The Trillion-Dollar Industries AI Is Disrupting: Voice, Law & the End of the Billable Hour
频道: All-In with Chamath, Jason, Sacks & Friedberg
视频: https://allinchamathjason.libsyn.com/the-trillion-dollar-industries-ai-is-disrupting-voice-law-the-end-of-the-billable-hour
原文语言: en
统计: 共 69 轮 · Jason Calacanis 26 · Mati Staniszewski 24 · Max Junestrand 19
[0:00] Jason Calacanis
You're on a bit of a heater, huh?It's the best time to be building.And revenue has surged, but you face really intense competition.Let's go right at that to start.If you were building a global financial system from first principles today,you wouldn't build it on 50-year-old legacy rails.You'd build Airwallux, one AI-native platform for global accounts, cards, and payments.It's designed to make the entire world feel like a local market.Others are bolting AI onto broken infrastructure,but Airwallux was built for the intelligent era from day one.Stop paying the legacy tax and start building the future at airwallux.com slash all-in.Airwallux, built for the future.$350 million in, what, two or three years?
你最近手气很旺啊？现在是做产品最好的时候。收入暴涨，但你们面对的竞争也极其激烈——我们一上来就直接聊这个。
【广告】如果今天让你从第一性原理出发去搭一套全球金融体系，你不会把它建在有 50 年历史的老旧管道上。你会用 Airwallex——一个 AI 原生的全球账户、卡与支付平台，它的目标是让整个世界都像本地市场一样好做生意。别人是把 AI 硬栓在已经坏掉的基础设施上，而 Airwallex 从第一天起就是为智能时代而生。别再交「老系统税」了，去 airwallex.com/all-in 开始建设未来。Airwallex，为未来而建。【广告结束】
3.5 亿美元，用了多久？两三年？
[0:47] Jason Calacanis
And I'm hearing numbers $500 or $600 million now.Tell us about the revenue ramp of the company from the moment you released the software to today,that the product's been in market for 40 months, 50 months?
而且我现在听到的数字已经是 5 亿或者 6 亿美元了。跟我们讲讲公司的收入爬坡吧——从你们把软件放出去那一刻算到今天。产品在市场上大概 40 个月？50 个月？
[0:59] Mati Staniszewski
You tell me.Spot on.We started company 2022.First year was all about building the research and the product to really kickstart the work.We built the first text-to-speech model that finally could sound human.Released it in 2023, beginning of 2023.Then it took us roughly 20 months to get to the first $100 million in ARR.Roughly 10 months to get to $200.Five months to get to $300.And that's how we closed end of the last year.And now we are at $600.You're at $600 million in revenue.This is just extraordinary.How many employees now?
（Mati）你说呢。／（Jason）说得没错。／（Mati）我们 2022 年创立公司。第一年全部投在做研究和产品上，把这件事真正启动起来。我们做出了第一个终于能听起来像真人的文本转语音（text-to-speech）模型，2023 年初发布。之后大约花了 20 个月做到第一个 1 亿美元 ARR（年度经常性收入）。再用大约 10 个月到 2 亿。5 个月到 3 亿——去年年底就收在这个数上。而现在我们是 6 亿。／（Jason）你们收入到 6 亿美元了。这太不寻常了。现在有多少员工？
[1:37] Jason Calacanis
Because the company has obviously hit incredible valuations.But you have to fill in that valuation and you're competing at a very high level for talent.So tell us about how many employees you have now and how you maintain the culture of the company.When revenue is ripping, investors are throwing money at you, showing up at your doorstep.I mean, quite literally.But you've got to run the company.You've got to build a culture.So how many employees now and how are you dealing with these competing priorities?
因为公司显然已经拿到了非常高的估值。但你得把这个估值填实，而且你在很高的层面上抢人才。所以跟我们说说你们现在有多少员工、你又是怎么维持公司文化的。收入猛涨的时候，投资人拿着钱往你身上砸，直接堵到你家门口——我说的是字面意义上的。可你还是得把公司运转起来，还得把文化建起来。所以现在多少人？你怎么处理这些互相打架的优先级？
[2:10] Mati Staniszewski
Yeah, that's the key element of how you all...For us, the element of how we can maintain the culture despite the quick growth is kind of critical.And how we optimize both the interview cycle, how we are bringing people on board, how we onboard them.We have 600 people today.So also very quick growth on that people side.And as a company, we combine research and product.So we are building a communication platform for AI.On the research side, this includes everything across audio.Generating speech, transcribing speech, orchestrating speech for interactions.On the product, this is how we can complete the entirety of the customer journey.From marketing and creating assets and localizing them internationally.Through customer support with voice agents.To proactive enablement of how voice agents can help in operations, training, and sales.So this requires a lot of different talent.And a part of that revenue growth is actually a reflection of the functions we've grown over time.So from the original team, very research, very engineering heavy.From the first 10 people, we had zero attrition.Everybody is still at the company from those core research and engineering talent building together with us.
对，这正是关键的一环……对我们来说，怎么在快速增长中仍然保住文化，是极其关键的一件事。也包括我们怎么优化面试流程、怎么把人招进来、怎么做入职。我们今天有 600 人，所以人员这一侧也是很快的增长。作为一家公司，我们把研究和产品捏在一起。我们在做的是一个面向 AI 的通信平台。研究这一侧，涵盖音频的一切：生成语音、转写语音、为交互编排语音。产品这一侧，是怎么把客户旅程完整跑通——从做营销、做素材、把素材做国际本地化，到用语音 agent 做客户支持，再到主动赋能：语音 agent 怎么在运营、培训和销售里帮上忙。所以这需要非常多不同类型的人才。而那条收入增长曲线，其实有一部分就是我们这些年长出来的职能的反映。所以，从最早那支非常偏研究、非常偏工程的团队开始——最早的 10 个人，零流失。那批核心研究和工程人才到今天还全都在公司，一直和我们一起建。
[3:19] Mati Staniszewski
So, so far, we've been able to outcompete.And I think the common thread, and credit to my co-founder, who is an incredible researcher himself.We've been able to assemble the team that is truly excited about solving audio, solving interaction, and building that research.And if they are looking for an opportunity out there and looking for a company to join and solve that, we are one of the leading, if not the leading, place to do that.And you started before AI was so impactful at making software.Right.So when you were starting four years ago, five years ago, and working on this, building software was limited to low percentage of the population of planet Earth.You know, the number of people could write code.And now here we are, you know, when from Vibe, we had a no code moment, then Vibe coding.And now we actually have people building production code who are not developers.You have developers going 10x and token maxing.How has building software changed internally?
所以到目前为止我们一直能竞争得过。我觉得共同的那条线——这要归功于我的联合创始人，他本人就是个了不起的研究者——是我们组起了一支真正为「解决音频、解决交互、把这套研究做出来」而兴奋的团队。如果有人在外面找机会、找一家公司去解决这件事，我们就算不是唯一，也是最领先的那批地方之一。
（Jason）而且你们起步的时候，AI 还没有对「做软件」这件事产生这么大的影响。／（Mati）对。／（Jason）所以四五年前你开始做这个的时候，写软件还只是地球上很小一部分人能干的事——会写代码的人就那么多。而现在我们先经历了 no code（无代码）时刻，然后是 vibe coding（凭感觉写代码）。现在真的有不是开发者的人在写生产环境代码了。开发者本人则效率翻 10 倍、疯狂烧 token。你们内部「做软件」这件事发生了什么变化？
[4:23] Mati Staniszewski
And how do you deal with making sure that the code is really high quality?Because people are paying you this money, but they're going to demand really high quality product since they're spending so much money with you.Yeah, it's also true that 2022 was still the year where topics of the day were crypto and metaverse.So the building there was also the best time to start because we could actually take a bit of time to focus on what we thought is the future.But the way we are structured is a lot of small teams, especially across the product engineering, but also in how we think about go to market optimized for specific industries, telco, financial services, healthcare.So every unit is very tightly knit together.And we do that across the company.So it's usually five to 10 people teams that run ahead.And inside of each of those teams, the decision we took, which is slightly different than how it's usually structured, we embedded engineers in every place.And even in the places which aren't engineering.So our talent team will have an engineer.Our legal team will have an engineer.Our revenue engineering or go to market engineering have engineers embedded all across.And those people have two roles.
（Jason）以及你怎么保证代码质量真的够高？因为客户付你这么多钱，他们花这么多钱就一定会要求非常高质量的产品。
（Mati）是啊。另外 2022 年那会儿，当时的话题还是加密货币和元宇宙。所以在那时候动手其实是最好的时机——因为我们真的可以花点时间，专注在我们认为是未来的那件事上。至于我们的组织结构，是大量的小团队，尤其在产品工程这一侧；还有我们按行业（电信、金融服务、医疗）去做优化的 go-to-market（进入市场／销售打法）。所以每个单元内部都咬合得非常紧，整个公司都这么干，通常是 5 到 10 人的团队往前冲。而在每一个这样的团队内部，我们做了一个跟通常结构略有不同的决定：我们在每一个地方都嵌入了工程师，连那些本来不属于工程的地方也嵌。我们的人才团队里会有一名工程师，我们的法务团队里会有一名工程师，我们的 revenue engineering（收入工程）或者 go-to-market engineering（进入市场工程）里到处都嵌着工程师。而这些人有两个角色。
[5:33] Mati Staniszewski
One is, of course, creating automations and bringing the software inside of that team.But second is actually helping everybody else do what you said, which is make sure that people are adopting AI.But also there's a security check for everything they deploy.Because ultimately, if you're not using a lot of the coding software, a lot of the co-working software, then you're probably in the wrong spot.If you're using too much of it, that is also a flag because you maybe are not doing that in the right way.And of course, as you start bringing that into the sites of the organizations that never were exposed, they frequently can create but not necessarily review whether that's actually doing behind the scenes all the secure ways.So that's an essential role in the company.Yeah, it's fantastic that everyone can build software until you put it into production and you have a leak.Or that person leaves the company and people forget they built that software and it's just deprecating on its own.The other thing that seems to have changed is management.When you had 10 developers in your pod or six, you had a UX designer, you might have a pure graphic designer, you'd have a product manager, they rolled up.
