's Podcast 速读：Anthropic 首位 technical PM 如何在 token maxing 与 jagged edge 中做 AI 产品

Dianne Penn 解释 Anthropic 如何把模型跃迁转成产品价值，以及为什么 evals、亲手使用和人的判断仍是 AI 产品工作的核心。

先给结论

这是一场关于「AI 产品到底怎样被做出来」的访谈，不是岗位替代率预测。Lenny Rachitsky 对话 Dianne Penn，讨论 Anthropic 如何从早期只有五名产品工程师的团队，走到同时做前沿模型、Claude Code、MCP、Skills、computer use、tool use 和 reasoning 的组织。Dianne 2023 年加入 Anthropic，成为公司的第一位 technical product manager，目前负责 AI Research 与 Labs 团队的产品工作。节目时长约 1 小时 34 分钟。1

如果你只想知道一个结论，可以先记住这一句：前沿模型的能力增长不会自动变成用户价值，中间还需要产品把新能力变成可探索的体验，需要 evals 把模糊的用户抱怨变成研究团队能处理的失败类型，也需要人持续验证模型到底做对了什么。Dianne 把这条链路讲得很具体，谈到了模型、产品、研究、组织和个人工作方式之间如何互相牵引。

这期尤其适合三类读者：正在做 AI 产品的人，想理解 research PM 到底在做什么的人，以及已经被 AI 工作节奏压得喘不过气、却还没有找到有效用法的人。它没有提供一份「未来 PM 技能清单」，而是展示了一个团队怎样把实验、反馈、评估和交付连在一起。

这期访谈在回答什么

这是一场观点访谈，主问题是：当模型每次升级都可能改变用户能做什么，产品团队应该怎样决定现在做什么、怎样验证做得好不好？谈话从 Anthropic 的早期创业状态开始，经过 coding 成为关键使用场景、Opus 3 和 Claude Code 的相互促进，再进入 Labs 的孵化方式、研究团队的工作语言和 eval-driven development。后半段转向产品经理的招聘与管理、如何实际使用 Claude、过度依赖 AI 的风险、模型为什么需要适时反驳用户，以及人在写作、判断和关系中的位置。

节目页面给出的嘉宾介绍还包括她此前在 Amazon 参与 Alexa AI、在 JP Morgan Chase 交易高收益债券的经历。这段经历解释了她为什么同时熟悉技术、产品和高风险决策，但这期的重点仍然是她在 Anthropic 看到的工作方法，而不是个人履历本身。1

1. Anthropic 的早期优势不是规模，而是共同下场

Dianne 回忆，自己 2023 年加入时，Anthropic 的产品团队只有五名工程师，API 业务甚至只有一名工程师。那时外部很容易觉得公司没有机会：OpenAI 已经领先，Anthropic 似乎来得太晚。她认为后来真正留下来的东西，是那种「一起把问题做出来」的文化。研究员、产品、工程和设计不是在各自的职能边界里交接，而是一起观察模型到底能做什么，再判断这些能力如何给用户和社会带来价值。

她举了 Golden Gate Claude 的例子。Anthropic 早期的可解释性研究发现，模型内部存在与特定主题相关的 feature；把 Golden Gate Bridge 这个 feature 调高后，Claude 会在几乎每个回答里谈到金门大桥。用户问意大利面食谱，模型也会把面条的颜色和金门大桥的国际橙联系起来。这个结果当然不是一个严肃的产品功能，但它让研究成果变成了大众可以直接体验的东西。

团队在大约 24 小时内把这个体验做进 Claude.ai，研究、产品、工程和设计一起完成，最终大约触达 2,000 人。数字不大，意义在于团队看见了自己的工作方式：研究可以用一个轻量体验被公开验证，产品不必等所有东西都成熟，工程和设计也可以主动投入一个尚未写进路线图的想法。Dianne 把它视为 Anthropic 找到自身产品身份的早期转折点。关于这项可解释性实验的公开背景，可参见 Anthropic 的 Golden Gate Claude 介绍。

2. Coding 的转折，来自模型和产品一起到位

Dianne 说，2023 年她加入时，没有人会把 Anthropic、Claude 和 coding 放在同一句话里。模型当时已经可以补全代码，但用户开始尝试让模型直接写长篇代码，这暴露出一个产品机会：不只是把模型训练得更强，还要把这种能力包装成用户真正能使用的工作方式。

她把 Opus 3 视为一个重要的内部里程碑。团队当时还不到 200 人，却在训练、测试、推理、微调和预训练之间形成了共同目标。研究负责人和产品团队一起在假期里盯训练结果，逐渐建立了跨团队信任。Opus 3 也让 Anthropic 找到了一批早期开发者用户，因为它在 coding 上提供了当时很多人没有想到的价值。

更大的变化发生在 Opus 4.5 和 Claude Code 之间。Dianne 的说法是，团队终于同时拥有了模型和「vehicle」，也就是让前沿智能真正进入用户工作流的产品。Claude Code 需要 Opus 4.5 才能加速采用，Opus 4.5 也需要 Claude Code 才能让用户感受到它的 agentic 能力。模型发布与产品体验不是先后关系，而是互相放大的组合。1

3. Token maxing 不是多花钱，而是把成本换成发现

节目中提到的 token maxing，来自这样一个判断：如果今天愿意一年花 10 万美元购买模型 token，就等于提前生活在别人几年后才会普遍拥有的工作方式里。YC 对这个词的公开解释也把重点放在「一个人借助 AI agent 完成过去需要整支工程团队才能完成的工作」上。2

Dianne 没有把它理解成单纯增加 token 预算。她更关心 token spend 带来的输出，也就是实验。Anthropic 内部最有创造力、最擅长原型的人，往往会反复使用每个新研究模型，先把它摸清楚，再从「能用」走到「好用」，最后形成产品想法。模型变化很快时，不亲手使用技术，就很难写出可信的策略。

她还描述了早期 Anthropic 的一个 Slack 频道：几乎全公司的人一起测试 Claude，分享让模型改文章、写邮件或处理其它任务的尝试。一个人先提出用法，其他人很快用不同变体验证，十来次请求后，某个新的使用场景可能就浮现出来。实验不是个人运动，公共试用会让组织共同发现模型的能力边界。

这里有一个很实际的工作建议：不要把 token maxing 简化成「多调用模型」。先把目标定成实验和发现，再决定哪些 token 支出值得保留。对个人来说，选一两个真实问题深入做，通常比同时试几十个工具更容易获得稳定收益。

4. Labs 的价值是押注非连续机会

Anthropic Labs 负责把那些可能不适合放进核心路线图的想法继续往前推。Dianne 对 Labs 的定义是：识别一个不连续的大赌注，沿着线索追下去，先判断「这里到底有没有东西」，再问它能不能放大到 10 倍、100 倍或 1,000 倍。

Labs 团队对主题可以强烈坚持，对具体原型则保持较弱的承诺。工程师被鼓励自驱实验，一个原型即使没有立即发布，只要帮助团队理解了未来一两代模型可能出现的机会，也有价值。某个想法现在做不成，不代表永远不做，团队可能在一两次模型迭代后重新回来看它。

这套方法和传统创新部门最大的区别，是它不把「实验失败」当成无效产出。真正的选择发生在后面：哪些 bet 继续投入，哪些要关闭，哪些只能暂存。Dianne 也承认，亲手推动一个想法却被关闭很难受，所以 Labs 需要挑选能承受这种零到一不确定性的人，同时用小团队保持速度。Anthropic 对 Labs 的正式介绍将其定位为孵化前沿 Claude 能力实验产品的团队，产品与规模化组织也在同一套结构中被区分开。3

5. Research PM 的工作，是把模糊失败翻译成可行动问题

普通用户说「Claude hallucinated」，对研究团队来说还不够具体。Dianne 的团队要继续追问：模型当时有没有调用工具？调用了正确的工具吗？看了正确的文档却抓错了事实吗？问题属于 tool use、搜索、知识整合，还是 alignment？只有把反馈还原成失败轨迹，研究人员才知道下一步该改什么。

研究团队同时承担两种时间尺度的工作。一端是日常训练、数据、算法和 eval 迭代，另一端是更大胆的未来设想，例如怎样让 Claude 使用电脑、怎样让 AI 导航屏幕。Dianne 把研究员形容为带有 founder-like energy 的人：他们既要在细节里看训练运行、底层数据和 eval，也要能对未来提出大胆描述。

她观察到，优秀研究员通常有几项共同点：能用第一性原理推理问题，对自己的研究领域足够着迷，能保持对训练细节的兴趣，同时敢于把目标想得很大。更重要的是，他们会对方向保持执着，对实现路径保持松弛，因为模型能力变化得太快，原来的做法可能很快失效。

6. Evals 正在变成 AI 产品的测试驱动开发

Dianne 说，团队里有一句话：「evals are the new PRDs」。这不是说 PRD 已经没用，而是说在模型产品中，真正代表用户价值的，不再只是写出一份描述愿景的文档，而是把用户痛点变成一组可以反复运行的测试。

她用 Claude 早期不擅长遵循特定 schema 的问题举例。用户说「Claude 不会遵循指令」，团队继续追问具体场景、输入和输出，发现早期反馈中大约 80% 的问题其实是模型没有生成正确的 JSON。接着，团队整理出 30 到 40 个具体例子，形成 eval set：每个样本包含 prompt、模型 response 和期望的 golden answer。新版本模型发布时，团队重新运行这组测试，检查问题是否真的改善。

这套工作有几个关键环节：先读取真实用户轨迹，再判断失败是否可复现、是否持续出现、是否值得修复；然后把它抽象成研究团队能消费的测试；最后用同一组测试衡量新版本。Dianne 把它称为 PM 的 test-driven development，也就是先写清「什么叫做对」，再推动模型向这个标准靠近。

但 evals 不是 PRD 的全面替代品。已知问题适合用 eval 快速描述；模型发布仍需要 PRD 来让产品、工程、法务、安全和其它利益相关者对齐。对于 computer use 这种尚未充分验证的方向，PRD 里的产品愿景仍然有用，因为团队要先探索一个还没准备好服务所有人的技术，怎样先为某一类用户创造连贯体验。

7. AI 时代的 PM 和经理必须亲手交付

Dianne 对管理者的要求很直接：不能只停留在讨论 AI，要亲自用它、亲自交付。她会为自己保留一到两个工作流，持续观察模型怎样变化，借此保持对模型能力的「理论心智」和对速度的感知。她认为，即使是经验丰富的 PM，新加入团队时也要从理解用户、阅读反馈、和客户交流、亲手构建开始。

原因不是人人都要变成工程师，而是没有亲手做过，就很难判断一个 AI 产品的好坏，也很难理解团队遇到的真实限制。她把「sweat the tokens」和过去 PM 关注像素放在一起：今天既要看界面，也要看模型在上下文、调用和输出上经历了什么。

她给不适应 AI 的人一个不太鸡汤的建议：不要独自寻找完美用法。找一个已经对这项技术有兴趣的人，围绕自己真正关心的问题一起试。Anthropic 内部的原型分享会让更多人发现「原来现在已经能做到这个」，这种共同发现比被迫打卡更可能带来乐趣。

8. 先有自己的判断，再让 Claude 反驳你

Dianne 会用 Claude 帮自己准备困难对话。她把《Crucial Conversations》里的方法做成 skill，在和同事进行高难度沟通前，让 Claude 帮她检查是否说得足够具体、是否应该更直接、是否遗漏了对方的反应。她不一定采用建议，但模型可以帮助她快速换一个角度思考。

她同时强调不能把全部思考交给模型。面对需要个人判断的议题，她会先形成自己的观点，再把 Claude 当作 sparring partner，要求它推动、质疑和完善这个观点；对于月度业务复盘等标准化工作，她则希望未来把写作尽量交给 Claude，自己负责审核和验证。

这里的区分很重要：AI 的价值不只是替人完成任务，也在于判断它是否做了正确的事。Dianne 认为，Claude 的 constitution、alignment 和 safety 训练之所以让产品更有趣，恰恰是因为模型不会对所有想法都点头。一个只会同意用户的助手，不能真正改善用户的判断；适时的 pushback 才像一个有用的同事。

9. Jagged edge 让「谁来验证」比「谁来写」更重要

Dianne 用 jagged edge 描述 AI 能力的不均匀：模型可能在某一类复杂任务上表现惊人，却在另一个看似普通的任务上暴露粗糙边缘。写作就是她提到的一条边界。即使语言模型大量学习过优秀文本，用户仍能从固定的套路里认出 AI 味道；当模型在 agentic 能力、工具调用等方面取得进展后，写作可能又成为更明显的短板。

她没有把这个问题归结成「AI 写得不像人」。她把关注点转向可验证性：谁在检查输出，谁在签字确认，谁承担发布结果。对月度业务复盘，写作本身也许可以委托给 Claude，但信息的验证、判断和最终签字不能一起委托出去。

这和前面 evals 的逻辑是一致的。AI 产品的工作不是追求每一次输出都可预测，而是把「什么算是好结果」说得足够清楚，并用测试、复核和责任边界守住那些模型还没有跨过去的边缘。

10. 值不值得完整收听

值得，尤其适合想理解 AI 产品内部工作方式的人。节目最有价值的部分不是某个新工具名称，而是把几件经常被分开讨论的事放到同一条链上：模型能力会突然跃迁，产品要快速把跃迁变成体验，用户反馈要被拆成可定位的失败，evals 要让研究能测量改进，经理又必须亲自使用技术来保持判断。

它也保留了不舒服的部分：技术变化快到让人疲惫，AI 的能力不是均匀增长，越强的模型越需要安全和发布护栏，个人如果只追逐工具数量反而可能失去方向。Dianne 给出的办法并不复杂，却需要持续投入：选一两个真实问题，亲自做，和别人一起做，先形成自己的判断，再让 AI 加速、反驳和复核。

节目元信息

节目： Lenny's Podcast: Product | Career | Growth
单集： Anthropic’s first technical PM on token maxing, the jagged edge, and living in the future | Dianne Penn
嘉宾： Dianne Penn，Anthropic AI Research 与 Labs 团队产品负责人，Anthropic 首位 technical PM
发布时间： 2026 年 7 月 26 日 20:32（北京时间）
时长： 约 1 小时 34 分钟
官方页面： 本期节目与音频

逐字稿说明

官方节目页面提供本期音频、简介和章节，但没有可读取的完整文字稿。以下内容来自完整音频转写，英文原文按 ASR 结果保留，包含广告、重复、听不清或专名识别错误，不对原始英文做校正或删减。由于音频没有可靠的 speaker diarization，附录使用「讲话人」作为中性标签；中文译文紧跟对应英文段落。时间轴每 10 分钟标注一次。

完整中英对照逐字稿

00:00:00

英文原文（完整 ASR）

讲话人：In 2023,when I started,

讲话人：nobody said anthropic and clod and coding

讲话人：in the same sentence.I want to go back

讲话人：to the beginning of anthropic.I remember

讲话人：feeling,man,

讲话人：these guys have no chance.OpenAI is

讲话人：so far ahead.At the time,

讲话人：I saw people were starting to use these

讲话人：models not just for code autocomplete,

讲话人：but actually writing long-form code.

讲话人：It's not an opportunity for us to train

讲话人：Opus 3 to be better at.That was the

讲话人：inflection.I always think about Opus

讲话人：4.5 a year later during winter break

讲话人：when everyone was home,

讲话人：able to code.What was magical about

讲话人：Opus 4.5 is we also now not just had

讲话人：a model,but a vehicle,

讲话人：a great product experience like Cloud

讲话人：Code.Opus 4.5 wouldn't have had that

讲话人：moment without a product like Cloud

讲话人：Code.And Cloud Code wouldn't have had

讲话人：that type of adoption accelerated without

讲话人：Opus 4.5.I want to talk about how the

讲话人：product role is changing.For my team,

讲话人：the way to drive user value is to figure

讲话人：out the right user feedback,

讲话人：the evals.We actually have a saying

讲话人：on the team of evals are the new PRDs.

讲话人：Something Gary Tan's been talking about,

讲话人：if you're willing to spend$100,000

讲话人：a year right now in tokens,

讲话人：you are living the way somebody in 2028

讲话人：is going to live.You have to sweat the

讲话人：tokens as much as you sweat the pixels.

讲话人：You have to be using the models to come

讲话人：up with good and great and better ideas.

讲话人：And there's no substitute for that.

讲话人：People need to be more ambitious with

讲话人：AI tools these days because they're

讲话人：just capable of so much.One thing I

讲话人：ask the team is,let's say,

讲话人：Cloud 8 comes around.What changes in

讲话人：what users do?

讲话人：What does that mean for how you're building

讲话人：today?

讲话人：Today my guest is Diane Penn,

讲话人：head of product for the AI research

讲话人：and labs teams at Anthropic.She joined

讲话人：Anthropic as the first technical product

讲话人：manager over three years ago,

讲话人：which is a lifetime in AI time.When

讲话人：the product team was just five engineers,

讲话人：she's helped ship every model at Anthropic

讲话人：from CLAW2 through Fable.She's also

讲话人：helped incubate and launch CLAW Code,

讲话人：MCP,Skills,CLAW Design,

讲话人：and also core capabilities like computer

讲话人：use,tool use,

讲话人：and reasoning.It is always such a treat

讲话人：and so mind-expanding to get to talk

讲话人：to someone who's at the very center

讲话人：of AI and product management.It's hard

讲话人：to imagine someone who has seen more

讲话人：of where things are going than the head

讲话人：of product for Anthropix Research and

讲话人：Labs teams.Before we get into it,

讲话人：don't forget to check out Lenny's Product

讲话人：Pass.com for a year free of the hottest

讲话人：and most beautifully crafted AI products

讲话人：in the world,

讲话人：available exclusively to Lenny's newsletter

讲话人：subscribers.With that,

讲话人：I bring you Diane Penn.Diane,

讲话人：thank you so much for being here.Welcome

讲话人：to the podcast.Thank you,

讲话人：Lenny.It's so nice to see you again.

讲话人：I want to go back to the beginning of

讲话人：Anthropic,

讲话人：the early days.I remember when Anthropic

讲话人：first launched.This was,

讲话人：I don't know,

讲话人：the first model when it launched years

讲话人：ago,three years ago,

讲话人：something like that.It was.Three years.

讲话人：I remember just like feeling that,

讲话人：man,

讲话人：these guys have no chance.OpenAI is

讲话人：so far ahead.They're just like,

讲话人：what are they thinking?

讲话人：How is this possible?

讲话人：OpenAI has won.It's too late.Things

讲话人：are very different now.The latest number

讲话人：I saw was Anthropic was making like,

讲话人：I don't know,

讲话人：$50 billion in ARR.That's like what

讲话人：companies used to go public at.Like

讲话人：very successful companies went public

讲话人：at$50 billion in valuation.Anthropic

讲话人：reportedly is making that every single

讲话人：year.You joined as one of the earliest

讲话人：PMs.There were something like five engineers

讲话人：when you joined.The model hadn't even

讲话人：launched when you joined.What was it

讲话人：like in those early days of Anthropic?

讲话人：What's something that might surprise

讲话人：people about what it was like at the

讲话人：beginning?

讲话人：I think a big part of what's made Anthropic

讲话人：today actually has been very much the

讲话人：core of even the early days.So I joined

讲话人：in 2023.Like you said,

讲话人：we had five product engineers.There

讲话人：was one engineer for the entirety of

讲话人：our API business,

讲话人：if you believe.And I think a big portion

讲话人：of it was the culture was really strong.

讲话人：And I think this is something I emphasize

讲话人：for folks who are interested in the

讲话人：company,

讲话人：really do walk the walk of the mission

讲话人：and the culture and the values.And the

讲话人：energy was very much like a startup.

讲话人：And I think you're right.We were very

讲话人：much trying to find our identity in

讲话人：the early years.I think there's one

讲话人：piece around the technology,

讲话人：but how does that technology bring value

讲话人：to users,bring value to society,

讲话人：and what could it possibly be?

讲话人：And I think the early years were us

讲话人：exploring that in different ways.Like

讲话人：we did start with like Cloud.ai,

讲话人：another chatbot,chat assistant,

讲话人：and evolving into things like tool use.

讲话人：I think one of the moments where Really,

讲话人：we started to get into our groove was

讲话人：shipping things like Golden Gate Clawed.

讲话人：I don't know if you remember that.No.

讲话人：So this was actually up for about 24

讲话人：hours or so.We had just published one

讲话人：of our early interpretability research

讲话人：in early 2024.Yeah.of the examples was

讲话人：essentially you could have what's called

讲话人：like features of the model within the

讲话人：layers,

讲话人：which express certain types of thematics.

讲话人：So one of the themes that the researchers

讲话人：was able to identify was,

讲话人：let's say,

讲话人：bullet point writing.Another one was

讲话人：people and places.And one that really

讲话人：came up frequently that resonated was

讲话人：the Golden Gate Bridge.And so when you

讲话人：actually essentially dialed up that

讲话人：feature,

讲话人：Claude would obsess about the Golden

讲话人：Gate Bridge.So meaning in every one

讲话人：of its responses,

讲话人：it would come back and talk about the

讲话人：Golden Gate Bridge.So if you said like,

讲话人：give me a recipe for making spaghetti,

讲话人：it would say,

讲话人：here is a recipe.And the orange color

讲话人：is just like international red that

讲话人：the Golden Gate Bridge looked like.

讲话人：And so it was like really quirky.And

讲话人：we very much wanted to, in that situation,

讲话人：just bring that user,

讲话人：bring it to the masses and bring it

讲话人：to people who are starting to use Cloud.

讲话人：And so the entire experience actually,

讲话人：we spun up on our Cloud.ai website within

讲话人：24 hours.And that took engineering,

讲话人：product,design,

讲话人：our research teams all working together.

讲话人：And we were really,

讲话人：really proud of it.I think it maybe

讲话人：reached only 2,000 people,

讲话人：to be honest.But it made us feel like,

讲话人：oh,we can actually bring new user experiences,

讲话人：showcase our research in a way that's

讲话人：different and authentic to us.And in

讲话人：a very startup-y pace,right?

讲话人：to me,

讲话人：was one of those maybe hidden inflection

讲话人：points of we were starting to find our

讲话人：identity,that we could build products,

讲话人：build experiences that were different

讲话人：from what our competitors had seen,

讲话人：what was already out there.And I think

讲话人：that,obviously,labs,cloud code,

讲话人：et cetera,

讲话人：we then started to identify ourselves

讲话人：as,

讲话人：would we actually think the world is?

讲话人：think about AI,

讲话人：how to bring that closer to the public.

讲话人：But it was a very bottoms up culture.

讲话人：And so that entire experience was very

讲话人：bottoms up.I see engineers,

讲话人：I see designers donating time to work

讲话人：on.And so I like to always use that

讲话人：as an example of like what the early

讲话人：days were like.But the culture and the

讲话人：values have very much,I think,

讲话人：stayed the same since those early days.

讲话人：This episode is brought to you by our

讲话人：season's presenting sponsor,

讲话人：WorkOS.What do OpenAI,Anthropic,

讲话人：Cursor,Vercel,Replit,Sierra,

讲话人：Clay,

讲话人：and hundreds of other winning companies all

讲话人：have in common?

讲话人：They are all powered by WorkOS.If you're

讲话人：building a product for the enterprise,

讲话人：you've felt the pain of integrating

讲话人：single sign-on,skim,RBAC,

讲话人：audit logs,

讲话人：and other features required by large

讲话人：companies.WorkOS turns those deal blockers

讲话人：into drop-in APIs with a modern developer

讲话人：platform built specifically for B2B

讲话人：SaaS.Literally every startup that I'm

讲话人：an investor in that starts to expand

讲话人：upmarket ends up working with WorkOS.

讲话人：And that's because they are the best.

讲话人：Whether you are a seed stage startup

讲话人：trying to land your first enterprise

讲话人：customer or a unicorn expanding globally,

讲话人：WorkOS is the fastest path to becoming

讲话人：enterprise-ready and unblocking growth.

讲话人：It's essentially Stripe for enterprise

讲话人：features.Visit WorkOS.com to get started

讲话人：or just hit up their Slack where they

讲话人：have actual engineers waiting to answer

讲话人：your questions.WorkOS allows you to

讲话人：build faster with delightful APIs,

讲话人：comprehensive docs,

讲话人：and a smooth developer experience.Go

讲话人：to WorkOS.com to make your app enterprise

讲话人：ready today.What are some of the other

讲话人：big inflection moments as you think

讲话人：about just Anthropic going from just

讲话人：this like lab that's trying to compete

讲话人：with this juggernaut of open AI at that

讲话人：point to what it is today,

讲话人：what are some moments that stick out

讲话人：of like,wow, that really changed things?

讲话人：Definitely when we were training and

讲话人：testing Opus 3,

讲话人：I think that was the moment when the

讲话人：company,

讲话人：I think we were less than 200 people

讲话人：still at that point.And it was very

讲话人：clear that we needed and wanted to create

讲话人：a frontier model.And that was very important

讲话人：in terms of like our ability to reach

讲话人：like users,consumers,

讲话人：and to showcase our research.And we

讲话人：were looking for ways for also,

讲话人：why should somebody choose Claude?

