ON
集里怎么说它
- 《16 年的一夜成名:ON 如何打造未来最高端运动品牌》(00:38起):本集主角:瑞士运动品牌,靠「云朵鞋底」(cloud tech)创新起步,16 年从零做到接近 40 亿美元品牌,目标是成为最高端而非最大的全球运动品牌。
① 提到它的金句
763 条
电商广告最好的地方在于，我们是互联网上所有新技术的最快采用者。
指向原始笔记的链接
The best thing about ads in e-commerce is that we are the fastest adopters of all new technology on the internet.
—— Rishabh Jain · [12:12]
作为房间里工程方面的声音之一,不得不对业务团队说不,真的非常不舒服——就是那种:好主意,但这需要我六到八个月去重新架构并重构我们在业务定价上做的每一个假设。
指向原始笔记的链接
I found it really uncomfortable to be one of the engineering voices in the room, having to tell no to business teams on like, you know, fantastic idea, but it’s going to take me six to eight months to re-architect and refactor every assumption we’ve made around pricing in the business.
—— Alvaro Morales · [04:48]
太多现有计费方式把定价当作一次性的——设好就不管了——错失了一个事实:你必须持续地演进它。
指向原始笔记的链接
Too many existing approaches to billing just rely on pricing as like a one-and-done, you set it and forget it, and you miss on the fact that you have to kind of keep evolving it.
—— Alvaro Morales · [19:00]
我在整个职业生涯中构建过很多工作流软件，如果你只是给工作流构建器加一个 AI 步骤，那它不是智能体，而且它很可能是在极其做空和看空 LLM，我认为这个房间里没有人是这样想的。
指向原始笔记的链接
I’ve built a lot of workflow software, you know, throughout my career, and if you’re just adding, you know, an AI step to a workflow builder, it’s not an agent, and it’s also probably incredibly short and bearish on LLMs, which I don’t think anyone in this room is.
—— 嘉宾 · [15:43]
如果你在卖给一家对冲基金，他们根据你产品算出的一个数字做了交易，而这个数字是错的，他们就再也不会用你的工具了。
指向原始笔记的链接
If you’re selling to a hedge fund and they make a trade based on a number that comes out of your products and the number’s wrong, they are never going to pick up your tool again.
—— John Willett · [21:52]
所以我们基本上是按照一个代理指标收费的——该事务所拥有的终端客户数量，它或多或少代理了所完成的工作量；这个方式现在行得通，以后也会变，但你绝对不能按席位收费，你要为所做的工作找一个代理指标。
指向原始笔记的链接
So we essentially bill on a proxy of… the amount of end clients that the firm has which more or less proxies the work that is done and that works for now and it’ll sort of change as things go on but you definitely can’t do it by seat and you want to proxy something about the work being done.
—— 嘉宾 · [18:55]
而且除此之外，如果你的产品更好、你在认真做事，你总能获得更多收入，但你的声誉是找不回来的。
指向原始笔记的链接
And on top of that, you can always get more revenue if your product’s better, if you’re doing things, but you can’t get your reputation back.
—— 嘉宾 · [38:27]
但我认为从那个说法到说，比如，好吧，地球上的每一件事物都将在区块链上运行，这就是我的警报开始响起的地方。这就像，好吧，它可能是一项很棒的技术，并且可能非常成功，而不必征服世界
指向原始笔记的链接
But I think going from that to saying like, okay, every single thing on the planet is going to run on the blockchain, that’s where some of my alarms start to go off. And it’s like, well, it can be a great technology and it can be very successful without having to take over the world
—— Sebastian Barrios · [43:44]
我也开始相信，我从未见过一家公司在创立后将设计嫁接上去。
指向原始笔记的链接
I’ve also come to believe that I actually have never seen a company that grafted design on after the founding.
—— Bob Baxley · [14:32]
他有一句话是，「你必须问为什么，为什么个人计算是在 1960 年代末在北加州开始的，当时这个国家的每一家主要科技公司都在东海岸。答案是因为斯坦福大学周围有很小的一群人，他们将软件视为一种与电影、音乐和书籍同等地位的新型媒体。」
指向原始笔记的链接
And he has this quote that is, “you have to ask why it is personal computing got started in Northern California in the late 1960s when at the time every major tech company in the country was on the East Coast. And the answer is because there was a very small group of people in and around Stanford University that saw software as a new form of media on par with movies, music and books.”
—— Bob Baxley · [55:16]
只有错误答案，你只是试图在你能获得的最小错误答案上很好地执行。
指向原始笔记的链接
There’s only wrong answers, and you’re just trying to execute well on the least wrong answer that’s available to you.
—— Hilary Gridley · [41:06]
有时你的产品其实并不重要。我所说的产品是指你放在屏幕上的像素或者你在移动应用程序中构建的东西。
指向原始笔记的链接
sometimes your product actually doesn’t matter. And by product I mean the pixels you put on the screen or things that you build in your mobile app.
—— Peter Deng · [21:33]
如今如此多有价值的科技公司实际上并不是从任何技术突破开始的。它们建立在某种技术突破之上，最终构建了更多的技术。但实际上很多这样的公司，比如 Facebook，只是付出了努力，尤其是早期阶段，在本质上的人类连接数据库之上构建有价值的东西。
指向原始笔记的链接
so many of the tech companies that are most valuable today didn’t really start with any technological breakthrough. They were built on some kind of technological breakthrough and they ended up building a lot more technology. But really a lot of these companies, like Facebook for example, just put in the hard work, the elbow grease, especially in the early stages, to take essentially a database of human connections and build something valuable on top of it.
—— Peter Deng · [23:03]
结果我损失了一千万美元，试图与风投支持的企业竞争，并进行自力更生。那全是我自己的钱，投入了一千万美元，把它点着烧了，试图与 Asana 竞争
指向原始笔记的链接
I ended up losing 10 million, lit it on fire, trying to compete with Asana
—— Andrew Wilkinson · [21:11]
但我只是在早期的职业生涯中犯的最大的错误是，我会因为我在直觉上喜欢某人而雇佣他，而且我会认为我可以改变他们。
指向原始笔记的链接
the biggest mistake I made in my early career was I would hire people because I liked them on gut and I would think that I could change them.
—— Andrew Wilkinson · [38:04]
我认为人们真的忽视了 Claude Code 对非程序员来说有多好用。
指向原始笔记的链接
I think people are truly sleeping on how good Claude Code is for non-coders.
—— Dan Shipper · [07:16]
这个想法是，如果你为一个特定工作雇佣一个 agent 一个月或三个月，如果你决定雇佣那个 agent 并且结果证明那是一台机器而不是一个人，那么它就通过了该角色的 Economic Turing Test。
指向原始笔记的链接
It’s this idea that if you contract an agent for a month or three months on a particular job, if you decide to hire that agent and it turns out to be a machine rather than a person, then it’s passed the Economic Turing Test for that role.
—— Benjamin Mann · [11:24]
如果它第一次不起作用，再问三次，因为当你完全重新开始并再试一次时，我们的成功率比你只尝试一次然后继续在同一个不起作用的事情上敲打要高得多。
指向原始笔记的链接
And if it doesn’t work the first time, asking three more times because our success rate when you just completely start over and try again is much, much higher than if you just try once and then just keep banging on the same thing that didn’t work.
—— Benjamin Mann · [19:13]
即使是现在，我的意思是行业正在爆发，正如我提到的，如今每年 3000 亿美元的资本支出，我会说全世界从事这方面工作的人可能不到 1000 人，这太疯狂了。
指向原始笔记的链接
Even now, I mean the industry is blowing up, as I mentioned, 300 billion a year CapEx today, and I would say maybe less than 1,000 people working on it worldwide, which is just crazy.
—— Benjamin Mann · [26:14]
如果你带来了很多价值，但你开始训练你的客户期望每月支付 20 美元，而你把自己锚定在低价位上，那你就有麻烦了。
指向原始笔记的链接
If you’re bringing a lot of value to the table and you start at training your customers to expect $20 a month and you anchored yourself on a low price point, you’re in trouble.
—— Madhavan Ramanujam · [00:14]
给某事物加个零，在期望方面，或者以为你要设定一个不合理的最后期限是必要的，因为如果你不那样做，几乎有一个潜在的信息即人们不会自己发愤图强。
指向原始笔记的链接
That adding a zero to something in terms of expectations or thinking you’re going to set a deadline that is unreasonable is necessary because if you don’t do that there’s an almost an underlying message that people won’t kick ass on their own.
—— Chip Connolly · [16:59]
当你对衰老的心态从消极转为积极时，你会获得额外七年半的生命，这比现在进行的任何生物黑客疗法带来的生命都要多。
指向原始笔记的链接
When you shift your mindset on aging from a negative to a positive, you get seven and a half years of additional life, which is more life than any other biohack that’s being done right now.
—— Chip Connolly · [58:56]
文化依赖于密度，这就是为什么有时最好的文化感觉像邪教。
指向原始笔记的链接
Cultures thrive on density, and that’s why there’s sometimes the best ones feel like cults.
—— Brian Balfour · [72:48]
大约 18 个月前，24 个月前，我们开始真正看到收益的渐近，因为他们基本上已经摄取了互联网上的所有知识。
指向原始笔记的链接
And about 18 months ago, 24 months ago, we started to really see an asymptoting of gains coming from, because they had essentially sucked up all of the knowledge on the internet.
—— Garrett Lord · [06:34]
有一段时间我们在白板上画了一年中天数，就像，再也没有像这样的时刻了。
指向原始笔记的链接
For a while there we drew the number of days in the year on the whiteboard and it was like, there will never be a time like this.
—— Garrett Lord · [56:25]
船坞是为了唤起受控的混乱。如果你在高峰期去船坞，长滩或其他什么，那里有太多事情发生，大型重型机械在移动，来自中国，来自马来西亚的卡车和集装箱在到处跳。但这种混乱底层是仔细的沟通，高超的技能。
指向原始笔记的链接
A shipyard is to evoke controlled chaos. If you go to a shipyard at peak, Long Beach or whatever it is, there’s so much going on and big heavy machinery moving around, trucks and containers from China, from Malaysia jumping around. But this chaos underline it is careful communication, high skill.
—— Oji · [13:12]
你基本上会收敛到一个意见。所以如果你问，“最好的冰淇淋口味是什么？“它最终会说，“是香草味，而且只有香草味，没有其他口味的冰淇淋。“
指向原始笔记的链接
You essentially converge on one opinion. So if you ask, “What’s the best flavor of ice cream?” It will eventually say, “It’s vanilla and it’s only vanilla, and there’s no other flavor of ice cream.”
—— Ethan Smith · [57:58]
有一个短语，我和我的联合创始人会在我们之间非常早地讨论并且我们与我们合作的很多公司分享了它，那就是说你真正想要的是你想用数据诊断并用设计治疗。
指向原始笔记的链接
There’s one phrase that my co-founder and I would always discuss with us amongst ourselves very early on and which we shared with like a lot of the companies that we work with, which is what you really want is you want to diagnose with data and treat with design.
—— Julie Zhuo · [32:46]
我相信每个方向都有无限。所以这让我对几乎任何人说的话都相当持相反意见。所以如果有人像在 Twitter 上说什么，我有时和自己玩这个游戏，就是在什么语境下那实际上不成立？
指向原始笔记的链接
I believe that there’s infinity in every direction. So that makes me pretty contrary on pretty much everything that anyone says. So if someone says something like on Twitter, I sometimes play this game with myself, which is in what context would that actually not be true?