（Mati）第一，当然是做自动化、把软件带进那个团队。第二，其实是帮所有其他人做你刚说的那件事：确保大家真的在用 AI。但同时也要对他们部署的每一样东西做安全检查。因为归根到底，如果你没在大量使用那些写代码的软件、那些协同工作的软件，那你多半站错位置了；而如果你用得太多，那也是一个警示信号——因为你可能没在用对的方式做。当然，当你开始把这套东西带进组织里那些从来没接触过的角落，他们常常是能「造」、但不一定能「审」——审不出这东西背后到底有没有用安全的方式在跑。所以这在公司里是一个不可或缺的角色。
（Jason）是啊，「人人都能造软件」很棒，直到你把它放进生产环境、然后出了数据泄露；或者那个人离职了、大家忘了这软件是他建的，它就自己在那儿慢慢烂掉。另一件看起来变了的事是管理。以前你的小组里有 10 个开发、或者 6 个，你会配一个 UX 设计师，可能还有一个纯平面设计师，你会有一个产品经理，大家往上汇报。
[6:46] Jason Calacanis
And then suddenly, you know, we watched over the past three years.Oh, hey, this is pretty good at summarizing what happened on the call.Oh, it is actually creating action items and it's telling us what to do next.Oh, and it's, you know, doing all the different stories in our Kanban board.And now how do you think about product managers and management as the CEO and as the co-founder?
然后突然之间，过去这三年我们眼看着：噢，这玩意儿总结会议内容还挺好用。噢，它居然能生成待办事项，还告诉我们下一步该干什么。噢，它还把我们看板（Kanban）里那些 story 全写了。那么现在，作为 CEO 和联合创始人，你怎么看产品经理、怎么看管理这件事？
[7:11] Mati Staniszewski
Yeah, we...You fired them all, right?We don't have any PMs.Right.Did you ever or did you have...Never.You never did.Never did.It's a little bit of what you mentioned also before the True AI Impact started,which was ideal person in that role, can code, can understand the customer, can understand design.Of course, that's very hard to find.There is no truly that many people that are experts in any...All of those fields at the same time.So we optimize for profiles that are experts in at least one of those fields,but understand at least one other field really well.To your point, what we are seeing now, there is...If you can do a little bit of all with AI,you can maybe step change from being an amateur to being an advanced level,maybe not an expert level.So suddenly, you are not bottlenecked on all the other functions to do your work.In growth, phenomenal.With growth engineering, a person can design experiments, ship an experiment,it's working and bringing it back.We also have the privilege where we are using a lot of our product ourselves.So to be able to do that, ultimately to help everybody else create voice agents,we ourselves need to create voice agents too.So we are seeing that also in the non-traditional functions,
（Mati）是这样，我们……／（Jason）你把他们全裁了，对吧？／（Mati）我们没有任何 PM（产品经理）。／（Jason）明白。以前有过吗？还是说……／（Mati）从来没有。／（Jason）你们从来没有过。／（Mati）从来没有。
这有点像你刚才在真正的 AI 冲击开始之前提到的那一点：这个岗位上的理想人选，是既能写代码、又能理解客户、又能理解设计的人。当然这种人非常难找。真正在这几个领域同时都是专家的人，本来就没那么多。所以我们优化的是这样一种画像：至少在其中一个领域是专家，同时对另外至少一个领域理解得非常好。
顺着你的思路，我们现在看到的是：如果你借助 AI 能把每样都干一点，你也许能实现一次台阶式跃迁——从业余水平跳到进阶水平，也许到不了专家水平。但突然之间，你做自己的工作就不再被其他所有职能卡脖子了。在增长这块，效果好得惊人：有了 growth engineering（增长工程），一个人就能设计实验、把实验发出去、跑通了再把结果带回来。我们还有一个便利条件，是我们大量使用自己的产品——为了能帮所有人做出语音 agent，我们自己首先就得做出语音 agent。所以我们在那些非传统职能里也看到了同样的现象，
[8:22] Mati Staniszewski
even in go-to-market.You need to be able to create a version of thatif we are offering that to the customers too.And we do.We created our inbound AI SDR agent.In addition to the form that you fill on the website,you have an agent that you can call.And people, of course, can give all the information in a much easier and quicker way.But the second thing that happened is people also leave a lot more information,so you can get connected to the right problem and right person a lot quicker.So we are seeing that kind of phenomenon all the time,where actually using a lot of tooling makes you yourself better in your job overall,and in 11 labs, in our specific tooling that we are solving for customers.Yeah, it seems like the use case of calling on the phone and talking to a computeror previously going through voice gel,and it was incredibly arduous and painful and annoying.It made you just say operator and hit the zero button as fast as possible.But now it seems to have turned a corner where talking to a human,I almost feel bad talking to a human where I'm like,I am so sorry I'm wasting your time with this.And the AI is just so much more precise.And the fidelity is so great that when you tell them what you're looking to do
（Mati）连 go-to-market 也一样。如果我们把这东西卖给客户，我们自己就得能做出一个版本来。我们确实做了：我们做了自家的 inbound AI SDR agent（入站 AI 销售开发代表 agent）。除了网站上那个填写表单之外，你还可以打电话给一个 agent。人们当然可以用轻松得多、也快得多的方式把信息给出来。但发生的第二件事是，人们还会留下多得多的信息，所以你能更快地把他们对接到正确的问题和正确的人那里。所以我们一直在看到这种现象：大量使用工具这件事本身，会让你自己在工作上整体变得更强——在 ElevenLabs，就是在我们为客户解决问题的那套工具上。
（Jason）是啊，看起来「打电话跟电脑说话」这个用例——以前走那种语音菜单，简直是折磨、痛苦又让人烦躁，逼得你只想赶紧喊「人工」、猛按 0 键。但现在好像拐过弯了：跟真人说话时，我几乎会有负罪感，心想「太抱歉了，我拿这么点小事浪费你时间」。而 AI 就是精确得多。而且保真度这么高，当你告诉它你想办什么事、
[9:38] Jason Calacanis
and you cut them off, you don't feel bad.You don't have to make small talk.Is that what you're seeing in your customer basein terms of the ability in real time to interrupt the agent,to interrupt the conversation and just move faster,has made consumers and companies basically embrace the technology?
然后你把它打断，你不会觉得不好意思。你不用寒暄。你在自己的客户群里看到的是这个吗——实时打断 agent、打断对话、直接快进的能力，基本上让消费者和企业都接受了这项技术？
[9:58] Mati Staniszewski
Yeah, it's slowly becoming that you will be asking for a give me an AI agenteffectively on the call.Give me the agent.AI operator.But we are seeing a transition where suddenly,and that's the biggest fuel of the recent growth for us,is enterprises, sales team just doing incredible work.But then finally the product combines the reliability that's corewith the orchestration for a lot of the AI models,but also the knowledge and the integrations to provide you the right experience.And yeah, I think it was a step change in the last 12 monthsand especially in the last six of how good that experience becamewhere it's like this golden era for the consumers out there,customers and other customers is comingwhere you can actually open a websiteand call an agent and have the agent have informationfrom your past interactions and deliver that help.And I think we'll see this kind of interesting phenomenacombining your previous question and thiswhere now, of course, you are reaching frequentlywhen you have a problem and you're asking for help,but ultimately, A, the whole interface will change and morphdepending on how you are operating with that interfacewith voice helping you in the background find that information.
是的，正在慢慢变成：你在电话里会主动要求「给我转 AI agent」。／（Jason）「给我转 agent。」AI 接线员。／（Mati）但我们看到的是一个转折——突然之间（这也是我们最近增长最大的燃料），企业客户起来了，销售团队干得非常漂亮。但最终还是产品把几样东西合在了一起：作为内核的可靠性、对一大堆 AI 模型的编排（orchestration），再加上知识和各种集成，共同给你正确的体验。是的，我认为过去 12 个月、尤其是最近 6 个月，那种体验好到发生了台阶式变化——外面的消费者、客户和客户的客户，那个「黄金时代」正在到来：你真的可以打开一个网站、给一个 agent 打电话，而这个 agent 掌握你过去所有互动的信息，然后把帮助给到你。我觉得把你上一个问题和这个问题结合起来，我们会看到一个有意思的现象：现在你当然常常是遇到问题、需要帮助时才去联系；但最终，第一，整个界面本身会随着你怎么使用它而改变、变形，语音在后台帮你把信息找出来。
[11:07] Mati Staniszewski
It will shift from reactive to proactive to help you get that helpbefore you potentially ask for it.And we are seeing those examples, those examples too.It seemed to me that speech to text had a major blocker.Again, in fidelity, 10 years ago, lawyers would put on Dragon Dictate,if you remember that terrible software, they get a headset.And it seemed like the big blocker was you felt like an idiottalking to a computer in an office, right?
（Mati）它会从被动响应转向主动出击——在你可能还没开口之前，就把帮助送到你面前。我们已经看到这样的例子了，确实有。
（Jason）在我看来，语音转文字（speech to text）曾经有一个大障碍。同样还是保真度问题：10 年前，律师会装 Dragon Dictate（听写软件）——你还记得那个糟糕的软件吗——戴上耳麦。而当时最大的障碍好像是：在办公室里对着电脑说话，你会觉得自己像个傻子，对吧？
[11:37] Jason Calacanis
And so people who did it quietly in their office,you know, they kind of got away with it.But now we see something very different.The whisper in the office, people very quietly talking to their computer,giving it a prompt, you know, and talking to their agents.And now there's a ring out, you can press it.And I use a really cool product called Whisperflow.I don't know if they use 11 labs on the back end.They use us and a few others as well.They are doing phenomenal work too.Whisperflow is just a tremendous product.And then I got a pedal.Does anybody here use a pedal on their computer?
（Jason）所以那些在自己办公室里偷偷用的人，还算蒙混过关了。但现在我们看到的完全不一样了：办公室里的「低语」——人们非常小声地对着电脑说话，给它下 prompt，跟自己的 agent 说话。现在还有一个戒指，按一下就行。我在用一个很酷的产品叫 Wispr Flow，不知道他们后端是不是用的 ElevenLabs。／（Mati）他们用我们，也用另外几家。他们干得也非常漂亮。／（Jason）Wispr Flow 真是个了不起的产品。然后我还搞了个脚踏板。在座有人用电脑脚踏板吗？
[12:12] Jason Calacanis
Raise your hand if you're a...There's one dork, two dorks.Any others?Raise it high.Oh, she's half dork.Okay, so there's about three and a half dorks here.Next year, this is going to be...Do you have a pedal?