讲话人：And that was like a core question.And

讲话人：that was a core question we were getting

讲话人：asked in the early days.And I think

讲话人：with Opus 3,

讲话人：it launched I think early March,

讲话人：2024,but there was many,

讲话人：many months of various teams across

中文译文（机器翻译，按本时间段分段）

演讲人：2023 年，当我开始的时候，

说话人：没有人说人类、土块和编码

说话人：同一句话。我想回去

说话人：到人类的开始。我记得

说话人：感觉，伙计，

说话人：这些家伙没有机会。OpenAI 是

说话人：遥遥领先。当时，

说话人：我看到人们开始使用这些

说话人：模型不仅仅用于代码自动完成，

说话人：但实际上是在写长代码。

说话人：这不是我们训练的机会

说话人：Opus 3 更擅长。那是

说话人：语调变化。我总是想起 Opus

说话人：一年后寒假期间 4.5

说话人：当大家都在家的时候，

讲话人：会编码，有什么神奇之处

演讲人：Opus 4.5 是我们现在也不只是拥有的

说话人：一个模型，但是一辆车，

说话人：像云一样很棒的产品体验

说话人：Code.Opus 4.5 就不会这样

说话人：没有像云这样的产品的时刻

说话人：Code.And Cloud Code 就没有了

演讲人：这种类型的采用加速了没有

说话人：Opus 4.5.我想谈谈如何

演讲人：产品角色正在发生变化。对于我的团队来说，

演讲人：驱动用户价值的方法是数字

说话人：输出正确的用户反馈，

说话人：the evals.我们其实有一句话

演讲人：评估团队中有新的 PRD。

演讲人：Gary Tan 一直在谈论的事情，

说话人：如果你愿意花 10 万美元

说话人：现在用代币来说已经一年了，

演讲人：2028 年你正在像某人一样生活

说话人：要活下去，你必须付出汗水

说话人：代币和你流汗的像素一样多。

讲话人：你必须使用模型才能来

演讲人：提出好的、伟大的、更好的想法。

演讲人：这是无可替代的。

讲话人：人需要更有野心

演讲人：现在的人工智能工具是因为它们

说话人：能力就这么多了。有一件事我

说话人：问团队的是，比方说，

说话人：Cloud 8 来了，有什么变化

说话人：用户做什么？

说话人：这对于你的构建方式意味着什么

说话人：今天？

演讲人：今天我的嘉宾是黛安·佩恩 (Diane Penn)，

演讲人：人工智能研究产品负责人

说话人：还有 Anthropic 的实验室团队。她加入了

演讲人：人类作为第一个技术产品

讲话人：三年多前的经理，

说话人：AI 时代就是一辈子。

演讲人：产品团队只有五位工程师，

说话人：她帮助运送了 Anthropic 的每个模型

说话人：从 CLAW2 到 Fable。她也是

演讲人：帮助孵化并推出了 CLAW Code，

演讲人：MCP,技能,CLAW 设计,

讲话人：还有计算机等核心能力

说话人：使用，工具使用，

说话人：和推理。这总是一种享受

说话人：所以要扩大思维才能说话

说话人：对处于中心位置的人

说话人：人工智能和产品管理，很难

说话人：想象一个见过更多东西的人

说话人：比头脑更重要的是事情进展的方向

说话人：Anthropix Research 的产品和

演讲人：实验室团队。在我们开始讨论之前，

说话人：别忘了看看 Lenny 的产品

说话人：Pass.com 最热免费一年

说话人：还有最精美的 AI 产品

说话人：在世界上，

说话人：Lenny's newsletter 独家提供

讲话人：订阅者。这样，

讲话人：我给你带来了黛安·佩恩。黛安，

说话人：非常感谢您来到这里。欢迎光临

说话人：到播客。谢谢，

说话人：Lenny.很高兴再次见到你。

说话人：我想回到开头

说话人：人类，

言语人：早期。我记得当人类

说话人：第一次启动。这是，

说话人：我不知道，

说话人：推出时的第一款车型

说话人：以前，三年前，

说话人：大概是这样吧，那是，三年了。

说话人：我记得就像感觉一样，

讲话人：伙计，

说话人：这些家伙没有机会。OpenAI 是

说话人：太遥遥领先了。他们就像，

说话人：他们在想什么？

说话人：这怎么可能？

演讲人：OpenAI 赢了，已经太晚了。事情

说话人：现在很不一样了。最新的数字

说话人：我看到 Anthropic 正在做这样的事情，

说话人：我不知道，

说话人：500 亿美元的 ARR。就像这样

讲话人：公司过去常在以下地点上市：Like

演讲人：非常成功的公司上市了

说话人：估值 500 亿美元。Anthropic

说话人：据说每一个

说话人：年。你是最早加入的人之一

说话人：PM 们，大概有五个工程师

说话人：你加入的时候，模特还没有

演讲人：你加入时就推出了，那是什么

言语人：就像 Anthropic 早期那样？

说话人：有什么事情可能会让你感到惊讶

说话人：人们谈论当时的情况

说话人：开始？

说话人：我认为 Anthropic 的一个重要组成部分

说话人：今天其实已经很

说话人：甚至是早期的核心。所以我加入了

说话人：2023 年，就像你说的，

演讲人：我们有五名产品工程师。

说话人：整个过程都是一名工程师

演讲人：我们的 API 业务，

说话人：如果你相信的话。而且我认为很大一部分

演讲人：因为文化真的很强大。

说话人：我想这就是我强调的

演讲人：对于那些对此感兴趣的人

说话人：公司、

说话人：真正践行使命

演讲人：还有文化和价值观。

讲话人：能源公司很像一家初创公司。

说话人：我认为你是对的。我们非常

讲话人：非常努力地寻找我们的身份

讲话人：早年。我认为有一个

演讲人：围绕技术展开，

演讲人：但是这项技术如何带来价值

演讲人：为用户、为社会带来价值、

说话人：那可能是什么？

讲话人：我认为早年是我们

说话人：用不同的方式探索这一点。比如

说话人：我们确实是从 Cloud.ai 开始的，

说话人：另一个聊天机器人，聊天助手，

讲话人：并演变成诸如工具使用之类的事情。

说话人：我认为其中一个时刻真的，

说话人：我们开始进入最佳状态是

说话人：运送像金门爪这样的东西。

说话人：我不知道你还记不记得。不记得了。

讲话人：所以这实际上是大约 24 小时

讲话人：几个小时左右。我们刚刚发表了一篇

演讲人：我们早期的可解释性研究

演讲人：2024 年初。是的。其中的例子是

说话人：基本上你可以得到所谓的

说话人：喜欢模型内的特征

说话人：层层，

言语人：表达某种类型的主题。

演讲人：所以研究人员的主题之一

说话人：能够识别出是，

说话人：比方说，

讲话人：要点写作。另一个是

说话人：人和地方。而且一个真正

说话人：经常出现引起共鸣的是

说话人：金门大桥。所以当你

讲话人：实际上基本上是拨号的

说话人：特点，

说话人：克劳德会痴迷于金球奖

说话人：Gate Bridge.每个人都有这样的意思

说话人：就其回应而言，

说话人：它会回来谈论

说话人：金门大桥。所以如果你说，

说话人：给我一份做意大利面的食谱，

说话人：它会说，

说话人：这是一个食谱。还有橙色的颜色

说话人：就像国际红那样

说话人：金门大桥的样子。

讲话人：所以这真的很奇怪。而且

演讲人：我们非常希望，在那种情况下，

讲话人：只要带上那个用户，

讲话人：把它带给群众并带来它

演讲人：对于刚开始使用云的人。

说话人：所以整个经历实际上，

演讲人：我们在 2017 年推出了 Cloud.ai 网站

演讲人：24 小时。这需要工程，

演讲人：产品、设计、

发言者：我们的研究团队都在共同努力。

说话人：我们真的，

说话人：真的很自豪。我想也许吧

演讲人：只达到了 2000 人，

说话人：说实话。但这让我们觉得，

演讲人：哦，我们其实可以带来新的用户体验，

演讲人：以这样的方式展示我们的研究

说话人：对我们来说不同且真实。

讲话人：非常有创业精神，对吧？

说话人：对我来说，

讲话人：可能是隐藏的语调变化之一

说话人：我们开始寻找我们的要点

演讲人：身份，我们可以制造产品，

演讲人：打造不同的体验

说话人：从我们的竞争对手所看到的来看，

说话人：已经有什么了。我认为

说话人：那，显然，实验室，云代码，

说话人：等等，

说话人：然后我们就开始表明自己的身份

说话人：如，

说话人：我们真的会认为世界是这样的吗？

演讲人：想想人工智能，

演讲人：如何让这一点更贴近公众。

演讲人：但这是一种自下而上的文化。

说话人：所以整个经历非常

说话人：自下而上。我看到工程师，

说话人：我看到设计师们捐出时间来工作

说话人：on.所以我喜欢一直用这个

说话人：举个例子，比如早期的

说话人：日子过得很好，但是文化和

说话人：价值观有很多，我认为，

讲话人：从早期开始就一直保持不变。

说话人：本期节目是我们的

演讲人：本季的主讲赞助商，

演讲人：WorkOS.OpenAI、Anthropic、

说话人：Cursor,Vercel,Replit,Sierra,

说话人：粘土，

演讲人：以及其他数百家获奖公司

说话人：有共同点吗？

说话人：它们都是由 WorkOS 提供支持的。如果你是

演讲人：为企业打造产品，

说话人：你已经感受到融入的痛苦

说话人：单点登录、略读、RBAC、

说话人：审计日志，

说话人：以及大客户要求的其他功能

演讲人：companies.WorkOS 扭转了那些交易障碍

演讲人：与现代开发人员一起了解嵌入式 API

演讲人：专为 B2B 打造的平台

演讲人：SaaS。实际上是我所参与的每一个初创公司

演讲人：开始扩张的投资者

演讲人：高端市场最终选择了 WorkOS。

说话人：那是因为他们是最好的。

演讲人：无论您是种子期初创公司

演讲人：尝试落地你的第一家企业

演讲人：客户或全球扩张的独角兽，

演讲人：WorkOS 是成为的最快途径

演讲人：企业就绪，畅通增长。

讲话人：本质上是企业用的 Stripe

说话人：features.访问 WorkOS.com 开始使用

说话人：或者只是打开他们的 Slack

说话人：有实际工程师等待回答

发言者：您的问题。WorkOS 允许您

讲话人：使用令人愉快的 API 更快地构建，

演讲人：综合文档，

说话人：以及流畅的开发者体验。Go

演讲人：到 WorkOS.com 让您的应用程序成为企业

说话人：今天准备好了，还有哪些其他的

演讲人：如你所想的大拐点

演讲人：关于 just Anthropic 从 just

说话人：这就像试图竞争的实验室

说话人：有了开放人工智能的霸主

演讲人：指出今天的情况，

说话人：有哪些印象深刻的时刻

说话人：就像，哇，这真的改变了事情吗？

说话人：肯定是我们训练的时候

说话人：测试 Opus 3，

说话人：我想那一刻

说话人：公司、

演讲人：我想我们不到 200 人

讲话人：还是那个时候。而且非常

发言者：明确我们需要并且想要创建

演讲人：前沿模型。这非常重要

说话人：就我们达到的能力而言

说话人：比如用户、消费者、

演讲人：并展示我们的研究。而且我们

讲话人：也在寻找方法，

演讲人：为什么有人要选择克劳德？

演讲人：这就像一个核心问题。

演讲人：这是我们收到的一个核心问题

说话人：早期就问过，而且我认为

说话人：用 Opus 3，

说话人：我想是三月初推出的，

说话人：2024 年，但是有很多，

发言者：多个团队跨越多个月

00:10:00

英文原文（完整 ASR）

讲话人：inference,across research,

讲话人：fine tuning,

讲话人：pre-training that rallied at different

讲话人：points and towards a common goal.And

讲话人：I think everybody that was involved

讲话人：was like really proud.I remember being

讲话人：the PM,us,the research leads,

讲话人：myself,we were all in our,

讲话人：this was around December.So we were

讲话人：all at home in our various parents'

讲话人：homes and seeing everybody's background

讲话人：of like their childhood room.And everybody

讲话人：was working really hard.to figure out

讲话人：the,like,

讲话人：what are we training the model for?

讲话人：Is it showing up the right way?

讲话人：So I think that was really powerful

讲话人：in terms of just building a lot of trust.

讲话人：And a lot of our research leads have

讲话人：actually,from that time,

讲话人：are now,like,leading reinforcement learning,

讲话人：leading our character work,

讲话人：alignment work.So that foundational

讲话人：trust,I think,

讲话人：also helped us work well.now with any

讲话人：of our production models across product

讲话人：and research,

讲话人：because we were working just so much

讲话人：in the trenches together in the early

讲话人：days,

讲话人：And then I think there were things like

讲话人：identifying that coding was important,

讲话人：right?

讲话人：In 2023,when I started,

讲话人：nobody said anthropic and clod and coding

讲话人：in the same sentence.I think competitor

讲话人：models like GPT-4 at the time was used

讲话人：a bit for coding,

讲话人：but it was one of many use cases.And

讲话人：one thing that,for example,

讲话人：I saw was...People are starting to use

讲话人：code,these models, not just for code,

讲话人：not just like code autocomplete,

讲话人：but actually writing long form code.

讲话人：And it's not an opportunity for us to

讲话人：train Opus 3 to be better at.And it

讲话人：ended up being a relatively smaller

讲话人：change from a training perspective,

讲话人：but it ended up helping us differentiate

讲话人：in the early days competitively for

讲话人：users.and actually bring a lot of the

讲话人：very early cloud enthusiasts and developers

讲话人：because we were providing a value that

讲话人：they didn't really think was possible

讲话人：at the time.It's so interesting you

讲话人：talk about Opus 3,

讲话人：like that's so long ago and just like,

讲话人：it's hard to think that was a big inflection.

讲话人：And so this is really interesting to

讲话人：hear that that was internally a big

讲话人：milestone.It almost feels like this

讲话人：confidence you all built that,

讲话人：wow,

讲话人：we could really ship a frontier model,

讲话人：which is now today.So not great.If you

讲话人：compare it to what we've got today,

讲话人：what I always think about is Opus 4

讲话人：.5,which was an interestingly,

讲话人：like a year later,

讲话人：also during winter break,

讲话人：when everyone was home able to code,

讲话人：uh,was that another big milestone?

讲话人：Yeah,

讲话人：Opus 4.5 was definitely another large

讲话人：moment.I think what was magical about

讲话人：Opus 4.5 is we also now not just had

讲话人：a model,but a vehicle,

讲话人：which is like a great product experience

讲话人：like Cloud Code.One thing we say a lot

讲话人：on the team is you need frontier products

讲话人：in order to have frontier models and

讲话人：for people to feel the magic.of frontier

讲话人：models.And I think,you know,

讲话人：we felt the magic of cloud code for

讲话人：many months before that.But the fact

讲话人：that the model essentially got to a

讲话人：level of intelligence where at a very

讲话人：broad level,

讲话人：users can experience both frontier intelligence

讲话人：in new use cases,

讲话人：allow it to run things end to end in

讲话人：an agentic manner.I think that was the

讲话人：inflection.It was actually both.I think

讲话人：Opus 4.5 wouldn't have had that moment

讲话人：without a product like Cloud Code.And

讲话人：Cloud Code,I think,

讲话人：wouldn't have had that type of adoption

讲话人：accelerated without Opus 4.5.So kind

讲话人：of speaking on this thread,

讲话人：Dario,interestingly,

讲话人：if you look back at all his predictions,

讲话人：he's just like,okay,

讲话人：coding is going to be solved.There's

讲话人：100%in like a year or something like

讲话人：that.He kept talking about how we're

讲话人：going to do,

讲话人：like AI is going to do all our code.

讲话人：And I remember everyone being like,

讲话人：there's no way this is way too complicated.

讲话人：How is AI ever going to get really good

讲话人：at this very complex thing that humans

讲话人：do?

讲话人：No,

讲话人：this is going to be humans for a long

讲话人：time.He was completely right.Something

讲话人：else that he talks a lot about is this

讲话人：exponential that we're now,

讲话人：now that we're on,

讲话人：that's the way he describes it.Now we're

讲话人：like,

讲话人：we're on the exponential curve.I remember

讲话人：not long ago,

讲话人：we were new models were being released

讲话人：and everybody was like,okay,

讲话人：we're done.There's no more upside.It's

讲话人：plateauing.It's over.There's no more

讲话人：room to grow.And now it's like the opposite.

讲话人：Now we're inside.Like if you think about

讲话人：the curve of the exponential,

讲话人：we're like inside of the exponential

讲话人：now,

讲话人：which by definition means every improvement

讲话人：is a massive jump because we're like

讲话人：on that hockey stick part.What's it

讲话人：like just being on the inside of this

讲话人：crazy historic moment when AI is improving

讲话人：so fast,so much is being unlocked?

讲话人：What is it like and how should people

讲话人：prepare for the coming season acceleration

讲话人：of more and more improvement for me.

讲话人：I,

讲话人：one thing I like to say on the team is

讲话人：most of us weren't like actively working

讲话人：yet when the internet transition from

讲话人：this novelty to something that everyone

讲话人：can use.And it feels like that's just

讲话人：taking humans.I think analogies are

讲话人：helpful.And so like the analogy of that

讲话人：is,I think a couple of things.Um,

讲话人：number one is adaptability.Thank you.

讲话人：TAG and others.But it's very hard to

讲话人：predict the exact moment or the exact

讲话人：model.And so the adaptability of when

讲话人：you're faced with new information,

讲话人：how do you then make better decisions

讲话人：versus keeping the same plan?

讲话人：And so like that agility is really important.

讲话人：I think another piece is with that,

讲话人：how do you actually be thinking very

讲话人：first principles and reason through

讲话人：what's next?

讲话人：What's the so what?

讲话人：How do we invest in new products?

讲话人：How do we invest in explaining the differences

讲话人：to users?

讲话人：So a lot of the experiences,

讲话人：I think,of being in that exponentialist,

讲话人：that pace,

讲话人：understanding how you operate and make

讲话人：better decisions and then applying that

讲话人：first principles thinking to then–do

讲话人：something that maybe we pull up a plan

讲话人：that we would,

讲话人：we're expecting a few months from now,

讲话人：but now the model can actually do and

讲话人：work on and actually bring that to users.

讲话人：So this is things like cowork skills,

讲话人：tag,you know,

讲话人：it's a very positive self-reinforcing

讲话人：loop.And I think a big part of it also

讲话人：is just having the like trust in each

讲话人：other,like making sure we have like,

讲话人：We're thinking through the right decision

讲话人：making.We're bringing folks along.Some

讲话人：teams might see the exponential,

讲话人：feel it faster than others.So how do

讲话人：we kind of have the grace to bring the

讲话人：organization,

讲话人：the growing organization and company

讲话人：along on that?

讲话人：So what I'm hearing here is you almost

讲话人：don't know what will be possible with

讲话人：every model release.And so the important

讲话人：things to focus on is being adaptable

讲话人：as things emerge.To your point,

讲话人：the product itself has to stay up to

讲话人：has to catch up to what is possible.

讲话人：To your point again,

讲话人：just like it can do so much,

讲话人：but people may not understand how to

讲话人：do it and may not be able to do it.

讲话人：So the product making it easy and even

讲话人：just like telling you here's something

讲话人：you could do feels like an important

讲话人：part.Is that roughly what you're describing?

讲话人：I think so.I think there's some really

讲话人：interesting graphs in the original scaling

讲话人：law papers.And I think folks are very

讲话人：familiar with the scaling laws in the

讲话人：lens of as you add in more compute and

讲话人：data,

讲话人：what's called loss.AKA the loss from

讲话人：next token prediction goes down.And

讲话人：so it's a very smooth linear curve of

讲话人：like the models get more intelligent

讲话人：as you scale them up.What's actually

讲话人：also interesting in that paper is there

讲话人：are these like very different emerging

讲话人：capability graphs.And so,

讲话人：for example,

讲话人：as you add in more data and you train

讲话人：the models with more compute,

讲话人：you essentially see these actually discontinuous

讲话人：emergent capabilities jump.So the models

讲话人：go from one plus one being a thing that

讲话人：it can't calculate to a thing that it

讲话人：can reliably calculate.And so these

讲话人：emerging capabilities,

讲话人：this like some nature of like predictability

讲话人：is not necessarily everyone knows the

讲话人：exact moment.Like you need the evals

讲话人：to be able to assess that has actually

讲话人：always been a part of how this technology

讲话人：works.And also what makes things like

讲话人：safety harder,

讲话人：because unless you have the evals,

讲话人：unless you have the systems to test,

讲话人：these jumps might actually happen and

讲话人：you don't know.Hmm.That's so interesting

讲话人：that you may have developed this like

讲话人：AI brain that can do something you're

讲话人：not even aware of.And so part of the

讲话人：job is just uncovering.Wow,

讲话人：it just got really good at this thing.

讲话人：What can we do with that?