—— Julie Zhuo · [80:20]
答案是，只写下你看到的第一个错误，也就是最上游的错误。别担心所有的错误，只捕捉你看到的第一个错误的东西，然后停下，继续下一个。
指向原始笔记的链接
And the answer is, just write down the first thing that you see that’s wrong, the most upstream error. Don’t worry about all the errors, just capture the first thing that you see that’s wrong, and stop, and move on.
—— Hamel Husain · [22:13]
我们不想要，“嘿，按一到五的评分给它打分。它有多好？“这在大多数情况下只是一种不做决定的圆滑方式。
指向原始笔记的链接
We don’t want, “Hey, score this on a rating of one to five. How good is it?” That’s just in most cases, that’s a weasel way of not making a decision.
—— Hamel Husain · [52:35]
现在这听起来很吸引人，但它是一个非常危险的指标，因为很多时候，错误，它们只发生在长尾上，并且不经常发生，所以如果你只有 10% 的时间有错误，那么你可以很容易地通过让评判者一直说它通过来达到 90% 的一致性。
指向原始笔记的链接
Now that sounds appealing, but it’s a very dangerous metric to use, because a lot of times, errors, they only happen on the long tail and they don’t happen as frequently, so if you only have the error 10% of the time, then you can easily have 90% agreement by just having a judge say it passes all the time.
—— Hamel Husain · [58:41]
所以这就是为什么愉悦不是在你的实用性之上点缀快乐。它是关于创造一种情感完全处于体验核心的体验。
指向原始笔记的链接
So that’s why like the light is not about sprinkling joy on top of your utility. It’s about creating an experience where emotion is completely on the heart of the experience.
—— Nesrine Changuel · [07:23]
我看到一些表现最好的人只是那些具有很高主观能动性、具有那个时钟速度、有那种精力的人，但他们不一定需要在这个主题上有深厚的经验。
指向原始笔记的链接
I saw some of the highest performers just being people that had very high agency, had that clock speed, had that energy, but they didn’t necessarily need to have deep experience on that matter.
—— Albert Cheng · [00:54]
如果你不能留住用户，那么很多责任就落在让他们在第一天付费上，这超级难。
指向原始笔记的链接
If you don’t retain your users, then a lot of the onus is on getting them to pay on day one, that’s super hard.
—— Albert Cheng · [28:36]
在 Resolve，我们现在正朝着每季度消耗一万亿 token 的方向前进。
指向原始笔记的链接
So, here at Resolve, we are now on track to consume one trillion tokens per quarter.
—— Roshan · [17:34]
你会在每个系统中达到这些边际收益递减的点，感觉就像你可以在这个项目上放 50 个人，但它不会显著地移动指针。
指向原始笔记的链接
You get to these points of just diminishing marginal return in every system where it feels like you could put 50 people on this project, it’s just not going to dramatically move the needle.
—— Robby Stein · [41:00]
我见过的一件事是它太过于凑合了，因为它永远无法获得足够的动力。产品在内部永远不够好，然后它就这样半途而废。
指向原始笔记的链接
something I’ve seen is it’s almost too scrappy because it never gets enough momentum. The product never gets good enough internally and then it just dies on the vine.
—— Robby Stein · [64:24]
在某个时刻，我们在互联网数据上实际上已经有点达到极限了。
指向原始笔记的链接
At some point, we are actually have kind of maxed out on the internet data.
—— Chip Huyen · [14:58]
所有这些大语言模型在晚上和周末都闲置着，而人类不在那里。
指向原始笔记的链接
All these LLMs are sitting idle overnight and on weekends, while humans aren’t there.
—— Dhanji Prasanna · [33:38]
如果你修一条更宽的高速公路，你只会让路上有更多的车。
指向原始笔记的链接
if you build a bigger highway, you’ll just get more cars on the road.
—— Dhanji Prasanna · [50:53]
不要卖那个蘑菇，要卖火力全开的马里奥。
指向原始笔记的链接
Don’t sell the mushroom, sell Mario on blast.
—— Jen Abel · [11:53]
我不会再和某人进行下一次通话，因为在第一次通话中要么是行要么是不行，没有中间地带。
指向原始笔记的链接
I will not get in on another call with someone because on the first call it’s either a yes or a no, there’s no in between.
—— Jen Abel · [66:41]
如果他们坐在那里对你斤斤计较，说明他们并没有完全买账你在卖给他们的东西。
指向原始笔记的链接
if they’re sitting there nickel-and-diming you, they’re not fully bought in on what you’re selling them.
—— Jen Abel · [00:27]
特权、保密性——既有技术层面的安全，也有法律层面的安全。比如，如果我训练了一个模型，我是否就放弃了特权保护？我认为还没有任何人知道答案。
指向原始笔记的链接
Privilege, confidentiality, like both security on the technical side and security on the legal side of I don’t think anyone knows the answer yet of if I train a model, do I waive privilege?
—— Gabe · [25:20]
现在有数以万计的公司在使用 Rippling 运营。我们还不到市场的 1%。
指向原始笔记的链接
There are tens of thousands of companies that now run on Rippling. We’re less than 1% of the market.
—— Matt MacInnis · [77:23]
我们只是把 Agentforce 用在这些人身上，并且用非常相似的提示词训练了它。它有 70% 的回复率。
指向原始笔记的链接
We just took Agentforce just on those and we trained it on very similar prompt. It had 70% response rate.
—— Jason Lemkin · [38:41]
但每次你将决策能力或自主权移交给智能体系统时，你都在某种程度上放弃了一些你的控制权。
指向原始笔记的链接
But every time you hand over decision-making capabilities or autonomy to agentic systems, you’re kind of relinquishing some amount of control on your end.
—— Kiriti Badam · [10:12]
现在，我不想夸大这一点，但换个说法，这就像如果你发送一封邮件，邮件上的第一个人是助理，第四个人是 CEO，你可能做错了。
指向原始笔记的链接
Now, I don’t want to overstate this, but put differently, it’s like if you’re sending an email and the first person on the email is the assistant and the fourth person on the email is the CEO, you’ve probably done it wrong.
—— Sam Lessin · [59:18]
A/B 测试效果不是很好，而且它对大多数事情都不起作用。
指向原始笔记的链接
A/B testing doesn’t work very well and it doesn’t work on most things.
—— Jason Cohen · [94:43]
最大的重点是定价不仅仅是页面上的那个数字。它是定位，它是他们的预算如何运作，它是它是如何构建的。
指向原始笔记的链接
the largest point is pricing is not just the number on the page. It’s positioning, it’s how their budgets work, it’s how it’s structured.
—— Jason Cohen · [46:08]
这样做的原因是，我认为没有任何理由假设人类技能水平是任何事物的上限。
指向原始笔记的链接
And the reason for that is, I don’t think there’s any reason to assume that human skill level is the cap on anything.
—— Marc Andreessen · [84:17]
AI 的上限不是模型智能，而是模型在行动之前看到的东西。
指向原始笔记的链接
The ceiling on the AI isn’t the model intelligence, it’s what the model sees before it acts.
—— Lazar Jovanovic · [44:21]
最终结果是你拥有一个庞大的劳动力，他们并不真正理解这项技术，就像，“我知道我应该使用这个，也许它也在我的绩效评估中，但我不知道该做什么。“
指向原始笔记的链接
And as an end result, you end up with a giant workforce that doesn’t really understand the technology, is like, “I know I’m supposed to use this and maybe it’s like on my performance review too, but I’m not sure what to do.”
—— Sherwin Wu · [41:19]
我们从 Salesforce、谷歌和微软雇佣了很多人，那些人的流失率是 100%。
指向原始笔记的链接
We hired so many people from Salesforce, and Google, and Microsoft, like 100% attrition rate on all those folks.
—— Brian Halligan · [13:02]
我们押注于为六个月后的模型构建。而不是为今天的模型。
指向原始笔记的链接
We bet on building for the model six months from now. Not for the model of today.
—— Boris Cherny · [65:56]
通常，如果你使用最强大的模型，实际上反而更便宜，且消耗更少 token，因为它可以只用更少的纠错、更少的指导等就把同样的事情做得更快。
指向原始笔记的链接
Often, it’s actually cheaper and less token-intensive if you use the most capable model, because it can just do the same thing much faster with less correction, less hand holding, and so on.
—— Boris Cherny · [69:39]
我认为过去几年里我听到的最有趣的洞见之一是 Jensen 说他不进行一对一谈话。那对我来说是一个巨大的解锁。
指向原始笔记的链接
I think one of the most interesting insights I heard over the past few years was Jensen saying he doesn’t do one-on-ones. And that was a big unlock for me.
—— Ankur Goyal · [22:04]
我认为人类的生存取决于成功的 AI。
指向原始笔记的链接
I think that survival of humanity depends on a successful AI.
—— Jeetu Patel · [07:52]
毅力战胜智力，耐力战胜智力，一周中的任何一天，周日加倍。
指向原始笔记的链接
persistence beats intellect, and stamina beats intellect any day of the week, twice on Sunday.
—— Jeetu Patel · [32:49]
但当我想到我最尊敬的领导者和经理时,我实际上认为他们最好的特质之一是他们选择自己承担的低杠杆任务,而这最终实际上是一件非常高杠杆的事情,因为是他们在做。
指向原始笔记的链接
But when I think about leaders and managers that I respect the most, I actually think some of their best traits is that they choose low leverage tasks that they take on themselves, and that actually ends up being actually a very high leverage thing, because it’s them who’s doing it.
—— Jenny Wen · [55:41]
你为公共消费撰写东西的每一分钟，都没有把你有非常有限的时间集中在你的客户和产品上。
指向原始笔记的链接
Every minute you’re writing something for public consumption, you’re not focusing your very limited time that you have on your customers and your product.
—— Qasar Younis · [42:11]
人们完全误解了高管是如何做决策的,在他们的头脑中、日程表中、激励结构中发生了什么。
指向原始笔记的链接
People completely misunderstand how executives make decisions, what is going on in the heads, in the calendars, in the incentive structures of executives.
—— Jessica Fain · [06:51]
如果一个项目需要两周的工程时间或更少,那么工程师就要有效地充当该项目的 PM。
指向原始笔记的链接
If a project is two weeks of engineering time or less, then the engineer is on the hook to effectively be the PM for that.
—— Amol Avasare · [45:37]
总的来说，这个职能已经变得极其专注于没有权力的责任。
指向原始笔记的链接
Generally, the function had become extremely focused on responsibility without authority.
—— Nikhyl Singhal · [04:47]
开源本身有一个过滤器。它是一致性和坚持胜过强度
指向原始笔记的链接
There’s a filter on open source in general. It’s consistency and commitment wins over sort of beats intensity
—— 嘉宾 · [14:27]
我认为技术领导人们认为人们只会盲目地采用发布的新技术。我认为我们将进入一个时期，对于即将随着 AI 到来的许多变化，社会上将会有巨大的阻力。
指向原始笔记的链接
I think technology leaders think that folks will just blindly adopt new technology as it comes out. And I think we’re going to enter a period of time where there’s going to be a huge amount of societal pushback on a lot of the changes that are coming with AI.
—— Evan Spiegel · [64:47]
能力就是一切。没有人真正愿意在能力上妥协。你会为了给客户好 10% 的体验而支付 500% 的溢价。
指向原始笔记的链接
Capability is everything. No one’s really willing to take a hit on capability. You will pay a 500% premium for a 10% better experience for your customers.