是的话请举手。有一个呆子，两个呆子。还有别人吗？举高点。哦，她算半个呆子。好，这儿大概有三个半呆子。明年这就会变成……（转向嘉宾）你有脚踏板吗？
[12:27] Mati Staniszewski
I don't.Have you considered a pedal?I should consider a pedal.I love the devices that you can wear.I have the plod.It's incredible.Plod, pocket, phenomenal.Like, so good.And especially in events like this,I feel if you pre-preempted that you are recording, of course.But how incredible would it be?
（Mati）我没有。／（Jason）你考虑过脚踏板吗？／（Mati）我应该考虑一下脚踏板。我很喜欢那些能穿戴的设备。我有 Plaud，特别厉害。Plaud 那个口袋款，太棒了，好用得不行。尤其在这种活动场合——当然前提是你事先说明了自己在录音。但那会有多厉害啊？
[12:47] Jason Calacanis
All the signal, all the conversations that otherwise disappear.You maybe tap a few notes here and there to try to get signal afterwards.If you can just have that automatically fill your specific notes and make sure you do yourfollow-ups, phenomenal.All right.So let me make the case for the pedal.Okay.I have three pedals under the desk.And I think I'm trying to figure out what the company is.But with Whisperflow, you press down, it turns on, and you talk, and then you let it go.And one of the annoying parts of working with an LLM is typing.And you're kind of like exhausted when you're giving it the prompt.So you stop prompting.But if you're a professional bulls**t artist like me and a talker, this is like incredible.Because when I press the pedal down, I just give a stream of consciousness now.And it turns out what these LLMs actually do really well with is taking a massive streamof consciousness where you just keep talking and talking and talking.So I'll give it a one to two minute prompt.Then I let go.And it has changed everything.Everything.It's, you know, like the whole experience is changing so much.A similar version of what we see happen is, you know how you have, you want to say a thought
（Mati）所有那些信号、那些本来会消失掉的对话。你可能只能这儿记两笔那儿记两笔，事后再想从里面捞点信号。如果这些能自动填进你自己的笔记里、还确保你把该跟进的事都跟进了，那太厉害了。
（Jason）好，那我来给脚踏板辩护一下。我桌子底下有三个脚踏板。我还在琢磨做这个的公司到底是哪家。配上 Wispr Flow，你踩下去它就打开，你说话，然后松开。跟 LLM 打交道最烦人的一部分就是打字——你给它写 prompt 的时候会写到累，于是你就不写了。但如果你像我一样是个职业忽悠大师、话痨，这简直神了。因为我一踩下踏板，就可以来一大段意识流。而事实证明，这些 LLM 特别擅长处理的，恰恰是那种你一直说一直说一直说的大段意识流。所以我会给它一到两分钟的 prompt，然后松开。这彻底改变了一切。一切。整个体验变化太大了。
（Mati）我们看到的类似情况是，你知道那种——你想说一个想法，
[14:04] Mati Staniszewski
and then you're like, okay, I actually want to change and say something else.Now you have those two contexts combined.And the experience you get as an answer is so much better.So we already see that as an experience.But even the previous example of like people are adjusting how they speak to AI versus howthey speak to human.People are.How so?
然后你又想「等等，我其实想改一下、说点别的」。现在这两段上下文被合在了一起，而你得到的回答体验就好得多。所以我们已经把这当成一种新体验在观察了。但还有前面那个例子：人们说话的方式，对 AI 和对人是不一样的。真的在变。／（Jason）怎么个不一样？
[14:22] Jason Calacanis
Yeah.How should you speak to the LLM?We saw Sergey Brin say threaten it with bodily harm.It's a very effective technique if you haven't tried it.But what are the things that are different when you're talking to the LLM?
对。你该怎么跟 LLM 说话？我们看到 Sergey Brin（谢尔盖·布林，Google 联合创始人）说要用人身伤害威胁它。如果你没试过，这招其实挺有效。但跟 LLM 说话的时候，有哪些地方是不一样的？
[14:34] Mati Staniszewski
The specific emotional example.We work with a lot of financial services companies, Revolut, Klarna, PagBank.And some of the frequent case, not in all of them, is of course how you remind people aboutpayment or that you collect that from the people that aren't answering.And frequently people would naturally feel ashamed of telling the real situation.With AI, people are much more open to share what actually happened, give the information.And suddenly this emotional block of like in front of other human, I don't want to beable to say all of that, is very different.So that's different.Usually people are more snappy with AI voice agent.It's like quick responses.Yeah.You don't mind cutting it off.Exactly.So you can like kind of go through to the point you want much quicker, which you neededto like change a little bit of the interaction model too, which is working.But we'll work on the pedal and whether we should do an integration there.Let's talk a little bit about celebrities on the platform.You have some celebrities who are on there.You also have an issue with impersonation.Um, I know this because somebody was like, oh my God, I love your bulldog videos.Many people know I'm a big fan of bulldogs.
（Mati）说一个具体的、情绪层面的例子。我们和很多金融服务公司合作——Revolut、Klarna、PagBank。其中一个常见场景（不是全部，但很常见），当然就是怎么提醒人还款、或者怎么向那些不接电话的人催收。人们天然会因为羞耻而不愿说出真实处境。而面对 AI，人们要开放得多，愿意讲实际发生了什么、把信息给出来。突然之间，那种「在另一个人面前我说不出口」的情绪障碍，就完全不一样了。这是一个区别。另外，人们对 AI 语音 agent 通常更干脆——都是短平快的回应。／（Jason）对，你不介意打断它。／（Mati）正是。所以你能快得多地直奔你要的那个点，这也需要把交互模型改一改，而这套是跑得通的。不过我们会研究一下脚踏板，看要不要做个集成。
（Jason）我们聊聊平台上的名人吧。你们平台上有一些名人，同时你们也有冒名顶替的问题。我知道这个，是因为有人跟我说：「天哪，我超爱你那些斗牛犬视频。」很多人知道我是斗牛犬的铁粉。
[15:53] Jason Calacanis
I currently have three.Um, and I said, I'm sorry, I don't know what you're talking about.And they sent me a channel where somebody had created a bunch of dogs telling jokes and theymade one.And I guess they were looking for a podcaster.So they used the This Week in Startups archive and 11 lab to create my voice and do thishuge channel.And I contacted them and I said, oh my God, it's very flattering.How did you do this?
我现在养了三只。我说：不好意思，我不知道你在说什么。然后他们发来一个频道，有人在里面做了一堆讲笑话的狗，还做了一只（用我的声音）。我猜他们当时在找一个播客主持人的声音，于是就用《This Week in Startups》的存档加上 ElevenLabs 克隆了我的声音，做出了这个流量很大的频道。我联系了他们，说：天哪，我太荣幸了，你们是怎么做出来的？
[16:19] Jason Calacanis
This is like a year or two ago.And they said, oh, I used 11 labs.So I think I emailed you about it.I'm like, how do you protect against this?In advertising, in the law in the United States, I'm not sure about here in France.I'm sure they have 17 laws for this.We have one.You guys are great at regulations and laws.No offense.And the French guy over here is like, oh, mon dieu.Chekal.The, that's my French angry developer.I cannot smoke in the Louvre.This is crazy.And so it's super like interesting with this right to privacy.And I think you've got a quick education on this because you've had a couple people, I'msure, write you a legal letter.What it basically means is you can't take somebody's voice and use it to, you know, docommerce in the world.You can use it for parity.There is fair use.I can do a Donald Trump impersonation up here if I like.We're going to take about 5% of 11 lab stock.Is it okay with you to put them in Trump accounts?
这大概是一两年前的事。他们说：哦，我用的 ElevenLabs。所以我记得我还就这事给你发过邮件。我当时就想：你们怎么防这个？在广告领域、在美国的法律里——法国这边我不确定，我猜他们有 17 部法管这事，我们只有 1 部。你们（欧洲人）在监管和立法这块很在行，无意冒犯。这边那位法国老哥的表情就是「哦，我的天」。（模仿）这个……那是我的法国暴躁开发者：「我在卢浮宫居然不能抽烟。这也太离谱了。」
总之，这里面涉及的隐私权非常有意思。我猜你在这方面被快速教育了一轮，因为我相信肯定有几个人给你写过律师函。它基本的意思是：你不能拿别人的声音去做商业用途。你可以用于滑稽模仿（parody），那属于合理使用（fair use）。我要愿意的话，可以在台上模仿唐纳德·特朗普。／（模仿特朗普）我们要拿走 ElevenLabs 5% 的股份。把股份记在特朗普账上，你没意见吧？
[17:27] Jason Calacanis
Sounds good.Okay.And for that, you have to come to the White House.Great.Okay.Thank you.Nasty guy.Wouldn't give 5%.Loves socialism, but not America.What's the problem with the Nordics?Nasty, nasty socialism.
（模仿特朗普）听着不错。行。为这事你得来白宫一趟。太好了。行。讨厌的家伙。5% 都不肯给。热爱社会主义，就是不爱美国。北欧人到底怎么回事？讨厌，讨厌的社会主义。
[17:45] Jason Calacanis
Then I noticed when my guys wanted to clone my voice so that they could fix the ads whereI mispronounce something or I do the wrong promo code, use the code JCal20.They were like, it's 25, dummy.And I'm like, okay, I have dyslexia.And then they redid it.And it was like, I'm sorry, you cannot clone Jason's voice.And then it's like, I have to go in there and do it.And you put a bunch of protections in there.So explain what's happening in that regard in terms of people's, you know, concerns aroundthis.And then the other side, which is the opportunity, because I think you got Jamie Foxx and someother folks actually that you paid for their voices.Yeah.No, the voice is identity and IP.It's like, you know, when you speak a certain way, people recognize it, can feel that emotion.And, you know, to some extent, it could be a problem, could be opportunity before.I mean, as you did impersonation of the President Trump, it's, of course, similarly, somethingthat is possible even with a human, not specifically AI.But for us on the safeguard side, you know, over last years, we took the role as we areleading another development.We also need to lead another of the safeguards.So that's like a critical element.