讲话人：I think there's like product overhang

讲话人：and user overhang,

讲话人：like to maybe put it in our PM language,

讲话人：even on today's models.And I think there's

讲话人：like a lot that we could be exploring

讲话人：on like our current opuses and definitely

讲话人：with like Fable,

讲话人：for example.And that discovery is actually

讲话人：another part of what's been in the early

讲话人：days of Anthropix DNA.And I think it's

讲话人：also continuing to be a big part of

讲话人：how we operate in product,

中文译文（机器翻译，按本时间段分段）

说话人：推理，跨研究，

说话人：微调，

演讲人：在不同的场合集会的预训练

说话人：指向一个共同的目标。并且

发言者：我想所有参与其中的人

说话人：真的很自豪。我记得当时

讲话人：总理、我们、研究负责人、

说话人：我自己，我们都在我们的，

讲话人：那是在十二月左右。所以我们是

讲话人：在我们不同的父母家里都感到很自在

讲话人：回家看看每个人的背景

说话人：就像他们童年的房间。还有大家

讲话人：我真的很努力地想弄清楚

说话人：那个，喜欢，

演讲人：我们训练模型的目的是什么？

说话人：它的显示方式正确吗？

说话人：所以我认为这真的很强大

讲话人：就建立大量信任而言。

演讲人：我们的很多研究负责人都

说话人：其实，从那个时候开始，

说话人：现在，比如，领导强化学习，

讲话人：领导我们的品格工作，

演讲人：协调工作。这样才能奠定基础

说话人：信任，我认为，

讲话人：也帮助我们很好地工作。现在与任何

演讲人：我们跨产品的生产模式

演讲人：和研究，

说话人：因为我们工作太多了

讲话人：早期一起在战壕里

说话人：天，

演讲人：然后我认为有类似的事情

讲话人：认识到编码很重要，

说话人：对吗？

演讲人：2023 年，当我开始的时候，

说话人：没有人说人类、土块和编码

讲话人：同一句话。我认为竞争对手

说话人：当时用的是 GPT-4 这样的模型

讲话人：一点编码，

说话人：但这只是众多用例之一。而且

说话人：有一件事，例如，

说话人：我看到...人们开始使用

讲话人：代码，这些模型，不只是为了代码，

说话人：不只是像代码自动完成，

讲话人：但实际上是在写长格式的代码。

讲话人：这不是我们的机会

说话人：训练 Opus 3 变得更好。而且它

讲话人：最终变得相对较小

讲话人：从培训角度改变，

说话人：但它最终帮助我们区分

演讲人：早期竞争激烈

讲话人：用户。实际上带来了很多

说话人：非常早期的云爱好者和开发者

演讲人：因为我们提供的价值是

说话人：他们真的认为不可能

说话人：当时，你真有趣

说话人：谈谈 Opus 3，

说话人：就像是很久以前的事了，就像，

讲话人：很难认为这是一个很大的变化。

说话人：所以这真的很有趣

讲话人：听说那是内部的一个大事件

说话人：里程碑。差不多就是这样的感觉

讲话人：相信你们都做到了，

说话人：哇，

演讲人：我们真的可以推出前沿模型，

说话人：就是现在。所以不太好。如果你

演讲人：与我们今天的情况相比，

说话人：我一直想到的是 Opus 4

说话人：.5，这是一个有趣的，

说话人：就像一年后，

说话人：也是在寒假期间，

讲话人：当每个人都在家可以编码时，

演讲人：呃，这是另一个重大里程碑吗？

说话人：是的，

演讲人：Opus 4.5 绝对是又一个大型作品

说话人：那一刻，我觉得有什么神奇之处

演讲人：Opus 4.5 是我们现在也不只是拥有的

说话人：一个模型，但是一辆车，

演讲人：这就像一次很棒的产品体验

说话人：就像云代码一样，我们经常说的一件事

演讲人：团队里你需要的是前沿产品

演讲人：为了有前沿的模型和

说话人：让人们感受边疆的神奇

说话人：模特。我认为，你知道，

演讲人：我们感受到了云代码的魔力

讲话人：在那之前好几个月了。但事实是

演讲人：该模型本质上达到了

说话人：智力水平非常高

讲话人：宽泛的层次，

演讲人：用户可以同时体验前沿智能

演讲人：在新的用例中，

讲话人：允许它从头到尾地运行事情

说话人：一种代理人的态度。我认为这就是

讲话人：语调变化。实际上两者都是。我想

说话人：Opus 4.5 不会有那个时刻

说话人：没有像云码这样的产品。而且

说话人：云码，我想，

说话人：不会有这样的收养

说话人：没有 Opus 4.5 就加速了。非常好

演讲人：在这个话题上发言时，

说话人：达里奥，有趣的是，

说话人：如果你回顾一下他所有的预测，

说话人：他就像，好吧，

说话人：编码要解决了。有

说话人：100%在一年之内

说话人：那个。他一直在谈论我们怎么样

说话人：要做，

讲话人：就像人工智能将完成我们所有的代码一样。

说话人：我记得每个人都这样，

说话人：这不可能太复杂。

演讲人：人工智能怎样才能变得真正优秀

说话人：在人类这个非常复杂的事情上

说话人：做什么？

说话人：不，

说话人：这将是人类很长一段时间

说话人：时间。他完全正确。有事

讲话人：他常说的其他就是这个

说话人：我们现在呈指数级增长，

说话人：现在我们开始了，

说话人：他就是这么描述的。现在我们

说话人：比如，

说话人：我们正处于指数曲线上。我记得

说话人：不久前，

讲话人：我们正在发布新型号

说话人：每个人都说，好吧，

说话人：我们已经完成了。没有更多的好处了。

说话人：稳定。结束了。没有了

说话人：成长空间。而现在却恰恰相反。

说话人：现在我们在里面。就像你想想

说话人：指数曲线，

讲话人：我们就像在指数内部

说话人：现在，

说话人：顾名思义就是每一次进步

说话人：这是一个巨大的跳跃，因为我们就像

说话人：在曲棍球杆部分。那是什么

说话人：就像在里面一样

演讲人：人工智能进步的疯狂历史性时刻

说话人：这么快，解锁了这么多？

说话人：这是什么样的，人们应该怎样做

讲话人：为即将到来的赛季加速做好准备

说话人：对我来说越来越有进步。

说话人：我，

说话人：我在团队里喜欢说的一件事是

说话人：我们大多数人都不喜欢积极工作

演讲人：然而当互联网从

说话人：这个新奇的东西大家都知道

说话人：可以用，而且感觉就这样

讲话人：以人类为例。我认为类比是

说话人：很有帮助。就像这样的类比

说话人：我认为有几件事。嗯，

说话人：第一是适应性，谢谢。

说话人：TAG 等人，但是很难

讲话人：预测确切的时刻或确切的时间

讲话人：模型。等等时的适应性

说话人：你面临着新的信息，

说话人：那么你如何做出更好的决定

说话人：还是保持同样的计划？

说话人：所以敏捷性真的很重要。

说话人：我认为另一件事是这样的，

说话人：你实际上是如何思考的？

说话人：首要原则和推理

说话人：接下来怎么办？

说话人：那又怎样？

演讲人：我们如何投资新产品？

演讲人：我们如何投资解释差异

说话人：对用户？

演讲人：很多经历，

演讲人：我认为，处于指数论者之中，

说话人：那个步伐，

说话人：了解你如何操作和制作

演讲人：更好的决定然后应用

说话人：先思考然后做的原则

说话人：也许我们可以制定一个计划

说话人：我们会，

说话人：我们预计几个月后，

演讲人：但是现在模型实际上可以做到并且

演讲人：致力于并将其实际带给用户。

演讲人：这就是协作技能之类的事情，

说话人：标签，你知道，

说话人：这是一种非常积极的自我强化

说话人：循环。我认为其中很大一部分也是

说话人：就是对彼此有同样的信任

说话人：其他，比如确保我们有类似的，

演讲人：我们正在考虑正确的决定

讲话人：制作。我们带着人们一起。一些

演讲人：团队可能会看到指数增长，

说话人：感觉比别人快，那怎么办

讲话人：我们有幸带来

演讲人：组织，

演讲人：不断成长的组织和公司

说话人：就这样吗？

说话人：所以我在这里听到的几乎就是你

说话人：不知道会发生什么

讲话人：每个模型发布。所以重要的

说话人：要注重的是随机应变

说话人：随着事情的发展。就你的观点而言，

演讲人：产品本身要跟上

言语人：必须赶上可能的事情。

发言者：再次强调你的观点，

说话人：就像它能做这么多一样，

讲话人：但人们可能不明白如何

讲话人：做到了也不一定能做到。

说话人：所以这个产品让它变得简单甚至均匀

说话人：就像告诉你这里有件事

说话人：你可以做的事情感觉很重要

说话人：部分，你所描述的大致是这样吗？

说话人：我想是的。我认为确实有一些

讲话人：原始缩放中的有趣图表

说话人：法律论文。我认为人们非常

说话人：熟悉尺度法则

说话人：当你添加更多的计算和

说话人：数据，

说话人：什么叫做损失。AKA the loss from

讲话人：下一个令牌预测下降。并且

说话人：所以这是一条非常平滑的线性曲线

说话人：就像模型变得更聪明一样

说话人：当你放大它们时，实际上是什么

说话人：那篇论文也很有趣

说话人：这些是非常不同的新兴吗？

演讲人：能力图。所以，

讲话人：例如，

说话人：当你添加更多数据并进行训练时

演讲人：具有更多计算能力的模型，

说话人：你本质上看到这些实际上是不连续的

说话人：紧急能力跳跃。所以模型

讲话人：从一加一开始

说话人：它无法计算出它所知道的事情

讲话人：可以可靠地计算。所以这些

演讲人：新兴能力，

说话人：这就像可预测性的某种本质

讲话人：不一定每个人都知道

说话人：正是时候。就像你需要评估一样

说话人：能够评估实际上已经

演讲人：一直是这项技术的一部分

讲话人：有效。还有什么让事情变得像这样

说话人：安全更难，

说话人：因为除非你有评估，

说话人：除非你有系统可以测试，

说话人：这些跳跃实际上可能发生并且

说话人：你不知道。嗯。这很有趣

说话人：你可能已经这样发展了

说话人：人工智能大脑可以做你想做的事情

讲话人：甚至没有意识到。所以部分

说话人：工作只是揭露。哇，

说话人：它在这件事上变得非常擅长。

说话人：我们能用它做什么呢？

说话人：我认为产品存在悬而未决的情况

说话人：和用户悬而未决，

说话人：也许可以用我们的 PM 语言来表达，

说话人：即使在今天的模型上。我认为有

说话人：就像我们可以探索的很多东西一样

说话人：就像我们目前的做法一样，而且绝对如此

说话人：像寓言一样，

说话人：举个例子。那个发现实际上是

说话人：早期的另一部分内容

说话人：Anthropix DNA 的日子。我认为这是

演讲人：也继续成为重要组成部分

演讲人：我们如何运作产品，

00:20:00

英文原文（完整 ASR）

讲话人：in labs,

讲话人：and across research.This makes me think

讲话人：about something Gary Tan's been talking

讲话人：about,

讲话人：president of YC.I don't know what his

讲话人：title is.He had this interesting point

讲话人：that if you're willing to spend$100

讲话人：,000 a year right now on tokens,

讲话人：you are living the way somebody in 2028

讲话人：is going to live.Because by then it'll

讲话人：be really cheap.Everyone can work this

讲话人：way.But if there's this alpha opportunity

讲话人：right now to just live in the future,

讲话人：go crazy on token spend.And so there's

讲话人：a big opportunity for people to learn

讲话人：what the future is like and also just

讲话人：build much faster.Thoughts on this idea

讲话人：and the value of token maxing,

讲话人：let's call it.Yeah,

讲话人：I think I take more of like a almost

讲话人：product lens.It's almost like token

讲话人：spin is more the input.And really the

讲话人：output is what you described of experimentation.

讲话人：And I think if we were orienting like

讲话人：goals around experimentation,

讲话人：I feel like that is might be the better

讲话人：framing of the outcomes.And therefore,

讲话人：there might be different ways of achieving

讲话人：that outcome.I will say internally,

讲话人：some of the most creative thinkers,

讲话人：the best like prototypers,

讲话人：do spend a lot of time with Claude,

讲话人：with every new version of a research

讲话人：model that we have.And so there is something

讲话人：around,

讲话人：you have to be like using the models

讲话人：to then come up with good,

讲话人：then great,

讲话人：then better ideas.And there's no substitute

讲话人：for that.It's very hard to come up with

讲话人：a perfect strategy without touching

讲话人：the technology when it's moving this

讲话人：quickly.At the same time,

讲话人：I think there's other things that we

讲话人：could be doing.Like,

讲话人：so one thing that we do a lot is actually

讲话人：working in public and internally within

讲话人：Anthropic.And so in the early days when

讲话人：we had less product surfaces,

讲话人：There was a Slack channel where everyone,

讲话人：almost the entire company,

讲话人：was testing early versions of Claude

讲话人：and trying different use cases.Like,

讲话人：people were not calling them use cases,

讲话人：but you might be asking it to edit an

讲话人：essay or to come up with the right way

讲话人：to send this email.Like,

讲话人：they were all different use cases.We

讲话人：all worked in public and the way you

讲话人：would see magically is different users

讲话人：or different different folks on the

讲话人：team coming up with an idea and then

讲话人：other people trying different variations

讲话人：of that idea.And then within maybe 10

讲话人：or so requests,

讲话人：there was something magical or potentially

讲话人：in a use case that emerges.And I think

讲话人：there's a lot in not just individuals

讲话人：figuring out by themselves how to use

讲话人：this technology.I think we could be

讲话人：doing more to actually bring like that

讲话人：communal discovery when we do experimentation.

讲话人：Like experimentation is not always necessarily

讲话人：an individual sport.It's so interesting.

讲话人：Yeah,this idea that we're just,

讲话人：we're not sure what this is capable

讲话人：of or what we could do with it.And it

讲话人：takes all this poking around and people

讲话人：trying things,

讲话人：hearing what other people are trying

讲话人：to figure out what's possible.Such an

讲话人：interesting,I don't know,

讲话人：technology.We're just like,

讲话人：okay,here's what,oh,

讲话人：I figured out I could do this thing.

讲话人：What are you going to do with that?

讲话人：I think at a broad theme,

讲话人：we know,right?

讲话人：We know that the models write great

讲话人：essays or can write long form writing.

讲话人：But individual pain points of what can

讲话人：you actually solve with that and bring

讲话人：it to like a user level that people

讲话人：can use,I think,

讲话人：is something that is more exploration

讲话人：or experimentation based.Following this

讲话人：thread,

讲话人：you oversee product for the labs team,

讲话人：right?

讲话人：extremely cool.We've had Ben Mann on

讲话人：the podcast,Mike Rieger,

讲话人：whom both work on labs now.Talk about

讲话人：labs.What is labs?

讲话人：What's come out of labs?

讲话人：Many people have heard of these things.

讲话人：And how do they work that enables them

讲话人：to create such innovative ideas outside

讲话人：of even the core product team?

讲话人：The thesis of labs in many ways is identifying

讲话人：and pulling the thread on the thread

讲话人：of discontinuous large bets that might

讲话人：not be in the core roadmap.And figuring

讲话人：out is there a there there?

讲话人：And also,what is the 10x,

讲话人：100x,1000x of the there there?

讲话人：And so,for example,

讲话人：things like clog code,

讲话人：I think.I've heard of it.Things like

讲话人：cloud code,things like skills,

讲话人：and most recently,cloud design,

讲话人：MCP.The thing that we really try to

讲话人：emphasize within the teams is,

讲话人：especially right now,

讲话人：there are so many things that could

讲话人：be built.What does it mean then to have

讲话人：a discontinuous SPED?

讲话人：I think one approach that we're taking

讲话人：this year is it can be very strongly

讲话人：held opinion about the theme or the

讲话人：area.and then more weakly held about

讲话人：the exact prototype.And so there is

讲话人：a culture of experimentation.There's

讲话人：a lot of the bottoms up,

讲话人：like engineers on the team are very

讲话人：self-enabled,

讲话人：self-driven to test out different ideas.

讲话人：And sometimes we have a thesis and it

讲话人：might not work yet.And so we then might

讲话人：revisit it in one to two model generations.

讲话人：And so this idea of like these prototypes

讲话人：that actually end up just helping us

讲话人：learn,like that's also valuable,

讲话人：even if it doesn't lead to something

讲话人：immediately shipping.And so I think

讲话人：that allows the incubation and like

讲话人：the charter of labs to really accelerate

讲话人：and see around corners more broadly

讲话人：for anthropic.It's so funny to think

讲话人：about a labs within an Anthropic,

讲话人：which was already so innovative and

讲话人：creative and just,you know,

讲话人：shipping like crazy,

讲话人：that there's value to still creating

讲话人：a labs team within Anthropic.What enables

讲话人：labs to work as well as it has?

讲话人：Because you listed all these products

讲话人：and it's like,

讲话人：what else has Anthropic shipped?

讲话人：It feels like all the biggest wins almost.

讲话人：I'm sure there are many that I'm not

讲话人：thinking about right now.What's kind

讲话人：of core to creating a successful labs

讲话人：org within a larger company?

讲话人：I think the team culture,

讲话人：like similar to broadly at Anthropic,

讲话人：I think the team culture is very valuable.

讲话人：I think Ben sets an incredible vision

讲话人：and pushes people to think about the

讲话人：10x,100x of the idea.And,

讲话人：you know,

讲话人：the pods within labs is small.Sometimes

讲话人：these ideas start with one engineer,

讲话人：right?

讲话人：And I think,

讲话人：Sometimes when there's almost really

讲话人：large teams pursuing very ambiguous,

讲话人：large ideas,

讲话人：you end up actually being slowed down

讲话人：because of that.So I think it's culture.

讲话人：I think we actually also select for...

讲话人：folks who actually want to do that zero

讲话人：to one experimentation and it's not

讲话人：easy there's a lot of bets that we end

讲话人：up turning down or turning off um and

讲话人：maybe you know we revisit them in the

讲话人：future uh but that's hard that's hard

讲话人：when you pour your heart and soul you're

讲话人：acting as a founder for a bet and it's

讲话人：not working yet um so i think it's like

讲话人：that type of selecting for that type

讲话人：of personality,

讲话人：folks who are really passionate and

讲话人：deep about the zero to one.So you lead

讲话人：product for the research team.You work

讲话人：with the researchers at Anthropic.A

讲话人：lot of people kind of get a sense of

讲话人：what is research,

讲话人：what researchers do.I think a lot of

讲话人：people don't totally understand these

讲话人：very valuable people at all the AI labs.

讲话人：The way I think about it,

讲话人：and I want to help people understand,

讲话人：help me understand just what are the

讲话人：researchers doing all day?

讲话人：What I imagine is they have a hypothesis

讲话人：for how to improve the model.They find

讲话人：data.They tweak some algorithms.They

讲话人：adjust how it's trained.And they test

讲话人：it,see how it did,keep iterating,

讲话人：and keep trying to find ways to improve

讲话人：the model.Is that roughly right?

讲话人：Slash,

讲话人：help us understand what researchers are

讲话人：doing all day.That's really,

讲话人：I think that's a lot of maybe the more

讲话人：day-to-day.I think one piece around

讲话人：researchers and research organizations

讲话人：like Anthropic is there's also a vision

讲话人：of the future more broadly.So,

讲话人：for example,

讲话人：things like...I think even at the founding

讲话人：of the company,

讲话人：researchers were talking about how do

讲话人：we get Claude to use a computer?

讲话人：How do we get AI to like navigate a

讲话人：screen,right?

讲话人：So there's a lot of actually very founder

讲话人：-like energy is how I describe it within

讲话人：researchers are really bold and ambitious

讲话人：researchers.And we have a ton of those

讲话人：at Anthropic.So there's one layer of

讲话人：vision of what this technology can go.

讲话人：And then I think on this other side

讲话人：of the loop,

讲话人：there's also now that this technology

讲话人：or cloud is in people's hands,

讲话人：how do we make it better today?

讲话人：So it's a medium and long-term and a

讲话人：lot of energy thinking about that lens

讲话人：of the future.And also in the immediate

讲话人：and short term,

讲话人：what are the improvement areas we can

讲话人：make?

讲话人：And so like,

讲话人：I think you're describing a really good

讲话人：sense of how do we make iterative improvements

讲话人：on different versions of Cloud.The way

讲话人：that like my team works with researchers

讲话人：is kind of being very integrated and

讲话人：embedded in those loops,

讲话人：particularly areas where there's a lot

讲话人：of impact on users.So this is things

讲话人：like vision,computer use.Coding,

讲话人：agentic coding,tool use,

讲话人：test time compute,

讲话人：things where there's a direct user impact.

讲话人：And then figuring out what are the ways

讲话人：to bring the user feedback and ground

讲话人：it in a level that is understandable

中文译文（机器翻译，按本时间段分段）

说话人：在实验室里，

演讲人：跨研究。这让我思考

说话人：关于 Gary Tan 一直在谈论的事情

说话人：关于，

说话人：YC 总裁，我不知道他的是什么

说话人：标题是。他有一个有趣的观点

说话人：如果你愿意花 100 美元

说话人：，现在一年 000 美元，

演讲人：2028 年你正在像某人一样生活

说话人：会活下去，因为到那时它就会

讲话人：真的很便宜。每个人都可以做到这一点

说话人：好吧。但是如果有这个阿尔法机会

说话人：现在只为活在未来，

说话人：对代币花费疯狂。所以有

讲话人：人们学习的大机会

说话人：未来是什么样子，也只是

说话人：构建得更快。对这个想法的思考

说话人：以及 token maxing 的价值，

说话人：就这么称呼吧。是啊，

说话人：我想我更像是一个几乎

说话人：产品镜头，几乎就像令牌一样

说话人：旋转更多的是输入。而且真的

说话人：输出就是你所描述的实验结果。

说话人：我想如果我们像

演讲人：围绕实验的目标，

说话人：我觉得这样可能会更好

演讲人：结果的框架。因此，

说话人：可能有不同的实现方式

演讲人：那个结果。我会在内心说，

演讲人：一些最有创造力的思想家，

说话人：最好的原型师，

说话人：确实花了很多时间和克劳德在一起，

演讲人：每一个新版本的研究

说话人：我们有模型。所以有一些东西

说话人：周围，

讲话人：你必须喜欢使用模型

说话人：然后想出好的，

说话人：那就太好了，

演讲人：那就有更好的想法。而且没有替代品

说话人：对于这个，很难想出

演讲人：不碰触的完美策略

说话人：移动这个的技术

说话人：快点。同时，

说话人：我认为我们还有其他事情

说话人：可能正在做。比如，

说话人：所以我们经常做的一件事实际上是

讲话人：在公共场合和内部工作

言语人：人类。所以在早期的时候

演讲人：我们的产品表面较少，

说话人：有一个 Slack 频道，每个人，

发言者：几乎整个公司，

讲话人：正在测试 Claude 的早期版本

说话人：并尝试不同的用例。比如，

演讲人：人们并没有称它们为用例，

说话人：但是你可能会要求它编辑一个

说话人：论文或想出正确的方法

说话人：发送这封电子邮件。比如，

说话人：他们都是不同的用例。我们

说话人：所有在公共场合工作的人以及你的方式

讲话人：会神奇地看到不同的用户

说话人：或者不同的人

演讲人：团队提出一个想法，然后

说话人：其他人尝试不同的变化

演讲人：这个想法。然后大概 10 天内

说话人：或者这样的要求，

讲话人：有一些神奇的或潜在的东西

说话人：在出现的用例中。我认为

演讲人：不只是个人，还有很多事情

说话人：自己弄清楚如何使用

说话人：这项技术。我想我们可以

说话人：做更多事情才能真正做到这一点

演讲人：我们做实验时的共同发现。

说话人：就像实验并不总是一定的

说话人：个人运动，很有趣。

说话人：是的，我们只是这个想法，

说话人：我们不确定这能做什么

说话人：我们可以用它做什么。还有它

说话人：接受所有这些打探和人们

说话人：尝试事物，

说话人：听别人在尝试什么

讲话人：弄清楚什么是可能的。

说话人：有趣，我不知道，

说话人：技术。我们就像，

说话人：好的，这就是，哦，

说话人：我发现我可以做这件事。

说话人：你打算用它做什么？

演讲人：我认为从一个广泛的主题来看，

说话人：我们知道，对吧？

说话人：我们知道模特写得很棒

言语人：散文或可以写长篇文章。

演讲人：但是个别痛点可以

说话人：你实际上解决了这个问题并带来了

说话人：喜欢那个人的用户级别

说话人：可以用，我想，

说话人：是更多探索的东西

说话人：或者基于实验。如下

说话人：线程，

演讲人：你负责监督实验室团队的产品，

说话人：对吗？

讲话人：非常酷。我们邀请了本·曼 (Ben Mann)

说话人：播客，Mike Rieger，

说话人：他们现在都在实验室工作。谈谈

演讲人：labs.什么是 labs？

演讲人：实验室出来了什么？

说话人：很多人都听说过这些事情。

说话人：他们是如何工作的？

说话人：在外面创造出这样的创新想法

演讲人：甚至是核心产品团队？

演讲人：实验室的论文在很多方面都在识别

说话人：然后把线拉到线上

说话人：不连续的大赌注

讲话人：不在核心路线图中。并且计算

讲话人：那里有那里吗？

演讲人：还有，10x 是多少？

说话人：100x、1000x 的还有吗？

演讲人：那么，例如，

讲话人：诸如木屐代码之类的东西，

说话人：我想，我听说过，比如

说话人：云代码，技能之类的东西，

说话人：还有最近的云设计，

说话人：MCP.我们真正尝试做的事情

演讲人：团队内部强调的是，

说话人：尤其是现在，

说话人：有很多事情可以

讲话人：bebuilt.What does itmeans then to have

说话人：不连续的 SPED？

演讲人：我认为我们正在采取的一种方法

说话人：今年可以说是非常强烈了

演讲人：对主题或议题持有意见

言语人：区域。然后更弱地举行

讲话人：确切的原型。所以有

说话人：一种实验文化。有

说话人：很多自下而上的，

演讲人：团队里的工程师都非常

说话人：自我启用，

演讲人：自我驱动，尝试不同的想法。

说话人：有时我们有一篇论文，它

说话人：可能还不行。所以我们可能

说话人：在一到两代模型中重新审视它。

说话人：所以这个想法就像这些原型

说话人：这实际上最终只是帮助我们

说话人：学习，这样也有价值，

说话人：即使没有什么结果

讲话人：立即发货。所以我认为

说话人：允许孵化之类的

演讲人：实验室章程真正加速

说话人：更广泛地观察周围的角落

说话人：for anthropic.想想真有趣

演讲人：关于人类内部的实验室，

演讲人：这已经非常有创新性了

说话人：富有创意且公正，你知道，

演讲人：疯狂运送，

讲话人：仍然创造是有价值的

说话人：Anthropic 内的一个实验室团队。什么使

演讲人：实验室工作还有吗？

说话人：因为你列出了所有这些产品

说话人：就像，

说话人：Anthropic 还发货了什么？

说话人：感觉几乎都是最大的胜利。

说话人：我确信还有很多我不是的

说话人：现在想一下，怎么样

演讲人：创建成功实验室的核心

演讲人：大公司内部的组织？

演讲人：我认为团队文化，

说话人：与 Anthropic 的大致相似，

演讲人：我认为团队文化非常有价值。

说话人：我认为本设定了一个令人难以置信的愿景

演讲人：并促使人们思考

说话人：10x,100x 的想法。而且，

说话人：你知道，

说话人：实验室里的豆荚很小。有时

演讲人：这些想法始于一位工程师，

说话人：对吗？

说话人：我认为，

说话人：有时几乎真的

说话人：大团队追求的很暧昧，

说话人：大想法，

说话人：你最终实际上被放慢了速度

说话人：正因为如此，所以我认为这是文化。

演讲人：我想我们实际上也选择了...

说话人：那些真正想做零的人

演讲人：进行一项实验，但事实并非如此

说话人：很简单，有很多赌注我们会结束

言语人：上转下或关闭嗯和

说话人：也许你知道我们会在

说话人：未来呃但是那很难那很难

说话人：当你倾注心血时

讲话人：作为一个打赌的创始人，这是

说话人：还没有工作，嗯，所以我想这就像

言语人：那种类型的选择适合那种类型

说话人：个性的，

说话人：那些真正充满热情和热情的人

说话人：关于零到一的深奥，所以你领先

讲话人：研究团队的产品。你工作

演讲人：与 Anthropic.A 的研究人员

说话人：很多人都有这样的感觉

演讲人：什么是研究，

说话人：研究人员做什么。我想了很多

说话人：人们并不完全理解这些

演讲人：所有人工智能实验室都非常有价值的人。

说话人：我的想法是，

演讲人：我想帮助人们理解，

说话人：帮我理解到底是什么

演讲人：研究人员整天都在做什么？

说话人：我想他们有一个假设

演讲人：关于如何改进模型。他们发现

说话人：数据。他们调整了一些算法。他们

讲话人：调整训练方式，然后进行测试

说话人：它，看看它是怎么做的，不断迭代，

说话人：并不断努力寻找改进的方法

说话人：模型，大概是这样吧？

说话人：斜杠，

演讲人：帮助我们了解什么是研究人员

讲话人：整天都在做。那是真的，

讲话人：我认为这很多，也许更多

说话人：日常。我想一件事情

演讲人：研究人员和研究组织

说话人：就像人类一样，也有一个愿景

讲话人：更广泛的未来。所以，

讲话人：例如，

讲话人：诸如此类的事情……我想即使是在成立之初

讲话人：公司的，

演讲人：研究人员正在谈论如何做

说话人：我们让克劳德使用电脑吗？

演讲人：我们如何让人工智能喜欢导航

说话人：屏幕，对吧？

说话人：所以有很多实际上非常创始人

说话人：我是这样形容它的——就像能量一样

演讲人：科研人员真是大胆又雄心勃勃

说话人：研究人员。我们有很多这样的人

说话人：在 Anthropic。所以有一层

演讲人：展望这项技术可以走向何方。

说话人：然后我想到了另一边

说话人：循环的，

说话人：现在还有这个技术

言语人：或者说云在人们的手中，

说话人：今天我们怎样才能做得更好呢？

演讲人：所以这是一个中长期的

说话人：很多精力都在思考那个镜头

讲话人：未来的。而且也是现在的

演讲人：短期而言，

演讲人：我们可以改进的地方有哪些

说话人：做？

说话人：所以就像，

说话人：我认为你描述的非常好

演讲人：了解我们如何进行迭代改进

说话人：关于不同版本的 Cloud.The way

演讲人：就像我的团队与研究人员合作一样

说话人：是一种非常整合和

说话人：嵌入这些循环中，

说话人：特别是人多的地方

说话人：对用户的影响。所以这就是事情

说话人：喜欢视觉、计算机使用、编码、

说话人：代理编码、工具使用、

说话人：测试时间计算，

讲话人：对用户有直接影响的事情。

说话人：然后弄清楚有什么方法

讲话人：带来用户反馈和地面

说话人：处于可以理解的水平

00:30:00

英文原文（完整 ASR）

讲话人：for researchers and also actionable

讲话人：for researchers.And I think that's the

讲话人：second piece is actually a big part

讲话人：of the job and sometimes a hard part

讲话人：of the job.So,for example,

讲话人：we might get feedback on cloud.ai,

讲话人：cloud hallucinated,right?

讲话人：It's very vague.If you bring that to

讲话人：a researcher and you say,

讲话人：please fix Claude from being hallucinated,

讲话人：it's not very actionable.And so part

讲话人：of the time of the team is understanding,

讲话人：okay,

讲话人：what's the trajectory of why that user

讲话人：gave that feedback?

讲话人：And it's like consented.And so we look

讲话人：at,okay,what's the trajectory?

讲话人：cloud have called tools in that moment,

讲话人：or from its current knowledge,

讲话人：or it called the right,

讲话人：looked at the right document,

讲话人：but it looked at the wrong facts.In

讲话人：the first case,

讲话人：that would have been a failure on tool

讲话人：use.On the second case,

讲话人：it would have been a failure on let's

讲话人：say search or knowledge and search and

讲话人：search synthesis.or it could be something

讲话人：around alignment.And so bring that level

讲话人：of detail to researchers,

讲话人：coming up with like,

讲话人：is this a big enough problem?

讲话人：Figure out things like evals to then

讲话人：describe what we've improved it.Like

讲话人：those are the levels of actionability

讲话人：and it's the day-to-day language of

讲话人：the researchers.And so we try to stay

讲话人：very close to how to bring that in an

讲话人：actionable manner.between users to the

讲话人：core model training and the research

讲话人：development loop.I was talking to someone

讲话人：the other day about how it feels like

讲话人：research,

讲话人：AI research is the place to be now if

讲话人：you want to be very successful in life.

讲话人：What does it take to become a really

讲话人：successful researcher from what you

讲话人：can tell?

讲话人：You know,

讲话人：not everyone can get in.Not everyone's

讲话人：brain is going to work this way.But

讲话人：just say people are like,

讲话人：hey,

讲话人：I want to explore this career path from

讲话人：what you've seen.What does it take to

讲话人：make it there?

讲话人：Researchers generally are research and

讲话人：product managers working with researchers,

讲话人：or both.Let's do both.But the researchers,

讲话人：like,you know,

讲话人：PMs working with researchers are also

讲话人：going to be very successful.But it feels

讲话人：like everyone's trying to,

讲话人：you know,

讲话人：poach all the top researchers across

讲话人：every company.So just,

讲话人：I know you're not an AI researcher,

讲话人：but just from what you've seen,

讲话人：just like,

讲话人：what does it take to make it in that

讲话人：career path?

讲话人：Yeah,

讲话人：I think a lot of the most successful researchers

讲话人：and research leadership at Anthropic

讲话人：are folks who are really strong first

讲话人：principles thinkers about problems.