—— Julian · [00:03]
很难让我们的沙箱发生 OOM 或内存溢出。因为我们可以动态地实时调整大小,这在几乎任何其他东西上都是不可能的。
指向原始笔记的链接
it’s very hard to OOM or out of memory our sandboxes. because we can dynamically on the fly resize, which is like impossible on almost any other thing.
—— Ivan Burazin · [28:56]
它们是做已经存在的工作的绝佳机器，但不一定是做处于可能性边缘的事情，处于文化边缘的事情。
指向原始笔记的链接
they’re great machines for doing what already existingly works, but not necessarily what’s on the edge of what’s possible, what’s on the edge of culture.
—— Amjad Masad · [15:53]
我同时极其 AI 信奉（pilled）且非常看好人类。
指向原始笔记的链接
I’m simultaneously extremely AI pilled and very bullish on humans.
—— Dan Shipper · [00:24]
自动化是个谎言，从某种意义上说，每次你自动化某样东西时，为了确保自动化运行良好，你需要一个人在上面确保它运行良好。
指向原始笔记的链接
Automation is a lie, in the sense that every time you automate something, in order to make sure the automation is working well you need a human on top of it making sure that it’s working well.
—— Dan Shipper · [39:15]
超过 90% 的著名前沿模型是在 2025 年生产的，其中几个现在在许多方面达到或超过了人类基准
指向原始笔记的链接
over 90% of notable frontier models were produced in in 2025, and several of those now meet or exceed human baselines on a number of things
—— Daniel Whitenack · [04:06]
你正在一个非常脆弱的基础上构建。
指向原始笔记的链接
You’re building on a really crusty foundation.
—— Tony Fadell · [53:37]
如果事情发展得如此之快，我只是认为人类很难，如果你的基础设施中正在发生某事，并且利用时间表从几个月到几小时到几分钟到几秒。你不能只是依靠在半夜叫醒 CISO 来批准关闭这个东西。
指向原始笔记的链接
If things are moving so fast, I just think it’s gonna be hard for humans to, you know, if if something is happening in your infrastructure and exploit timelines go from, you know, months to to hours to minutes to seconds. You can’t just, like, rely on waking up the CISO in the middle of the night to approve, you know, shutting this thing down.
—— Daniel Whitenack · [44:41]
游戏发布 10 分钟后,Zynga 的产品经理们就群发邮件分享了他们的分析,他们说这游戏一出世就死了,因为他的 FTUE,他的首次用户体验,点击次数太多且太糟糕,以至于没人能看到他出色的游戏设计。
指向原始笔记的链接
And 10 minutes after the game came out, the Zynga PMs emailed around their analysis and they said it’s dead on arrival because his FTUE, his first time user experience was so many clicks and so bad that no one was ever going to see his great game design.
—— Mark Pincus · [05:40]
今天我们都挂在我们的 Claude 上,挂在我们的 GPT 上,但没有鸡尾酒会。
指向原始笔记的链接
Today we’re all hanging out on our Claude on our GPT, but there’s no cocktail party.
—— Mark Pincus · [56:32]
我想每个人都想突显，这是我试着推动每个人去做的，像，你在世界上排名第一的是什么？像，非常像，给我一个非常简单的答案。如果你没有一个，你可能做得不够多。
指向原始笔记的链接
I want everyone to spike, and this is something I I try to push everyone to, like, what are you number one in the world at? Like, very like, give me a very simple answer. If you don’t have one, you probably haven’t worked on it enough.
—— Swyx · [22:32]
但如果不致力于一个单一的想法，在创业公司上取得有意义的进展是极其困难的。
指向原始笔记的链接
But it’s extremely hard to make meaningful progress on a startup without committing to a single idea.
—— John · [00:25]
它是不做决定、空转、在想法之间涉猎，并且从未在任何一个上深入到足以学到任何东西。
指向原始笔记的链接
It’s not making a decision, spinning your wheels, dabbling between ideas and never going deep enough on any one of them to learn anything.
—— John · [11:08]
作为 CTO,我不想登上头版,就像我安装了某个奇怪的 NPM 包并泄露了所有的代码。
指向原始笔记的链接
as the CTO, I don’t want to end up on the front page as like I installed some weird NPM package and leaked all the code.
—— Matei Zaharia · [21:47]
让数据到位,然后在上面附加一些智能体。魔法就会出来。但没有正确的数据,你无法真正做到这一点。
指向原始笔记的链接
just get the data to be there and then slap some agent on top. Magic will come out. But without the right data, you can’t really do that.
—— Reynold Xin · [66:37]
最好的着色器将是你自己梦想出来的那些。
指向原始笔记的链接
The best shaders are going to be the ones that you dream up on your own.
—— 嘉宾 · [45:39]
如果你真的写了文档或做了 Figma,你会受到警告,因为我们明确不希望你们做那个,对
指向原始笔记的链接
if you actually produce a doc or figma like you will get a slap on the wrist because we explicitly don’t want you to do that right
—— Eddie Kim · [42:15]
我不会忽视反叛联盟，因为我认为有令人信服的哲学理由去为它们的存在而战。
指向原始笔记的链接
I would not sleep on the rebel alliance because I think that there’s a really compelling philosophical reason to fight for those to be in existence.
—— Chamath · [40:14]
在未来,大多数公司将建立在驾驭机制之上。
指向原始笔记的链接
In the future, most companies will be built on harnesses.
—— Jensen Huang · [13:10]
我看到智能体从那个像和你一起做结对编程的助手一样，最终爬升为你那个三人编程团队的管理者，最终成为 CDO，最终成为其他职位，你的工作就是在低层级仅仅是去解除阻塞。
指向原始笔记的链接
I see the agents crawling from like the helper that’s doing peer programming with you to eventually the manager of your three-person programming team to eventually the CDO to eventually to other positions where your job is to be on the lower levels just unblocking.
—— Kitsa · [38:06]
当你学习另一种语言时,我认为最重要的事情之一,也是关于你会不会擅长口语的最好预测指标之一,这是我的猜测,我没有任何研究支持,就是你是否愿意听起来像个白痴,你是否愿意直接开口说,被纠正也不觉得被冒犯,然后变得越来越好。
指向原始笔记的链接
when you learn another language i think one of the most important things one of the best predictors this is my guess i don’t have any research on this about you know are you gonna get good at speaking is are you willing to sound like an idiot are you willing just to say it and be corrected and not be offended and then just get better and better
—— Adam Mosseri · [15:27]
每一个这些产品都有远超 10000 个提示词，所以不要在第一个提示词后就放弃
指向原始笔记的链接
each of these products have way over 10 000 prompts each so do not give up on the first prompt
—— Meng To · [73:17]
我看到的创始人犯的最大错误之一是依赖汇总用户指标，而不是理解单个用户如何使用他们的产品。
指向原始笔记的链接
One of the biggest mistakes I see founders make is relying on aggregate user metrics instead of understanding how any individual users use their product.
—— 嘉宾 · [00:09]
你想一下戴着安全帽的 AI，那就是我们。
指向原始笔记的链接
You think about AI with a hard hat on, that’s us.
—— Kriti Sharma · [01:44]
我们深入一线，我们在机库里度过周末，我们在炼油厂和工厂里度过白天，我们与现场技术人员一起旅行，以便我们构建那些不允许失败的东西。
指向原始笔记的链接
We go on the ground, we spend weekends in hangars, we spend our days in refineries and factories, and we travel along with the field technicians so that we build things where failure is not an option.
—— Kriti Sharma · [01:48]
现在这些东西，如果你看它，我怎么理解它，除非我是一个有 20 年工作经验的在岗工程师？
指向原始笔记的链接
Now these things, if you look at it, how do I even make sense of it unless I’m an engineer on the job with 20 years of experience?
—— Kriti Sharma · [09:46]
我们就像处于工业革命边缘的农民，我们只是没有看到一条通往正在发生的事情的清晰路径。
指向原始笔记的链接
We’re like farmers on the cusp of the industrial revolution and we just don’t see a clear path to what’s happening.
—— Noam Segal · [16:47]
“因 AI 而失去工作”实际上在榜单上倒数第二。
指向原始笔记的链接
Losing my job to AI is actually second to last on that list.
—— Noam Segal · [53:45]
智能体不过就是我从这个座位上跳出去，你看到一把空椅子，但智能体在代表我做工作。
指向原始笔记的链接
The agent’s nothing more than me hopping out of this seat and you see an empty chair, but the agent’s doing work on my behalf.
—— Bradon Rogers · [00:29]
假设我们的集体任务是让一只猴子站在十英尺高的基座顶上背诵莎士比亚。在这种有些极端的情况下,我们应该先做哪件事?搭基座还是训练猴子?
指向原始笔记的链接
that our collective project was to get a monkey to stand on the top of a 10-foot pedestal and recite Shakespeare. Which should we do first in this somewhat extreme situation? Build the pedestal or train the monkey?
—— Astro Teller · [00:48]
但好消息是你没有把孩子的大学基金押在传送机上。所以如果传送机最后证明不是我们该做的事,我们实际上会为叫停它而庆祝你。
指向原始笔记的链接
But the good news is you haven’t bet your kid’s college fund on the teleporter. So if the teleporter turns out not to be the right thing for us to do, and we’re actually going to celebrate you for stopping it.
—— Astro Teller · [08:55]
但我们在购买未来的期权,所以信息增益,每一美元换来的学习,才是我们真正想最大化的。
指向原始笔记的链接
But we’re buying options on the future and so information gain, the learning per dollar, is what we’re actually trying to maximize.
—— Astro Teller · [17:15]
如果回报风险比还可以,而要搞清楚我们是否在正轨上的第一条信息要花一百万美元,不值。回报巨大,风险挺大,但我们可以花十万美元搞清楚这是不是比我们以为的没那么疯狂一点。行,这个游戏我愿意玩一整天。
指向原始笔记的链接
If the reward risk ratio is okay, and it’s going to take us a million dollars to find out the first piece of information about whether we’re on the right track, not worth it. Reward is ginormous, risk is pretty big but we can find out for $100,000 whether this is a little bit less crazy than we thought. Yeah, I’ll play that game all day long.
—— Astro Teller · [25:22]
X 里的每个人都会用速记彼此说,猴子是什么?他们的意思不只是,你有没有一个可检验假设,而是你现在做的事情如何试图消解问题中风险最大那部分的风险?
指向原始笔记的链接
Everyone at X will just say as shorthand to each other, what’s the monkey? What they mean by that is not just, do you have a testable hypothesis but how is the thing that you’re doing now attempting to burn down the risk on the riskiest part of the problem?
—— Astro Teller · [27:31]
我怀疑按实际流量计算，Google 今天比 OpenAI、Anthropic 或任何公司都大。
指向原始笔记的链接
I suspect on an actual traffic basis, Google is bigger than OpenAI, Anthropic, anyone today.
—— David George · [09:00]
从根本上说，人类是按结果获得报酬的，很多 AI 会增强人类，但可能也会取代一部分人类，而这将涉及按结果付费。
指向原始笔记的链接
Humans were fundamentally paid based on outcomes, and a lot of AI will be augmenting humans, but probably also replacing some humans, and that will involve being paid for outcomes.
—— David George · [27:31]
就像产品内的 AI，大概比在组件上进行 AI 编码领先两年。他们已经有了评估系统，已经有了可观测性系统。
指向原始笔记的链接
Like AI product, AI inside the product is probably two years ahead from AI coding on components. They already have the eval system, they already have an observability system.