（Jason）后来我注意到，我的团队想克隆我的声音，好去修那些广告口播里我念错的地方，或者我把优惠码念错了——「用 JCal20」，他们说「是 25，笨蛋」。我说好吧，我有阅读障碍。然后他们重做，结果系统说：「抱歉，你不能克隆 Jason 的声音。」于是变成我得亲自进去弄。你们在里面放了一堆保护措施。所以说说这方面在发生什么、人们对这件事的担忧是什么。还有另一面，也就是机会——因为我记得你们签下了 Jamie Foxx（杰米·福克斯）和另外一些人，是真金白银买了他们的声音。
（Mati）对。声音就是身份，也是知识产权（IP）。你用某种方式说话，人们能认出来、能感受到那种情绪。某种程度上，它可能是个问题，也可能是个机会。就像你刚才模仿特朗普总统一样，这种事其实一个真人也做得到，并不是 AI 独有的。但在防护这一侧，过去这些年我们的态度是：既然我们在引领这项技术的发展，我们也需要引领防护措施的发展。所以这是极其关键的一环。
[19:02] Mati Staniszewski
We do three things.One, trace everything that's generated so we can take action when needed.Two, now we moderate both on the voice and text level.So if you were to input something that would be commercial in nature or would try to scamsomeone, that gets flagged.We can block it.And now, three, because over the last years, we've seen the development of those modelsmore broadly.How can we create systems for the wider world so people can upload a sample and get information,whether it's AI or not, immediately?
我们做三件事。第一，追踪（trace）生成出来的每一样东西，这样需要时我们能采取行动。第二，现在我们在声音和文本两个层面都做审核——所以如果你输入的内容带商业性质、或者试图去骗人，就会被标记出来，我们可以拦掉。第三，因为过去这些年我们看到这类模型在更大范围内发展起来了：我们能不能为更广的世界建一套系统，让任何人都能上传一段样本，立刻知道它是不是 AI 生成的？
[19:30] Mati Staniszewski
And we do it for 11 Labs, but we also do it for other open source models.The interesting part, given that it's such a good IP and part of your element, it opensup new opportunities.So we partnered with Matthew McConaughey on creating a world cap.All right, all right, all right.And across languages.And it's the first...I haven't gotten paid a lot of money for these independent films, but, oh, 11 Labs stockis juicy.Yum, yum.Could you do it in Spanish?
（Mati）我们不只对 ElevenLabs 这么做，对其他开源模型也做。有意思的地方在于，正因为声音是这么好的 IP、是你身份的一部分，它反而打开了新的机会。所以我们和 Matthew McConaughey（马修·麦康纳）合作做了一个「全球声音」项目。／（Jason）好啊好啊好啊。／（Mati）而且是跨多种语言的。这是第一个……／（Jason）我拍那些独立电影可没赚到什么钱，不过 ElevenLabs 的股票很香啊，好吃好吃。你能用西班牙语说一遍吗？
[19:57] Mati Staniszewski
It's a fugazi, a fugazi.But the crazy thing with AI technology open is that now the voice can be created not onlyin English, but also in Spanish and Italian and Portuguese.And you can still have exactly that element of emotions coming through.So that's kind of a good example there.But we've seen that with Masterclass.What do you pay these guys?
（Jason 模仿麦康纳）这是假的，全是假的（fugazi）。
（Mati）但 AI 技术打开的疯狂之处在于，现在这个声音不只能用英语生成，还能用西班牙语、意大利语、葡萄牙语，而且那种情绪的传递依然原封不动。所以这是个不错的例子。不过我们在 Masterclass 上也看到了同样的事。／（Jason）你们给这些人付多少钱？
[20:19] Mati Staniszewski
What does it cost to get Matthew McConaughey?Is this like an eight-figure deal, seven-figure deal?You give them a little equity?Always depends.So like, you know, the Masterclass, for example, is a good example where they worked with talentdirectly.And here you have previously a static content that you would learn from.Now you have interactive content.So you have Gordon Ramsay teaching you how to cook in the kitchen.He can scream at you if you're not doing...F***ing raw!
（Jason）请到 Matthew McConaughey 要花多少钱？这是八位数的合同还是七位数的？还是给点股权？
（Mati）总是看情况。比如 Masterclass 就是个好例子，他们直接和艺人合作。以前你学的是静态内容，现在你有的是可交互的内容。所以你会有 Gordon Ramsay（戈登·拉姆齐）在厨房里教你做菜，你要是没做好他还能冲你吼。／（Jason 模仿）「他妈的还是生的！」
[20:44] Mati Staniszewski
Scallops are raw!So that is definitely...So they're doing characters now, or AI instances, using 11 Labs so you can interact with them aspart of your subscription.Exactly.But as a company, what we now do, and this from the beginning, we created a marketplacewhere people can create their voice.We authenticate it.You can share it.And you earn money.Today we paid back over $22 million back to the community of talent.Really?
（Jason 模仿）「扇贝是生的！」／（Mati）所以这确实是……／（Jason）所以他们现在在做角色、或者叫 AI 实例，是用 ElevenLabs 做的，你订阅之后就能跟他们互动。／（Mati）正是。但作为一家公司，我们现在做的、而且从一开始就在做的，是建了一个市场（marketplace）：人们可以在上面创建自己的声音，我们做认证，你可以把它分享出去，然后赚钱。到今天我们已经向创作者社区付回了超过 2200 万美元。／（Jason）真的？
[21:09] Jason Calacanis
So those voiceover actors now who got paid as hourly workers, sometimes they get a littleback-end if they were doing a commercial or something.Now they can spend an hour reading, create an 11 Labs voice, and then license it out?
所以那些以前按小时计酬的配音演员——有时候如果是广告之类的还能拿一点后端分成——现在他们可以花一个小时朗读，做出一个 ElevenLabs 声音，然后把它授权出去？
[21:25] Mati Staniszewski
100%.Do they get to pick their price or you pick the price?Depends on the model.We do both.So you can either give it a default that lets us distribute that slightly more optimally,or you can pick yours and the use case is going to be different.And like you said, opens up a set of incredible opportunities in a dynamic context in otherlanguages.But maybe a last one on that, like voice is such a big part of identity, and probably ourmost important work was actually working with people that lost their voice due to ALS, dueto throat cancer, and working on bringing that voice back.So I worked with congresswomen in the U.S., Jennifer Wexton, who lost it, and wanted tocontinue to inspire others that you can do incredible work despite that, and was the firstspeech delivered in Congress.Or more recently, I think this was the most heartwarming story.There was this woman that wanted to get married, lost her voice before she could get married.Oh, wow.And then they decided to redo the marriage together.Do the vows again?
（Mati）百分之百。／（Jason）价格是他们定还是你们定？／（Mati）看模式，两种我们都做。你可以给一个默认价，让我们能稍微更优地去分发；也可以自己定价，用例本身也会不一样。就像你说的，这在动态场景和其他语言里打开了一大批很棒的机会。
不过关于这个再说最后一点：声音是身份中如此重要的一部分，而我们可能最重要的工作，其实是和那些因为 ALS（渐冻症）、因为喉癌而失去声音的人合作，把他们的声音带回来。我和美国的一位女众议员 Jennifer Wexton 合作过，她失去了声音，但希望继续激励别人——即使这样你依然能做出了不起的事。那是（用这种方式）在国会发表的第一次演讲。还有更近的一件，我觉得是最暖心的故事：有位女士想结婚，但在结婚前失去了声音。／（Jason）哇。／（Mati）后来他们决定两个人再把婚礼重办一次。／（Jason）重念一次誓词？
[22:27] Mati Staniszewski
Do the vows.Oh.And you could see the whole family just for the first time hearing the vows.It was just, you could feel the emotions that you couldn't see in any other way, becausethe voice is such a connecting thing.Yeah.And you've done it for some iconic voices.My understanding is the estate of James Earl Jones.I'm not sure if they, did he pass?
（Mati）重念誓词。／（Jason）哦。／（Mati）你能看到全家人第一次听到那段誓词。那种情绪你能真切感受到，是任何别的方式都换不来的，因为声音是这么强的一种连接。
（Jason）是的。而且你们也为一些标志性的声音做过。我理解是 James Earl Jones（詹姆斯·厄尔·琼斯）的遗产管理方。我不确定他们……他去世了吗？
[22:49] Jason Calacanis
Is James Earl Jones alive?Can somebody ask?He passed, right?Yes.But before he passed, I think he did a deal with Disney, and he said, listen, for my family,I would like to license the Darth Vader voice for all time to Disney.They gave him some incredible deal, and then they were left with, well, how do we actuallydo this?
James Earl Jones 还在世吗？有人能查一下吗？他去世了，对吧？／（回应）是的。／（Jason）但在他去世之前，我记得他和迪士尼做了个协议，他说：听着，为了我的家人，我想把达斯·维达（Darth Vader）这个声音永久授权给迪士尼。他们给了他一笔非常可观的钱，然后迪士尼就面临一个问题：那我们具体怎么做？
[23:10] Jason Calacanis
Do we get a voice impersonator?But instead, they went to you.Talk a little bit about that deal and how it went down.And is that what they used recently in some of the new films with Darth Vader?There's a new Darth Maul series where they have Darth Vader, and did you power that?
我们去找个声音模仿者吗？结果他们没有，他们找了你们。讲讲那笔交易，以及它是怎么谈成的。最近几部新片里的达斯·维达用的就是这个吗？有一部新的《达斯·摩尔》剧集里出现了达斯·维达，那是你们做的吗？
[23:27] Mati Staniszewski
I don't know what I can say about the new things, but definitely the big use case thatdid that big, completely new experience was in the gaming space where Fortnite, so EpicGames, not game, Fortnite, launched Darth Vader, which people and players could interactwith live in partnership with the estate, in partnership with Disney.So every player, after reaching a certain stage, could have a Darth Vader interact and help yousolve the missions.And we are seeing that kind of mode coming up more and more often of how you can effectivelyextend your likeness, your, like you said, publicity into interactive use cases, bringit across the world together.So that was exactly that model.And now we are working on, one of the public ones is Headspace.So Headspace has a great meditation.Yes, this is the second greatest meditation app right behind Calm.Which you are an investor of.Oh, I am?
（Mati）新的东西我不知道能说什么，但确实有一个把这件事做得很大、体验完全崭新的用例，是在游戏领域：《堡垒之夜》（Fortnite）——是 Epic Games 做的——上线了达斯·维达，玩家可以实时和他互动，这是和遗产管理方、和迪士尼一起合作的。所以每个玩家打到一定阶段之后，就能让达斯·维达跟你互动、帮你完成任务。我们看到这种模式越来越频繁地出现：你可以把自己的形象、你的——像你说的——公开人格，延伸到可交互的用例里，把它带到全世界。所以那就是这个模式。
现在我们在做的，公开的一个是 Headspace。Headspace 有很好的冥想内容。／（Jason）对，这是仅次于 Calm 的第二好的冥想 App。／（Mati）而你是 Calm 的投资人。／（Jason）我是吗？
[24:24] Jason Calacanis
I didn't realize.You're right.I did invest in Calm.But it was a $4 million company.But Calm is incredible.I think their team...But anyway, you were working with the second place.Exactly.Not exactly the second place.Not exactly to the working part.So they are localized a lot of the content.And Calm, I think, is trying some of the interactive elements.Could you have a meditation lesson that's personalized to you?