讲话人：Like they reason through problems really

讲话人：well,

讲话人：who are just passionate about their research

讲话人：area and have a bold description of

讲话人：what that could look like.And then who

讲话人：are actually close to the details And

讲话人：so our leadership, our chief scientists,

讲话人：our heads of fine tuning and RL,

讲话人：folks are actually really close to the

讲话人：training runs and actually look at things

讲话人：like how the training run is going,

讲话人：evals,

讲话人：looking at the underlying data.So actually

讲话人：staying really close and be excited

讲话人：to be in the details,

讲话人：I think have been a sign of really strong

讲话人：researchers and developing taste.And

讲话人：I think like another piece is just like

讲话人：their ability to think big over time

讲话人：and be like very ambitious,

讲话人：right?

讲话人：Like the Dario,

讲话人：like we can transform software engineering

讲话人：and,and,and the,and I think,

讲话人：uh,going in that direction,

讲话人：you learn so much,you get,

讲话人：you had to shoot for the stars in,

讲话人：in many ways across,uh,

讲话人：your ideas.I think in order to be a,

讲话人：uh,a successful researcher,

讲话人：I love just this meme of just be more

讲话人：ambitious comes up so often now,

讲话人：which is so hard.Like it's it's easy

讲话人：to say that it's hard to actually just

讲话人：like how big can you think?

讲话人：And how that's so much of what AI now

讲话人：unlocks.Just be more ambitious.Yeah.

讲话人：Yeah.I think it's thinking through it

讲话人：once or twice end to end and then being,

讲话人：I think,

讲话人：stubborn about the area and maybe more

讲话人：loose around the exact approach.It is

讲话人：a question we challenge ourselves with.

讲话人：the technology is moving so quickly.

讲话人：And so how do you make sure what you're

讲话人：building is actually forward compatible?

讲话人：And so it's also actually part of like,

讲话人：I think the core product development

讲话人：loop to think bigger,right?

讲话人：One thing I ask the team frequently

讲话人：or how I think about when we're building

讲话人：a product is let's say cloud eight comes

讲话人：around,what changes in what users do?

讲话人：And then what should,

讲话人：what does that mean for how you're building

讲话人：today?

讲话人：Is it going to be forward compatible

讲话人：to that experience?

讲话人：Right.So like just grounding,

讲话人：it's,

讲话人：I think being ambitious is very broad.

讲话人：And so trying to like ground it in,

讲话人：in some ways of describing,

讲话人：describing that.And also,

讲话人：yeah,

讲话人：everything heading in a direction that

讲话人：all is cohesive and makes sense versus

讲话人：just ambitious in a completely different

讲话人：direction.Speaking of ambition and cloud

讲话人：eight,

讲话人：Fable slash Mythos recently feels like

讲话人：hit this very new kind of tipping point

讲话人：with models where it used to be you

讲话人：have an awesome model.Release it.Hey,

讲话人：everyone.Welcome.Opus 4.5 is out.Everyone

讲话人：can use it.Nithos went in a very different

讲话人：direction.It got blocked.There was a

讲话人：lot of scrutiny,

讲话人：a lot of concern about what it was capable

讲话人：of.All the companies had to go make

讲话人：sure it wasn't going to hack into all

讲话人：their systems.And it feels like now

讲话人：every model,

讲话人：because they continue to get better,

讲话人：will now have a lot more scrutiny and

讲话人：there will be more restrictions on who

讲话人：can use them,

讲话人：which feels like a big deal.How do you

讲话人：think about that?

讲话人：How does that change the way you operate.

讲话人：I'm going to maybe leave the policy

讲话人：and the expert control side to focus

讲话人：on that and work on that.I think the

讲话人：product question and how we interact

讲话人：with these internally is,

讲话人：I think,as you mentioned,

讲话人：as frontier models become more capable,

讲话人：the safeguards and the ways of red teaming

讲话人：and testing and the pre-release process

讲话人：also needs to evolve and adapt quickly

讲话人：to address that.And so one example is,

讲话人：you know,before Fable models,

讲话人：we didn't have a strong of,

讲话人：let's say,

讲话人：fallback UXs and systems because our

讲话人：goal is to make sure that like there

讲话人：is asymmetrical benefit for this technology

讲话人：and to minimize like the downside or

讲话人：like a severe risk of of it.And so we

讲话人：ended up building fallback systems so

讲话人：that users will still get a great response

讲话人：from Opus 4.8 immediately.And so I think

讲话人：there's a piece around as we evolve

讲话人：and improve safety systems,

讲话人：how do we continue to develop and deliver

讲话人：great user experiences I think there's

讲话人：more that we can do on both sides.And

讲话人：so you'll see us innovating,

讲话人：improving on what we called out the

讲话人：model safeguards package more and more

讲话人：in the coming weeks and months.What's

讲话人：really interesting and just like unexpected

讲话人：here is creates this really interesting

讲话人：advantage for Anthropic where you have

讲话人：access to the latest stuff.And this

讲话人：is going to happen at every lab.Everyone's

讲话人：going to keep improving.And it creates

讲话人：this unfair advantage within the labs

讲话人：to have access to the best stuff that

讲话人：other people can't yet outside of your

讲话人：control.You'd prefer everyone use it.

讲话人：So it's a really interesting this new

讲话人：feedback loop that's going to start

讲话人：where models that are so advanced are

讲话人：only accessible to certain companies.

讲话人：And that's going to be a whole new unexpected.

讲话人：It's like a second order effect of all

讲话人：these restrictions.Our goal is to develop

讲话人：these systems and the models to be as

讲话人：inclusive as possible.I think our goal

讲话人：is to not have that happen for the general

讲话人：purpose,

讲话人：general use technologies and to make

讲话人：it more accessible.I think this is one

讲话人：of our top priorities right now to kind

讲话人：of reduce what we're seeing there.Yeah,

讲话人：that makes sense.I would imagine you'd

讲话人：want as many customers and people using

讲话人：this thing as possible.This episode

讲话人：is brought to you by Mercury.Radically

讲话人：different banking loved by over 300

讲话人：,000

讲话人：entrepreneurs.And now with Command.I've

讲话人：been a customer of Mercury's for over

讲话人：six years.I have never once thought

讲话人：about leaving.Mercury is basically what

讲话人：happens when banking is built by product

讲话人：people,

讲话人：not by bankers.They make it so easy.

讲话人：Dare I say fun?

讲话人：Yeah.send invoices, move money around,

讲话人：set up virtual cards for folks on my

讲话人：team.Does your bank have an API,

讲话人：a terminal-native CLI,

讲话人：or an AI-ready MCP server?

讲话人：I don't think so.And just recently they

讲话人：launched Command,

讲话人：a conversational interface built directly

讲话人：into Mercury,

讲话人：which acts as your financial operator.

讲话人：I've been using Command to transfer

讲话人：money around to figure out what categories

讲话人：I've been spending the most money in,

讲话人：analyze my cash flows.And just today,

讲话人：I used it to find out how much I've

讲话人：made from a specific sponsor over the

讲话人：past year.I just asked,

讲话人：how much have I made from X over the

讲话人：past year?

讲话人：10 seconds later,

讲话人：I have an answer.It is so freaking cool.

讲话人：Visit Mercury.com to learn more and

讲话人：apply online in minutes.Mercury is a

讲话人：fintech company,

讲话人：not an FDIC-insured bank.Banking services

讲话人：provided through Choice Financial Group

讲话人：and Column NA,

讲话人：members FDIC.I want to talk a little

讲话人：bit about how the product role is changing

讲话人：and who is doing well in this new world

讲话人：now that AI is such a core part of our

讲话人：life.When you're hiring PMs,

讲话人：product people,

讲话人：when you're looking at people that do

讲话人：well in today's world,

讲话人：what are some things that you notice?

讲话人：What are you looking for most?

讲话人：What are you looking for more?

中文译文（机器翻译，按本时间段分段）

演讲人：对于研究人员来说，也是可操作的

演讲人：对于研究人员来说。我认为这就是

讲话人：第二段其实是很大一部分

讲话人：工作的一部分，有时是一个困难的部分

说话人：关于工作。所以，例如，

说话人：我们可能会得到有关 cloud.ai 的反馈，

说话人：云出现幻觉了吧？

说话人：这很模糊。如果你把它带到

说话人：一位研究员，你说，

说话人：请修正克劳德的幻觉，

说话人：这不是很有可操作性。所以一部分

说话人：团队的时间就是理解，

说话人：好的，

说话人：那个用户的轨迹是什么

演讲人：给出了反馈吗？

说话人：这就像是同意了，所以我们看起来

说话人：at，好吧，轨迹是怎样的？

讲话人：云在那一刻调用了工具，

说话人：或者从目前的知识来看，

说话人：或者叫右边，

说话人：看了正确的文件，

说话人：但它着眼于错误的事实。

说话人：第一种情况，

说话人：那工具就失败了

讲话人：用。对于第二种情况，

说话人：如果我们这么做的话就会失败

讲话人：说搜索或知识并搜索和

讲话人：搜索综合。或者它可能是某事

说话人：围绕对齐。所以要达到那个水平

演讲人：向研究人员详细介绍，

说话人：想出类似的，

说话人：这是一个足够大的问题吗？

说话人：弄清楚诸如评估之类的事情

说话人：描述一下我们改进了什么。比如

演讲人：这些是可操作性的级别

说话人：这是人们的日常语言

说话人：研究人员。所以我们尽量留下来

说话人：非常接近如何将其带入

说话人：可操作的方式。用户之间

演讲人：核心模型训练与研究

说话人：开发循环。我正在和某人说话

演讲人：前几天谈论的感觉如何

说话人：研究、

演讲人：人工智能研究是现在的地方，如果

说话人：你想在生活中取得非常成功。

说话人：怎样才能成为一个真正的

说话人：成功的研究员来自于你

说话人：看得出来吗？

说话人：你知道，

讲话人：不是每个人都能进去。不是每个人的

说话人：大脑就是这样工作的。但是

言语人：只是说人们就像，

说话人：嘿，

演讲人：我想从

说话人：你所看到的，需要什么才能

说话人：能到那儿吗？

演讲人：研究人员一般都是研究和

演讲人：与研究人员一起工作的产品经理，

说话人：或者两者都做。让我们两者都做吧。但是研究人员，

说话人：就像，你知道，

演讲人：与研究人员一起工作的项目经理也

演讲人：将会非常成功，但是感觉

说话人：就像每个人都在努力的那样，

说话人：你知道，

说话人：挖走世界各地的顶尖研究人员

讲话人：每个公司。所以，

说话人：我知道你不是人工智能研究员，

说话人：但仅从你所看到的来看，

说话人：就像，

讲话人：需要什么才能做到这一点

说话人：职业道路？

说话人：是的，

演讲人：我认为很多最成功的研究人员

演讲人：Anthropic 的研究领导力

说话人：首先是真正强大的人

演讲人：思考问题的原则。

说话人：就像他们真的在推理问题一样

说话人：嗯，

发言者：对他们的研究充满热情的人

讲话人：区域并有一个大胆的描述

讲话人：那可能是什么样子。然后是谁

讲话人：实际上很接近细节并且

演讲人：所以我们的领导层，我们的首席科学家，

演讲人：我们的微调和强化学习负责人，

说话人：人们其实很接近

演讲人：训练跑步并实际观察事物

说话人：比如训练进行得怎么样，

发言者：evals,

演讲人：看底层数据，所以实际上

说话人：保持非常亲密并感到兴奋

说话人：要讲细节，

说话人：我认为这是真正坚强的标志

说话人：研究人员和开发品味。并且

说话人：我觉得另一首曲子就像

说话人：他们随着时间的推移思考远大目标的能力

说话人：而且要非常有野心，

说话人：对吗？

说话人：像达里奥一样，

讲话人：就像我们可以改变软件工程一样

讲话人：并且，并且，以及，并且我认为，

说话人：呃，朝那个方向走，

说话人：你学到了很多东西，你得到了，

说话人：你必须为其中的星星而努力，

说话人：在很多方面，呃，

说话人：你的想法。我认为为了成为一个，

说话人：呃，一位成功的研究员，

讲话人：我喜欢这个 meme，“just be more”

说话人：雄心勃勃现在经常出现，

说话人：这很难，但好像很容易

讲话人：说实际上很难

说话人：比如你能想到多大？

说话人：这就是现在人工智能的主要内容

说话人：解锁。更有野心。是的。

说话人：是的，我想它正在考虑清楚

说话人：一次或两次端到端，然后是，

说话人：我认为，

说话人：对这个地区很固执，甚至更多

说话人：围绕确切的方法松散。

说话人：我们挑战自己的一个问题。

演讲人：技术发展如此之快。

说话人：那么你如何确定你是什么

演讲人：建筑实际上是向前兼容的吗？

说话人：所以它实际上也是一部分，

演讲人：我认为核心产品开发

说话人：循环思考更大的事情，对吧？

演讲人：我经常问团队的一件事

说话人：或者说我们建造时我的想法

说话人：一个产品就是云八来的

演讲人：周围，用户的行为有哪些变化？

演讲人：然后呢，

说话人：这对于你的构建方式意味着什么

说话人：今天？

说话人：会向前兼容吗

说话人：那段经历？

讲话人：对。所以就像接地一样，

说话人：是，

讲话人：我认为雄心勃勃的范围很广。

说话人：所以试着把它融入进去，

说话人：在某些描述方式中，

讲话人：描述这一点。而且，

说话人：是的，

说话人：一切都朝着一个方向发展

说话人：所有内容都是有凝聚力且有意义的

说话人：只是在一个完全不同的领域雄心勃勃

说话人：方向。说志气与云

说话人：八、

说话人：最近感觉寓言斜线神话

说话人：达到了这种全新的临界点

说话人：与曾经是你的模特

说话人：有一个很棒的模型。发布它。嘿，

讲话人：大家欢迎。Opus 4.5 已经出来了。大家

说话人：可以用它。Nithos 采取了一种非常不同的方式

讲话人：方向。被挡住了。有一个

说话人：经过大量的审查，

说话人：非常关心它的能力

发言者：of.All the Companies had to go make

说话人：当然它不会侵入所有

讲话人：他们的系统。感觉就像现在

说话人：每个模特，

说话人：因为他们不断变得更好，

演讲人：现在将会有更多的审查和

说话人：对谁会有更多限制

说话人：可以使用它们，

说话人：这感觉像是一件大事。你好吗？

说话人：想一下吗？

讲话人：这如何改变你的运作方式？

说话人：我可能会放弃这个政策

演讲人：与专家控方聚焦

讲话人：就这个问题并努力解决这个问题。我认为

演讲人：产品问题以及我们如何互动

说话人：这些内部是，

说话人：我认为，正如你提到的，

演讲人：随着前沿模型变得越来越有能力，

演讲人：红队的保障措施和方式

演讲人：以及测试和预发布过程

说话人：也需要快速进化和适应

演讲人：为了解决这个问题。举个例子，

说话人：你知道，在《神鬼寓言》模型出现之前，

说话人：我们没有很强的，

说话人：比方说，

演讲人：后备用户体验和系统，因为我们

说话人：目标是确保像那里一样

演讲人：这项技术的利益是不对称的

讲话人：并尽量减少不利因素或

说话人：就像有严重的风险一样。所以我们

说话人：最终建立了后备系统，所以

演讲人：用户仍然会得到很好的回应

讲话人：立即从 Opus 4.8 开始。所以我认为

说话人：随着我们的发展，总会有一些事情发生

演讲人：并改善安全系统，

演讲人：我们如何持续开发和交付

演讲人：我认为有很棒的用户体验

说话人：我们双方都可以做更多的事情。并且

演讲人：所以你会看到我们创新，

说话人：改进我们所说的

演讲人：模特保障套餐越来越多

说话人：在接下来的几周和几个月里。什么

讲话人：真的很有趣，就像出乎意料一样

演讲人：这里创造了这真的很有趣

说话人：你所拥有的 Anthropic 的优势

说话人：获取最新的东西。还有这个

说话人：每个实验室都会发生。每个人

演讲人：将不断改进。它创造了

演讲人：实验室内部的这种不公平优势

说话人：能够接触到最好的东西

讲话人：其他人还不能脱离你的范围

说话人：控制。你希望每个人都使用它。

说话人：这个新东西真的很有趣

说话人：反馈循环即将开始

讲话人：这么先进的模型在哪里

讲话人：仅对某些公司开放。

演讲人：这将是一个全新的意外。

讲话人：这就像所有的二阶效应

讲话人：这些限制。我们的目标是发展

演讲人：这些系统和模型是

讲话人：尽可能包容。我认为我们的目标

说话人：就是不让将军发生这种事

说话人：目的，

演讲人：一般使用技术和制作

讲话人：更容易理解。我认为这是一个

演讲人：我们现在的首要任务是

说话人：减少我们在那里看到的东西。是的，

说话人：这是有道理的。我想你会

讲话人：希望尽可能多的客户和人使用

讲话人：这件事尽可能。这一集

说话人：是 Mercury.Radically 给你带来的

说话人：超过 300 人喜爱的不同银行业务

说话人：,000

说话人：企业家。现在是指挥部。我已经

说话人：已经是 Mercury 的客户超过

说话人：六年了，我从来没有想过

讲话人：关于离开。水星基本上是什么

说话人：当银行业是由产品构建时就会发生

讲话人：人们，

说话人：不是银行家。他们让事情变得如此简单。

说话人：我敢说有趣吗？

说话人：是的。发送发票，转移资金，

说话人：为我的朋友设置虚拟卡

演讲人：团队。你们银行有 API 吗？

说话人：终端原生 CLI，

演讲人：或者支持 AI 的 MCP 服务器？

说话人：我不这么认为。而且就在最近他们

说话人：发起命令，

说话人：直接构建的对话界面

说话人：进入水星，

讲话人：充当您的财务操作员。

讲话人：我一直在使用 Command 进行转接

说话人：钱周围要弄清楚什么类别

说话人：我花最多的钱在，

说话人：分析一下我的现金流。就在今天，

说话人：我用它来了解我有多少钱

讲话人：由特定发起人在

说话人：去年。我刚才问，

说话人：这段时间我从 X 赚了多少钱

说话人：过去一年？

说话人：10 秒后，

说话人：我有一个答案，太酷了。

说话人：访问 Mercury.com 了解更多信息

讲话人：几分钟内在线申请。Mercury 是

演讲人：金融科技公司、

说话人：不是 FDIC 承保的银行。银行服务

讲话人：通过 Choice Financial Group 提供

讲话人：和 NA 专栏，

讲话人：FDIC 成员。我想谈谈

演讲人：关于产品角色如何变化的一些信息

说话人：在这个新世界里谁过得好

演讲人：现在人工智能已经成为我们的核心部分

说话人：生活。当你招聘产品经理时，

讲话人：产品人员，

说话人：当你看着这样做的人时

说话人：在当今世界，

说话人：您注意到哪些事情？

说话人：你最在寻找什么？

说话人：你还想知道什么？

00:40:00

英文原文（完整 ASR）

讲话人：What's trending up and what you find

讲话人：is important and what's trending down?

讲话人：We actually on my team have not changed

讲话人：our hiring loop for three years now.

讲话人：So what we actually look for and the

讲话人：traits and how we evaluate generalists,

讲话人：like PNs,generalists,

讲话人：like research product managers have

讲话人：actually been the same.So I think some

讲话人：of those traits,

讲话人：number one is first principles thinking.

讲话人：And this is really,

讲话人：rather than pattern matching what you

讲话人：used to do in,let's say,

讲话人：consumer product or B2B SaaS,

讲话人：but actually figuring out in this moment

讲话人：for this user group with this technology,

讲话人：what is the user value?

讲话人：Is there an example that a lot of people

讲话人：hear first principles thinking,

讲话人：they're like,yes,

讲话人：I got it.I'm good at this.What is it?

讲话人：What's an example of someone having

讲话人：really demonstrated really good first

讲话人：principles thinking?

讲话人：I think one example is,

讲话人：I think you think of a product manager

讲话人：as I own product strategy and delivering

讲话人：user value,

讲话人：but I demonstrate day-to-day by writing

讲话人：a PRD or writing a product vision doc.

讲话人：And for my team,

讲话人：as like research product managers,

讲话人：the way to drive user value is to figure

讲话人：out the right user feedback,

讲话人：the evals.right,

讲话人：that then can be a personification of

讲话人：that user need.So like we do write some

讲话人：product documents and PRDs,

讲话人：but we actually have a saying on the

讲话人：team of evals are the new PRDs.Because

讲话人：in order to deliver that user value,

讲话人：it's not that exact artifact that people

讲话人：used to write in the last one to two

讲话人：decades.It's a new way of working.And

讲话人：so the first principles thinking would

讲话人：be,

讲话人：let me figure out what is the thing I

讲话人：should do to achieve my goals,

讲话人：rather than here is a set of activities

讲话人：that I've done and therefore I will

讲话人：continue to do.So the idea here is,

讲话人：it used to be,have kind of an idea,

讲话人：create a PRD.talk to people about it,

讲话人：align on the plan,design it,

讲话人：build it,ship it,see how it goes,

讲话人：iterate.What I'm hearing here is it's

讲话人：like,okay,

讲话人：here's some feedback about something

讲话人：that's wrong or an opportunity.Step

讲话人：one is the eval is now how you define

讲话人：what the work is versus a PRD.Maybe

讲话人：step one would be understanding the

讲话人：user pain point.And so the way to even

讲话人：access that user pain point is different,

讲话人：right?

讲话人：In the past,

讲话人：we might do a user interview.I think

讲话人：if you go like deep enough,

讲话人：you might have the user walk you through

讲话人：their user flow,the pixels.Here,

讲话人：you have to sweat the tokens as much

讲话人：as you sweat the pixels.And so one activity

讲话人：we have on the team is reading the transcripts

讲话人：and understanding what was the trajectories

讲话人：that failed.very deeply to then say,

讲话人：was this like a hallucination?

讲话人：Was this cloud being overconfident?

讲话人：So like the theme of the failure actually

讲话人：has a lot of nuance.And then that allows

讲话人：you to build a description,

讲话人：a like sustained description of that

讲话人：pain point.So that could be essentially

讲话人：in a new eval.And is it eval on distribution,

讲话人：right?

讲话人：Is it capturing both the positive situations

讲话人：where this is failing and also areas

讲话人：when it should actually not fail?

讲话人：And then bring that back to,

讲话人：let's say,

讲话人：research so that we can make the improvements

讲话人：and actually measure the quality of,

讲话人：okay,when we have Opus 5.5,

讲话人：is this area improving or not?

讲话人：Is Claude now able to identify the right

讲话人：places in the document?

讲话人：and pull the right synthesis out.So

讲话人：it's just the actionability and shorting

讲话人：the distance to actionability for our

讲话人：stakeholders and partner teams like

讲话人：researchers to take action on.Is there

讲话人：an example of something like this where

讲话人：you found an issue or opportunity and

讲话人：then wrote the eval?

讲话人：And what is the eval looking like in

讲话人：most cases?

讲话人：When people wanna picture an eval,

讲话人：what is that?

讲话人：What do they picture?

讲话人：We actually pioneered this concept within

讲话人：Anthropic.So one of the early examples

讲话人：is the early cloud models were not very

讲话人：good at following specific schemas,

讲话人：so like things like outputs and JSON.

讲话人：And now that is fundamental to cloud

讲话人：being able to be a good agent,

讲话人：right?

讲话人：If you can't output a certain format,

讲话人：you don't know how to like access APIs,

讲话人：you can't call tools,

讲话人：et cetera.And so the initial end-to

讲话人：-end was I was hearing feedback around,

讲话人：you know,

讲话人：Cloud 2 days.Cloud was not very good

讲话人：at following instructions.So then digging

讲话人：in with users,

讲话人：what do you mean by Cloud is not good

讲话人：at following instructions?

讲话人：Give me what situations this was happening.

讲话人：Like,what's the exact situation?

讲话人：what did you ask?

讲话人：What was Claude's response?

讲话人：Going to like that level of detail.

讲话人：And what I saw was something like 80

讲话人：%of what people meant in the early days

讲话人：for this failure was Claude would not

讲话人：write the right JSON.And so then,

讲话人：okay,

讲话人：let's generate maybe to start just 30

讲话人：to 40 examples of when Claude was not

讲话人：doing this thing correctly.And then

讲话人：that actually is your eval set.And you

讲话人：can have essentially a prompt and a

讲话人：response.And if that is not working

讲话人：in the right golden answer that you

讲话人：might have,

讲话人：then that means that the eval essentially

讲话人：is beneficial because it's identifying

讲话人：a pain point consistently.And so then

讲话人：we added that to our...repositories

讲话人：for evals.And when we have versions

讲话人：of cloud,

讲话人：we actually run that eval and just check.

讲话人：I think at this point it's always 100

讲话人：%or like 99.9.And so it's no longer

讲话人：a pain point.But in the early days it

讲话人：was taking the user feedback,

讲话人：figuring out actually what they mean.

讲话人：Can we reproduce it?

讲话人：Is it consistent?

讲话人：Is it a big issue?

讲话人：And then figuring out how to,

讲话人：standardize it in a way that can be

讲话人：consumable for researchers.It's basically

讲话人：test-driven development for PMs is the

讲话人：world we're living now,

讲话人：where you write the test first.So is

讲话人：this just a core part of the product

讲话人：management job now at Anthropic Writing

讲话人：Evals?

讲话人：I think so.I also think it's something

讲话人：I've talked to other PMs at other companies

讲话人：about,

讲话人：and I think it's also more and more

讲话人：of the skill set more broadly,

讲话人：because a lot of the products that we're

讲话人：building is at the intersection of models

讲话人：with harnesses,

讲话人：with a set of contexts for a set of

讲话人：users.And so having things like evals

讲话人：actually is a way not just for...folks

讲话人：working on models,

讲话人：but generally within product,

讲话人：to get to better user experiences.Because

讲话人：you can't improve what you can't measure.

讲话人：And a lot of this is very still tactile

讲话人：based,

讲话人：it's still very judgment based.And so

讲话人：you have to stay close to the details.

讲话人：And also very non-deterministic,

讲话人：which is a big part of this.Just like,

讲话人：it's not going to give you the same

讲话人：answer every time.So you got to describe

讲话人：it kind of more broadly.It's not going

讲话人：to be an exact match.So this is a really

讲话人：interesting change in the way product

讲话人：happens and will happen is evals writing

讲话人：evals versus PRDs is a big part of this.

讲话人：Do you guys still do PRDs?

讲话人：Is there still like a one pager describing

讲话人：a problem or is it replay?

讲话人：Okay.Now you're shaking your head.Yes.