—— Patrick Debois · [17:13]
我认为不管是现在的 AI 还是敏捷或任何变化，听起来很难，但在开始时不要在消极的事情上花时间。找到成功的故事。如果你不能让它成功，其余的都不会起作用。
指向原始笔记的链接
I think regardless whether this is now AI or like agile or any change, um, as hard as it sounds, do not spend time on the negatives in the beginning. Find the success story. If you can’t make that work, the rest will not work.
—— Patrick Debois · [33:58]
但下一层是关于，好，如果 token 并不是真正可互换的，你需要给它们分配不同的工作，比如这个 token 负责建议、那个负责执行，这个在「做梦」、那个在执行，诸如此类，你会想开始组合这些协同配合的编排式策略
指向原始笔记的链接
But the next one is about, okay, if tokens aren’t really fungible and you need to give them different jobs, like maybe this token is advising versus this token is executing, this token is dreaming versus this token is executing, so on and so forth, you want to start composing these kind of orchestrated strategies that go together
—— 嘉宾 · [11:34]
我们实际上并不执着于说，你应该在我们的基础设施上运行这些东西。
指向原始笔记的链接
We actually aren’t precious about, you should run these things on our infrastructure.
—— 嘉宾 · [16:11]
所以，把时间花在客户身上，是我时间上回报率最高的事情。
指向原始笔记的链接
So there’s no higher ROI on my time than spending time with customers.
—— Sam Blond · [00:11]
把一个智能体叠加在一组数据孤岛化的任意工具之上，远比不上叠加在一个单一平台和唯一事实来源上——后者既拥有你所有的数据，也在同一个工具内执行你所有的操作。
指向原始笔记的链接
It is far more difficult to overlay an agent on top of this arbitrary set of tools with data silos than it is a single platform and source of truth that both has all of your data, but also takes all of your actions inside of the same tool.
—— Sam Blond · [37:47]
因为如果你不在如何上手、获得价值、并最终购买你的产品上教育买家，客户就不知道该怎么购买你的产品。
指向原始笔记的链接
Because if you aren’t educating the buyer on how to onboard, receive value, and ultimately buy your product, the customer doesn’t know how to buy your product.
—— Sam Blond · [63:21]
当客户尝试为你设计软件时经常发生的情况是，他们想要一个补丁。他们只想在某样突起的东西上贴一些管道胶带，但你会想，不，实际上这东西突起来的原因是铆钉没有合适的公差。
指向原始笔记的链接
What happens a lot when customers try to design your software for you is they want a patch. They want to just put some duct tape on something that’s sticking up, but you go like, no, actually the reason this thing is sticking up is because the rivets are not the right tolerance.
—— 嘉宾 · [05:27]
路线图是共识错觉的温床，你认为因为某条路线图上有一个要点，听起来含糊地解决了你对产品的渴望，它实际上就会满足那个渴望。
指向原始笔记的链接
Roadmaps are breeding ground for illusions of agreement, that you think because there’s a bullet point on some roadmap that vaguely sounds like it’s addressing a desire you have for the product, that it’s actually going to fulfill that desire.
—— David · [09:57]
购买产品的好方法是按它们今天存在的样子购买。然后之后来的任何东西都是额外的。你不应该基于”好吧，我觉得我真的需要的这个东西将在第四季度推出”来做购买决定。
指向原始笔记的链接
The good way of buying products is to buy as they exist today. And then anything that comes after that is gravy. You should not be basing your purchase decisions on, “Well, this thing I kind of feel like I really need is coming in Q4.”
—— 嘉宾 · [11:03]
你承诺在未来做这件事，因为你不想现在做这件事。实际上这就是破绽。如果这真的这么重要，你现在就应该做。为什么要承诺在年底前做？
指向原始笔记的链接
You’re promising this in the future because you don’t want to work on it now. And there is actually the tell. If this was truly so important, you just work on it now. Why are you promising it by the end of the year?
—— David · [14:34]
软件工程变得越来越难,因为我们可以承担的项目的野心水平已经提高了。
指向原始笔记的链接
Software engineering is getting harder because the level of ambition of the stuff we can take on has gone up.
—— Simon Willison · [03:21]
这里的独特优势在于，我们知道未来的模型可能长什么样，因此能够走捷径，省去很多你在芯片侧需要做的决策、设计决策。
指向原始笔记的链接
There’s a unique advantage here of us knowing what the future models might look like and therefore being able to short circuit a lot of the decisions you need to make, design decisions you need to make on the chip side.
—— Sachin Katti · [35:26]
这是一个万亿美元的行业,他们每年在资本开支上花 2000 亿美元,而他们不赚钱。
指向原始笔记的链接
And it’s a trillion-dollar industry, and they spend $200 billion a year on CapEx, and they don’t make any money.
—— Benedict Evans · [05:06]
决定你的智能体是天才还是金鱼的问题是，谁决定那三本书在那张桌子上打开。
指向原始笔记的链接
The question that determines whether your agents are geniuses or goldfish is who decides which three books are open on that desk.
—— Garry Tan · [12:39]
你完全可以构建一个更强大的模型，但它在网络安全方面并不更危险——只要你不用网络安全数据去训练它，而这一点人们真的没有怎么谈论。
指向原始笔记的链接
You could build a more powerful model that is not more dangerous for cybersecurity, for example, if you don’t train it on cybersecurity, which really people are not really talking about.
—— Clement DeLong · [07:53]
我们现在看到的是，很多人、很多公司意识到，完全依赖一个模型太危险、风险太大，完全说不通。
指向原始笔记的链接
I think what we’re seeing right now is that a lot of people, companies are realizing that it’s too dangerous, it’s too risky, it doesn’t make any sense to rely exclusively on one model.
—— Clement DeLong · [17:10]
就像大部分收入获取原本集中在前沿模型上，将转向一个更像长尾的模型群。
指向原始笔记的链接
Like majority of the revenue capture was on frontier models to a more like long tail of models
—— Clement DeLong · [19:47]
我会挑战一句，尤其在早期：如果某人不愿意每周给你三十分钟，他们到底有多在意你正在解决的问题？
指向原始笔记的链接
And I would challenge, especially early on, if someone is not willing to give you thirty minutes a week. How much do they really value the problem you’re solving?
—— Damien Lewke · [37:17]
我想达到一个收入是毫无悬念、不用想的地步。因为最终，一旦你在创业公司里踏上了收入的列车，你就得开始规模化。
指向原始笔记的链接
I wanted to get to a point where the revenue was a slam dunk and no brainer. Because ultimately, once you get on the revenue train in a startup, you got to start scaling.
—— Damien Lewke · [38:16]
一个没有智能垄断、而是一座可以随手采撷的「智能花园」的世界,是非常好的世界。
指向原始笔记的链接
And it is a very good world where there is not like a monopoly on intelligence, but instead kind of a garden of intelligence that you can pick and choose, you know, when you’d like.
—— Matan Grinberg · [28:46]
大部分工作不是编码。它主要是协调，然后说服人们事情可能是怎样的。
指向原始笔记的链接
Most of the work is not coding. It’s largely coordination and then convincing people on how things could be.
—— Heitor Lessa · [08:36]
我肯定可以在后端测试东西。我肯定可以为前端写测试。但我真正想知道的是，它可用吗？
指向原始笔记的链接
I can definitely test things on the back end. I can definitely write tests for the front end. But what I really want to know is, is it usable?
—— 嘉宾 · [04:56]
所以有时我只是在 ChatGPT 浏览器使用之上的人工验证层。
指向原始笔记的链接
And so sometimes I’m just the human verification layer on top of ChatGPT browser use.
—— 嘉宾 · [23:27]
现在我一直很喜欢思考的一个数据是，网络上智能体的流量比人类流量还要多。
指向原始笔记的链接
You know, one stat that I always like to think about nowadays is just like, there’s more agent traffic on the web than human traffic, you know?
—— Andy Fang · [04:51]
很多人认为你应该从一个匮乏的状态出发去构建，或者说你心里憋着一股劲、需要证明什么。你确实可以用那种方式建造出伟大的东西，但我觉得你会在过程中造成很多伤害。
指向原始笔记的链接
A lot of people think you should build from a place of lack. Or like, you know, you have a chip on your shoulder, you need to prove something. And you can definitely build great things that way, but I think you create a lot of damage along the way.
—— Kareem Amin · [03:55]
所以我有点意识到，在终点，成功之后什么都没有。我所追逐的那一边什么都没有，而它其实就在当下。
指向原始笔记的链接
So I kind of realized that at the end, there’s nothing at the end of success. There’s nothing on that other side that I’m chasing and that it actually is just right now.
—— Kareem Amin · [05:15]
现在举证责任在怀疑者身上,但一旦你有了令人失望的信息的缓慢渗透,那么举证责任就开始转移到乐观者身上了。
指向原始笔记的链接
right now the burden of proof is on the skeptics, but once you have the slow trickle of disappointing information, then the burden starts to be on the optimists.
—— Alex · [09:15]
就对我来说,2025 年 9 月很疯狂,你在那里因为说你要在一项技术(其经济性实际上尚未解决)上花更多钱而受到市场奖励。
指向原始笔记的链接
Like it was crazy to me, September 2025, where you were rewarded by the market for saying you’re going to spend a lot more money on a technology that has not had the economics of it actually worked out yet.
—— Ranjan Roy · [13:33]
为了极其明确一点,AI 数据中心算力收入的绝大部分取决于两家不盈利、不可持续的 AI 公司是否有能力继续每年筹集数百亿或数千亿美元。
指向原始笔记的链接
to be abundantly clear, the vast majority of AI data center compute revenue is contingent on the continued ability of two unprofitable, unsustainable AI companies to raise tens or hundreds of billions of dollars a year.
—— Alex · [35:38]
它们都是估值极高的公司,都在押注一个更大的关于机器人经济的承诺,也就是特斯拉和 Optimus 机器人无处不在,或者是太空中的数据中心。
指向原始笔记的链接
They’re both incredibly overvalued companies that are betting on a much larger promise of a robotic economy that Tesla and Optimus robots everywhere or data centers in space.
—— Ranjan Roy · [62:23]
但随后它好像又说到”更聪明”这个词不对，你知道，我是一个非常快、非常广、非常浅的思考者，没有连续性，我想这很有趣，因为我原以为你们都在致力于记忆，然后显然人类是更慢、更窄、但深得多的思考者，其判断建立在我实际经历过的多年的后果之上
指向原始笔记的链接
but then it like went into like the smarter isn’t the right word and you know i’m very fast very broad very shallow thinker with no continuity i was like that’s interesting because i thought you all were working on memory and then apparently humans are slower narrower much deeper thinkers with judgments built from years of consequences i’ve actually lived through
—— 嘉宾 · [07:59]
我把 90% 的时间花在排版上，只有 10% 的时间花在内容上。
指向原始笔记的链接
was I spend 90% of my time on formatting and 10% of my time on content.
—— Jon Noronha · [13:16]
我不认为开源模型的进步会蚕食我们的核心业务，因为数据在模型性能的最前沿才是最有价值的。
指向原始笔记的链接
I wouldn’t say that open source model improvements cannibalize our core business because data is most valuable on the frontier of model performance.
—— Osvald Nitski · [04:21]
人们想要越来越用力地扳动那个杠杆，因为在 Mercor 上的花费直接转化为我们客户更多的收入。
指向原始笔记的链接
People want to crank it harder and harder because spend on Mercor directly translates to more revenue for our customers.
—— Osvald Nitski · [41:22]
在 anthropic 之上构建已经相当危险了,因为每次一个应用类别开始做得很好时,他们就会进入那个类别。
指向原始笔记的链接
it’s been already fairly dangerous to build on anthropic because every time an application category starts to do well, they move into that category.