（Jason）我都没意识到。你说得对，我确实投过 Calm，不过那时候它才是个 400 万美元估值的公司。但 Calm 太棒了，我觉得他们团队……总之，你合作的是第二名。／（Mati）也不完全是第二名。也不完全是……在合作那部分上不完全是。他们把很多内容做了本地化。而 Calm 我觉得是在尝试一些可交互的元素。能不能有一堂为你个人定制的冥想课？
[24:48] Jason Calacanis
Which we would love to...That would be amazing.And like, imagine just, you know, so many voices.David Sachs is defending Trump.Take a deep breath in.Breathe out.Breathe in.Breathe out.Maybe you should license the voice to Calm.I mean, that would be interesting.Let's talk a little bit about being up against some of the greatest entrepreneurs ever whowant to take your business from you.Specifically, Dario and Anthropic.Sam from OpenAI.They want your business.They've been pretty clear about it.And I think you have used the frontier models in your product.But you must be thinking, my lord, am I enabling my own demise by partnering with them?
（Mati）那我们会非常乐意……／（Jason）那就太棒了。而且想象一下，那么多种声音。「David Sacks（大卫·萨克斯）正在为特朗普辩护。深吸一口气。呼气。吸气。呼气。」也许你该把这个声音授权给 Calm，那会挺有意思。
我们聊聊——你正在和一些史上最厉害的创业者正面撞上，他们想把你的生意抢走。具体说就是 Anthropic 的 Dario，还有 OpenAI 的 Sam。他们想要你这门生意，这点他们说得很明白。而且我知道你们产品里用了前沿模型（frontier models）。你心里一定在想：天哪，我跟他们合作是不是在给自己掘墓？
[25:39] Mati Staniszewski
And there's all these open source models.So how do you think about your partnerships with those type of frontier models and the factthat they want to kill your company?So on the first part, given we create a platform, we try to provide all our lamps out there soour customers can pick.Anthropic, OpenAI, open source, Google models.And that agnostic, being agnostic to a specific model is actually helpful because customers,can they make sure that they build a harness, build the agent orchestration, create a voiceelement of how that agent interacts with the world, how the marketing way interacts withthe world.But they are not dependent on any model.So for us, that part is actually good because we can provide that to the customers.On the kind of the second big part of like, of course, the space is overlapping.Increasingly, models are platform, platform are application.Everything is becoming a little bit more fuzzy.For us, there's still the defining piece was focusing on that one layer of like, how doesinteraction look like?
（Jason）而且外面还有那么多开源模型。所以你怎么看和这类前沿模型的合作关系、以及他们想干掉你公司这件事？
（Mati）先说第一部分：因为我们做的是平台，我们尽量把外面所有的大模型都提供出来，让客户自己挑——Anthropic、OpenAI、开源模型、Google 的模型。这种对具体模型保持中立（agnostic）的做法其实是有帮助的，因为客户可以放心去搭 harness（外层脚手架）、去做 agent 编排、去创造这个 agent 怎么和世界交互、营销侧怎么和世界交互的语音层，而不被任何一个模型绑死。所以对我们来说，这部分反而是好事，因为我们能把这个能力提供给客户。
至于第二个大问题：当然，这个空间是重叠的。而且越来越是——模型即平台，平台即应用，一切都变得有点模糊。对我们来说，定义性的那一块仍然是聚焦在那一层：交互到底长什么样？
[26:43] Mati Staniszewski
How does communication look like?And we've been able to out-compete them on voice models, both on text-to-speech, speech-to-text,on the turn-taking, on music.And here, our research team is a set of magicians that are able to continuously do it time andtime again.And I think part of the reason is it's on the research side.It's the architecture that matters, not the scale.You really need to change how the model operates.Two, you need very specific data that there's, of course, a wide set of data out there, butit's unlabeled data and where we spend a lot of time.So you build an internal team of over 1,000 contractors that label all those audio assetsto make them good.So that's on the research side.And then as we think about the rest of product stack, we want to create a fully verticalizedsolution for that communication angle.The product understanding the right workflow in financial services is very different tohealthcare, very different to telcos.We spend all of our product team to figure out how that works and those companies done.And then ultimately, last piece is the ecosystem.Can you build the wider set of integrations, voices that you use, templates?
沟通到底长什么样？而我们在语音模型上一直能竞争得过他们——文本转语音、语音转文本、轮次交接（turn-taking，即人机对话中谁该开口的节奏控制）、音乐都是。这里我们的研究团队是一群魔术师，能一次又一次持续做到。我想部分原因在研究侧：重要的是架构，不是规模。你真的需要改变模型的运作方式。第二，你需要非常特定的数据——外面当然有海量数据，但那些是没标注的数据，而我们在这上面花了大量时间。所以我们建了一支超过 1000 人的内部标注（外包）团队，把所有那些音频素材标好、变成可用的。这是研究侧。
然后再看产品栈的其余部分：我们想为「沟通」这个角度做一套完全垂直化的解决方案。产品上，金融服务里正确的工作流和医疗完全不同，和电信也完全不同。我们把整个产品团队都投进去，搞清楚这些怎么跑通、那些公司怎么做。最后一块是生态：你能不能建起更广的一整套集成、可用的声音、模板？
[27:49] Mati Staniszewski
For the agent authentication that you can benefit from instead of starting from scratch.And so far, we've been able to create a new model for that.Certainly, though, you must be concerned about, hey, the reinforcement learning, the data leakage.They say they're not using your data, but they're kind of using your data.And so do you have an open source project internally as the, like, in case of glass, we've got to break this?
（Mati）还有 agent 认证（agent authentication），让你能直接受益，而不用从零开始。到目前为止，我们已经为此建出了一套新模式。
（Jason）不过你肯定还是会担心：嘿，强化学习（reinforcement learning）、数据泄漏。他们说他们没在用你的数据，但其实多少在用。所以你内部有没有一个开源项目，就像「紧急时砸碎玻璃」那种备胎？
[28:15] Mati Staniszewski
And when do you think you'll be able to discontinue working with them if you had to?We know that some companies are continuously trying to figure out how to distill and use the data.So that is an existing problem.And we have a few mechanisms to stop it or slow it down, not stop it.But on the open source question, or, like, creating our own versions, we are looking a little bit closer on, like, how we could use our expertise of how does, like, you know, we won't focus on knowledge work.We won't focus on coding.But any interaction and how you can combine all those pieces together and make sure this is great, we won't own.So we are spending more time there.But it's also just great to be in the arena and compete with those guys and every so often show that we can do it and do it better.Yeah, it's pretty clear in my estimation that that's where you will wind up.And the ability to make your own language model today, especially with all these great models out there that are now open sourced, it's going to be pretty easy for a company with your level of resources.So why wouldn't you?
（Jason）以及你觉得到什么时候，如果非要断，你就能不再和他们合作了？
（Mati）我们知道有些公司一直在琢磨怎么蒸馏（distill）、怎么用这些数据。所以这个问题是真实存在的。我们有几种机制去阻止它、或者说拖慢它——不是彻底阻止。至于开源那个问题，或者说做我们自己的版本：我们在更仔细地看，怎么把我们的专长用起来。我们不会去做知识工作，不会去做写代码。但任何「交互」，以及怎么把这些拼在一起、把它做好——这个我们不会让出去。所以我们在这上面花的时间更多。不过说到底，能进到这个竞技场里和那几位同台竞争，本身就很棒，而且时不时能证明我们做得到、而且做得更好。
（Jason）是的，在我看来很明显，你最后会走到那一步。而且今天要做自己的语言模型，尤其外面这么多优秀模型都开源了，对一家像你们这种资源水平的公司来说会相当容易。所以你为什么不做呢？
[29:32] Jason Calacanis
At least offering it as an option.And then I guess there's cost.I mean, you must be shipping tens of millions of dollars to the Frontier models every year?Ship a good amount.We are good partners.We're good partners with them.But it's ultimately showing up in the value we can create, too.So a lot of what we spoke at the beginning of how we can elevate ourselves as an organization, too, is definitely helpful.So I think they've done tremendous work on building.It's almost crazy that each of us has a Turing.If you were to chat with an agent now, it feels like the Turing test will be completed.It's as smart as another human.And we hope this year we'll do that same thing for voice, where any conversation feels like you are speaking with another human.Yeah, I think you're there.It just depends on the application and what question you ask.But it definitely passes.I mean, if we were to look at the tests that were created to define artificial general intelligence or just to define artificial intelligence, we passed all of those.These were tests that were created 30 or 40 years ago.We need a new set of tests right now.I think the new test is like, can this be more intelligent than every single person on the planet times 10?
（Jason）至少可以作为一个选项提供出来。然后我猜还有成本问题。你们每年得给前沿模型付出好几千万美元吧？／（Mati）付得不少。我们是好伙伴。／（Jason）你们是好伙伴。／（Mati）但归根到底，这也体现在我们能创造的价值上。我们一开始聊的那些「怎么把我们这个组织自身抬升上去」的做法，绝对是有帮助的。所以我认为他们在建设这件事上做得非常出色。几乎有点疯狂的是，我们每个人手里都有一台「图灵机」。如果你现在去和一个 agent 聊天，感觉图灵测试（Turing test）已经通过了——它和另一个人一样聪明。而我们希望今年能在语音上做到同样的事：任何一段对话，感觉都像在和另一个真人说话。
（Jason）是的，我觉得你们已经到了。这取决于具体应用、取决于你问什么问题，但它确实通过了。我是说，如果我们回头看当年为定义通用人工智能、或者只是为定义人工智能而设计的那些测试，我们已经全部通过了。那些测试是 30 年、40 年前设计的。我们现在需要一套新的测试。我觉得新的测试大概是：这东西能不能比地球上每一个人都聪明 10 倍？
[30:53] Jason Calacanis
And if we get anything less than that, we're kind of like, oh, yeah, it's not smart.I mean, these things, we're kind of there on AGI, don't you think?That we've kind of achieved it.We just haven't deployed it.There are definitely places where we did achieve it.Yeah, for sure.All right.Continued success.Let's give it up for Mati from 11 Left.Well done.Thanks so much.Thanks for coming out.The AI companies Building the Future run on Oracle Cloud Infrastructure, training and deploying at scale on one of the world's largest AI infrastructures.