讲话人：We are,

讲话人：we do.I think when there's a very defined

讲话人：problem,

讲话人：I think things like evals might be almost

讲话人：a shorthand.I think there's other cases

讲话人：where PRDs are really valuable.PRDs

讲话人：are great vehicles for getting a very

讲话人：large group of people aligned on a set

讲话人：of sources of truth about experience

讲话人：and set of goals.So when we do have

讲话人：a model,we actually, for every model,

讲话人：we do have a PRD,

讲话人：less necessarily for our researchers,

讲话人：but more for our growing product surfaces,

讲话人：for our engineering teams,

讲话人：for our stakeholders like legal and

讲话人：safety and others,

讲话人：as just a source of truth of putting

讲话人：together what we're aiming to achieve

讲话人：so that a big group of people can row

讲话人：in the same direction.The other place

讲话人：where I do think PRDs are valuable are

讲话人：on the more ambiguous problems and opportunities.

讲话人：So if we haven't shipped a thing like

讲话人：computer use,

讲话人：we don't necessarily have a set of like

讲话人：user specific pain points always.And

讲话人：I think there's value in the product

讲话人：vision portions of a PRD to explore

讲话人：what could,

讲话人：even if a technology is not yet ready

讲话人：to work for everyone.How do you get

讲话人：it to work well for some group?

讲话人：So you can explore the value.You can

讲话人：actually bring something that is coherent

讲话人：to a user group.So we do have PRDs.

讲话人：I think the application's a little different

讲话人：Okay,this is great.I just had Andrew,

讲话人：he's the head of the Codex app at OpenAI,

讲话人：and you guys are aligned.PRD is not

讲话人：dead,

讲话人：still very useful for specific projects

讲话人：and ideas.Great.Okay,

讲话人：we've closed the book on PRD is still

讲话人：kicking.Okay,

讲话人：so we've been talking a bit about just

讲话人：what kind of skills are kind of emerging

讲话人：for product people.Is there anything

中文译文（机器翻译，按本时间段分段）

演讲人：什么是流行趋势以及你发现了什么

说话人：很重要，什么正在下降？

说话人：我们其实我的团队没有变

演讲人：我们的招聘循环已经三年了。

演讲人：那么我们真正寻找的是什么？

演讲人：特征以及我们如何评价通才，

说话人：如 PN、通才、

说话人：就像研究产品经理一样

说话人：其实都是一样的，所以我觉得有些

说话人：在这些特征中，

讲话人：第一是第一性原理思维。

说话人：这确实是，

说话人：而不是模式匹配你的内容

说话人：过去常做，比方说，

演讲人：消费产品或 B2B SaaS，

说话人：但实际上此刻正在弄清楚

说话人：对于这个拥有这项技术的用户群来说，

演讲人：什么是用户价值？

说话人：有没有一个例子，很多人

说话人：听到第一原则的思考，

说话人：他们就像，是的，

说话人：我明白了。我很擅长这个。这是什么？

说话人：某人有什么例子

演讲人：首先表现得非常好

说话人：原则思维？

演讲人：我认为一个例子是，

说话人：我想你想到的是产品经理

说话人：因为我有自己的产品策略和交付

演讲人：用户价值，

说话人：但我日常通过写作来展示

说话人：PRD 或正在撰写产品愿景文档的人。

发言者：对于我的团队来说，

演讲人：就像研究产品经理一样，

演讲人：驱动用户价值的方法是数字

说话人：输出正确的用户反馈，

发言者：the evals.right,

言语人：那么可以是以下的拟人化

说话人：用户需要的，就像我们写一些

演讲人：产品文件和 PRD，

讲话人：但实际上我们有一句话

演讲人：评估团队是新的 PRD。因为

演讲人：为了传递用户价值，

说话人：这并不是人们所认为的那么精确的人工制品

说话人：用来写最后一到二

说话人：几十年了。这是一种新的工作方式。而且

演讲人：所以首要原则思维是

说话人：是，

说话人：让我弄清楚我到底在做什么

说话人：为了实现我的目标应该做什么，

讲话人：这里不是一组活动

讲话人：我已经做到了，因此我会这样做

说话人：继续做。所以这里的想法是，

说话人：以前是有一个想法，

说话人：创建一个 PRD。与人们谈论它，

演讲人：对齐计划，设计它，

说话人：建造它，运送它，看看它进展如何，

说话人：迭代。我在这里听到的是

说话人：就像，好吧，

说话人：这里有一些关于某事的反馈

讲话人：那是错误的，或者是一个机会。步骤

讲话人：一个是 eval，现在你如何定义

演讲人：与 PRD 相比，工作是什么。也许

演讲人：第一步是理解

说话人：用户痛点。等等的方法

演讲人：接入用户痛点不同，

说话人：对吗？

说话人：过去，

说话人：我们可能会做一次用户访谈。我想

说话人：如果你说得足够深入，

说话人：你可能会让用户引导你完成

说话人：他们的用户流量，像素。这里，

讲话人：你必须为代币付出同样多的汗水

说话人：当你为像素出汗时。等等一项活动

讲话人：我们团队里的人正在朗读笔录

讲话人：了解轨迹是什么

讲话人：那失败了。非常深刻地说，

说话人：这像是幻觉吗？

说话人：云是不是太自信了？

说话人：其实很喜欢失败的主题

说话人：有很多细微差别。然后这就允许

说话人：你来建立一个描述，

说话人：类似的持续描述

说话人：痛点。所以本质上可能是

讲话人：在一个新的评估中。它是关于分布的评估吗？

说话人：对吗？

讲话人：是否同时捕捉到了积极的情况

演讲人：失败的地方以及领域

说话人：什么时候它实际上不应该失败？

说话人：然后回到，

说话人：比方说，

讲话人：研究以便我们能够做出改进

说话人：并实际衡量质量，

说话人：好的，当我们有 Opus 5.5 时，

说话人：这方面有进步吗？

说话人：克劳德现在能辨别正确的了吗？

说话人：文件中的位置？

说话人：并拉出正确的综合。所以

说话人：只是可操作性和不足

说话人：我们与可采取行动的距离

发言者：利益相关者和合作伙伴团队，例如

说话人：研究人员要采取行动吗？

说话人：类似这样的例子，其中

讲话人：您发现了一个问题或机会并且

说话人：然后写了 eval？

说话人：eval 是什么样子的

演讲人：大多数情况？

说话人：当人们想要想象一个评估时，

说话人：那是什么？

说话人：他们想象什么？

演讲人：我们实际上是在

说话人：Anthropic.So 早期的例子之一

说话人：就是早期的云模型不太好

说话人：善于遵循特定图式，

说话人：就像输出和 JSON 之类的东西。

演讲人：现在这是云的基础

说话人：能够成为一名优秀的代理人，

说话人：对吗？

说话人：如果你不能输出某种格式，

讲话人：你不知道如何访问 API，

说话人：你不能调用工具，

说话人：等等。所以最初的结尾

说话人：-最后我听到了周围的反馈，

说话人：你知道，

说话人：Cloud 2 天。Cloud 不太好

言语人：按照以下指示。然后挖掘

说话人：与用户一起，

说话人：你说的云不好是什么意思

说话人：按以下指示？

说话人：请告诉我这是什么情况发生的。

说话人：比如说，具体情况是怎样的？

说话人：你问什么？

演讲人：克劳德的反应是什么？

说话人：我会喜欢这种程度的细节。

讲话人：我看到的是 80 左右

说话人：% of the people in the Early days

说话人：对于这次失败，克劳德不会

讲话人：写出正确的 JSON。然后，

说话人：好的，

讲话人：让我们生成也许从 30 开始

演讲人：举出 40 个克劳德不在场的例子

说话人：正确地做这件事，然后

说话人：那实际上是你的评估集，而你

说话人：本质上可以有一个提示和一个

说话人：回应。如果这不起作用

演讲人：你的正确黄金答案

说话人：可能有，

说话人：那么这意味着评估本质上是

说话人：是有益的，因为它可以识别

说话人：始终是一个痛点。然后

说话人：我们将其添加到我们的...存储库中

讲话人：用于评估。当我们有版本时

说话人：云的，

讲话人：我们实际上运行了该评估并进行了检查。

说话人：我想现在总是 100

说话人：%or like 99.9.所以不再是了

说话人：一个痛点。但是在早期它

说话人：正在接受用户反馈，

说话人：弄清楚他们的真正意思。

说话人：我们可以重现一下吗？

说话人：前后一致吗？

说话人：这是一个大问题吗？

说话人：然后弄清楚如何，

说话人：以一种可以的方式标准化它

说话人：研究人员的消耗品。基本上是

演讲人：测试驱动开发对于 PM 来说是

说话人：我们现在生活的世界，

说话人：你先在哪里写测试。所以是

说话人：这只是产品的核心部分

讲话人：现在在 Anthropicwriting 担任管理职务

发言者：埃瓦尔斯？

说话人：我也这么认为，我也觉得是这样的

演讲人：我和其他公司的其他 PM 谈过

说话人：关于，

说话人：而且我觉得也越来越多

演讲人：就更广泛的技能而言，

说话人：因为我们的很多产品

演讲人：建筑处于模型的交叉点

说话人：带着安全带，

说话人：用一组上下文来表示一组

说话人：用户。所以有像评估这样的东西

说话人：其实不只是...人的一种方式

说话人：从事模特工作，

说话人：但一般在产品内部，

说话人：为了获得更好的用户体验。因为

演讲人：你无法改进你无法衡量的东西。

说话人：其中很多都是非常静止的触觉

说话人：基于，

说话人：这仍然是非常基于判断的。所以

说话人：你必须注重细节。

讲话人：而且也非常不确定，

说话人：这是其中很重要的一部分。就像，

说话人：它不会给你同样的东西

说话人：每次都回答，所以你必须描述一下

说话人：有点更广泛。不会的

讲话人：要完全匹配。所以这真是一个

演讲人：产品方式的有趣变化

说话人：发生和将会发生是评估写作

演讲人：评估与 PRD 是其中的一个重要部分。

说话人：你们还做 PRD 吗？

说话人：还有像单页传呼机那样描述吗？

讲话人：有问题还是重播？

说话人：好吧，现在你摇头了。是的。

说话人：我们是，

说话人：我们愿意。我认为当有一个非常明确的

说话人：问题，

说话人：我认为像 evals 这样的事情可能就差不多了

说话人：简写，我想还有其他情况

演讲人：PRD 真正有价值的地方。PRD

说话人：是获得非常好的机会的好工具

说话人：一大群人排成一排

演讲人：关于经验的真实来源

讲话人：还有一系列目标。所以当我们确实有

说话人：一个模特，我们其实，对于每一个模特，

演讲人：我们确实有一个 PRD，

说话人：对于我们的研究人员来说不一定，

演讲人：但更多的是为了我们不断增长的产品面，

演讲人：对于我们的工程团队来说，

发言者：对于我们的利益相关者，如法律和

说话人：安全及其他，

说话人：只是作为事实的来源

演讲人：共同努力实现我们的目标

说话人：这样一大群人就可以划船

说话人：同一个方向，另一个地方

演讲人：我认为 PRD 有价值的地方是

讲话人：关于比较模糊的问题和机会。

说话人：所以如果我们还没有发货这样的东西

说话人：电脑使用，

说话人：我们不一定有一套类似的

说话人：用户总是有特定的痛点。并且

演讲人：我认为产品有价值

演讲人：PRD 中需要探索的愿景部分

说话人：什么可以，

说话人：即使技术还没有准备好

讲话人：为每个人工作。你如何得到

说话人：它对某些群体有效吗？

说话人：这样你就可以探索价值。你可以

讲话人：实际上带来连贯的东西

讲话人：针对一个用户组。所以我们确实有 PRD。

说话人：我认为应用程序有点不同

说话人：好的，这太棒了。我刚刚和安德鲁在一起，

演讲人：他是 OpenAI Codex 应用程序的负责人，

讲话人：你们是一致的。PRD 不是

说话人：死了，

讲话人：对于具体项目还是很有用的

说话人：和想法。太好了。好的，

演讲人：我们已经关掉了有关珠三角的书，但仍然

讲话人：踢。好吧，

说话人：所以我们一直在谈论一些关于

演讲人：正在出现什么样的技能

说话人：针对产品人员，有什么吗

00:50:00

英文原文（完整 ASR）

讲话人：else that you find is shifted in what

讲话人：patterns are common across people that

讲话人：are doing well in this new AI world

讲话人：in terms of product managers and folks

讲话人：on the product teams?

讲话人：Is there anything else that you're like,

讲话人：okay,

讲话人：there's something you got to shift or

讲话人：something you look for more people?

讲话人：I think maybe specifically for folks

讲话人：who might be mid-career or folks who

讲话人：have been more in a managerial product

讲话人：leadership seat.One thing that I think

讲话人：I feel pretty strongly about is in order

讲话人：to be good managers of teams and PMs

讲话人：working with this technology,

讲话人：you have to be really hands-on,

讲话人：right?

讲话人：and has spent not just time tinkering,

讲话人：but actually shipping with this technology

讲话人：and,and,and again,

讲话人：being in the details and sweating the

讲话人：tokens along with your PMs and your

讲话人：engineers and your teams.And so.Even

讲话人：for folks that I hire who have more

讲话人：tenured PM experience,

讲话人：the onboarding plans are exactly the

讲话人：same as somebody who is more early career.

讲话人：And it's around understanding users,

讲话人：reading content and user feedback,

讲话人：talking to customers.being able to like

讲话人：understand what to do with this,

讲话人：what good looks like and having developed

讲话人：that in a very hands-on manner that's

讲话人：important.It's not necessarily easy

讲话人：for someone to agree or be able to see

讲话人：what a good or great AI product or AI

讲话人：feature could look like if they haven't

讲话人：kind of experienced building themselves.

讲话人：So I think there is a,

讲话人：I do feel pretty strongly that like,

讲话人：you know,if you're a manager,

讲话人：you have to be hands-on,

讲话人：you have to spend a portion of your

讲话人：time actually shipping.You have to kind

讲话人：of walk in the shoes of your teams.

讲话人：And that's,

讲话人：I always try to carve out a portion

讲话人：of time to actually like own one to

讲话人：two work streams when we have models

讲话人：in order to keep like,

讲话人：keep my theory of mind,

讲话人：keep my sense of how the models are

讲话人：moving,

讲话人：how quickly it's improving.So I can

讲话人：help the team make decisions and make

讲话人：better decisions.What I'm hearing here

讲话人：is if you're not,

讲话人：no matter where you are in the ladder

讲话人：of hierarchy at a company,

讲话人：if you're not building yourself,

讲话人：if you're not actually talking to Clyde,

讲话人：talking to Codex,building stuff,

讲话人：you're not going to make it.And you

讲话人：should have fun working with his technology.

讲话人：I think that's the other piece.I think

讲话人：the folks that would be most successful,

讲话人：regardless of their level,

讲话人：are the people that are going to be

讲话人：the most successful.who love working

讲话人：with AI and are exploring and experimenting

讲话人：and carving out the time,

讲话人：not just for the experimentation,

讲话人：but actually hands-on shipping end to

讲话人：end,getting the user feedback,

讲话人：I think has to be fundamental for everyone.

讲话人：I a hundred percent know what you mean

讲话人：there.Just like me sitting on my newsletter

讲话人：and this podcast,

讲话人：just talking about stuff and like,

讲话人：yeah,yeah,

讲话人：that sounds great.Like every time I

讲话人：actually built something and I tinker

讲话人：with all kinds of little projects,

讲话人：you just like,okay,

讲话人：I see what's happening here and you

讲话人：just get so much.It's like hard to exactly

讲话人：describe what you're,what you,

讲话人：what you experience actually working

讲话人：with the models and building stuff,

讲话人：but it's like a whole different world

讲话人：of like,okay, I see.Here's what the look,

讲话人：here's what they're talking about.Computer

讲话人：use.Here's what they're talking about

讲话人：with this limitation of this UX situation.

讲话人：Yeah.Yeah.So it's just like,

讲话人：and you made this really interesting

讲话人：point that you have to have fun with

讲话人：it,

讲话人：which is not easy for a lot of people because

讲话人：they're pushed to use AI or they just

讲话人：don't know exactly what to do with it.

讲话人：For people that are just like,

讲话人：I don't know,

讲话人：it's just so annoying.I just have to

讲话人：do this.I don't know what,

讲话人：just like,

讲话人：I hate this frigging thing.Why do I

讲话人：have to work with this?

讲话人：Things are changing so much.I'm tired.

讲话人：Advice for helping people find that,

讲话人：find that joy in this work.I think maybe

讲话人：I'll reemphasize something I said earlier

讲话人：around just that experimentation is

讲话人：not an individual sport.Some of the

讲话人：moments where I think I've touched practically

讲话人：every version of research models across

讲话人：20 plus versions of production cloud

讲话人：at this point.I think part of the joy

讲话人：comes from seeing other people discover

讲话人：use cases too.And so maybe one idea

讲话人：here would be pairing with somebody

讲话人：who is excited and seeing what on a

讲话人：use case that you care about and working

讲话人：together versus identifying or trying

讲话人：to figure out the perfect use case yourself.

讲话人：Because that might feel like work,

讲话人：working with others feels like joy a

讲话人：lot of the time.And is there more that

讲话人：we could do to bring that,

讲话人：bring other people along?

讲话人：That's something like a lot of times

讲话人：internally we have somebody who is like

讲话人：very curious and them sharing an idea

讲话人：of a new prototype actually brings a

讲话人：ton more people who are like,

讲话人：oh,

讲话人：I didn't know this could work now with

讲话人：Claude.And so there's just some virtuous

讲话人：cycles here and ways of,

讲话人：yeah,

讲话人：continue to have joy with this technology.

讲话人：That's such a good point.I think that's

讲话人：also why Twitter is so useful for a

讲话人：lot of this is you see other people

讲话人：sharing what they've done.And it inspires

讲话人：you to come up with your own little

讲话人：ideas.And also,

讲话人：it's just like fun to share your own

讲话人：thing that you've done.So that's a really

讲话人：good point.Just like find other people

讲话人：to kind of play around with and share.

讲话人：for use cases.The thing I've also heard

讲话人：a lot is just find like a problem you

讲话人：want to solve in your life or work and

讲话人：just open up cloud code,

讲话人：tell it here's what I want to do.And

讲话人：it's incredible how far you can get

讲话人：just with like a vague idea of a problem

讲话人：you want to solve.Yeah.I think it gets

讲话人：hard in that there's so many different

讲话人：things that you could try.And so you

讲话人：just like narrowing in on either pairing

讲话人：with someone,

讲话人：working with somebody who is,

讲话人：who have a lot of joy about this technology

讲话人：or figuring out something that you could

讲话人：immediately find value.Like either of

讲话人：those things allow you to go deeper

讲话人：rather than like more high level about

讲话人：too many things.I find it hard to keep

讲话人：pace with the number of prototypes or

讲话人：products that are out there.And so my

讲话人：lens has been,

讲话人：how do I go deep in one to two of them

讲话人：myself?

讲话人：That's so interesting you say that,

讲话人：because that's exactly it.We just had

讲话人：the survey that I ran with my colleague

讲话人：Noam,

讲话人：asking my readers just how they're feeling

讲话人：about all the things going on in the

讲话人：tech right now and AI.And one of the

讲话人：most interesting takeaways we had was

讲话人：to find that happiness is exactly what

讲话人：you said,

讲话人：is go deep in a couple things.versus

讲话人：trying to just ton of little things,

讲话人：find a couple of things to really solve

讲话人：well,

讲话人：and then go deep.And that is a source

讲话人：because a lot of the happiness people

讲话人：feel is when they finally unlocked a

讲话人：way for AI to actually make their lives

讲话人：better versus just like a couple messed

讲话人：up,

讲话人：broken half working things.Yeah.It's

讲话人：how do you go from this being a check

讲话人：the box,right?

讲话人：And so like us as product people,

讲话人：it's then an exercise of product prioritization

讲话人：of your time and your energy.And if

讲话人：the goal is to experiment with joy,

讲话人：then how do you,

讲话人：what are the inputs that you need for

讲话人：that?

讲话人：Yeah.But yeah,I think a lot of the,

讲话人：I think the secret sauce of Anthropic

讲话人：is the culture and the bottoms of nature

讲话人：of how people work and this like experimenting

讲话人：in public.And by doing that,

讲话人：it's very much about how to bring other

讲话人：people along.That ends up being,

讲话人：I think,really valuable.Yeah,

讲话人：I've heard this so many times from all

讲话人：the labs,

讲话人：just like no one's exactly sure how

讲话人：some of this is going to be used.And

讲话人：a lot of it is just putting stuff out

讲话人：early,seeing how people use it,

讲话人：seeing what's possible,

讲话人：and then using that information to build

讲话人：the actual product to lean in.Yeah.

讲话人：I'm curious how,

讲话人：kind of on this thread of finding ways

讲话人：for AI to help you in your work in life,

讲话人：are there any interesting ways you've

讲话人：been using Cloud lately in your work

讲话人：as a PM?

讲话人：I think there's a lot of things with,

讲话人：you know,

讲话人：Fable and things like Tag.So I think

讲话人：Tag is in the very,like,

讲话人：early days,

讲话人：I think there's something around how

讲话人：you work in a different paradigm of

讲话人：allowing an agent to go off and work

讲话人：and then bring back product experiences

讲话人：to you.So I think one area that it's

讲话人：not more recent,

讲话人：but one that...I bring up a lot with

讲话人：the team and I think we could do more

讲话人：on using AI is just like how to use

讲话人：it to also be more,uh,

讲话人：to have better conversations with each

讲话人：other,

讲话人：to be better managers.I don't think

讲话人：it's necessarily just about raising

讲话人：the IQ of like experiences we build,

讲话人：but also I use it a lot and actually

讲话人：like prepping for how to have better

讲话人：conversations in the moment during like

讲话人：crucial conversations.So I love that

讲话人：book.And so I actually have a skill

讲话人：that helps me figure out Am I going

讲话人：in the right level of detail given the

讲话人：situation at hand and actually helping

讲话人：me be a better manager and better supporter

讲话人：for the team?

讲话人：So for managers on the team,

讲话人：that's actually a thing that I've been

讲话人：sharing more with our managers.Okay,

讲话人：how do you actually use Claude to make

讲话人：you a better coach?

讲话人：Because it's hard sometimes to find

讲话人：the right perfect words.And the model

讲话人：have a lot of perfect and right words.