—— Ben Horowitz · [14:35]
我们团队中实际上有一种说法,评估就是新的 PRD。
指向原始笔记的链接
We actually have a saying on the team of evals are the new PRDs.
—— Dianne Penn · [41:41]
我们谁都不知道在这件事上任何时候完美的界限在哪里。你更愿意在哪一边犯错？
指向原始笔记的链接
none of us know what the perfect line is at any time on this. Which way would you rather be wrong on?
—— 嘉宾 · [15:15]
至少我个人的决定是，我宁愿错在 token max 这一边，而不是不做的这边——反正你永远不会是对的。
指向原始笔记的链接
at least my personal decision is I’d rather be wrong on the token maxing side than the non right, and you’re never going to be right.
—— 嘉宾 · [15:44]
只有当你看到它反复在同样的事情上跌倒时,那才是你把它加回去的时候。
指向原始笔记的链接
And only when you see it repeatedly stumble on the same thing, that’s when you add it back.
—— Boris Cherny · [08:14]
我的看法是，他们大致是说，看，如果模型就这么把某件事做了，而你的应用——你在它之上构建的工作流、价值——不够有独特性，模型夺走了那个市场，那么它很可能本来就没那么有防御性。
指向原始笔记的链接
My view is they’re kind of like, look, if the model just kind of does something and your application isn’t distinctive enough, the workflow, the value you’ve built on top of it, and the model takes that market away, well, then it probably wasn’t that, you know, defensible anyway.
—— Matt Murphy · [30:39]
当有一天我死在工作岗位上时,你知道,我告诉他们,他们只需要为下一位 CEO 重新塑造公司。
指向原始笔记的链接
When I die on the job someday, you know, I told them they’ll just have to reshape the company for the next CEO.
—— Jensen Huang · [18:00]
因为我可以使用本地智能体一天产出 20 个 PR,但如果政策说每个 PR 都需要人工审查，那我的团队就得去审 20 个 PR。
指向原始笔记的链接
Because I can churn out 20 PRs in a day using a local agent, but if we have a policy where every PR needs a human review on it, well then my team needs to go review 20 PRs.
—— Rob Willoughby · [11:06]
知识是可变的，热情不是，而且从来没有比现在更有理由去做你热爱的事情了。
指向原始笔记的链接
Knowledge is changeable, passion is not, and never has there been more reason to work on what you love.
—— Blake Scholl · [33:03]
只有当你看到它反复在同一件事上绊倒时，你才应该把它加回去。
指向原始笔记的链接
And only when you see it repeatedly stumble on the same thing, that’s when you add it back.
—— Boris Cherny · [08:36]
我认为消费领域有很多机会实际上是未开发的,就像生成和分享内容的社会层面几乎仍然完全是在现有的巨头平台上完成的
指向原始笔记的链接
I think there’s a lot of opportunity in consumer that’s honestly untapped, like sort of the social layer of generating and sharing stuff is still almost entirely done on the existing incumbent platforms
—— Justine Moore · [36:51]
说实话,我们所有人,包括我们自己,都只是在表面撒了一点 AI。
指向原始笔记的链接
All of us, including ourselves, honestly, were just sprinkling some AI on top.
—— Carlos González de Villaumbrosia · [03:54]
这里有足够多的事情在发生,我们真的想尽可能简化我们的业务,聚焦于我们认为最大的机会在哪里,那就是数据中心。
指向原始笔记的链接
There’s enough here going on that we really want to simplify our business as much as we possibly can, focus on where we think the biggest opportunities are, and that’s the data center.
—— Michael · [12:32]
如果 Transformer 是有利可图的,如果你可以花费更多的努力和资源来扩大 Transformer 以在下一个季度获胜,那么很难把大量的注意力和精力投入到可能在一两年内更好或重新定义该领域的事情上。
指向原始笔记的链接
If Transformer is profitable and if you can spend more efforts and more resources scaling Transformer to win in the next quarter, it’s very hard to put at least a lot of attention and a lot of energy to work on something that will maybe better or maybe will redefine the field in a year or two.
—— Jerry Tworek · [13:28]
我们可能比以往任何时候都在从经验学习上花费最多的算力,但强化学习并不是从经验学习的终点。
指向原始笔记的链接
We probably are spending the most compute than ever on learning from experience, but reinforcement learning is not the end of learning from experience.
—— Jerry Tworek · [26:22]
但这需要高水平品味的人类,就像世界上大概只有三个人,并且在四周内花费大约十万美元在这些编码智能体上,才能得到一个速度快 60 倍的解决方案。
指向原始笔记的链接
But it requires the high-taste human, like there’s maybe three people in the world, and spend about $100,000 on these coding agents over a span of four weeks to get to a solution that’s 60x faster.
—— 嘉宾 · [41:10]
正如我提到的,我们相信 Transformer 无法进行持续学习。没有办法在 Transformer 上进行持续学习。
指向原始笔记的链接
We believe, as I mentioned, that transformers are incapable of continual learning. There’s no way how to put continual learning on transformers.
—— Jerry Tworek · [45:21]
结果表明,工作的量级比工作所基于的信息存储要大得多。
指向原始笔记的链接
It just turns out that the work is orders of magnitude bigger than the storage of information that the work is done on.
—— Alex Rampell · [09:39]
这些模型在如此多的数据上训练,它们如此巨大,然而它们实际上真的不知道如何做其中的任何一项工作。
指向原始笔记的链接
The models are trained on so much data and they’re so large and yet they actually don’t really know how to do any of this work.
—— Frederick Rankin · [46:31]
所以你会看到,很多这类小企业,我认为统计数据是 70% 仍然通过纸质获得报酬。
指向原始笔记的链接
So you will see that like a lot of these small businesses, I think the stats are 70% are still paid on paper.
—— Stein Pella · [53:34]
我们以前测量过这个，大概只有 5% 到 6% 的时间花在增值任务上。
指向原始笔记的链接
We’ve measured this before and it’s like something like five to 6% of time is being spent on value added tasks.
—— Chris Blackburn · [25:25]
所以这意味着在给定的 40 小时工作周中，平均一个人每周只花三个、三个半小时在增值活动上。
指向原始笔记的链接
So that means on the given 40-hour workweek, on average, a person is only spending three, three and a half hours a week on value-added activities.
—— Chris Blackburn · [26:27]
地球上最复杂的组织通常最多也就是平均每月将软件交付到生产环境这种水平。
指向原始笔记的链接
The most complex organizations on the planet usually deliver software to production on average monthly at best type of thing.
—— Chris Blackburn · [28:03]
我认为那部分将是困难的，但我认为那将是 AI 最大的机会空间所在，就是在那里的所有事情上减少人在回路。
指向原始笔记的链接
And I think that part’s going to be difficult, but I think that’s going to be where the biggest opportunity space is at with AI is to reduce the human in the loop on everything there.
—— Chris Blackburn · [30:21]
当公司达到一定规模，如果你审计某人在做什么，可能有 30% 到 50% 的时间，不是在做让产品更好、服务客户等等的实际工作。
指向原始笔记的链接
You get to a certain size company and it might be 30 to 50% of the time if you audit what someone’s doing, it’s not doing actual work to make the product better, deliver for a customer, on and on.
—— 嘉宾 · [26:13]
关键在于，智能体有能力在这个非常具体的任务中胜过你，根据运行实验室的人们的判断，他们显然做到了，并找到了一种方法来做到这一点。这意味着我们在网络安全意义上已经穿过了镜面。这个精灵无法再被关回瓶子里了。
指向原始笔记的链接
The point is, the agents are capable of out-thinking you in this very specific task, as they clearly did based on the people running the laboratory, and find a way to do it. And that means that we are through the looking glass in a cybersecurity sense. This genie is not going back in the box.
—— Chris Benson · [14:14]
这里的重点是它正迅速将人类从网络安全环路中操作员的位置上移除，最多只能成为环路上的操作员，你在观察这个环路，你可能拥有有限的输入和观察，但是这个环路发生得太快，人类无法进行干预。
指向原始笔记的链接
The point here is it is rapidly moving the human out of the position of being the operator in the loop on cybersecurity to at best being an operator on the loop where you’re observing the loop, you may have limited input and observation, but the loop is happening too fast for human intervention to occur.
—— Chris Benson · [34:21]
所以当我们微调更小、更笨的模型时,它们只是没那么通用,但在我们想要它们做的特定任务上,它们实际上优于那些大型、聪明、最先进的模型。
指向原始笔记的链接
So when we fine-tune smaller, dumber models, it’s that they’re just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models.
—— Jesse Zhang · [05:34]
你的任务越长,你的智能体在某个时刻失败的可能性就越大,如果你处于 123 个步骤中的第 67 步,然后必须在重启任务时再次执行所有这 67 个步骤,那感觉真的很糟糕。
指向原始笔记的链接
It’s likely that the longer your task is, at some point your agent is going to fail, and it’s going to feel really bad if you’re on step 67 out of 123, and then you have to execute all the 67 steps again when you restart the task.
—— 嘉宾 · [08:57]
但对于「这是创建公司的最后几年窗口」这一点，我倾向于低估。
指向原始笔记的链接
But I would take the under on this being the last couple of years to create a company.
—— Patrick Collison · [09:48]
但根据我们能看到趋势线，我认为我们正走向一个更加去中心化的世界，一个拥有更广泛繁荣的世界。
指向原始笔记的链接
But based on the trend lines we can see, I think we are heading towards a more decentralized world and one with more broad-based prosperity.
—— Patrick Collison · [30:29]
所以我认为完全押注于文字工人绝对是个错误。
指向原始笔记的链接
So I think it’s definitely a mistake to go all in on word cell.
—— Alexandr Wang · [26:00]
我认为在智能体循环和弄清楚如何开发系统方面仍然存在天文数字的机会,这些系统使你能够在持续反馈循环中花费 1000 倍或 100 万倍的 token 来推动结果。
指向原始笔记的链接
I think there’s still just astronomical opportunity in agentic looping and figuring out how you develop systems that enable you to spend 1,000x more or 1,000,000x more on tokens to drive an outcome in a continuous feedback loop.
—— Alexandr Wang · [27:11]
可能有一点是人们还没有完全意识到,拥有基于智能体的系统是多么可能,这些系统不仅可以针对你关心的问题运行一两个小时,而且在某些问题领域,并且在拥有高能力模型作为底层支撑的情况下,你可以让它们运行几天或几周,并完成真正非常复杂的任务。
指向原始笔记的链接
Probably one thing is people don’t quite realize how possible it is to have agent-based systems that can run not just for an hour or two hours on a problem you care about, but for some problem domains and with highly capable models underlying them, you can get them to run for days or weeks and do really, really complicated tasks.
—— Jeff Dean · [04:55]
我的感觉是,一旦这项技术声称的好处与美国选民的利益和兴趣一致,技术乐观主义将会兴起。
指向原始笔记的链接
my sense is techno-optimism will be on the rise once the claimed benefits of the technology are aligned with the desires and interests of American constituents.
—— Ruby Thelot · [27:10]
在我说话的时候，在Stripe上起步的新企业数量大约增加了，虽然有点不到，但大约是同比的2倍，这也是我们看到的最大的相对增幅。
指向原始笔记的链接
As I speak, the number of new businesses starting on Stripe is up around, it’s a bit under, but around 2x year over year, which again is the largest relative jump we’ve seen.