（Jason）如果达不到这个，我们就会说「嗐，也就那样，不算聪明」。我是说，这些东西——我们在 AGI 上基本已经到了，你不觉得吗？我们其实已经实现了，只是还没部署出去。／（Mati）确实有些地方我们已经做到了。／（Jason）当然。
好，祝继续成功。让我们把掌声送给 ElevenLabs 的 Mati！干得漂亮。／（Mati）非常感谢。／（Jason）谢谢你来。
【广告】那些正在建设未来的 AI 公司，都跑在 Oracle Cloud Infrastructure 上，在全球最大的 AI 基础设施之一上做训练和大规模部署。
[31:25] Jason Calacanis
The same Oracle AI platform gives enterprises access to leading models, AI grounded in their own data, and the security to move from pilot to production.Learn more at oracle.com slash AI or experience it live at Oracle AI Experience Live.You're growing also at a very significant clip.Exponentially.Is it exponential?
【广告续】同一套 Oracle AI 平台还让企业能用上领先模型、用自有数据为 AI 打底，并具备把项目从试点推到生产的安全能力。了解更多请访问 oracle.com/ai，或者到 Oracle AI Experience Live 现场体验。【广告结束】
（Jason）你们的增长速度也非常快。／（Max）是指数级的。／（Jason）真是指数级吗？
[31:50] Max Junestrand
No, it's not exponential.Oh, it's sustained 50% quarter over quarter for the last seven quarters.50% quarter over quarter last seven quarters.Yeah, that's pretty darn fast.So I think we actually just became, as of the close last week on Tuesday, one of the fastest enterprise companies with a direct sales motion to go from one to 150, bidding Sierra with one quarter.Amazing.Amazing.And so people, I mean, there's a couple of things in life that people really hate.And paying lawyers is like way up on the top of the list.With your tools, obviously, you got your contemporary and Harvey and people.Sorry, who?
（Max）不，不是指数级。／（Jason）哦。／（Max）是连续七个季度保持每季度环比 50% 的增长。／（Jason）连续七个季度环比 50%。是的，那是相当快了。／（Max）所以我认为，截至上周二收盘，我们刚刚成为采用直销模式的企业软件公司里、从 100 万做到 1.5 亿（美元 ARR）最快的公司之一，比 Sierra 还快了一个季度。／（Jason）了不起。了不起。
所以说——人生里有几件事是人们真的深恶痛绝的，给律师付钱绝对排在最前面。有了你们的工具，当然，你还有同行 Harvey 那些人。／（Max）不好意思，谁？
[32:32] Jason Calacanis
It's a small company in the States.And then you also have, I guess, Claude and other folks also want to be in your business.So this is a big prize to take, I don't know, 80% of what we pay lawyers for and compress it by 90%.Like what is the realistic power law here in terms of making for startups in the audience, your legal bills dramatically drop in costs.Yeah.And I'm seeing it already in the startup space.I had one firm, one startup that hit a million in revenue.Yeah.They had closed multiple rounds of funding, multiple, obviously, large number of employees, a decent couple dozen employees.They didn't have a corporate lawyer.And I said, well, you think at a million dollars in revenue, like somebody should review the contracts?
（Jason）美国的一家小公司而已。（笑）然后我猜还有 Claude 之类的也想进你们这门生意。所以这是一块很大的蛋糕——把我们付给律师的钱拿走 80%、再把成本压掉 90%。对台下的创业公司来说，你们的法务账单会大幅下降——现实中的幂律（power law，即少数玩家吃掉绝大部分收益的分布）会长什么样？
（Max）是的。／（Jason）我在创业圈已经看到了。我碰到过一家公司，收入做到 100 万美元，融了好几轮，员工也有不少了、好几十号人，却没有公司法务。我说：你觉得做到 100 万美元收入，是不是该有人来审一下合同？
[33:29] Jason Calacanis
And they're like, chat GPT, bruh.And I'm like, well, what about the cap table?They're like, chat GPT, bruh.And I was like, okay.And HR?And they're like, same thing, bruh.And I'm like, okay.There's got to be a fun diligence target one day.Well, that's what I said.I said, hey, when you do the Series A, they're going to ask some of this stuff to be reviewed.Like, do you guys have, like, IP assignments?
他们说：「ChatGPT 呗，兄弟。」我说：那股权表（cap table）呢？「ChatGPT 呗，兄弟。」我说：行吧。人力资源呢？「一样，兄弟。」我说：好吧。／（Max）这以后会是个很有意思的尽调标的。／（Jason）我就是这么说的。我说：等你们做 A 轮，人家会要求审这里面的一些东西。比如你们有没有做知识产权归属（IP assignment）？
[33:54] Jason Calacanis
They're like, yeah.I'm like, how did you know to do IP assignments?First-time founders are like, we asked chat GPT.And I'm like, okay, wow, I just turned into Unc.Like, I guess.Yeah.So take us through what you think is happening out there.This is not uncommon, right?
他们说：「有啊。」我问：你们怎么知道要做 IP 归属的？这些第一次创业的人说：「我们问了 ChatGPT。」我就想：好吧，哇，我一下变成「老登」了。／（Max）是啊。／（Jason）那你说说你看到外面在发生什么。这种情况不少见吧？
[34:07] Max Junestrand
What scenario I'm talking about?But a seed stage startup operates very differently from, you know, one of the biggest banks in the U.S.And so the way to think about the market, or at least the way that we like to, is you have this enormous bucket of legal services, which today is being done manually.It's a trillion dollars every year into legal services, which is very fragmented.But the software spend into legal technology is about 40 billion.So it means there's 4% software, 96% service, which is bananas.The software piece should be much bigger than that.And so the software piece naturally will grow into the service revenue.But also legal is a very supply-constrained market.The demand for legal services is much larger than what there are lawyers or legal services available.And so many of the legal service providers are now using technology to serve new use cases, new market segments, and to actually package new products.And you will not make...What's an example of that?
我说的那个场景？种子期创业公司的运作方式，和美国最大的几家银行完全不同。所以看这个市场的方式，或者至少我们喜欢的看法是：你有一个巨大的法律服务池子，而它今天是靠人工在做。每年有 1 万亿美元流进法律服务，而且非常分散。但流进法律科技的软件支出大约是 400 亿美元。也就是说 4% 是软件、96% 是服务，这太离谱了。软件那块本该大得多。所以软件这块自然会吃进服务那块的收入。而且法律是一个供给严重受限的市场：对法律服务的需求远大于现有律师或法律服务的供给。因此现在很多法律服务提供方都在用技术去服务新的用例、新的细分市场，甚至打包出新的产品。而你不会……／（Jason）举个例子？
[35:15] Max Junestrand
So an example of that is Cooley, actually.They started serving startup founders directly with a sort of software platform that you just log onto the platform.They pumped it full with their material and their precedent.And then you have the startup material there.And they've embedded workflows that reviews the contracts.And what I think is interesting by that is it starts to break this model where you charge out associates for very high hourly rates.And you have a billable hour model.And actually, if you look in law firms, the way that that business model works is you overcharge for the associates.And you actually undercharge for the partners.I don't know if they're undercharged.I mean, I got a bill recently and it was $1,800 an hour.Right.But for a senior person, I think the associates were $800 an hour.Well, you know, Kirkland can go up to $4,000 an hour.But the thing is, when a Kirkland, so let's say, you know, 30 minutes of a Kirkland partner's time when it really matters can be worth a lot more than that.Like a lot more than that.If it's bet the company litigation or you avoid a pitfall that would have costed the company tens of millions of dollars.Well worth it.Yeah.Right.Exactly.
（Max）比如 Cooley（科律律所）就是一个例子。他们开始直接服务创业公司创始人，做了一个软件平台，你登录进去就能用。他们把自己的资料和判例先例（precedent，即以往可参照的成案与合同范本）灌了进去，创业公司要的材料也都在里面，然后他们把工作流嵌进去、自动审合同。我觉得有意思的地方在于，这开始打破那种模式——你按很高的小时费率把 associate（初级律师）的时间卖出去，靠计费小时（billable hour）赚钱。而且实际上你去看律所，这门生意的模式是：你对 associate 收费过高，对合伙人收费反而偏低。
（Jason）我不确定合伙人算不算收低了。我最近收到一张账单，1800 美元一小时。／（Max）对。／（Jason）但那是资深的人，初级律师我记得是 800 美元一小时。／（Max）Kirkland（凯易律所）能到 4000 美元一小时。但问题是，当 Kirkland 的合伙人在真正关键的时刻花掉 30 分钟，那 30 分钟的价值可以远远超过这个数——远远超过。如果那是一场生死攸关的诉讼、或者你因此避开了一个会让公司损失几千万美元的坑，那太值了。／（Jason）是的。对。正是。
[36:32] Max Junestrand
And but the only way they know how to price that is to overcharge for the associates.But as you're saying, the enterprises are looking at this and they're going, huh, we're spending a lot of dollars on legal services.Let's take this in-house.Oh, really?
（Max）但他们唯一知道怎么给这件事定价的办法，就是对 associate 收费过高。／（Jason）不过就像你说的，企业方看着这个在想：嗯，我们在法律服务上花了好多钱，把这块拿到内部来做吧。／（Jason）真的吗？
[36:45] Max Junestrand
Absolutely.I mean, we're doing this partly at Legora.We acquired four businesses so far this year.We did the diligence in-house with our own tool.And the fastest transaction we did was 12 days from LOI to closing.Because your motivation as the founder is to get the deal done.Right.The motivation of the lawyer is to not have you sue them if they f*** up the deal.Right.And to make as much money as possible.Which means to drag it out.Which means their incentive is to, even if they don't say it explicitly, it is to drag it out.Your incentive is to close it as quick as possible.Yeah.Yeah.And so, you know, I think a lot of law firms are also experimenting with different pricing models where you do a fixed fee for a transaction or for a fundraise.In litigation, you can take a part of the success fee when you win the deal.Yeah.Or win the case.And so, I think it's just very interesting how, you know, one of the biggest industries in the world now is being completely transformed and reshapen as a consequence of the tech.And are those law firms feeling like they're being disrupted or this is a huge opportunity?