讲话人：And I think there is something about

讲话人：how it can actually augment us from

中文译文（机器翻译，按本时间段分段）

说话人：否则你发现转移了什么

说话人：模式在人们中是常见的

说话人：在这个新的人工智能世界里表现得很好

讲话人：就产品经理和人员而言

发言者：关于产品团队？

说话人：你还有什么喜欢的吗？

说话人：好的，

说话人：有件事你必须改变一下或者

说话人：你找的人多吗？

讲话人：我想也许是专门针对人们的

讲话人：可能处于职业生涯中期的人或正在从事工作的人

说话人：更多的是在管理产品

演讲人：领导席。我认为的一件事

说话人：我强烈地认为这是正确的

演讲人：成为优秀的团队管理者和 PM

演讲人：使用这项技术，

说话人：你必须亲自动手，

说话人：对吗？

讲话人：不仅花时间修修补补，

说话人：但实际上是用这项技术发货的

说话人：还有、还有、再一次、

说话人：注重细节，挥汗如雨

讲话人：代币以及您的 PM 和您的

说话人：工程师和你的团队。等等。甚至

说话人：针对我雇用的拥有更多资源的人

演讲人：终身 PM 经验，

讲话人：入职计划正是

说话人：与职业生涯较早的人相同。

演讲人：围绕着理解用户，

说话人：阅读内容和用户反馈，

言语人：与顾客交谈。能够喜欢

说话人：明白该怎么做，

说话人：好的是什么样子的以及已经发展起来的

说话人：以一种非常实际的方式

说话人：重要，但不一定容易

言语人：表示某人同意或能够看到

说话人：多么好的或伟大的人工智能产品或人工智能

说话人：如果他们没有的话，功能可能看起来像

说话人：有点经验丰富的自我建设。

说话人：所以我认为有一个，

说话人：我确实强烈地感觉到，

说话人：你知道，如果你是经理，

说话人：你必须亲自动手，

说话人：你必须花费你的一部分

讲话人：实际发货时间。你得客气

说话人：站在你团队的角度思考。

说话人：那是，

说话人：我总是试图分出一部分

讲话人：有时间真正喜欢自己的人

演讲人：当我们有模型时有两个工作流程

说话人：为了保持喜欢，

说话人：保持我的思想理论，

说话人：让我了解模型的情况

说话人：感动，

说话人：改善得有多快，所以我可以

演讲人：帮助团队做出决策并做出

说话人：更好的决定。我在这里听到的

说话人：如果你不是，

讲话人：无论你在梯子的哪个位置

说话人：公司的等级制度，

说话人：如果你不去打造自己，

说话人：如果你实际上不是在和克莱德说话，

讲话人：与法典、建筑材料交谈，

讲话人：你不会成功的，而你

说话人：使用他的技术应该会很有趣。

说话人：我想那是另一件事。我想

说话人：最成功的人，

说话人：无论水平高低，

说话人：是那些将要成为的人

说话人：最成功、热爱工作的人

说话人：用 AI，正在探索和实验

说话人：并挤出时间，

说话人：不只是为了实验，

讲话人：但实际操作运输结束

演讲人：结束，获取用户反馈，

演讲人：我认为这对每个人来说都是最基本的。

说话人：我百分百明白你的意思

讲话人：那里，就像我坐在新闻通讯上一样

说话人：还有这个播客，

说话人：只是谈论一些事情之类的，

说话人：是啊，是啊，

说话人：听起来很棒，就像我每次

说话人：实际上建造了一些东西，但我修补了一下

演讲人：各种小项目，

说话人：你只是喜欢，好吧，

说话人：我明白这里发生了什么，你也明白

说话人：就这么多了，很难准确地

说话人：描述一下你是什么，你是什么，

说话人：你实际工作中的感受

说话人：有了模型和建筑材料，

说话人：但这就像一个完全不同的世界

说话人：就像，好吧，我明白了。这就是样子，

讲话人：这就是他们在谈论的内容。计算机

说话人：use.Here's what they're talk about

说话人：由于这种用户体验情况的限制。

说话人：是的，是的，所以就像，

说话人：你让这件事变得非常有趣

说话人：指出你必须玩得开心

说话人：它，

说话人：这对很多人来说并不容易，因为

说话人：他们被迫使用人工智能，或者他们只是

说话人：不知道该怎么办。

说话人：对于那些就像这样的人，

说话人：我不知道，

说话人：这太烦人了，我不得不这样做

说话人：这样做。我不知道什么，

说话人：就像，

说话人：我讨厌这种该死的事情。为什么我要这么做

讲话人：必须用这个吗？

说话人：事情变化太大了，我累了。

演讲人：帮助人们发现这一点的建议，

说话人：在这项工作中找到乐趣。我想也许

说话人：我再强调一下我之前说过的话

说话人：围绕着实验就是

说话人：不是个人运动。

说话人：我认为我已经触到实际的时刻

演讲人：各个版本的研究模型

演讲人：20 多个版本的生产云

说话人：此时此刻，我觉得有一部分是喜悦

说话人：来自于看到别人发现

演讲人：用例也是如此，所以也许有一个想法

讲话人：这里将与某人配对

说话人：谁很兴奋，看到了什么

说话人：你关心和工作的用例

说话人：一起与识别或尝试

说话人：自己找出完美的用例。

说话人：因为这可能感觉像是工作，

说话人：和别人一起工作感觉很快乐

说话人：很多时候。还有更多吗？

讲话人：我们可以做到这一点，

说话人：带其他人一起去吗？

说话人：这样的事情很多次了

说话人：我们内部有一个人，就像

说话人：很好奇，他们分享一个想法

说话人：一个新的原型实际上带来了

讲话人：更多的人喜欢，

说话人：哦，

说话人：我不知道这现在可以用

说话人：克劳德，所以只有一些有道德的人

演讲人：这里的循环和方式，

说话人：是的，

演讲人：继续享受这项技术带来的乐趣。

说话人：这是一个很好的观点。我认为那是

说话人：这也是为什么 Twitter 对于一个人来说如此有用

说话人：很多都是你看到别人的

说话人：分享他们所做的事情。这鼓舞人心

讲话人：你想出你自己的小

说话人：想法。而且，

说话人：分享自己的观点就像很有趣

讲话人：你已经做过的事情，所以这真的是

说话人：说得好，就像找其他人一样

言语人：一种玩耍和分享的方式。

说话人：针对用例。我也听说过

说话人：很多只是发现像你这样的问题

说话人：想解决你生活或工作中的问题

说话人：只要打开云代码，

说话人：告诉它这就是我想做的。

说话人：你能走多远真是令人难以置信

说话人：只是对一个问题有一个模糊的想法

说话人：你想解决。是的。我想它得到了

说话人：很难，因为有这么多不同的东西

说话人：你可以尝试的事情，所以你

讲话人：就像缩小任一配对的范围一样

说话人：与某人，

说话人：与这样的人一起工作，

发言者：谁对这项技术感到非常高兴

说话人：或者想出一些你可以做的事情

说话人：立即找到价值。就像其中之一

说话人：那些东西可以让你走得更深

说话人：而不是喜欢更高层次的谈论

说话人：东西太多了，记不住了

讲话人：与原型数量或

说话人：产品就在那里。所以我的

说话人：镜头已经，

说话人：我如何深入其中一到两个

说话人：我自己？

说话人：你说的真有趣，

说话人：因为就是这样。我们刚刚

讲话人：我和同事进行的调查

说话人：诺姆，

说话人：询问我的读者感受如何

说话人：关于发生在里面的所有事情

演讲人：现在的科技和人工智能。其中之一

演讲人：我们最有趣的收获是

说话人：发现幸福到底是什么

说话人：你说，

说话人：深入探讨几件事。

说话人：试图做很多小事，

说话人：找到一些真正需要解决的事情

说话人：嗯，

说话人：然后深入。这就是来源

说话人：因为幸福的人很多

讲话人：感觉是当他们终于解锁了

说话人：人工智能真正创造生活的方式

说话人：比像一对搞砸的情侣更好

说话人：上，

讲话人：坏掉了一半的工作。是的。它是

说话人：你如何看待这成为一张支票

说话人：盒子，对吧？

演讲人：就像我们作为产品人员一样，

说话人：这就是产品优先级的练习

说话人：你的时间和精力。如果

说话人：我们的目标是带着快乐去尝试，

说话人：那你呢，

说话人：您需要什么输入

说话人：那个？

说话人：是的。但是，是的，我认为很多，

演讲人：我认为 Anthropic 的秘密武器

言语人：是文化，是自然的底蕴

讲话人：关于人们如何工作以及诸如实验之类的事情

讲话人：在公共场合。通过这样做，

讲话人：这非常关乎如何带其他人

演讲人：大家一起。最终结果是，

说话人：我认为，真的很有价值。是的，

说话人：我已经从所有人那里听过很多次了

发言者：实验室，

说话人：就像没有人确切知道如何做一样

讲话人：其中一些将被使用。并且

讲话人：很多只是把东西放出来

言语人：早期，看人们如何使用它，

说话人：看看有什么可能，

演讲人：然后使用该信息来构建

说话人：要靠的实际产品。是的。

说话人：我很好奇，

说话人：有点像在寻找方法

说话人：让 AI 在工作生活中帮助你，

说话人：你有什么有趣的方式吗？

说话人：最近工作中一直在使用云

说话人：作为总理？

说话人：我认为有很多事情，

说话人：你知道，

说话人：寓言之类的东西，所以我认为

讲话人：标签是在非常，喜欢，

说话人：早期，

说话人：我认为有一些事情是围绕如何

说话人：你的工作范式不同

说话人：允许代理人下班去工作

说话人：然后带回产品体验

说话人：对你来说，所以我认为这是一个领域

说话人：不是最近的，

说话人：但是……我提出了很多

发言者：团队和我认为我们可以做得更多

演讲人：使用 AI 就像如何使用

说话人：还可以，呃，

说话人：与每个人进行更好的对话

说话人：其他，

说话人：成为更好的管理者。我不认为

说话人：肯定只是为了提高

说话人：我们建立的类似经历的智商，

说话人：但我也经常使用它，而且实际上

说话人：就像准备如何过得更好

说话人：在喜欢的时候进行的谈话

说话人：关键对话，所以我喜欢这样

说话人：书。所以我其实有一技之长

说话人：这可以帮助我弄清楚我要去吗

说话人：在适当的细节水平上考虑到

说话人：眼前的情况和实际的帮助

讲话人：我是一个更好的管理者和更好的支持者

发言者：为了团队？

演讲人：所以对于团队中的管理者来说，

说话人：这实际上是我一直以来的事情

演讲人：与我们的经理分享更多。好的，

说话人：你实际上是如何使用 Claude 来制作

说话人：你是一个更好的教练吗？

说话人：因为有时很难找到

说话人：正确的完美话语。还有模型

说话人：有很多完美而正确的话语。

说话人：我认为有一些关于

说话人：它实际上如何增强我们的能力

01:00:00

英文原文（完整 ASR）

讲话人：an EQ perspective in addition to the

讲话人：Q.Oh,

讲话人：man.There's so much interesting stuff

讲话人：there.So just to understand what you're

讲话人：doing there.So you built a skill.You're

讲话人：just like,Claude,

讲话人：build a skill.pulling in lessons from

讲话人：Crucial Conversations,the book,

讲话人：which it knows enough about,

讲话人：you don't have to even give it the content.

讲话人：And then you use that skill to talk

讲话人：to Claude,hey,

讲话人：I have this very difficult conversation

讲话人：coming up with a colleague,

讲话人：give me some tips on how to approach

讲话人：it.Yeah,and it's a great,

讲话人：it's almost like,coaching,

讲话人：like individualized,

讲话人：personalized coaching of just how to

讲话人：make you...And there's so much context

讲话人：switching that we do all day.And having

讲话人：like Claude help me pair and help me...

讲话人：And maybe there are times where I end

讲话人：up not using suggestions from Claude,

讲话人：but it actually ends up being very helpful

讲话人：for just coming up and brainstorming.

讲话人：Am I thinking about...reactions in the

讲话人：right way?

讲话人：How do I actually go a bit deeper,

讲话人：faster,build trust faster,

讲话人：be more direct?

讲话人：Yeah,man,

讲话人：I have so many questions here.This is

讲话人：so interesting.One is just like there's

讲话人：concern people are going to start talking

讲话人：the way AI writes because they're talking

讲话人：to AI so much.And it's going to be like,

讲话人：Diane,it's not this,

讲话人：but it's that.I know that you're not

讲话人：doing that,

讲话人：but that's a concern people have.Let

讲话人：me just ask about that,

讲话人：I guess.Do you fear this?

讲话人：There's this brain rot,

讲话人：atrophy stuff.People talk about it.

讲话人：We're just so reliant on AI now and

讲话人：we stop.learning and thinking and,

讲话人：you know,overall AI thoughts on that,

讲话人：being so close to it and being so integrated

讲话人：with AI constantly.A lot of actually

讲话人：thinking process and writing process

讲话人：are tied together for me personally.

讲话人：And so I think There are ways where

讲话人：I use Claude to augment my thinking,

讲话人：but what I wanna make sure,

讲话人：and maybe this is what you're describing

讲话人：is,

讲话人：Claude doesn't take over all of my thinking

讲话人：for me.And so I think depending on the

讲话人：situation,

讲话人：depending on how much more personal

讲话人：judgment I wanna have in a situation,

讲话人：I might come up with my own POV first

讲话人：and then work with Claude through that.

讲话人：and making sure that like I maintain

讲话人：my sense and tone throughout.I think

讲话人：there are then other things like updates,

讲话人：right?

讲话人：We have like monthly business reviews.

讲话人：And then in those cases,

讲话人：it's much more,

讲话人：I actually want it to be standard.And

讲话人：I want it to be much more like it gets

讲话人：a crystallized information in the right

讲话人：way.And I have a skill and we're augmenting

讲话人：and improving our skill for that.But

讲话人：I want to get to a place where like

讲话人：the monthly business review,

讲话人：the writing of that is potentially–

讲话人：asymmetrically less valuable than the

讲话人：thinking.And so how do I get that piece

讲话人：delegated to Claude fully?

讲话人：And I'm more of a reviewer and a verifier

讲话人：of that information.So I think it depends

讲话人：on like what you're using Claude for

讲话人：and what you're trying to convey.And

讲话人：like,

讲话人：is there asymmetrical value in delegating

讲话人：more to Claude?

讲话人：What I'm also hearing,

讲话人：the first tip is really great,

讲话人：which was think first,

讲话人：have a point of view,

讲话人：and then kind of use Cloud as a sparring

讲话人：partner almost to evolve the idea,

讲话人：push back on the idea.Yeah.Yeah.And

讲话人：I think this is where things like actually

讲话人：our alignment research and safety research

讲话人：is helpful because what you don't want

讲话人：is like an AI that just agrees with

讲话人：you.Right.What you want is this technology

讲话人：to actually augment and grow and like

讲话人：get to a better outcome.And so sometimes

讲话人：it's having Claude push back makes me

讲话人：better.And so that's great.Like a coworker,

讲话人：I want somebody to push back when my

讲话人：ideas are not fully formed.I want to

讲话人：hear more about that.I've heard that

讲话人：when Ben Mann was on the podcast,

讲话人：he talked about the constitution that

讲话人：is built into Claude and how unintuitively

讲话人：the work and the focus on safety and

讲话人：alignment,as you said,

讲话人：and this constitution that describes

讲话人：how Claude should think and operate.

讲话人：You would think that would limit safety.

讲话人：of Claude and make it less fun and interesting.

讲话人：It's exactly the opposite.Claude is

讲话人：the most interesting personality.I hear

讲话人：that constantly.It's just like I much

讲话人：prefer talking to like OpenClaude famously

讲话人：was built on Claude and then People

讲话人：were forced to switch.We won't get into

讲话人：it.We're forced to switch to JAPI-T

讲话人：and they're like,

讲话人：this is so bad.This is not who I'm used

讲话人：to talking to.So that is,

讲话人：I think,

讲话人：a really interesting point.I just want

讲话人：to make sure we spend a little time

讲话人：on why is it why is that the case?

讲话人：Just this focus on alignment,

讲话人：safety,

讲话人：having this clear constitution.Why does

讲话人：that make Claude better and more interesting

讲话人：to talk to you also?

讲话人：Yeah.Claude as intelligent and as capable

讲话人：as possible,

讲话人：being able to have Claude actually push

讲话人：back in the right points and then add,

讲话人：it's like a yes or no and,

讲话人：actually helps you come to a better

讲话人：conclusion.Yeah.used Cloud to help with

讲话人：things like,

讲话人：are we making the right pricing decision

讲话人：on the next version of Cloud?

讲话人：It's a little bit meta,

讲话人：but using a research version of Opus,

讲话人：asking it to figure out how it should

讲话人：price and being able to come out with

讲话人：better outcomes is a goal at the end

讲话人：of the day.And so having AI not just

讲话人：be an assistant,

讲话人：not just be a doer and being delegated

讲话人：tasks,but figuring out,

讲话人：is it doing the right thing?

讲话人：That's actually very integrated with

讲话人：knowing when to push back,

讲话人：right?

讲话人：Right.That's part of knowing when you

讲话人：should be proactive.Proactivity is not

讲话人：as necessarily always doing a thing

讲话人：that you are scheduled to do.It is knowing

讲话人：when to come up with a new idea.And

讲话人：so in order for cloud to be more useful,

讲话人：the general approach has to be that

讲话人：it knows when to push back.It's a core

讲话人：part of the characteristics together

讲话人：of the models.That is so interesting.

讲话人：It's so interesting that that is what

讲话人：a big part of it,

讲话人：like it being less compliant is almost

讲话人：what makes it better and more useful

讲话人：because we need that.Like I've had so

讲话人：many people where they're like,

讲话人：hey,

讲话人：like AI told me I was right.And like,

讲话人：no,

讲话人：I wish I was to other people.Yeah.And

讲话人：it comes back to our earlier point around

讲话人：thinking.Right.How do you protect your

讲话人：thinking?

讲话人：If you have a AI that can be a thinking

讲话人：partner,

讲话人：a thinking partner.doesn't just agree

讲话人：with you,

讲话人：it should add to you.And you should

讲话人：come away at the end of the day having

讲话人：better ideas because you worked with

讲话人：Claude.That should be the hero goal,

讲话人：not just making your ideas 10%better.

讲话人：Yeah,

讲话人：I love this.It used to be think 10x.

讲话人：That used to be the way founders push

讲话人：people.Like,what if we 10x this?

讲话人：And I love what I keep hearing is like,

讲话人：it's like,

讲话人：how do we go 1000x from this idea?

讲话人：What is the most ambitious version of

讲话人：this?

讲话人：I want to come back to something that

讲话人：I was thinking about as we were talking

讲话人：about talking to Claude constantly.

讲话人：It's very clear when AI has written

讲话人：something still.It's funny that it's

讲话人：a large language model.You would think

讲话人：of all things.It would be very good

讲话人：at writing.And interestingly,

讲话人：just no AI is very good at writing.

讲话人：It's always very clear.This was AI written.

讲话人：Do you think we'll get to a place where

讲话人：we will not know this was AI?

讲话人：I think it depends on what's the goal

讲话人：that you're looking to achieve.What's

讲话人：the eval?

讲话人：Yeah,what's the eval?

讲话人：I actually do think there's more that

讲话人：we could be doing on making Claude write

讲话人：better.There's actually very active

讲话人：efforts on my team and on the research

讲话人：side about making Claude write better,

讲话人：just generally.I think it should be

讲话人：clear where an idea is being led by

讲话人：you or by Yuleni or me,

讲话人：Diane.really depends on what's the goal

讲话人：of that writing.Like for something like

讲话人：a monthly business review,

讲话人：I would actually love to have that end

讲话人：-to-end be written by cloth.And obviously,

讲话人：and not make it feel like it was written

讲话人：by a human.That's such an interesting

讲话人：point you're making.Like,

讲话人：is it actually better for us to know

讲话人：that it's AI versus not?

讲话人：Yeah.But it's also for maybe the lens

讲话人：is more around like verifiability or

讲话人：who's verifying the output.Right.Like

讲话人：who's signing off,

讲话人：maybe less around who's writing,

讲话人：but who's verifying who's signing off.

讲话人：That becomes like more what matters

讲话人：than who's writing it.Why do you think

讲话人：AI is not great at writing?

讲话人：Yeah.my guess is it has studied all

讲话人：of the best writing and all of humanity.

讲话人：It's figured out here's the best way

讲话人：to write.And now that we,

讲话人：and it's just,

讲话人：there's only so many ways to write.

讲话人：And so we've just recognized,

讲话人：okay,

讲话人：this is what AI does.It has these tropes.

讲话人：Is that the core of it?

讲话人：Is there something else that's keeping

讲话人：it from being a great writer?

讲话人：Ironically,

讲话人：being a large language model of all

讲话人：things,

讲话人：you think it'd be really great at language.

讲话人：I think part of this also,

讲话人：we need to invest more in training improvements

讲话人：to make AI continuously strong on areas

讲话人：like writing.I think it's also,

讲话人：Like the technology is jagged edged,

讲话人：like we mentioned.So sometimes when

讲话人：the models were good at writing,

讲话人：but not agentic,our thesis is,

讲话人：how do we make the models more agentic

讲话人：or call the right tools?

讲话人：Now that that's improved a bit,

讲话人：then it's,well,

讲话人：now these other areas actually become

讲话人：more of the rough edges.And so I think

讲话人：we're in one of those moments with writing

讲话人：where,

讲话人：yeah.we need to actually just focus

中文译文（机器翻译，按本时间段分段）

说话人：除了情商角度

说话人：Q.哦，

说话人：伙计，有趣的东西真多

讲话人：那儿。所以只是为了了解你在做什么

说话人：在那里做。所以你建立了一项技能。你

说话人：就像，克劳德，

演讲人：培养一项技能。汲取教训

演讲人：关键对话，这本书，

说话人：它对此足够了解，

讲话人：你甚至不必给出内容。

说话人：然后你用那个技巧去说话

说话人：对克劳德来说，嘿，

说话人：我的谈话非常困难

说话人：和一位同事一起来，

说话人：给我一些如何接近的技巧

说话人：是的，这很棒，

说话人：这几乎就像，教练，

说话人：喜欢个性化，

演讲人：个性化指导如何

说话人：让你...而且有这么多的背景

说话人：我们整天都在切换。

说话人：就像克劳德帮我配对，帮我……

说话人：也许有时候我会结束

说话人：不使用克劳德的建议，

说话人：但实际上最终非常有帮助

演讲人：刚刚上来进行头脑风暴。

说话人：我在想……人们的反应吗？

讲话人：正确的方式吗？

说话人：我实际上如何更深入一点，

说话人：更快，更快建立信任，

说话人：更直接一点吗？

说话人：是的，伙计，

讲话人：我这里有很多问题。这是

讲话人：真有趣。一个就像那里一样

说话人：担心人们会开始说话

说话人：人工智能写作的方式，因为他们在说话

演讲人：对人工智能非常感兴趣。它会是这样的，

说话人：黛安，不是这个，

说话人：但就是这样。我知道你不是

说话人：这样做，

说话人：但这是人们所关心的问题。让

说话人：我只是问一下这个，

说话人：我猜，你害怕这个吗？

说话人：这脑子烂了，

说话人：萎缩的东西。人们谈论它。

演讲人：我们现在非常依赖人工智能

说话人：我们停止学习和思考，

说话人：你知道，人工智能对此的总体想法，

说话人：如此接近，如此融为一体

说话人：不断地用 AI。其实很多

说话人：思维过程和写作过程

讲话人：对我个人来说是捆绑在一起的。

说话人：所以我认为有一些方法可以

说话人：我用克劳德来增强我的思维，

说话人：但是我想确定的是，

说话人：也许这就是你所描述的

说话人：是，

演讲人：克劳德并没有接管我所有的思考

说话人：对我来说。所以我认为取决于

说话人：情况，

说话人：取决于个人的程度

说话人：我想在某种情况下做出判断，

说话人：我可能会先想出我自己的观点

说话人：然后和克劳德一起完成这个工作。

说话人：并确保我保持这样的态度

说话人：我的感觉和语气贯穿始终。我认为

说话人：还有其他的事情比如更新，

说话人：对吗？

演讲人：我们每月都有业务回顾。

说话人：然后在那些情况下，

说话人：还有更多，

讲话人：我实际上希望它成为标准。并且

说话人：我希望它更像它得到的那样

说话人：右边的具体信息

说话人：方式。我有一项技能，我们正在增强

说话人：并为此提高我们的技能。但是

说话人：我想去一个地方

演讲人：每月的业务回顾，

说话人：这样的写法可能是——

说话人：不对称地不如

讲话人：思考。那我怎样才能得到那篇文章

讲话人：完全委托给克劳德？

说话人：而我更多的是一个审稿人和验证者

说话人：这个信息。所以我认为这取决于

说话人：就像你用克劳德做的那样

说话人：以及你想表达的意思。

说话人：比如，

说话人：授权是否存在不对称价值

说话人：还有更多关于克劳德的事吗？

说话人：我也听到了，

说话人：第一个提示真的很棒，

讲话人：这是首先想到的，

说话人：有一个观点，

演讲人：然后用云作为陪练

演讲人：合作伙伴几乎是为了发展这个想法，

说话人：反驳这个想法。是的。是的。还有

讲话人：我认为这就是事情的实际情况

演讲人：我们的对准研究和安全研究

说话人：很有帮助，因为你不想要什么

说话人：就像一个人工智能，只是同意

说话人：你，对，你想要的就是这个技术

说话人：实际上增强、成长和喜欢

说话人：得到更好的结果。所以有时

说话人：克劳德的反击让我很感动

说话人：更好。所以那太好了。就像同事一样，

说话人：当我的

说话人：想法还没有完全形成，我想

说话人：多听听。我听说过

说话人：当本·曼 (Ben Mann) 上播客时，

讲话人：他谈到了宪法

说话人：是克劳德内置的，多么不直观

讲话人：工作和关注安全

说话人：对齐，正如你所说，

说话人：这个宪法描述了

演讲人：克劳德应该如何思考和运作。

说话人：你可能会认为这会限制安全。

说话人：克劳德，让它变得不那么有趣和有趣。

说话人：恰恰相反，克劳德是

说话人：最有趣的个性。我听说

讲话人：一直这样。就像我一样

说话人：更喜欢和像 OpenClaude 这样著名的人说话

言语人：是建立在克劳德和人民的基础上的

说话人：被迫转换，我们不会进入

说话人：它。我们被迫切换到 JAPI-T

说话人：他们就像，

说话人：这太糟糕了，这不是我习惯的人

说话人：to talk to. So that is,

说话人：我认为，

讲话人：一个非常有趣的观点。我只是想

说话人：确保我们花一点时间

说话人：关于为什么会这样？

讲话人：就这个重点来说，

说话人：安全，

说话人：有这么明确的宪法，为什么

说话人：这让克劳德变得更好、更有趣

说话人：也跟你说话吗？

说话人：是啊，克劳德一样聪明，一样能干

说话人：尽可能，

说话人：能够让克劳德真正推动

演讲人：回到正确的观点，然后添加，

说话人：这就像是或否，并且，

说话人：实际上可以帮助你变得更好

说话人：结论。是啊。用了云来帮忙

说话人：比如，

演讲人：我们做出正确的定价决策吗

说话人：关于下一个版本的云？

说话人：有点元，

讲话人：但是使用 Opus 的研究版本，

说话人：要求它弄清楚应该如何做

说话人：价格和能够拿出来

演讲人：更好的结果是最终的目标

演讲人：of the day. 所以拥有人工智能不仅仅是

说话人：当助理，

演讲人：不只是做一个实干家和被委派

说话人：任务，但是搞清楚，

说话人：这样做正确吗？

说话人：这其实是非常结合的

说话人：知道什么时候该反击，

说话人：对吗？

讲话人：是的，这是了解你什么时候的一部分

说话人：应该主动，主动则不然

讲话人：必然总是做一件事

说话人：你预定要做的事，就是知道

说话人：什么时候想出一个新想法。

演讲人：所以为了让云变得更有用，

说话人：一般的做法必须是这样的

说话人：它知道什么时候该反击，它是一个核心

讲话人：部分特征组合在一起

演讲人：模特的。太有趣了。

说话人：太有趣了，原来是这样

说话人：很大一部分，

说话人：好像不那么顺从就差不多了

说话人：是什么使它更好、更有用

说话人：因为我们需要那个，就像我曾经经历过的那样

说话人：很多人都喜欢，

说话人：嘿，

说话人：就像人工智能告诉我我是对的一样，

说话人：不，

说话人：我希望我是对其他人来说的。是的。而且

说话人：又回到我们之前的观点了

讲话人：思考。对。你如何保护你的

说话人：思考？

演讲人：如果你有一个会思考的人工智能

说话人：合伙人，

说话人：一个有思想的伙伴。不只是同意

说话人：和你一起，

讲话人：它应该给你加分，而你应该

讲话人：一天结束时离开

演讲人：更好的想法，因为与你一起工作

说话人：克劳德，这应该是英雄目标，

演讲人：不仅仅是让你的想法变得更好 10%。

说话人：是的，

演讲人：我喜欢这个。以前是 think 10x。

说话人：这曾经是创始人推动的方式

说话人：人们。比如说，如果我们把这个放大 10 倍会怎样？

说话人：我喜欢我不断听到的声音，

说话人：就像，

演讲人：我们怎样才能把这个想法实现 1000 倍呢？

说话人：最雄心勃勃的版本是什么

说话人：这个？

说话人：我想回到一些事情

说话人：我们说话的时候我在想

说话人：关于经常和克劳德说话。

说话人：AI 写的时候就很清楚了

说话人：还是有些东西。有趣的是

说话人：一个大的语言模型。你会认为

说话人：在所有事情中，那就太好了

说话人：在写作时。有趣的是，

说话人：只是没有人工智能非常擅长写作。

说话人：总是很清楚，这是 AI 写的。

说话人：你认为我们会到达一个地方吗？

演讲人：我们不会知道这是人工智能吗？

说话人：我认为这取决于目标是什么

说话人：你想要实现的目标是什么

说话人：评估？

说话人：是的，评估结果是什么？

说话人：我实际上确实认为还有更多

说话人：我们可以让克劳德写信

说话人：好多了，其实很活跃

演讲人：我的团队和研究的努力

演讲人：关于让克劳德写得更好的方面，

说话人：一般般吧，我觉得应该是

说话人：清楚一个想法是由哪里引导的

说话人：你或者尤莱尼或者我，

说话人：Diane，这真的取决于目标是什么

讲话人：那种写作。就像这样的东西

演讲人：每月一次的业务回顾，

说话人：我其实很想有这样的结局

说话人：-到最后都是用布写的。显然，

说话人：不要让它感觉像是写下来的

说话人：由一个人说话。这真是太有趣了

说话人：你正在表达的观点。比如，

说话人：我们真的知道更好吗

说话人：人工智能与非人工智能？

说话人：是啊，不过也许也是为了镜头

说话人：更像是可验证性或

讲话人：谁在验证输出。对。喜欢

说话人：谁在签字，

说话人：也许不那么围绕谁在写，

说话人：但是谁在验证谁在签字。

说话人：这变得更重要了

说话人：比谁写的，你认为为什么

说话人：AI 不擅长写作？

说话人：是的，我猜它已经研究了所有

说话人：最好的写作和全人类。

说话人：我发现这是最好的方法

说话人：写。现在我们，

说话人：只是，

说话人：写作的方式就这么多。

说话人：所以我们刚刚认识到，

说话人：好的，

说话人：这就是人工智能所做的。它有这些比喻。

演讲人：这就是核心吗？

说话人：还有什么事情需要保留吗？

说话人：是因为他是一位伟大的作家吗？

说话人：讽刺的是，

说话人：成为所有人的大型语言模型

说话人：事情，

言语人：你认为它在语言方面真的很棒。

说话人：我认为这也是一部分原因，

演讲人：我们需要在培训改进上投入更多

演讲人：让 AI 在各领域不断强大

说话人：就像写作一样。我认为也是，

说话人：就像技术是锯齿状的边缘一样，

说话人：就像我们提到的，所以有时当

说话人：模特们都擅长写作，

说话人：但不是代理，我们的论点是，

说话人：如何让模型更具有代理性

说话人：或者调用合适的工具？

说话人：现在情况有所改善了，

说话人：那么，嗯，

说话人：现在这些其他领域实际上变成了

讲话人：更多的是粗糙的边缘。所以我认为

说话人：我们正处于写作的时刻之一

说话人：哪里，

说话人：是的，我们实际上需要集中注意力

01:10:00

英文原文（完整 ASR）

讲话人：and prioritize on training the models

讲话人：to be like great at this area and like

讲话人：that is an active a very active area

讲话人：for us so fun like you mentioned Okay,

讲话人：I'm glad.I'm glad and also it's going

讲话人：to be interesting once AI is so good.