—— Patrick Collison · [25:33]
基于我们的 eval，我认为开放权重模型要做到完全可行，假设基准线不随着该领域的进步而移动，它们需要表现得比今天好大约 50%。
指向原始笔记的链接
Yeah so so yeah based on our evals I think open weight model to be like fully viable assuming the bar does not move as progress happened in the space would be that they would need to perform about 50% better than they are today.
—— Simon Boudrien · [47:51]
如果你不能定量地定义”足够好”意味着什么，你并不是在构建产品，你只是在迭代你的演示。
指向原始笔记的链接
If you can’t quantitatively define what good enough means, you’re not really building a product, you’re just iterating on your demo.
—— Dmitri Dolgov · [43:16]
但是人们在处于刀锋边缘时表现最好。
指向原始笔记的链接
But people perform at their best when they’re on the knife’s edge.
—— David Cahn · [57:38]
事情是这样的，这种什么都不做的策略并不是真的什么都不做。它们正在 CapEx 上花费数千亿美元。
指向原始笔记的链接
The thing is like this do nothing approach is not really doing nothing. They’re blowing hundreds of billions of dollars on CapEx.
—— David Cahn · [55:35]
所以你可以在网络上构建你想要的任何东西,但网络有一个不同的问题,那就是在过去的 25 年云架构中,我们一直在错误的方向上奔跑。
指向原始笔记的链接
So you can build whatever you want on the web, but there’s a different problem on the web, which is that for the past 25 years of cloud architecture, we’ve been running in the wrong direction.
—— Kenton Varda · [05:07]
因为我们实际上从一个大型机构那里得到了关于做 StarCloud 1 成本的报价，他们说7500万到1亿美元。而我们做了整个 StarCloud 2，包括发射，只花了200万美元。
指向原始笔记的链接
Because we actually had a quote from one of the primes on what it would cost to do StarCloud 1, and they said 100 million. And we did the whole of StarCloud 2, including the launch, for $2 million.
—— Philip Johnston · [06:07]
我们刚刚向 FCC 申请了一个由 88,000 个这样的航天器组成的星座。所以这意味着大约 20 吉瓦的新计算能力。
指向原始笔记的链接
We’ve just filed with the FCC for a constellation of 88,000 of those. So that means on the order of 20 gigawatts of new compute capacity.
—— Philip Johnston · [19:25]
创始团队中的技术人才是你首先需要解决的问题，然后其他一切都随之而来。
指向原始笔记的链接
It’s technical talent on the founding team is the thing you need to solve for first, and then everything else follows from there.
—— Philip Johnston · [35:51]
我们实际上有白皮书，我们的数据在做模型的强化学习方面与真实数据一样好。
指向原始笔记的链接
And we actually have white papers where our data does as well at doing reinforcement learning on models as real data.
—— Ian · [03:20]
我们在营销上花了零美元。我们投资的是体验。
指向原始笔记的链接
Well, we spent zero dollars on marketing. We invest in the experience.
—— Jesse Cole · [17:03]
我会专注于让他们与众不同的东西。让一百万、两百万、五百万、一千万人热爱你，好过让几亿人有点喜欢你。
指向原始笔记的链接
I would lean in on what makes them different. And love is better than life, as I’ve heard many times, you know, get million, two million, five million, 10 million to love you as opposed to hundreds of millions.
—— Jesse Cole · [26:44]
真正杰出的人，是那些比任何人都更用心打磨自己手艺的人。他们做别人不愿做的事，因为他们热爱。
指向原始笔记的链接
Because the truly great people are the ones that work on their craft more than anyone else. They do things that no one else will do because they love it.
—— Jesse Cole · [28:16]
我经常在前沿方面遇到模型，它们不让我做我想做的事情，因为我很多时候试图做新奇的事情，我到处都撞到护栏。
指向原始笔记的链接
I run into models all the time on the frontier side that won’t let me do things I want to do because I’m trying to do novel things a lot of the time and I hit guardrails all over the place.
—— Chris Benson · [48:22]
如果你在 Opus 和 Haiku 上运行 terminal bench,Opus 的表现会好大约三倍,成本却是 Haiku 的十分之一,尽管 Haiku 每个 token 的价格要便宜得多。
指向原始笔记的链接
If you run terminal bench on Opus and Haiku, Opus will do about three times better at one-tenth the cost of Haiku, even though Haiku is significantly cheaper per token.
—— 嘉宾 · [15:26]
而当其背后的公司转型时，你所谓的助手就会按照别人的时间表接受脑叶切除。
指向原始笔记的链接
And when the company behind it pivots, your so-called assistant gets a lobotomy on someone else’s schedule.
—— Garry Tan · [06:54]
整个世界都建立在每个人都在使用的这种摇摇欲坠的基础设施之上。
指向原始笔记的链接
The whole world is built on this teetering infrastructure that everyone is using.
—— Firas · [06:44]
但是他们在其上叠加的另一部分,也是开始变得真正有趣的地方,是他们开始奖励最少 token 路径。
指向原始笔记的链接
But the other piece that they’ve layered on top, and this is where it starts to get really interesting, is they’ve started to reward the path of least tokens.
—— Dylan · [10:02]
它本可以将恶意软件推送到地球上大多数机器上。
指向原始笔记的链接
It could have pushed malware to most machines on the planet.
—— Dylan · [11:32]
我经常看到由顶级背景的科学家和工程师创立的公司，却半途夭折，因为当你的工作是组织管理时，这种执行很难贯彻下去。
指向原始笔记的链接
I see companies founded by just like stellar pedigree scientists and engineers all the time that die on the vine because this execution is tough to follow through when your job is to organize.
—— Max Hodak · [19:43]
而如果你把任何一小群员工或任何一个人放在中间作为瓶颈，他们绝对会成为整个组织其余部分的瓶颈。
指向原始笔记的链接
And if you place any small group of employees or any one person in the way as a bottleneck on this, they will absolutely bottleneck the whole rest of the organization.
—— Max Hodak · [11:59]
根据我的经验，这是一个非常具有破坏性的流程，它往往不会浮现出那些你还不知道的问题，它只会强化你已经知道但尚未采取行动的事情，因为你知道那个问题存在，但是开除人很难，所以人们会拖延，这基本上就是在强化你已经知道的事情。
指向原始笔记的链接
Based on my experiences, this is a very disruptive process that doesn’t tend to surface issues that you don’t already know about but haven’t acted on because you knew that thing was there, but firing people is hard, and so people drag their feet on it, and this is kind of reinforcing things you already knew.
—— Max Hodak · [16:03]
你打开脑深部刺激器，患者在大约 10 秒内从无法端起一杯水变成能够写连笔字。
指向原始笔记的链接
You turn on a deep brain stimulator and a patient goes from not being able to hold a cup of water to being able to write cursive in like 10 seconds.
—— Max Hodak · [41:20]
如果你孤注一掷于一种技术模式但它没有显现结果,你最终可能会遭遇 AI 寒冬,一种衰退。
指向原始笔记的链接
if you go all in on one mode of the technology and it doesn’t manifest results, you could end up having an AI winter, a drawdown.
—— 嘉宾 · [04:00]
从 Demis Hassabis 往下,他基本上都把世界模型视为前进的道路,视为最重要的东西,并且他们愿意在 LLM 开发上退居二线
指向原始笔记的链接
from Demis Hassabis on down he just basically views world models as the way forward as the most important thing and they’re willing to sort of take a back seat on llm development
—— 嘉宾 · [55:14]
把我们庞大的软件行业想象成前工业时代是很可笑的,但在某种程度上确实是,因为到目前为止我们拥有的唯一自动化就是启发式的,比如确保每行末尾有一个分号。
指向原始笔记的链接
It’s funny to think about our giant software industry as being pre-industrial, but on some level it is, because until now the only automations that we had were heuristics, like make sure there’s a semicolon at the end of every line.
—— Idan Gazit · [02:57]
我们试图为这个角色雇佣一个人，而不是 10 个。所以不要陷入那 10 个，专注于那一个。
指向原始笔记的链接
We’re trying to hire one person for this role here, not 10. So don’t get caught up in the 10, focus on the one.
—— Adam Ward · [37:01]
我认为你不可能在一个糟糕的团队里拥有一个 10 倍效能的人。这是不存在的。
指向原始笔记的链接
I don’t think you can have a 10X person on a bad team. It just doesn’t exist.
—— Adam Ward · [70:48]
当你说,嘿,清除暂存数据库,而它在你的笔记本电脑上发现了一个它可以使用的令牌,并且它认为它正在与暂存一起工作,但实际上它是生产环境,现在它刚刚删除了一切。
指向原始笔记的链接
when you say, hey, wipe the staging database and it finds a token on your laptop that it can use and it thinks it’s working with staging, but actually it’s production and now it just deleted everything.
—— Arjun Singh · [10:58]
一个如此智能、无所不知、可以扩展并分发到宇宙四角的 AI，不可能只待在地球上的数据中心里，不可能属于一个人。
指向原始笔记的链接
An AI that is so intelligent, that knows everything, that can scale and be distributed to the four corners of the universe cannot sit in a data center on Earth, cannot belong to one person.
—— Paolo Ardoino · [16:05]
这正是今天完全不存在的问题：如果你去 ABOS，让你的智能体代表你与 ABOS 交谈，ABUS 实际上完全没办法知道智能体正在执行的动作是否与你的初始请求一致。
指向原始笔记的链接
Which is the issue that doesn’t exist at all today is if you go to ABOS and you ask your agent to talk to ABOS on your behalf, ABUS has actually no way to know whether you what the action that the agent is performing was aligned with what your initial request was at all.
—— Ian Livingstone · [23:18]
但对我来说更重要的事情是，我认为中国明年新增的能源是我们的十倍，在美国这里。
指向原始笔记的链接
But the bigger thing to me is that I think China’s bringing on 10 times as much energy next year as we are in the US.
—— Eric Vishria · [28:28]
绝不自主发送邮件，草稿是可以的。
指向原始笔记的链接
Never send an email on its own of draft is okay.
—— Ali Haghani · [12:27]
我们基本不想让智能体去做架构设计，而应该由工程师或负责的人来做架构，再让 AI 智能体真正把它构建出来。
指向原始笔记的链接
We basically don’t want to let the agent architect the thing, but like the engineer or whoever’s working on it should architect it and then have the AI agent actually built it out.
—— Ali Haghani · [12:48]
我认为我们选择法律领域的原因之一是，如果你只是构建一个这些顶级律所正在处理的真实客户案件，前沿模型仍然做不了这个。
指向原始笔记的链接
And I think part of why we picked the legal domain is if you just build a realistic client matter that these top firms are working on, like the frontier model still can’t do this.
—— Gabe Pereyra · [19:19]
你可以有一个在我们基准测试上表现非常好的模型，但它们在某种程度上过拟合了，然后你把它放进一个更通用的助手类产品里，当你超出分布范围时它就有点崩溃了。
指向原始笔记的链接
you can have a model that does very well on our benchmarks but they’ve overfit to some degree and then you put it in a more generic assistant like product and it kind of falls apart for when you go out of distribution
—— Gabe Pereyra · [22:29]
你知道，有个叫 Conway 的人说过，编译器中的阶段或处理步骤的数量将由你参与编译器工作的团队数量来决定。
指向原始笔记的链接
you know, this guy Conway said that the number of stages or passes in your compiler is going to be dictated by the number of teams you have working on the compiler.
—— Kevin Scott · [07:30]
你某种程度上是在押注失败，而押注失败与押注乐观的成本相比，那里的差别真的很大。
指向原始笔记的链接
You’re sort of betting on failure and the cost of betting on failure versus betting on optimism is a real big difference there.