（Max）绝对的。我们在 Legora 内部就有一部分是这么干的。我们今年到目前为止收购了四家公司，尽调全部是用我们自己的工具在内部做的。最快的一笔交易，从签意向书（LOI）到交割只用了 12 天。／（Jason）因为作为创始人，你的动机是把交易做成。／（Max）对。／（Jason）而律师的动机是：别让你在他们搞砸交易时反过来告他们。／（Max）对。而且要尽可能多赚钱。／（Jason）也就是拖长时间。／（Max）也就是说，即使他们不明说，他们的激励就是把它拖长；而你的激励是尽快关掉。／（Jason）是的。
（Max）所以我觉得现在很多律所也在试不同的定价模式——按一笔交易或一次融资收固定费用；在诉讼里，可以在打赢时拿一部分成功费。／（Jason）或者赢下案子。／（Max）所以我觉得非常有意思的是，世界上最大的行业之一，正因为技术而被彻底改造和重塑。／（Jason）那些律所是觉得自己在被颠覆，还是觉得这是个巨大的机会？
[37:52] Max Junestrand
And did that switch at a certain point in time or has it switched for them?There's a lot of anxiety and a lot of fear.And, you know, these law firms are enormously profitable and big businesses.Kirkland Ellis turns around $10 billion a year.How many lawyers did it have?
（Jason）这个心态是在某个时间点切换的吗？还是他们已经切换过来了？／（Max）有很多焦虑，也有很多恐惧。而且这些律所本身是利润极高的大生意。Kirkland & Ellis 一年大约做 100 亿美元。／（Jason）它有多少律师？
[38:15] Max Junestrand
It's like 4,000 or 5,000.Wow.I mean, per partner, they make it between 5 and 10 million every year in profits.And so, when something like AI comes along, that poses existential threat and existential opportunity.And that's actually a big part of my job to help articulate with the leadership teams that we work with.Because we will only be as successful as our customers are.And so, we actually have a very unique role at Ligora as well, which is called the legal engineer.So, in the same way that Palantir has forward-deployed engineers, we have forward-deployed lawyers.And their job is to sit down with the Kirkland partners and help them transform their business from a pre-AI to a post-AI world.And it's sort of like document management and PCs were, but 20 or 30 years ago, when they were printing out and keeping drafts in a library and in a storage facility.And they had to sort of walk them through and handhold that.Absolutely.But I think the difference is...Revisions, right.Yeah.The difference is those were, you know, mild productivity gains.Yeah.This can do a lot of the work.And so, it's really reshaping what it also means to be a junior lawyer going into this occupation.What does it mean?
（Max）大概 4000 到 5000 人。／（Jason）哇。／（Max）平均下来，每位合伙人每年的利润在 500 万到 1000 万美元之间。所以当 AI 这种东西出现时，它同时构成生存威胁和生存级的机会。而这其实是我工作中很大的一块：帮我们服务的这些领导班子把这件事讲清楚。因为我们的成功上限就是客户的成功上限。
所以我们在 Legora 也有一个非常独特的角色，叫 legal engineer（法务工程师）。就像 Palantir 有 forward-deployed engineer（前线部署工程师）一样，我们有 forward-deployed lawyer（前线部署律师）。他们的工作就是坐到 Kirkland 的合伙人旁边，帮他们把生意从「AI 之前」改造成「AI 之后」。
（Jason）这有点像二三十年前文档管理系统和个人电脑刚进律所的时候——那会儿他们还在把草稿打印出来、存进资料室和仓库里，得有人手把手领着他们走一遍。／（Max）完全是。不过我觉得区别在于……／（Jason）改版本，对。／（Max）区别在于那时候是温和的效率提升；而这一次，它能把很多活儿直接干掉。所以这也在重塑「当一名初级律师」到底意味着什么。／（Jason）意味着什么？
[39:37] Max Junestrand
Are those jobs going to still exist?Or are a lot of the lawyers who are coming out of school going, oh my God, was this a good idea or a bad idea?The job will exist.The tasks will be different.Right?In order to have a partner-driven model, you need to bring people up the ranks.Right?
（Jason）那些岗位还会存在吗？还是说很多刚从法学院出来的律师正在想「天哪，我这个选择到底对不对」？／（Max）岗位会存在，但任务会不一样。对吧？要维持一个以合伙人为核心的模式，你必须让人一级一级往上长。对吧？
[39:56] Max Junestrand
In the same way as you do with software engineers.You need to have junior engineers so that one day you can have senior engineers who know what they're doing.But the way to get there is very different.The way of getting there today will not be lock yourself in the physical data room, read through every single document, mark the errors, and, you know, go fax it.Right?
就像软件工程师一样。你必须有初级工程师，将来才会有真正懂行的资深工程师。但走到那一步的路径会非常不同。今天的路径不会再是：把自己关在实体资料室里，把每一份文件读一遍、把错误标出来，然后拿去传真。对吧？
[40:18] Max Junestrand
And it's also no longer just looking at the virtual data room and control F.It's orchestrating the agent that will be doing that work.And when you look at that work, you have a global backdrop.Attorneys, obviously, very famously localized.Right?
（Max）而且也不再只是看虚拟资料室然后按 Ctrl+F，而是去编排（orchestrate）那个替你干活的 agent。
（Jason）而当你去看这些工作时，还有一个全球背景：律师这个职业出了名地高度本地化。对吧？
[40:37] Jason Calacanis
And is this going to create attorneys who can operate across borders in a way that didn't exist?And you're starting to see that.And is that something that's built into the product?So when you're doing, even in the United States, it's state-level certification, obviously.And doing a non-compete in the Northeast is very different than doing it in California.They're not very enforceable or enforceable at all in California, as people don't know.But they're quite enforceable if you're in Boston.Yeah.Exactly.So talk about that.Because that seems to be a place where there could be massive gains from AI.100%.And it's really two things.I mean, the data that Legora sits on top of is, on one hand side, the firms and enterprises own data.They're precedent.They're organizational data.And secondly, we do the hard work of gathering all the cases, all the legislation, all the regulatory updates for every jurisdiction in the world.And that is very painful.But once you start to do that at scale, it builds a real data mode.And so in the system, if you are the GC of a company in California and you just landed your first customer in South Africa, right?
（Jason）这会不会造就一批以前不存在的、能跨境执业的律师？你已经开始看到了吧？这是产品里内建的能力吗？因为哪怕在美国，执业资格也是按州发的。在东北部做竞业禁止（non-compete）和在加州做完全不一样——很多人不知道，在加州竞业禁止基本不可执行、甚至完全无效，但在波士顿就相当能执行。／（Max）对。／（Jason）正是。所以说说这个，因为那看起来是 AI 能带来巨大收益的地方。
（Max）百分之百。而且主要是两件事。Legora 底下坐着的数据，一方面是律所和企业自己的数据、他们的判例先例、他们的组织数据；另一方面，我们干的苦活是把全世界每一个法域的所有判例、所有立法、所有监管更新都收集起来。这非常痛苦。但一旦你开始规模化地做这件事，它就会长出一条真正的数据护城河。所以在系统里，如果你是加州一家公司的总法律顾问（GC），而你刚拿下第一个南非客户——
[41:50] Max Junestrand
Legora can be adapted to the local legislation in South Africa.And we actually had a case of this where, you know, instead of having to call a lawyer who then knows a lawyer in that region who will respond to the query, they can get an 80% accurate response immediately that they can start working out of.And the better that gets, the more interesting things I believe you can do because this data has really never been structured before.And there are so many people who are working with setting policy and building regulation.And this is an enormous inefficiency in society.And LexisNexis has been a juggernaut and the legacy player in, you know, all the case law and regulations.They have a massive data moat.But they only make a couple of billion dollars a year.And if you put your revenue and Harvey's revenue together, you guys are probably already just that, the two of you, you're both making hundreds of millions of dollars.So they must be looking in the review mirror at you like the Tyrannosaurus Rex in Jurassic Park and going, holy shit.Like, are they coming for our business?
（Max）Legora 可以适配到南非当地的立法。我们真的碰到过这样的案例：原本你得打电话找一个律师，他再认识那个地区的一个律师，那位律师再来回复你的问题；而现在他们能立刻得到一个 80% 准确的回答，拿着它就能开始干活。而且这一块做得越好，我相信能做的有趣的事就越多，因为这些数据以前从来没有被结构化过。有那么多人在制定政策、在起草监管规则，这是社会层面上一个巨大的低效。
（Jason）LexisNexis 一直是那个庞然大物、是判例法和法规这块的老牌玩家，他们有巨大的数据护城河，但一年也就赚个二十来亿美元。而如果把你们的收入和 Harvey 的收入加起来，光你们两家可能就已经到那个量级了——你们俩各自都在做几亿美元的收入。所以他们看着后视镜里的你们，一定像《侏罗纪公园》里回头看到暴龙一样，心想「我靠」——他们是不是要来抢我们的生意了？
[43:07] Jason Calacanis
And then here you are on stage saying, hey, we're doing all the manual hard work of getting that information into our, what I assume is a proprietary language model.We'll get to that in a second.Are you going to just try and buy LexisNexis?
而现在你站在台上说：嘿，我们正在干那些把信息弄进我们系统的人工苦活——我假设那背后是一个自研的语言模型，这个我们待会儿再说。你们是打算干脆把 LexisNexis 买下来吗？
[43:21] Max Junestrand
I know it's part of a larger enterprise.Or are you just going to kill it?Well, I think that some of the existing providers and the sort of legacy players have a really hard time pivoting into becoming AI native businesses.And they have a really hard time meeting and catching up to the tempo that we run at.They can't get the talent.They don't work our hours.And they're so political in their organizations that it's just hard to move.And I think at the outset of AI, many believed and made a bet that those organizations who had all the data was going to be the winners.As we're starting to see in the market, that's no longer the case.I think there's a real opportunity for us to partner with content providers.And we're already doing this in many of the smaller jurisdictions, like in Germany, in France, in Spain.The U.S. is peculiar because it's such a duopoly on legal research.Westlaw is the other one?