讲话人：We're like,

讲话人：I don't know who wrote that.But to your

讲话人：point,

讲话人：sometimes we actually want to know that

讲话人：it's AI.That's really interesting.I

讲话人：never thought of it that way.The other

讲话人：interesting part of this is that there's

讲话人：that comedian who was joking that we're

讲话人：like on a plane and the Wi-Fi is down

讲话人：and we're just like,what the hell,

讲话人：the Wi-Fi is not working on this plane.

讲话人：This sucks.How dare you?

讲话人：When you're like in a tube in the sky

讲话人：flying like a bird.And how dare you

讲话人：complain that the Wi-Fi doesn't work?

讲话人：Like your point is there's so much advancement

讲话人：and so much power.We can't fix it all.

讲话人：We can't make it all work the best possible.

讲话人：And so basically writing has been not

讲话人：the priority and it feels like there's

讲话人：more investment happening there.Yeah,

讲话人：I think like tone and character is a

讲话人：priority.I think it's this advancement

讲话人：of the technology is a work in progress.

讲话人：And so we made,

讲话人：we see a leap or emergence of like a

讲话人：jump in agentic behaviors.And so that

讲话人：is a new normal.And then these other

讲话人：capabilities need to continue like improving.

讲话人：And I think once we improve,

讲话人：let's say writing and like tone and

讲话人：character,we probably will say like,

讲话人：how do we have cloud be even more proactive?

讲话人：Like,

讲话人：proactivity is an opportunity.And that's

讲话人：human nature.Like,

讲话人：we want to make ourselves better.We

讲话人：want to make this technology better.

讲话人：So,yeah,

讲话人：and I think we're applying it to AI,

讲话人：which is the right thing.We should be

讲话人：making it better.coming.Um,

讲话人：one is where do you think human brains

讲话人：will continue to be most valuable over

讲话人：the years?

讲话人：I know Anthropix mission and,

讲话人：and vision is we'll reach a GI,

讲话人：a super intelligence.So in the future,

讲话人：maybe nowhere, but before we get there,

讲话人：where do you think human brains will

讲话人：continue to be most valuable as we approach

讲话人：that,that timeline?

讲话人：We started to talk about making Claude

讲话人：and models better at judgment.especially

讲话人：in the last year or so,

讲话人：I think judgment is one,

讲话人：is an area where it's a accumulation

讲话人：of so much nuance and so much experience,

讲话人：and these systems haven't experienced

讲话人：as much as humans have.And so I think

讲话人：that hard earned,like judgment,

讲话人：is a area for product leaders and just

讲话人：generally will continue to be really

讲话人：critical There are so many things AIs

讲话人：can build.Which one are the things that,

讲话人：you know,an org like Lab should build,

讲话人：right?

讲话人：A lot of that requires like human judgment,

讲话人：persistence.So,proactivity,

讲话人：these are all traits that are beyond

讲话人：just general capabilities,

讲话人：but just behaviors and characteristics

讲话人：of like people at that level of like,

讲话人：how do you get to the best solutions?

讲话人：How do you create the best experiences?

讲话人：So,

讲话人：I think those types of traits are actually

讲话人：the tactile traits that I think will

讲话人：be,

讲话人：continue to be important.I think there

讲话人：is also,

讲话人：still a lot of like capabilities and

讲话人：subject matter expertise as well.I think,

讲话人：you know,

讲话人：software engineering has been really

讲话人：transformed by AI.I think there's areas

讲话人：like biology,

讲话人：life sciences.These are all things that

讲话人：we're just kind of at like the foot

讲话人：of the exponential on,

讲话人：like maybe software engineering,

讲话人：we're on the exponential,

讲话人：on some of these other areas,

讲话人：we're not quite there yet.And so I think

讲话人：you're seeing us ship things like cloud

讲话人：science,investing in these areas,

讲话人：because those are areas that I think

讲话人：is just bring this technology to society

讲话人：and having a positive benefit for society.

讲话人：So I think there's a lot more to go

讲话人：there.Another question I want to ask

讲话人：is,

讲话人：as someone with kids...How do you think

讲话人：about what you are encouraging them

讲话人：to learn?

讲话人：Where do you think you're going to nudge

讲话人：them to be successful in this wild new

讲话人：world that we're entering?

讲话人：I actually think it's a lot of the same

讲话人：traits like you and I probably grew

讲话人：up with,which is curiosity for learning,

讲话人：persistence,

讲话人：believing in your own inner voice,

讲话人：developing and then believing in your

讲话人：own inner voice.Like I have a four-year

讲话人：-old,I have a eight-year-old,

讲话人：it's on us to help,

讲话人：it's on me to help them develop their

讲话人：inner voice and whether that's being

讲话人：opinionated and, Taking a stance to me,

讲话人：right?

讲话人：And developing that,

讲话人：encouraging that.I think that those

讲话人：types of skill sets are things that

讲话人：is important in the future.And I like

讲话人：having their own individual voice.That

讲话人：is so interesting.It's so related to

讲话人：the answer you had when asked about

讲话人：how to avoid a brain rot,

讲话人：essentially an over-relying in AI,

讲话人：which is just keep focused on your own

讲话人：point of view and your own perspective

讲话人：before you over-rely in AI and just

讲话人：this idea you're describing of building

讲话人：that in kids is,

讲话人：is really important.Uh,

讲话人：that is so interesting.And I love how

讲话人：all this,all this kind of connects judgment,

讲话人：persistence in a,

讲话人：a point of view of your own.Yeah.Both

讲话人：for kids and also adults.Yeah.Anything

讲话人：you think about for your...Oh,

讲话人：man.Well,

讲话人：like the question I'm thinking about

讲话人：is just when to get them on like some

讲话人：AI thing,you know,

讲话人：when I have a three-year-old,

讲话人：so it's pretty early for that.But,

讲话人：you know,how do you get,

讲话人：how do you onboard them to this crazy

讲话人：thing?

讲话人：I was at an event recently and a bunch

讲话人：of parents were talking about how they

讲话人：think about AI in their kids.And one

讲话人：person had a really interesting approach,

讲话人：which is keep them on the very early

讲话人：models of...So that they still have

讲话人：to struggle a bit and not get all the

讲话人：answers immediately.I thought that was

讲话人：interesting,

讲话人：like an open source local model,

讲话人：not fable.Yeah.Yeah.And curiosity is

讲话人：something I keep mentioning Ben Mann,

讲话人：but his answer actually to this question

讲话人：has always stuck with me,

讲话人：which is curiosity.And also just like

讲话人：he's a big fan of Montessori,

讲话人：which is what I'm encouraging for our

讲话人：kids.So there's something there.Maybe

讲话人：a last question,

讲话人：just kind of along these lines,

讲话人：something Fiona Fung actually suggested

讲话人：I ask you,who's recently on the podcast,

讲话人：how do you stay just recharged and not

讲话人：burn out,

讲话人：being in the center of this crazy storm

讲话人：of AI?

讲话人：As a mom working in,you know,

讲话人：we're seeing the research work at Anthropic.

讲话人：I just like,

讲话人：we're living through the most unprecedented

讲话人：time working at just like,

讲话人：being on the outside of Anthropic,

讲话人：it's crazy.I don't even know what it's

讲话人：like to be on the inside.What have you

讲话人：learned about avoiding burnout,

讲话人：staying recharged,

讲话人：staying sane during the middle of all

讲话人：this?

讲话人：In 2024,

讲话人：we shipped four models in the whole

讲话人：year,

讲话人：or four series of models.And I think

讲话人：we did more than that volume in just

讲话人：Q2 of this year.I think I've been really

讲话人：lucky with...the team that we've grown

讲话人：and built,

讲话人：both the stakeholders on the research

讲话人：side and within our research product

讲话人：management team.I think that one of

讲话人：the magical parts about approaching

讲话人：all of this is that it's not an individual

讲话人：sport.There's like a sense of radical

讲话人：ownership and team collaboration that

讲话人：I think Sometimes it does feel like

讲话人：a high-performance sport because you're

讲话人：in very critical decisions.There's new

讲话人：information about users,

讲话人：about training,

讲话人：and you have to make recommendations

讲话人：and judgments and decisions very quickly.

讲话人：And nobody can do that sustainably by

讲话人：themselves.And so I think what's really

讲话人：helped is having a team that is incredible,

讲话人：who looks out for each other,

讲话人：who,you know, the night before a launch,

讲话人：even if they're not the core DRI on

讲话人：that model,

讲话人：will stay up and help the DRI to review

讲话人：the blog post and make edits and come

讲话人：up with better demos and knowing to

讲话人：be each other's sort of extra hand.

讲话人：I think it's very easy if you take all

讲话人：of this change on your own shoulders

讲话人：to feel like you're alone and to feel

讲话人：like you have to do everything.But I

讲话人：think one of the magical parts of Anthropic

讲话人：is this ability for us to change.figure

讲话人：out what are those opportunities to

讲话人：help each other and actually then taking

讲话人：the next mile of like mind melding.

讲话人：We called it like entering the hive

讲话人：mind.There was an article about this.

讲话人：And I think like part of that is just

讲话人：that allows like the team to replenish.

讲话人：It's not that you–I was just on PTO

讲话人：in June.It's not just that you can take

讲话人：PTO and you come back to like 3X the

讲话人：amount of things to do.It's actually

讲话人：that you can take PTO and know the team

讲话人：can–figure out the right things to do,

讲话人：and that we individually can like watch

讲话人：out for each other.So I think that's

讲话人：a big part.I'm really lucky just personally.

讲话人：Also,

讲话人：my partner is really supportive.This

讲话人：is year six of me working in AI.So Amazon

讲话人：and then Anthropic.And so he sees how

讲话人：much I just love the technology and

讲话人：what this can do.And that really helps,

讲话人：I think,

讲话人：also from like a personal perspective as

讲话人：well.I love how many of these answers

讲话人：connect.So what I'm hearing here is

讲话人：just working with other people,

讲话人：relying on other people,

讲话人：helping each other out when things get

讲话人：crazy,

讲话人：which is a similar answer you had for

讲话人：just how to find the joy and fun in

中文译文（机器翻译，按本时间段分段）

演讲人：并优先训练模型

讲话人：在这个领域表现出色并且喜欢

讲话人：那是一个活跃的、非常活跃的区域

说话人：对我们来说很有趣，就像你提到的好吧，

说话人：我很高兴。我很高兴，而且一切都会过去

说话人：人工智能这么好了，就变得有趣了。

说话人：我们就像，

说话人：我不知道是谁写的，但是对你来说

说话人：点，

说话人：有时候我们其实想知道

语音人：这是人工智能，这真的很有趣。我

说话人：从来没有这样想过。另一个

说话人：有趣的是，有

讲话人：那个开玩笑说我们是

说话人：就像在飞机上一样，Wi-Fi 断了

说话人：我们就像，到底是什么，

说话人：这架飞机上没有 Wi-Fi。

说话人：这太糟糕了，你怎么敢？

说话人：当你就像在空中的管子里时

言语人：像鸟一样飞翔，你怎么敢

说话人：抱怨 Wi-Fi 不能用？

说话人：就像你的观点一样，有很多进步

说话人：还有这么大的力量。我们无法解决所有问题。

演讲人：我们无法让一切都发挥最佳效果。

说话人：所以基本上写作一直没有

讲话人：优先事项，感觉好像有

演讲人：那里正在发生更多的投资。是的，

说话人：我认为语气和性格是一个

讲话人：优先。我认为是这个进步

演讲人：该技术是一项正在进行中的工作。

说话人：所以我们做了，

讲话人：我们看到像一个飞跃或出现

说话人：跳入代理行为。

讲话人：是一种新常态。然后是其他这些

演讲人：能力需要不断提高。

演讲人：我认为一旦我们进步了，

说话人：比方说写作和喜欢的语气和

说话人：性格，我们可能会说，

演讲人：我们如何让云变得更加主动？

说话人：比如，

说话人：积极主动就是机会。那就是

说话人：人性。喜欢，

说话人：我们想让自己变得更好。我们

演讲人：想让这项技术变得更好。

说话人：所以，是的，

演讲人：我认为我们正在将其应用于人工智能，

说话人：这是正确的事情，我们应该

说话人：让它变得更好。来了。嗯，

说话人：一个是你认为人的大脑在哪里

演讲人：将继续成为最有价值的

说话人：那几年？

说话人：我知道 Anthropix 的使命，并且，

演讲人：我们的愿景是我们将成为一名美国军人，

讲话人：一个超级聪明的人。所以在未来，

说话人：也许无处可去，但在我们到达那里之前，

说话人：你认为人的大脑会在哪里

讲话人：当我们接近时，继续是最有价值的

说话人：那个、那个时间线？

说话人：我们开始讨论制作克劳德

说话人：并且在判断力方面表现得更好。尤其是

说话人：在过去一年左右的时间里，

说话人：我认为判断力是其中之一，

说话人：是一个聚集的区域

说话人：有如此多的细微差别和如此丰富的经验，

说话人：而且这些系统没有经历过

说话人：和人类一样多，所以我认为

说话人：那是来之不易的，就像判断力一样，

演讲人：是产品领导者和公正人士的区域

说话人：一般都会继续真的

说话人：关键人工智能的事情太多了

讲话人：可以建造。哪一个是，

演讲人：你知道，像实验室这样的组织应该建立，

说话人：对吗？

说话人：很多都需要人类的判断力，

说话人：坚持。所以，积极主动，

说话人：这些都是超越的特质

说话人：只是一般能力，

说话人：但只是行为和特征

说话人：在那个级别的同类人中，

演讲人：如何得出最佳解决方案？

演讲人：如何创造最佳体验？

说话人：那么，

说话人：我认为这些类型的特征实际上是

言语人：我认为会的触觉特征

说话人：是，

讲话人：仍然很重要。我认为有

说话人：也是，

说话人：还是有很多类似的能力和

演讲人：主题专业知识也是如此。我认为，

说话人：你知道，

演讲人：软件工程真的很

说话人：被 AI 改造了，我认为有一些领域

说话人：像生物学一样，

讲话人：生命科学。这些都是

说话人：我们只是有点像脚

说话人：关于指数，

说话人：比如软件工程，

演讲人：我们正处于指数级增长阶段，

演讲人：在其他一些领域，

说话人：我们还没到那一步，所以我想

说话人：你看到我们运送云之类的东西

演讲人：科学，投资这些领域，

说话人：因为这些是我认为的领域

说话人：就是把这个技术带给社会

演讲人：并为社会带来积极的利益。

演讲人：所以我认为还有很多事情要做

说话人：还有一个问题我想问

说话人：是，

说话人：作为一个有孩子的人……你觉得怎么样

说话人：关于你鼓励他们的事情

说话人：学习？

说话人：你认为你要推动哪里

说话人：他们要在这个疯狂的新领域取得成功

说话人：我们正在进入的世界？

说话人：其实我觉得很多都一样

讲话人：像你我这样的特质可能会成长

说话人：up with，这是对学习的好奇心，

说话人：坚持，

说话人：相信自己内心的声音，

说话人：发展然后相信你的

说话人：自己内心的声音。就像我有四年

说话人：-老，我有一个八岁的孩子，

说话人：我们有责任帮忙，

说话人：我有责任帮助他们发展他们的能力

说话人：内心的声音以及是否存在

说话人：固执己见，对我表明立场，

说话人：对吗？

讲话人：并且发展这一点，

说话人：令人鼓舞。我认为那些

演讲人：技能组的类型是指

说话人：将来很重要，而且我喜欢

说话人：有自己独特的声音。

说话人：太有趣了，这与

说话人：当被问到时你的答案

说话人：如何避免大脑腐烂，

说话人：本质上是对 AI 的过度依赖，

说话人：就是专注于自己的事情

说话人：观点和你自己的观点

演讲人：在你过度依赖人工智能之前

说话人：你所描述的这个建筑想法

说话人：对于孩子来说，

说话人：真的很重要。呃，

说话人：那太有趣了，我喜欢这样

说话人：所有这些，所有这些都与判断有关，

言语人：坚持，

讲话人：你自己的观点。是的。两者

讲话人：适合孩子和成人。是的。任何事

说话人：你想想你的...哦，

讲话人：伙计。嗯，

说话人：就像我正在考虑的问题一样

说话人：就是让他们像某些人一样相处的时候

说话人：人工智能的事，你知道，

说话人：当我有一个三岁的孩子时，

说话人：所以现在说还为时过早。但是，

说话人：你知道，你怎么得到，

说话人：你是如何让他们进入这种疯狂的状态的

说话人：东西？

说话人：我最近参加了一个活动，还有很多

演讲人：的父母正在谈论他们如何

说话人：想想他们孩子身上的人工智能。还有一个

说话人：这个人有一个非常有趣的方法，

讲话人：这让他们很早就开始

演讲人：……的模型，所以他们仍然有

讲话人：稍微挣扎一下，但没有得到全部

说话人：立即回答。我以为那是

说话人：有趣，

讲话人：就像一个开源的本地模型，

说话人：不是寓言。是的。是的。好奇心是

说话人：我一直提到本·曼（Ben Mann），

说话人：但他的回答实际上是针对这个问题的

说话人：一直困扰着我，

说话人：这是好奇心。而且也就像

说话人：他是蒙台梭利的忠实粉丝，

说话人：这就是我对我们的鼓励

说话人：孩子们。所以那里有东西。也许

说话人：最后一个问题，

说话人：就是这样，

说话人：Fiona Fung 实际上建议的一些事情

说话人：我问你，最近谁在播客上，

说话人：你如何保持充沛的精力而不是

说话人：精疲力竭，

讲话人：处于这场疯狂风暴的中心

说话人：人工智能？

说话人：作为一名工作妈妈，你知道，

说话人：我们正在看 Anthropic 的研究工作。

说话人：我只是喜欢，

讲话人：我们正在经历最前所未有的时期

说话人：工作时间就像，

言语人：处于人类之外，

说话人：太疯狂了，我什至不知道那是什么

说话人：喜欢在里面。你有什么

说话人：学会了如何避免倦怠，

说话人：保持充电，

说话人：在所有事情中间保持理智

说话人：这个？

说话人：2024 年，

说话人：我们一共发货了四种型号

说话人：年，

说话人：或者四个系列的型号。而且我认为

说话人：我们只用了超过这个量

说话人：今年第二季度，我觉得我真的很

说话人：幸运的是……我们成长的团队

说话人：并且建造，

演讲人：研究的双方利益相关者

演讲人：我们研究产品的侧面和内部

说话人：管理团队。我认为其中之一

说话人：接近的神奇之处

说话人：这一切都说明它不是一个人

说话人：运动。有一种激进的感觉

演讲人：主人翁精神和团队协作

说话人：我觉得有时候确实感觉像

演讲人：一项高性能运动，因为你

说话人：在非常关键的决定中。有新的

说话人：有关用户的信息，

演讲人：关于培训，

演讲人：你必须提出建议

说话人：判断和决定很快。

讲话人：没有人可以通过持续地做到这一点

说话人：他们自己。所以我想什么是真正的

演讲人：有帮助的是拥有一支令人难以置信的团队，

说话人：互相照顾的人，

讲话人：谁，你知道，在发射前一天晚上，

讲话人：即使他们不是 DRI 的核心

说话人：那个模特，

说话人：会熬夜帮助 DRI 审核

说话人：博客发帖并进行编辑就来了

说话人：有更好的演示并且知道

说话人：成为对方的助手。

说话人：我觉得如果你把所有的都拿走的话就很容易了

讲话人：这个变化要由你自己承担

说话人：感觉自己很孤独并且感觉

讲话人：好像你必须做所有事情，但我

说话人：想想《Anthropic》的神奇之处之一

说话人：这个能力是我们要改变的吗？

发言者：那些机会是什么？

说话人：互相帮助，然后实际采取

言语人：相似思想融合的下一英里。

说话人：我们称之为“进入蜂巢”

说话人：mind.有一篇文章是关于这个的。

说话人：我认为其中一部分只是

说话人：这样可以让团队补充。

说话人：不是你——我只是在 PTO

说话人：六月，不只是你能拿的

讲话人：PTO 和你回来喜欢 3X

讲话人：要做的事情很多，实际上是

说话人：你可以参加 PTO 并了解团队

说话人：can-找出正确的事情要做，

讲话人：我们个人可以喜欢看

说话人：互相帮助，所以我认为那就是

说话人：很大一部分。就我个人而言，我真的很幸运。

说话人：还有，

说话人：我的搭档真的很支持。这个

说话人：我在人工智能领域工作已经第六年了。所以亚马逊

说话人：然后是人性化的。所以他明白了

演讲人：我非常喜欢这项技术并且

说话人：这能做什么。这确实有帮助，

说话人：我认为，

说话人：也是从个人角度来看

说话人：嗯，我喜欢其中的很多答案

讲话人：连接。所以我在这里听到的是

讲话人：只是和其他人一起工作，

说话人：依靠别人，

说话人：遇到困难时互相帮助

说话人：疯了，

说话人：这和你的回答类似

演讲人：如何找到快乐和乐趣

01:20:00

英文原文（完整 ASR）

讲话人：this work.Just be inspired by other

讲话人：people,see what they're doing,

讲话人：work together.Yeah.And it's interesting

讲话人：when Fiona was on the podcast recently,

讲话人：I was asking her just like what's changed

讲话人：in the world of software engineering.

讲话人：And she pointed out it's a lot lonelier

讲话人：now because now we're working with agents

讲话人：instead of other humans.Teams are smaller.

讲话人：People are having all these fleets they're

讲话人：talking to constantly.And so this is

讲话人：just a reminder of just the power of

讲话人：just actual other humans around you.

讲话人：We're asked to work and make decisions

讲话人：on really big things because you have

讲话人：more scale from the technology.Right.

讲话人：And I think that's, Having individuals,

讲话人：having other folks more who can have

讲话人：some level of like mind meld with what

讲话人：you work on,how you approach,

讲话人：maybe not exactly every detail,

讲话人：but what are the first principles,

讲话人：what are the assumptions you make,

讲话人：then helps them,you know,

讲话人：back up for you or push your decision

讲话人：and sharpen your thinking.Um,

讲话人：so I think,you know,

讲话人：we really try to like,

讲话人：I really try to look for that when like

讲话人：building the team,growing the team,

讲话人：hiring,like,

讲话人：is this person going to care about their

讲话人：own ego and building out a big org,

讲话人：or are they going to care about contributing

讲话人：to Anthropic and contributing to the

讲话人：like impact of the team and orienting

讲话人：towards folks who are like low ego team

讲话人：oriented?

讲话人：Um,I think that's.Yeah,

讲话人：it's a big part of,I think,

讲话人：the sustainability.Yeah,

讲话人：just always a lot of it always just

讲话人：comes down back to culture and hiring.

讲话人：And I know I've heard a lot just the

讲话人：reason Anthropic is able to move so

讲话人：fast.I remember that moment when like

讲话人：something shipped every day of the month,

讲话人：like a calendar of launches.And people

讲话人：were talking about how is this possible?

讲话人：And what I heard a lot is just because

讲话人：everyone is so aligned around the mission

讲话人：and the values,

讲话人：it allows people to make decisions really

讲话人：quickly.Before we get to our very exciting

讲话人：lightning round, is there anything else,

讲话人：Dan,that you wanted to share,

讲话人：anything else you wanted to touch on,

讲话人：anything you want to maybe double down

讲话人：on of things we've talked about?

讲话人：This was actually really fun because

讲话人：I feel like your questions actually

讲话人：sharpened some of my thinking around

讲话人：how the thoughts kind of connect.I'm

讲话人：your real human clod over here.One thing

讲话人：that I really want to convey or have

讲话人：people take away is,I think one,

讲话人：in the ways of working,

讲话人：but also just two,

讲话人：that this is a lot of growth and change

讲话人：and having the joy in using this technology.

讲话人：And if you're feeling like in this moment,

讲话人：you don't have as much of that feeling

讲话人：of initial joy,

讲话人：how do you find people who do?

讲话人：if this is an area that you're excited

讲话人：and want to work on.And I think developing

讲话人：skill sets,

讲话人：replenishing skill sets in many ways

讲话人：of things like thinking from a first

讲话人：principles manner about what you solve.

讲话人：I think fundamentally,

讲话人：you didn't ask me this,

讲话人：but there is this question in the community

讲话人：of do we still need PMs when the models

讲话人：are so capable,

讲话人：when engineers are leaning in?

讲话人：I think the role of people who are user

讲话人：-centric,

讲话人：who go into the details of understanding

讲话人：what users are trying to accomplish,

讲话人：bubbling that up in an actionable manner

讲话人：and doing the relentless work to do

讲话人：that,

讲话人：that to me is the core of a product person.

讲话人：And I actually think we need more of

讲话人：that.I think we are becoming very technology

讲话人：-layered driven.And actually,

讲话人：to make that impactful,

讲话人：you have to go deep.You have to be curious.