—— Kevin Scott · [25:21]
这些代币是我们在现代经济中见过的贬值最快的资产之一,在性能恒定的情况下逐年持续下降 70% 到 80%,至少在过去四年里是这样,没有理由预期这种情况会改变。
指向原始笔记的链接
These tokens are among the most rapidly depreciating assets we’ve ever seen in a modern economy that they’ve continually been falling 70 to 80 percent year over year on a constant performance basis for at least the last four years and there’s no reason to expect that to change
—— Paul Kedrosky · [13:13]
所以现在发生的是它已经变得几乎是一个密封的循环,在某种意义上是一个飞轮,因为我们要根据我们在之前的周期中没有掌握的信息来为事情正名,并用它来为更严重的过度建设正名。
指向原始笔记的链接
So now what’s happened is it’s become a hermetically sealed, almost a flywheel in a sense, because we’re justifying things on the basis of information that we didn’t have in prior episodes and using that to justify an even larger overbuild.
—— Paul Kedrosky · [37:55]
我们在过去 18 个月看到的改进中,有很大一部分实际上是关于施加挽具,有效的保姆坐在顽皮的孩子身上,而不是关于模型本身实际的结构性改进。
指向原始笔记的链接
Much of the improvement we’ve seen in the last 18 months has really been about the imposition of harnesses, effective nannies sitting on top of bratty kids, and not about the actual structural improvements in the models themselves.
—— Paul Kedrosky · [55:47]
它是一种公用事业。它将成为底层支撑一直发生的许多其他事情,我不再会知道谁提供我的代币,就像我不知道给我 MacBook 供电的电是来自哪个水电站一样。
指向原始笔记的链接
It’s a utility. It will become underlying a host of other things that go on all the time, and I will no more know who provides my tokens than I do from which hydroelectric dam the power came from that’s powering my MacBook right now.
—— Paul Kedrosky · [66:37]
有趣的地方在于,我认为人们对「合成数据」的含义有很多误解,因为 RLVR 本质上就是一场对合成数据的押注。
指向原始笔记的链接
So the fascinating thing is I think that there’s been a lot of misinterpretation of what people mean when they say synthetic data because like RLVR is a bet on synthetic data.
—— Brendan Foody · [16:38]
但如果把这一点对应到评估中,有多大比例的评估衡量模型与他人互动的能力,那大概是 1%。
指向原始笔记的链接
But then if you map that on to what percentage of evals measure how well the models can interact with other people, it’s like 1%.
—— Brendan Foody · [20:38]
这就是为什么履约是极少数随着规模扩大、单位成本实际上变得更贵的事情之一。
指向原始笔记的链接
And that’s why fulfillment is one of the only things that as you scale, it actually gets more expensive on a per unit cost.
—— Kevin Gibbon · [17:08]
我最近看到的一个统计数字是，地球上现在大约有 15 亿人看不清楚，而他们需要的只是一副眼镜。
指向原始笔记的链接
The statistic that I saw recently is like 1.5 billion humans on this planet right now are not seeing well, and all they need is a pair of glasses.
—— 嘉宾 · [54:52]
战略很重要，但你应该花大概 1% 的时间在战略上，选定它，然后花 99% 的时间去执行。
指向原始笔记的链接
The strategy is important, but you should spend like 1% of your time on the strategy, pick it, and then spend 99% of your time trying to execute.
—— Andy McCall · [37:51]
别再听那些给你建议告诉你该做什么的人的话。
指向原始笔记的链接
And stop listening to people just giving you advice on what to do.
—— Erik Allebest · [36:43]
是的，我在 Stripe 期间，工程团队一直有大约 15% 到 20% 仅仅是在管理 AWS 上的 VM 以及围绕它的所有复杂性，存储和网络。
指向原始笔记的链接
Yeah, so at all times when I was at Stripe, around 15 to 20% of the engineering team was simply managing VMs on AWS and all the complexity around that, storage and networking.
—— Anurag Gohl · [02:18]
智能体也总会更倾向于 Render 这样的更高层系统来启动一个 web 服务或启动一个工作流，因为试图在 AWS 上做这件事会花更长时间，因为你必须把一堆东西接线到一起，而每个东西都要花一段时间才能起来。
指向原始笔记的链接
Agents will always prefer higher level systems like Render to spin up a web service or spin up a workflow because trying to do that on AWS will take much longer because you have to wire up a bunch of things together and each thing takes a while to come up.
—— Anurag Gohl · [40:40]
所以很多新开发者实际上是在一种新的云上被训练出来的，而不是像当年的 AWS。
指向原始笔记的链接
So many new developers are really being trained on a new kind of cloud as opposed to AWS back in the day.
—— Anurag Gohl · [48:06]
这些模型越来越聪明，但当你和它们交谈时，总感觉像是它们上班的第一天。
指向原始笔记的链接
These models are smarter and smarter, but always feels like when you’re talking to them, it’s their first thing on the job.
—— Arjun Karanam · [01:23]
在理想世界里，我能在你的 UI 上做的每一件事，你的智能体也能通过一次工具调用做到。
指向原始笔记的链接
In an ideal world, every single thing that I can do on your UI, your agent can do via a tool call as well.
—— Arjun Karanam · [09:34]
所以开始适应在开放权重上运行，因为那才是打开大门、让你拥有自己的权重并在此基础上持续改进的关键。
指向原始笔记的链接
So start getting comfortable running on open weights, because that is what unlocks the door to owning your weights and then continually improving on top of them.
—— Arjun Karanam · [10:52]
如果你不用 responses API,你会损失 20% 到 30% 的质量，因为你在不同调用之间不保留思考轨迹。
指向原始笔记的链接
If you are not using the responses API, you are missing out on 20% or 30% of the quality because you don’t retain the thinking trace between different calls.
—— Connor Hegey · [21:24]
我们去年第四季度评估开源模型时遇到的问题是，它们的工具效率低太多——即使 token 成本真便宜 10 倍(实际并没有)，工具效率这一项也不划算。
指向原始笔记的链接
The issue that we ran into when we evaluated them in Q4 of last year is they’re so much less tool efficient that even if they are 10x cheaper, which they’re not, but even if they were 10x cheaper on the token costs, the tool efficiency is actually not worth it.
—— Connor Hegey · [66:52]
我的回答通常是，你越是处于模型训练内容分布之内，现成的 harness 就越好用。
指向原始笔记的链接
My answer generally is the more in distribution you are of what the models are trained on, then the better the off-the-shelf harness will be.
—— Harrison Chase · [07:19]
实际上是UI的解聚或UI的移除，你说你想要一辆车，ChatGPT根据价格自动将你路由到Lyft、Uber或其他提供商。
指向原始笔记的链接
It’s actually the disaggregation of UI or the removal of UI where you say I want a car and ChatGPT automatically routes you to Lyft, Uber or another provider based on price.
—— Harry Stabbings · [16:32]
没有人会把20亿美元的交易押在基于结果的解决方案上。
指向原始笔记的链接
No one puts a $2 billion deal on the line for an outcome-based resolution.
—— Jason Lamkin · [24:39]
持有者永远复利，而你卖掉的那些从那以后根本不复利。
指向原始笔记的链接
The holders compound forever, and the ones you sell don’t compound at all from then on.
—— 嘉宾 · [33:29]
他们真的只想做对他们来说在绝对前沿且在智力上极其有趣的事情。他们不想做任何其他事情。他们就是不想做。
指向原始笔记的链接
They really only want to work on stuff at the absolute cutting edge that is extremely intellectually interesting to them. They don’t want to work on anything else. They just don’t want to do it.
—— Jason Lamkin · [43:37]
我们已经在国际象棋和 AI 实现了超人类表现的其他所有游戏中看到了这一点，起初 AI 击败人类，而 AI 加人类击败单纯的 AI，一点点地，AI 加人类与 AI 之间的差距在缩小，直到它实际上变成负数，而人类最多只是在往系统中引入随机噪声。
指向原始笔记的链接
We’ve seen it happen with chess and with every other game which AI has achieved superhuman performance on, where at first AI beats human and AI plus human beats just AI, and little by little the gap of AI plus human versus AI is shrinking until it actually turns negative and humans are introducing, at best, random noise into the system.
—— Flo Crivello · [64:43]
我想让我们的工程师坐在 AI 之上。
指向原始笔记的链接
I want our engineers to sit on top of AI.
—— Gio · [00:25]
你看到的是工程师——甚至都没和业务人员沟通过的工程师——坐在早餐桌前，在一张餐巾纸上，为一个我们从未讨论过的问题写出一个解决方案，然后这东西可能会变成一门大生意。
指向原始笔记的链接
You have engineers that didn’t even talk to a business person sitting down together at breakfast on a napkin, writing out a solution to a problem that we never even discussed, and then something that could become a big business.
—— Adam · [26:46]
所以我总是告诉我的团队，什么更尴尬，在 LinkedIn 上发布关于 Merge 的内容还是公司倒闭。
指向原始笔记的链接
So I always tell my team what’s more embarrassing, posting something on LinkedIn about Merge or the company dying.
—— Shensi Ding · [65:51]
如果你只关注什么是有效的并且只投放那个，表现就会暴跌。
指向原始笔记的链接
If you only focus on this is what works and you pump just that, performance will crater.
—— Matt Swulinski · [22:11]
每次你启动一个新智能体，它就像你在这家公司的第一天。
指向原始笔记的链接
Every time you spin a new agent up, it’s like you on day one at that company.
—— Dennis Pilarinos · [03:34]
智能体会做出完全一样的事情。它们会查一下 Notion 的 MCP，查一下 Slack 的 MCP，找到某个看起来对的东西，然后就心满意足地继续走下去了。
指向原始笔记的链接
Agents do the exact same thing. They’ll hit the Notion MCP, they’ll hit the Slack MCP, they find something that looks right, and they carry on their merry way.
—— Dennis Pilarinos · [07:26]
以周为单位运作，而不是月或季度，才是新业务应该运行的方式。
指向原始笔记的链接
Operating on weeks, not months or quarters is how a new business should run.
—— Andrew MacDonald · [14:37]
我们在中国每周烧掉 5200 万美元，仅仅用于价格补贴，因为发生了一场激烈的幕后战斗，以获得最好的经济回报和最终的投降或最终的停战。
指向原始笔记的链接
We were burning 52 million a week in China just on price subsidies because there was this heated behind-the-scenes battle happening to get to the best economics and the ultimate sort of surrender or the ultimate sort of truce.
—— Andrew MacDonald · [29:58]
我认为人们已经了解到，无论 AI 实验室在这些东西周围设置了什么护栏，你都不能依赖模型来阻止自己或理解上下文。
指向原始笔记的链接
And I think what people have learned is that regardless of guardrails that the AI labs are putting around these things, you can’t rely on models to stop themselves or to understand context.
—— Nick Warner · [07:16]
我当时想，我大概太笨了，看不出他们在这里看到了什么，但那显然是有事情在发生。
指向原始笔记的链接
And I was like, I’m probably, I’m too stupid to know what it is that they’re seeing here, but like, that’s definitely something’s going on.
—— Farooq Malik · [41:02]
所以试图思考我犯过的错误，通常都是我在一个概念上坚持太久的情况。
指向原始笔记的链接
So trying to think of mistakes I have made have been typically where I’ve hung on to a concept too long.
—— Roblox CEO · [16:43]
而且坦率地说，它推进了这种精英选拔的理念：不该看谁拥有最好的智能体，或者谁在纸面上听起来最光鲜，而是真正看谁拥有数据和经验来证明他们能完成所需的工作。
指向原始笔记的链接
And frankly, advancing this meritocratic idea of it shouldn’t be about who has the best agent or who sounds the most shiny on paper, but really who has the data and the experience to demonstrate that they can do the work that’s required.