（Jason）我知道它是一个更大集团的一部分。还是说你们打算直接把它干掉？
（Max）我觉得，现有的一些供应商、那些老牌玩家，要转型成 AI 原生（AI native）的生意非常难。他们也很难跟上、追上我们跑的节奏。他们招不到那样的人才，他们不会像我们这样工作；而且他们组织内部太政治化了，就是很难动。我还认为，在 AI 刚开始的时候，很多人相信并押注：那些手里握着全部数据的组织会是赢家。但正如我们在市场上开始看到的，事实已经不是这样了。我觉得我们和内容提供方合作是有真实机会的，在很多较小的法域我们已经在这么做了，比如德国、法国、西班牙。美国比较特殊，因为法律检索是一个双寡头垄断。／（Jason）另一家是 Westlaw？
[44:24] Max Junestrand
Westlaw and LexisNexis, exactly.But yeah, if you look at how their stock is doing, I think...Oh, are they getting priced in with the AI uncertainty?Yeah, that's one way of putting it.Yeah, they're getting crushed.And I would assume there's some power law here.You know, they might have an incredible breadth of, you know, old case law that they scanned in and went to the courthouses and did all that work on sent to India to be double blind, typed in.Like they literally...You're right.That's what you have to do.Yeah, they literally had two different people type in the cases or OCR them, then check them, look for the differences.I mean, because you can't get it wrong.Nope.But today with the AI tools, the AI tools are really good at doing what they did manually.Yes.You still have to ship the books because you have to physically scan.This is very strange in the U.S., but Westlaw basically has a monopoly with the American government to report on the cases.So they're not owned by the public in a way.They're owned by a company.You guys are very good at capitalism.Sometimes too good.Sometimes too good.I agree.Harvard has a project.There's the court law, court listener.They're trying.They're trying.
（Max）Westlaw 和 LexisNexis，正是。不过你要是看他们的股价表现，我觉得……／（Jason）哦，AI 带来的不确定性已经被计入股价了？／（Max）可以这么说。／（Jason）是的，他们被砸得很惨。
（Jason）我猜这里也有幂律。他们可能有极其广的老判例法库存，都是他们扫描进去的、跑到法院去一家家做出来的，寄到印度去做双盲录入。他们真的是……／（Max）你说得对，那就是你必须做的事。／（Jason）对，他们真的会让两个不同的人分别把案子录入进去，或者做 OCR，然后核对、找差异。因为这个不能出错。／（Max）不能。／（Jason）但今天有了 AI 工具，AI 工具做他们当年手工做的那些事非常在行。／（Max）是的。你仍然得把书运过去，因为你得实体扫描。美国这点很奇怪：Westlaw 基本上和美国政府之间有一个判例报告的垄断权，所以从某种意义上说，这些判例不归公众所有，而归一家公司所有。你们（美国人）搞资本主义真是太在行了。／（Jason）有时候是太在行了。／（Max）有时候确实太在行了，我同意。／（Jason）哈佛有一个项目（判例法开放获取），还有 CourtListener，他们在努力。／（Max）他们在努力。
[45:40] Max Junestrand
It doesn't work.Or rather, put it this way.You cannot build a legal research solution that doesn't have all of the data.Because if you go to Wachtel and a litigator at Wachtel, the best law firm in the world, says, I'm going to use this to go after Elon or do a billion dollar case, you better make sure you have all the cases.So it's the opposite of the power of law.You don't just need the top 80%.You actually need all of it.All of it.Which means you have to go to courthouses and ask them for a copy to print it out and pay them 10 cents a page?
（Max）但行不通。或者换个说法：你没法做出一个不包含全部数据的法律检索方案。因为如果你去 Wachtell（华赫德律所）——全世界最好的律所——那里的一位诉讼律师说「我要用这个去告 Elon、或者打一个十亿美元的案子」，你最好保证你手里有全部判例。所以这跟幂律恰恰相反：你不是只需要头部 80%，你需要的是全部。／（Jason）全部。也就是说你得跑到法院去，请他们给你复印一份，按每页 10 美分付钱？
[46:18] Max Junestrand
Well, there's other ways of getting it.But in practice, yes.You have to physically get the books all the way to India.You need to open them.You need to scan them.Oh, my Lord.You need to get what's called page citations.I never thought in college I would get this nerdy about legal data.But here we are.And what's interesting is that these previous generation of databases were very much search in the database, find the case, and then the lawyer does their work.Right.And what's really interesting about especially the agents following the release of Opus 4.5 and 4.6 is they can now start to do really intelligent case strategy.And they can actually start to combine the witness statements, the cases, and they can really do end-to-end work, which is, I think, moving us from a world where AI is just augmenting to AI is actually really doing things.And your job becomes to orchestrate and to manage those agents, as we're seeing in coding.And so you have partnerships with, I'm assuming, Anthropic and OpenAI, yes?
（Max）还有别的办法，但实际操作中，是的。你得把书实体运到印度，你得把它们一本本翻开，你得扫描。／（Jason）我的天。／（Max）你还得拿到所谓的页码引证（page citation）。／（Jason）我大学时从没想过自己会对法律数据这么较真，但我们就是聊到这儿了。
（Max）有意思的是，上一代这些数据库的用法基本是：在数据库里检索、找到判例，然后律师去干他的活。／（Jason）对。／（Max）而真正有意思的是，尤其在 Opus 4.5 和 4.6 发布之后，这些 agent 现在可以开始做真正有智力含量的案件策略了。它们真的能把证人证词、判例结合起来，做端到端的工作。我认为这把我们从「AI 只是在增强人」推向了「AI 真的在把事情做掉」。而你的工作就变成去编排和管理这些 agent——就像我们在写代码那边看到的一样。／（Jason）那我猜你们和 Anthropic、OpenAI 都有合作，对吧？
[47:33] Max Junestrand
And you spend millions or tens of millions of dollars on tokens.Absolutely.And they are also competing with you on the margins?They are not competing in our product category at all.For now.Well, from the outside, Claude has a legal offering, which is basically a bundling of markdown skills files and a couple of integrations.And so I think what's really helpful about that is that it illustrates to everyone how applicable AI is in law.What it also does is it drives a lot of initial usage there, and then you hit the ceiling.Or, you know, you understand how shallow it is, and then you call us.Right.And so it's actually a big pipeline generator for us, I think.Got it. So they start experimenting.And we were just talking with the CEO of Eleven Labs about, hey, building your own models is, you know, pretty doable these days.And every six months it gets easier and easier.So are you working on your own models using open source to then fork it and make your own models?
（Jason）而且你们在 token 上一年要花几百万甚至几千万美元。／（Max）绝对的。／（Jason）他们是不是也在边缘地带和你们竞争？／（Max）在我们的产品品类里，他们完全没有在竞争。／（Jason）目前是。
（Max）嗯，从外面看，Claude 有一个法律方向的产品，本质上是把一批 markdown 格式的 skill 文件加上几个集成打了个包。所以我觉得这件事真正有帮助的地方在于，它向所有人展示了 AI 在法律里有多适用。它同时还带来了大量初始使用量，然后你会撞到天花板——或者说，你会明白它有多浅，然后你就来找我们了。／（Jason）明白。／（Max）所以我觉得它其实是我们一个很大的销售线索来源。
（Jason）懂了，所以他们先去试。我们刚才还在和 ElevenLabs 的 CEO 聊——现在自己做模型是相当可行的，而且每过六个月就更容易一点。那你们会不会拿开源模型 fork 一份、做自己的模型？
[48:47] Max Junestrand
And is that the future for your firm?So I don't believe in fine tuning or building any general intelligence models.I think that's a total waste of time and money.I do believe in very narrow models for narrow use cases that you also drive a lot of scaling.So you can drive both cost and latency down.An example of this for us is we have a big feature called tabular review, which is basically the number of documents times the number of prompts.So 100 documents, 100 prompts, 10,000 API calls.If you make a fine tune model at extracting contract data, it's very applicable there.But it doesn't make sense to build a general legal intelligence model like some of our competitors are attempting.Yeah.And how do you mitigate against the data leakage issue with your customers?
（Jason）这是你们公司的未来吗？
（Max）我不相信微调（fine tuning），也不相信去建任何通用智能模型。我认为那是彻底浪费时间和金钱。我相信的是为窄用例做非常窄的模型，而且这些用例本身有很大的调用量，所以你能同时把成本和延迟都压下去。举个我们的例子：我们有个很大的功能叫 tabular review（表格式审阅），本质上就是文档数量乘以 prompt 数量——100 份文档、100 条 prompt，就是 10000 次 API 调用。如果你做一个专门抽取合同数据的微调模型，在这里非常适用。但去建一个「通用法律智能模型」——像我们某些竞争对手正在尝试的那样——没有意义。
（Jason）是的。那你们怎么防范客户的数据泄漏问题？
[49:43] Max Junestrand
These, you know, are highly regulated industries with a lot at stake.So putting in, you know, this recent case you're working on in a litigation, if any of that were to seep into a language model and then come out the other end, I mean, this is disastrous.You have a higher level of responsibility.Trust and compliance is our currency.And so it's actually one of the reasons why it's really hard to sell into law.There's a lot of legal AI companies and very few are making it through.And not because it's hard to build stuff.It's actually quite easy to understand where you can build value.But getting it to the customer is very hard.But that's something we cracked pretty early on.And once you're in, it's much easier to expand.So that's also one of the driving forces behind our M&A strategy.But yeah, I mean, we're hosting national secrets, weapons manufacturers with their contracts on Legora.And we work with governments.Does that mean you have to put it on-prem as well?
（Jason）这些都是高度监管的行业，赌注非常大。你把手上正在打的这场诉讼放进去，如果里面任何东西渗进了语言模型、再从另一头冒出来，那就是灾难。你承担的责任层级更高。
（Max）信任与合规就是我们的货币。这其实也是为什么卖进法律行业特别难：市面上有很多法律 AI 公司，但真正能跑通的很少。不是因为做东西难——其实要理解在哪里能创造价值是相当容易的——而是把它送到客户手里非常难。但这一点我们很早就打通了。而且一旦进去了，再扩展就容易得多，这也是我们并购策略背后的驱动力之一。不过说真的，我们身上托管着国家机密，有武器制造商的合同放在 Legora 上，我们也和政府合作。／（Jason）那是不是意味着你们也得做私有化部署（on-prem）？
[50:51] Max Junestrand
No, we don't do on-prem.That's on the roadmap or?No, I mean, you know, deploying in a VPC is very time-consuming.And it creates a lot of dependencies, which slow down your roadmap and the execution forward.All right.Continued success.Max, thanks for taking some time for us.Thank you.
（Max）不，我们不做 on-prem。／（Jason）那是在路线图上，还是……／（Max）不。我是说，部署到 VPC（虚拟私有云）非常耗时，而且会产生大量依赖，拖慢你的路线图和向前推进的执行速度。／（Jason）好的。祝继续成功。Max，谢谢你抽时间给我们。／（Max）谢谢。