讲话人：You have to be super hands-on.And those

讲话人：are things that I think are also traits

讲话人：that have,I think,

讲话人：helped Anthropic from a product development

讲话人：and model development perspective and

讲话人：as part of the culture.And hopefully,

讲话人：that's valuable for others as well.

讲话人：Amazing.What an inspiring way to end

讲话人：it.Oh,

讲话人：man.Yeah.And this is I've been saying

讲话人：this too,for a long time,

讲话人：just now that building is easy.The hard

讲话人：part becomes,as you said,

讲话人：what should we build and is the thing

讲话人：we have built correct and good and worth

讲话人：leaning into.And to me,

讲话人：that's what PMs do and what PMs are

讲话人：good at.Yeah,yeah,

讲话人：yeah.And it's getting into the details

讲话人：of the user.Yeah,empathy.Okay,

讲话人：great.PMs are gonna make it.Okay,

讲话人：PRD is not dead.All kinds of important

讲话人：lessons here.Diane,with that,

讲话人：we've reached a very exciting lightning

讲话人：round.I've got five questions for you.

讲话人：Are you ready?

讲话人：Yep.First question,

讲话人：what are two or three books that you

讲话人：find yourself recommending most to other

讲话人：people?

讲话人：One personal one,

讲话人：I really like How to Raise an Adult.

讲话人：So,

讲话人：I'm a mom.I think a lot about what is

讲话人：the things that I want to instill in

讲话人：my kids.And that book is really helpful

讲话人：for describing we're not trying to raise

讲话人：children.We're trying to raise adults.

讲话人：So just the framing of what does that

讲话人：mean and what does it mean?

讲话人：What are the characteristics that we

讲话人：want to hone and like harness and foster

讲话人：in our kids?

讲话人：The other book that I was listening

讲话人：to on Audible recently is Incorrigible

讲话人：by Eric Ries.Incorrup tible.Incorruptible.

讲话人：Yes,

讲话人：yes.Yeah.His recent podcast.Yeah.And

讲话人：I just I think the question of how to

讲话人：build great companies is important.

讲话人：Yeah.just been most fascinated with

讲话人：how to keep great teams and great companies

讲话人：going further.And it was very interesting

讲话人：to just kind of see his framing and

讲话人：reframing of the question.I loved some

讲话人：of the examples around having metrics

讲话人：around culture.If you can,

讲话人：if you only measure revenue and then

讲话人：that's kind of how you're going against,

讲话人：but if you have other better metrics,

讲话人：that's actually the way.to sustain the

讲话人：values you care about.I've been kind

讲话人：of trying to think about how to actually

讲话人：bring that to the team level of like,

讲话人：how do we better articulate,

讲话人：write our norms,

讲话人：a lot of the things we talked about

讲话人：on the team.So I think that's also a

讲话人：really good read.There you go.That'll

讲话人：be your next watch,everyone,

讲话人：as you're listening to this,

讲话人：the Eric Ries episode.Such a good episode.

讲话人：Yeah.And his book just came out,

讲话人：Incorruptible.And I think it was like

讲话人：a New York Times bestseller.Like it's

讲话人：actually doing incredibly well,

讲话人：which I was really happy to see.Yeah,

讲话人：exactly.Next question.Favorite recent

讲话人：movie or TV show?

讲话人：You really enjoy it.Most people at Anthropic

讲话人：do not have time to do what to watch

讲话人：things.But I'm curious if you have an

讲话人：answer.I would say during some time

讲话人：off last month,

讲话人：I did get to binge watch Fallout on

讲话人：Amazon Prime.So that was actually,

讲话人：I kind of like,it's kind of,

讲话人：have you heard of it?

讲话人：Yeah,yeah,

讲话人：it's based on the video game.Yes,

讲话人：it's based on the video game.I think

讲话人：it's a,it was really,it's witty,

讲话人：it's humorous,

讲话人：it's also like super action oriented,

讲话人：so highly recommend.Okay,

讲话人：next question.Do you have a favorite

讲话人：product you recently discovered that

讲话人：you really love?

讲话人：I really do think like CloudTag is very

讲话人：interesting in terms of a product experience.

讲话人：We actually have like different versions

讲话人：of this within Anthropic.And I think

讲话人：it's actually been really,

讲话人：really,really powerful tool.Yeah,

讲话人：it feels like I think some people are

讲话人：like,what's the big deal?

讲话人：The fact that everyone at Anthropic

讲话人：is like raving about it tells me something

讲话人：important is going on here.And I'm trying

讲话人：to actually get it working within my

讲话人：Slack community that I have for paid

讲话人：newsletter subscribers.How cool would

讲话人：that be?

讲话人：Yeah.Yeah.I'm trying to figure out how

讲话人：it works when it's not a company,

讲话人：when it's just a bunch of people that

讲话人：don't know each other and how that might

讲话人：work.But we're trying it out.Okay.Two

讲话人：more questions.Your favorite life motto

讲话人：that you find yourself often coming

讲话人：back to in work or in life?

讲话人：So I was actually raised by my grandparents

讲话人：for the first 10 years of my life.And

讲话人：my parents were immigrant college and

讲话人：master's students in the US.And my grandfather

讲话人：always says,

讲话人："'No matter how far you go,

讲话人：there's always another level."which

讲话人：is,I think,a really good way,

讲话人：though like a pretty intense way of

讲话人：describing his life or philosophy.But

讲话人：I go back to that whenever there's something

讲话人：new or unprecedented that we experience.

讲话人：And I think,you know,

讲话人：first half of this year,

讲话人：there was definitely a lot of that.

讲话人：Like there was a lot of new things that

讲话人：we were learning.I was learning.So just

讲话人：feeling like,

讲话人：There's always like another mountain,

讲话人：another opportunity to climb.Not good

讲话人：enough,

讲话人：Diane.We need to go better.We need to

讲话人：go bigger.Makes me think about actually

讲话人：another Ben Mann line from his podcast

讲话人：episode,

讲话人：that this is the most normal it's ever

讲话人：going to be.It's only going to get weirder

讲话人：and crazier.Yeah.Yeah.Oh my God.Okay.

讲话人：Final question.With Poconair on your

讲话人：LinkedIn.You were a high yield bond

讲话人：trader,JP Morgan Chase,

讲话人：early in your career.You had like,

讲话人：you have this like redacted a hundred

讲话人：million dollar trading portfolio of

讲话人：some kind.What did you learn from that

讲话人：time in your life that has stuck with

讲话人：you?

讲话人：And,

讲话人：or is there a crazy story from that period?

讲话人：It was four years of your life.I think

讲话人：I learned actually a lot that I apply

讲话人：here at Anthropic and other jobs thereafter.

讲话人：So when I was at J.P.Morgan,

讲话人：I,

讲话人：Trading floor you could kind of envision like

讲话人：sort of waffle Wall Street.That's very

讲话人：different Most traders I think are in

讲话人：front of a terminal.They're much more

讲话人：doing analyses on their computers but

讲话人：it's still very I would say like male

讲话人：-dominated and so I was the only woman

中文译文（机器翻译，按本时间段分段）

演讲人：这个作品。只是受到其他作品的启发

说话人：人们，看看他们在做什么，

讲话人：一起工作。是的。而且很有趣

说话人：Fiona 最近上播客的时候，

说话人：我问她好像有什么变化

演讲人：在软件工程领域。

说话人：她指出这样更孤独

讲话人：现在因为我们现在正在与代理商合作

说话人：而不是其他人。团队更小。

说话人：人们拥有这些舰队

说话人：不断地说话。所以这就是

说话人：只是提醒人们的力量

说话人：只是你周围真实的其他人。

演讲人：我们被要求工作并做出决定

说话人：谈论真正重要的事情，因为你有

演讲人：更多的规模来自于技术。对。

演讲人：我认为，拥有个人，

讲话人：拥有更多可以拥有的人

说话人：某种程度的相似思想与什么融合在一起

说话人：你的工作方式，你的方法，

说话人：也许不是每一个细节，

演讲人：但是首要原则是什么？

说话人：你做了什么假设，

说话人：然后帮助他们，你知道，

说话人：支持你或推动你的决定

说话人：并提高你的思维能力。嗯，

说话人：所以我认为，你知道，

说话人：我们真的很努力喜欢，

说话人：我真的很努力地寻找这样的东西

演讲人：建设团队，成长团队，

说话人：招聘，喜欢，

说话人：这个人会关心他们的事情吗？

说话人：拥有自我并建立一个大组织，

说话人：或者他们会关心贡献吗

演讲人：致 Anthropic 并为

演讲人：喜欢团队的影响力和导向

说话人：对那些低自我团队的人

说话人：面向？

说话人：嗯，我想是这样。是的，

演讲人：我认为这很重要，

讲话人：可持续性。是的，

讲话人：总是很多总是只是

演讲人：归根结底还是文化和招聘。

说话人：我知道我已经听过很多了

言语人：人类之所以能够如此移动的原因

说话人：快。我记得那一刻，就像

说话人：这个月每天都有东西发货，

讲话人：就像发布日历。还有人

说话人：我们正在谈论这怎么可能？

说话人：而我听到的很多只是因为

演讲人：每个人都围绕着自己的使命

讲话人：以及价值观，

说话人：它让人们真正做出决定

说话人：快点，在我们进入我们非常激动人心的话题之前

讲话人：闪电般的声音，还有什么吗？

演讲人：Dan，你想分享的，

说话人：还有什么你想谈的吗？

说话人：任何你想要的都可以加倍

说话人：关于我们谈论过的事情？

说话人：这实际上真的很有趣，因为

说话人：我感觉你的问题其实很像

说话人：锐化了我的一些思考

说话人：思想是如何联系起来的。我

说话人：这里是你真正的人类。有一件事

说话人：我真正想传达的或有的

说话人：人们带走的是，我认为一个，

说话人：在工作方式上，

说话人：但也只有两个，

说话人：这是一个很大的成长和改变

演讲人：并享受使用这项技术的乐趣。

说话人：如果你此时此刻感觉，

说话人：你没有那么多感觉

说话人：最初的喜悦，

说话人：你如何找到这样做的人？

说话人：如果这是一个让你感到兴奋的领域

说话人：并且想要继续工作。而且我认为正在发展

说话人：技能组，

说话人：多方面补充技能

讲话人：诸如从一开始就思考之类的事情

说话人：关于你所解决问题的原则方式。

说话人：我认为从根本上来说，

说话人：你没有问我这个，

说话人：但是社区里有这个问题

演讲人：我们还需要 PM 吗？

说话人：太能干了，

演讲人：工程师什么时候开始介入？

说话人：我认为人的角色是用户

说话人：以-为中心，

说话人：谁深入了解细节

说话人：用户想要完成什么，

说话人：以可行的方式将其冒泡出来

说话人：并坚持不懈地做该做的工作

说话人：那个，

演讲人：对我来说这就是产品人的核心。

演讲人：我实际上认为我们需要更多

说话人：那个。我认为我们正在变得非常科技化

说话人：-分层驱动。实际上，

演讲人：为了使其具有影响力，

说话人：你必须深入，你必须好奇。

说话人：你必须非常亲力亲为。还有那些

说话人：我认为这些东西也是特质

说话人：我认为，

说话人：帮助 Anthropic 进行产品开发

讲话人：以及模型发展的观点和

演讲人：作为文化的一部分。希望，

说话人：这对其他人来说也很有价值。

说话人：太棒了，多么鼓舞人心的结束方式啊

讲话人：哦，

说话人：伙计。是的。这就是我一直在说的

说话人：这也是，很长一段时间，

说话人：刚才说建造很容易，很难

说话人：正如你所说，部分变成了，

演讲人：我们应该构建什么，是什么

说话人：我们建立了正确、良好和有价值的

说话人：倾身而入。对我来说，

演讲人：这就是 PM 所做的事以及 PM 的本质

说话人：擅长。是啊，是啊，

说话人：是的，而且正在进入细节

说话人：用户的。是的，同理心。好的，

说话人：太好了。PM 们会成功的。好吧，

说话人：PRD 没有死。各种重要

说话人：这里上课。黛安，这样，

说话人：我们已经到达了非常激动人心的闪电

演讲人：圆形。我有五个问题要问你。

说话人：你准备好了吗？

演讲人：是的，第一个问题，

说话人：你读的两本书或三本书是什么？

说话人：发现自己向其他人推荐最多

说话人：人？

说话人：一个人的，

说话人：我真的很喜欢《如何养育一个成年人》。

说话人：那么，

说话人：我是一个妈妈，我想了很多关于什么是

说话人：我想灌输的东西

说话人：我的孩子们。那本书真的很有帮助

说话人：为了描述我们并不是想提出

说话人：孩子们。我们正在努力抚养成人。

说话人：所以只是框架是什么

讲话人：意思是什么？

说话人：我们有什么特点

说话人：想要磨练，喜欢驾驭和寄养

说话人：在我们的孩子身上？

说话人：我正在听的另一本书

说话人：最近在 Audible 上的声音真是不可救药

说话人：埃里克·里斯 (Eric Ries) 着。Incorrup tible.Incorruptible.

说话人：是的，

讲话人：是的。是的。他最近的播客。是的。还有

讲话人：我只是我觉得如何的问题

演讲人：建立伟大的公司很重要。

说话人：是的，只是最着迷

演讲人：如何留住优秀的团队和优秀的公司

讲话人：更进一步。而且非常有趣

讲话人：只是为了看看他的框架和

讲话人：重新组织问题。我喜欢一些

演讲人：关于度量的例子

演讲人：围绕文化。如果可以的话，

说话人：如果你只衡量收入，然后

说话人：这就是你反对的方式，

说话人：但是如果你有其他更好的指标，

说话人：这实际上是维持现状的方法

说话人：你关心的价值观。我一直很友善

说话人：试图思考如何实际

演讲人：将其带到团队层面，

说话人：我们如何更好地表达，

说话人：写下我们的规范，

说话人：我们谈论了很多事情

说话人：在团队中。所以我认为这也是一个

说话人：读得真好。就这样吧。

讲话人：大家，请成为你们的下一个守望者，

说话人：当你听这个的时候，

说话人：埃里克·里斯那一集。真是一集好节目。

说话人：是的。他的书刚刚出版，

说话人：廉洁。我认为这就像

说话人：《纽约时报》畅销书。就像这样

说话人：实际上做得非常好，

说话人：我很高兴看到。是的，

说话人：完全正确。下一个问题。最近最喜欢的

说话人：电影还是电视剧？

说话人：你真的很喜欢。Anthropic 的大多数人

说话人：没时间看什么

说话人：事情。但是我很好奇你是否有

说话人：回答。我会说在一段时间内

说话人：上个月休息，

说话人：我确实狂看了《辐射》

发言者：Amazon Prime。实际上，

说话人：我有点喜欢，有点，

说话人：你听说过吗？

说话人：是啊，是啊，

说话人：这是根据电子游戏改编的。是的，

说话人：这是根据电子游戏改编的，我想

说话人：这真是，这真是，这很机智，

说话人：很幽默，

说话人：这也像是超级行动导向，

演讲人：强烈推荐。好的，

说话人：下一个问题，你有最喜欢的吗

演讲人：您最近发现的产品

说话人：你真的爱吗？

说话人：我确实认为 CloudTag 非常好

说话人：就产品体验而言很有趣。

说话人：我们其实有不同的版本

说话人：这个在人类内部。我认为

说话人：实际上，

说话人：真的，真的很强大的工具。是的，

说话人：我感觉有些人是

说话人：就像，有什么大不了的？

说话人：事实上，Anthropic 的每个人

说话人：就像在胡言乱语一样，它告诉了我一些事情

说话人：这里发生了很重要的事情，我正在努力

演讲人：真正让它在我的范围内发挥作用

讲话人：我付费的 Slack 社区

说话人：时事通讯订阅者。多酷啊

说话人：那是？

说话人：是的，是的，我正在想办法

说话人：当它不是公司时它就有效，

说话人：当只有一群人的时候

说话人：彼此不认识，怎么可能

说话人：工作。但是我们正在尝试。好的。两个

说话人：更多问题。你最喜欢的人生格言

说话人：你发现自己经常来

说话人：回到工作中还是生活中？

说话人：所以我实际上是由我的祖父母抚养长大的

说话人：在我生命的前 10 年里。

说话人：我的父母是移民大学的学生

讲话人：美国的硕士生，还有我的祖父

说话人：总是说，

说话人：“‘无论你走多远，

说话人：总有另一个层次。”

说话人：我认为这是一个非常好的方法，

说话人：虽然像是一种非常激烈的方式

说话人：描述他的生活或哲学。但是

说话人：每当有什么事我都会回到那句话

说话人：我们经历过的新的或前所未有的事情。

说话人：我认为，你知道，

演讲人：今年上半年，

说话人：这样的事情肯定有很多。

说话人：好像有很多新东西

讲话人：我们正在学习。我正在学习。所以就这样

说话人：感觉，

说话人：总有另一座山，

说话人：又一个爬山的机会，不好

说话人：够了，

说话人：Diane.We need to go better.We need to

说话人：走得更大。让我思考实际上

说话人：本·曼播客中的另一句台词

演讲人：episode,

说话人：这是有史以来最正常的一次

说话人：会的，只会变得更奇怪

说话人：而且更疯狂。是啊，是啊，天啊，好吧。

讲话人：最后一个问题。您的 Poconair 已打开

说话人：LinkedIn.You was a high Yiy Yield Bond

演讲人：交易员，摩根大通，

说话人：在你职业生涯的早期。你曾经喜欢，

说话人：你这就像编辑了一百个一样

演讲人：百万美元的交易组合

说话人：某种，你从中学到了什么

讲话人：你生命中一直坚持的时间

说话人：你？

说话人：而且，

说话人：或者那个时期有什么疯狂的故事吗？

说话人：那是你生命中的四年。我想

说话人：我实际上学到了很多我应用的东西

讲话人：在 Anthropic 工作，之后还有其他工作。

演讲人：所以当我在摩根大通的时候，

说话人：我，

演讲人：你可以想象一下交易大厅

说话人：有点像华尔街的华夫饼。那非常

讲话人：不同我认为大多数交易者都处于

说话人：在终端前。他们还有更多

讲话人：在他们的计算机上进行分析，但是

说话人：我想说还是很像男性

说话人：-占主导地位，所以我是唯一的女性

01:30:00

英文原文（完整 ASR）

讲话人：I was the only Person with like my background

讲话人：on the trading desk and I learned that

讲话人：I that was a very good environment to

讲话人：kind of building one,

讲话人：my sense of authentic self and two,

讲话人：that even if I was the most junior person,

讲话人：even if I may look different,

讲话人：that the best ideas were,

讲话人：and having conviction in the best ideas,

讲话人：irregardless of all of those other factors,

讲话人：like it's the most important thing.

讲话人：And so I think just bringing that sense

讲话人：of how I show up more at work,

讲话人：I'm pretty vulnerable and authentic

讲话人：with my team.I try to really make sure

讲话人：that regardless of people's levels or

讲话人：tenures,if they have a great idea,

讲话人：how to help them pursue that.and to

讲话人：do also the same.So to like put the

讲话人：idea out there,

讲话人：to actually have conviction in it,

讲话人：to do the follow through,

讲话人：to do the like nitty gritty work to

讲话人：make it happen.So those were all things

讲话人：that I learned from trading.And yeah,

讲话人：I think applies to any job in many ways.

讲话人：That is beautiful.Where can people find

讲话人：you online if they wanna follow you?

讲话人：And how can listeners be useful to you?

讲话人：I don't have a large presence on social.

讲话人：I think the best way to find my work,

讲话人：my team's work is really the Anthropic

讲话人：blog.And when we're publishing new models,

讲话人：new product experiences,

讲话人：I think in terms of,

讲话人：useful for me.I think the best thing,

讲话人：number one,is your feedback.Like,

讲话人：we actually,

讲话人：if you thumbs up or thumbs down on any

讲话人：of our product surfaces,

讲话人：if you contact your salesperson with

讲话人：feedback about the model,

讲话人：it will make its way to me.We actually,

讲话人：with every research model,

讲话人：I actually get pretty close into understanding

讲话人：favorability and feedback.So giving

讲话人：us that feedback,pushing Claude,

讲话人：telling us where it's falling down,

讲话人：those help us make Claude better.The

讲话人：other thing is,like,

讲话人：if you have folks in your network who

讲话人：seem like this type of profile person

讲话人：that I just talked about,

讲话人：I'm hiring,

讲话人：the team is growing.We really will love

讲话人：just people who love this technology,

讲话人：who are deeply curious,

讲话人：first principles thinkers,

讲话人：who are fearless in questioning assumptions,

讲话人：and who have like a tinkering hackery

讲话人：spirit.Wow.What a dream job.So basically,

讲话人：open PM roles at Anthropic on the research

讲话人：team.Yes.And they apply,

讲话人：I assume,on the website,

讲话人：the careers page.Yes.Holy moly.All right,

讲话人：here we go.Enjoy the flood of resumes

讲话人：you're about to receive.Thank you,

讲话人：Lenny.Dan,

讲话人：thank you so much for being here.Thank

讲话人：you so much for having me.Thank you

讲话人：for really helpful,

讲话人：thought-provoking questions,

讲话人：helping me even connect the dots on

讲话人：how we work,

讲话人：how this whole technology is coming

讲话人：together and being product people in

讲话人：it.I really appreciate that.But thank

讲话人：you,Dan,for real.Okay,well,

讲话人：bye,

讲话人：everyone.Thank you so much for listening.If

讲话人：you found this valuable,

讲话人：you can subscribe to the show on Apple

讲话人：Podcasts,Spotify,

讲话人：or your favorite podcast app.Also,

讲话人：please consider giving us a rating or

讲话人：leaving a review,

讲话人：as that really helps other listeners

讲话人：find the podcast.You can find all past

讲话人：episodes or learn more about the show

讲话人：at lennyspodcast.com.See you in the

讲话人：next episode.

中文译文（机器翻译，按本时间段分段）

说话人：我是唯一一个和我背景一样的人

说话人：在交易台上我了解到

演讲人：我觉得这是一个非常好的环境

演讲人：一种建筑，

说话人：我对真实自我的感觉和两个，

说话人：即使我是资历最浅的人，

说话人：即使我看起来不同，

讲话人：最好的想法是，

说话人：并且坚信最好的想法，

说话人：不管所有这些其他因素，

说话人：好像这是最重要的事情。

说话人：所以我认为只是带来这种感觉

说话人：关于我如何在工作中表现得更好，

说话人：我很脆弱也很真实

说话人：和我的团队一起。我尽力确保

讲话人：无论人的水平或

演讲人：终身教职，如果他们有一个好主意，

说话人：如何帮助他们追求那个目标。

讲话人：也做同样的事情。所以要喜欢把

说话人：有想法，

说话人：要真正有信念，

讲话人：进行后续工作，

讲话人：做类似的实际工作

讲话人：让它发生。所以这些都是事情

讲话人：我从交易中学到的。是的，

演讲人：我认为在很多方面都适用于任何工作。

说话人：那很美，人们在哪里可以找到

说话人：你在线，他们想关注你吗？

演讲人：听众如何对你有用？

说话人：我在社交场合的影响力不大。

说话人：我认为找到工作的最好方法，

说话人：我团队的工作真的很人性化

说话人：博客。当我们发布新模型时，

演讲人：新产品体验、

说话人：我认为，

说话人：对我有用。我认为最好的事情，

说话人：第一，是你的反馈。比如，

说话人：我们实际上，

说话人：如果你对任何事情表示赞成或反对

说话人：我们的产品表面，

说话人：如果您联系您的销售人员

演讲人：关于模型的反馈，

说话人：它会传给我的。实际上，我们，

演讲人：对于每一个研究模型，

说话人：我实际上已经非常接近理解了

说话人：好感度和反馈。所以给予

说话人：我们反馈，推动克劳德，

说话人：告诉我们它在哪里掉下来，

说话人：那些帮助我们让克劳德变得更好。

说话人：另一件事是，比如，

说话人：如果你的人际网络中有人

说话人：好像是这种类型的人

说话人：我刚才谈到了，

演讲人：我正在招聘，

演讲人：团队正在成长。我们真的会喜欢

说话人：只是热爱这项技术的人，

说话人：非常好奇的人，

说话人：第一原则思想家，

说话人：敢于质疑假设的人，

说话人：谁有像修修补补的黑客

说话人：精神。哇。真是一份梦想的工作。所以基本上，

演讲人：在 Anthropic 就该研究开放 PM 职位

说话人：团队。是的。他们申请了，

说话人：我想，在网站上，

说话人：招聘页面。是的。天哪。好吧，

说话人：我们开始吧，享受海量的简历吧

说话人：您即将收到，谢谢，

演讲人：Lenny.Dan,

说话人：非常感谢你来到这里。谢谢

说话人：你非常感谢我。谢谢你

说话人：真的很有帮助，

演讲人：发人深省的问题，

说话人：甚至帮助我把这些点联系起来

说话人：我们如何工作，

演讲人：这整个技术是如何到来的

演讲人：在一起并成为产品人员

说话人：是的。我真的很感激。但是谢谢

说话人：你，丹，说真的。好吧，好吧，

说话人：再见，

讲话人：各位，非常感谢你们的聆听。如果

说话人：你发现这个很有价值，

说话人：你可以在 Apple 上订阅该节目

说话人：Podcasts,Spotify,

说话人：或者你最喜欢的播客应用程序。另外，

说话人：请考虑给我们评分或

说话人：留下评论，

说话人：因为这确实对其他听众有帮助

讲话人：找到播客，可以找到所有过去的内容

说话人：剧集或了解更多有关节目的信息

说话人：at lennyspodcast.com.再见

说话人：下一集。

Lenny's Podcast 速读：Anthropic 首位 technical PM 如何在 token maxing 与 jagged edge 中做 AI 产品

先给结论

这期访谈在回答什么

1. Anthropic 的早期优势不是规模，而是共同下场

2. Coding 的转折，来自模型和产品一起到位

3. Token maxing 不是多花钱，而是把成本换成发现

4. Labs 的价值是押注非连续机会

5. Research PM 的工作，是把模糊失败翻译成可行动问题

6. Evals 正在变成 AI 产品的测试驱动开发

7. AI 时代的 PM 和经理必须亲手交付

8. 先有自己的判断，再让 Claude 反驳你

9. Jagged edge 让「谁来验证」比「谁来写」更重要

10. 值不值得完整收听

节目元信息

逐字稿说明

完整中英对照逐字稿

00:00:00

00:10:00

00:20:00

00:30:00

00:40:00

00:50:00

01:00:00

01:10:00

01:20:00

01:30:00

References

Related content