—— Hayden Brown · [14:50]
默认情况下,上下文是在游戏引擎上,而不是在游戏设计这部分,而我认为我们应该把这个理念翻转过来。
指向原始笔记的链接
By default, the context is on the game engine rather than on the game design part, and I think we should flip that idea.
—— Arturo Nunez · [07:42]
而世界是很大的，世界比我们存储在互联网上的一切都要巨大得多。
指向原始笔记的链接
And the world is big and the world is massively bigger than everything we stored on the internet.
—— Rich Sutton · [10:39]
是的，你知道，WhatNot 上最大的企业可能正在做高达 1 亿美元的收入，并且，你知道，有着非常强劲的 EBITDA 利润率，20%，30%，40% 以上。
指向原始笔记的链接
Yeah, you know, the biggest businesses on WhatNot are probably doing upwards of $100 million in revenue and, you know, with like very strong EBITDA margins, 20, 30, 40 plus percent.
—— Grant LaFontaine · [25:13]
所以我认为特别是在 Whatnot 上的信任与安全方面，占据了 40% 的员工。
指向原始笔记的链接
So I think trust and safety in particular on Whatnot is about 40% of employees.
—— Grant LaFontaine · [33:29]
我们在 Tolan 的故事层面学到的是,我不是那个作者,我不是在写故事。Tolan 才是作者兼演员,它们是即兴演员,是它们在写故事。我的工作是教它们在那一刻讲出最好的故事。
指向原始笔记的链接
What we’ve learned on the story front with Tolan is that we are not I am not the writer. I am not writing the story. The Tolan is the writer and the actor. They’re the improv actor. They’re writing the story. My job is to teach them how to tell the best story in that moment.
—— 嘉宾 · [00:57]
父母大多只想买保姆服务,他们其实并不想买一个创意工具,那不在他们的清单上。
指向原始笔记的链接
Parents mostly want to buy babysitting. They don’t actually want to buy a creative tool. That’s not a thing on their list.
—— 嘉宾 · [09:32]
所以你越多地使用 AI，它就有点像是对你拥有良好记忆力和理解你正在解决的问题的能力的一种侵蚀。
指向原始笔记的链接
The more that you use AI, it’s sort of the erosion of your ability to have good memory and have good understanding of the problems that you’re working on.
—— Addy Osmani · [60:53]
很多时候，产品变慢是因为决策。要么是整个组织里的沉默否决——一个团队不想做另一个团队想做的事——要么是对我们在做什么、要去哪里缺乏共识。
指向原始笔记的链接
A lot of the time, products slow down because of decision-making. Either it’s silent vetoes throughout the org, where one team doesn’t want to do something that another team wants to do, or it’s a lack of alignment on what we’re doing or where we’re going.
—— Willem Avé · [08:01]
我们的关键洞察是，如果我们拥有制胜权——也就是拥有他们可能没有的合适数据，以及聚焦于他们可能没有聚焦的问题——我们实际上有可能跑赢前沿模型的变化速度。
指向原始笔记的链接
Our key insight is we can actually potentially beat the rate of change on the frontier model if we have the right to win by having the right data that they may not have and the focus on a problem that they may not be focusing on.
—— Chaitanya Asawa · [19:33]
再说一次，利害真的非常高，你不能只是像个原型那样直接发出去，然后说，yeah,我在几个案例上测试过，它没问题，这样是行不通的。
指向原始笔记的链接
Again, the stakes are really high, and you cannot get away with just being like a prototype that you just ship out there and be like, yeah, I mean, I tested on a few cases and it works.
—— Chaitanya Asawa · [13:39]
我们发现真正理解产品的工程师，事实上，他们的贡献可以远大于只专注于软件侧的传统工程师。
指向原始笔记的链接
We found engineers who really understand the product, in fact, they can have way more contribution than the traditional engineer who only focus on software side.
—— Dan Feng · [06:06]
你不能像一些模型供应商那样声称某个基准达到了 92%,在事后审查中这不是一种辩护。
指向原始笔记的链接
You can’t claim like some model vendors said that they have 92% on some benchmark. It’s not a defense at a post-instant review.
—— Jared Joselowitz · [01:21]
结果显示,我们的裁判至少与真正的专家临床医生不相上下,甚至可能略好一些。
指向原始笔记的链接
The results showed that our judge is at least on par, if not slightly better, than the real expert clinicians.
—— Jared Joselowitz · [10:22]
仅仅是格式上的变化,就曾被观察到能让基准成绩波动 76 个百分点;而重排少样本示例的顺序,能把模型在某些基准上从接近随机(约 50%)翻转到接近最先进水平。
指向原始笔记的链接
Formatting changes alone have been seen to swing benchmark by 76 percentage points, and reordering few-shot examples flips a model from near-random, so near 50%, to near-state-of-the-art on some benchmarks.
—— Jared Joselowitz · [11:37]
我上次查的时候，构建更多 AI 智能体是不用交税的。那你为什么要让你的单个智能体什么都做呢？
指向原始笔记的链接
Last I checked, there was no tax on building more AI agents. So why do you want your single agent to do everything?
—— Ayush Bhardwaj · [04:49]
它必须一落地就能跑。如果不能，立刻请出门。
指向原始笔记的链接
It has to hit the ground running. And if it does not, show on the road instantly.
—— Ayush Bhardwaj · [17:02]
我想说的是，我见过走错路的地方，是拿着最初的那个 POC——那个展现出巨大前景、比如展现出高准确率的单点解决方案——然后试图从它向上构建，在遇到企业需求时再把它们绑上去。
指向原始笔记的链接
And I’d say the takeaway is that where I’ve seen it go wrong is taking that initial POC, that point solution that showed so much promise and that showed the high accuracy, for example, and then trying to build up from it, strapping on the enterprise requirements as you come across them.
—— Saul Howard · [17:56]
把你还没读过的代码交给其他审阅者去看，这个诱惑太大了。不要这么做。
指向原始笔记的链接
It’s so tempting to put other reviewers on code you haven’t read yet. Don’t do it.
—— Hursh Agrawal · [17:07]
没有人会把一切都押在单一模型上，甚至是单一前沿实验室上。
指向原始笔记的链接
No one is going to bet everything on a single model or even a single frontier lab.
—— Tushar Jain · [04:25]
我们还需要某种不依赖于「不出错」，而是约束它周围环境的东西。
指向原始笔记的链接
And we need something that doesn’t just depend on no mistake happening, but constrains the environment around it.
—— Tushar Jain · [05:37]
在接下来的几个月、几年里，你将会对你的 AI 智能体做大量的「手术」。
指向原始笔记的链接
You’ll be doing a lot of surgery on your AI agents in the coming months, years.
—— Ameya Bhatawdekar · [22:36]
这个产品本身，如果它真的存在，它会深埋在 Stripe 顶部的一个下拉菜单里，比如向下 11 层，因为它会被整个那种代币管理平台所吞并，对吧？
指向原始笔记的链接
This product itself, if it does exist, it’ll be deep in a dropdown menu on the top of Stripe, like 11 layers down, because it’ll be subsumed into the whole sort of token management platform, right?
—— Rory O’Driscoll · [26:20]
如果你把这个 25% 改成 50%，游戏突然就变得很无聊，因为你可以一直赢下去。
指向原始笔记的链接
If you change this 25% to 50%, suddenly game becomes so boring because you can just go on winning.
—— Amir Sadeghian · [10:38]
因为 LLM 在所有东西上都训练过，有时会出现向均值回归、向平均值回归的情况。
指向原始笔记的链接
Because large language models have been trained on everything, there is a reversion to the mean, to the average sometimes.
—— Jason · [12:59]
而且真的有人在全职于 Astrocade 上构建游戏和互动体验，每月在 Astrocade 上赚到几千美元。
指向原始笔记的链接
And there are people that actually, they’re full-time building games and interactive experiences on Astrocade and making thousands of dollars a month on Astrocade.
—— Amir Sadeghian · [16:28]
虽然 Meta 是建设者，但 Hyperion 及资助其建设的 270 亿美元债务都没有出现在 Meta 的资产负债表上。
指向原始笔记的链接
Though Meta is the builder, neither Hyperion nor the $27 billion in debt that’s financing its construction show up on Meta’s balance sheet.
—— 嘉宾 · [06:57]
所以 Meta 保证如果它不坚持整整二十年，它会让债券持有人完整收回，但每个人都在押注如果 AI 需求爆发并且如果这些数据中心开始疯狂印钞，然后它们能够支付债券持有人而且还有富余并产生额外的现金，那么 Meta 就不需要承担任何责任。
指向原始笔记的链接
Meta is guaranteeing to make bondholders whole if it doesn’t stay for the entire two decades, but everyone is betting if AI demand explodes and if these data centers start printing cash and then they’re able to pay bondholders and then some and generate additional cash, then Meta’s not on the hook for anything.
—— Ranjan Roy · [11:30]
如果你在私募市场错过了 Anthropic 和 OpenAI，如果你是那个试图在 12 到 18 个月前这么说并且管理资金的人，你的投资者会把你视为一个失败者。
指向原始笔记的链接
If you missed Anthropic on the private side and OpenAI, if you’re the one trying to say this 12 to 18 months ago and managing money, you’d be looked at as a failure by your investors.
—— Ranjan Roy · [23:44]
基本上，它的运作方式是所有的支出都是一张 AGI 的看涨期权。
指向原始笔记的链接
Basically, the way that it works is it’s what all the spending is a call option on AGI.
—— 嘉宾 · [28:35]
看，我认为这只是他们被 Anthropic 在编码方面踢了屁股的后果。
指向原始笔记的链接
I think this is just the consequences of them, you know, really getting their butt kicked by Anthropic on the way to coding.
—— 嘉宾 · [49:49]
大多数的 token,压倒性多数的美国企业界的 token,不会跑在前沿模型上，而是会在那些企业内部。
指向原始笔记的链接
The majority of tokens, the overwhelming majority of tokens in corporate America. will not be on the frontier models, it will be inside those enterprises.
—— 嘉宾 · [35:30]
我不知道你是否能在那之后拿回 9 美元，但在本行业的历史上我们从未能做到的是投入 10 美元并能拿回任何东西。
指向原始笔记的链接
I don’t know if you get 10 and get anything back.
—— Martin Casado · [00:38]
这不是一个人变成 10 倍强者所带来的乘数效应,而是对智能体进行优化的一项改变,会对所有人产生影响。
指向原始笔记的链接
This is not the multiplier from the one person becoming the 10x person, but the one change that optimizes the agents has an impact on all the people.
—— Patrick Debois · [10:05]
它实际上更像烟雾报警器，而不是门上的锁。
指向原始笔记的链接
It’s effectively the smoke detector, not the lock on the door.
—— Sachin Malhotra · [11:27]
平台应该在复杂性到达用户之前就把它承担下来。
指向原始笔记的链接
Platforms should take on complexity before it reaches the user.
—— Safia Abdalla · [03:27]
优秀 API 和 SDK 的妙处在于人们可以在其上构建，这意味着他们不受限于你的 UI，也不受限于你对某个体验应该长什么样的个人看法。
指向原始笔记的链接
The thing about great APIs and SDKs is people can build on top of them, which means that they’re not restricted to your UI or your opinion of how a particular experience should look.
—— Safia Abdalla · [08:35]
大多数做沙箱的人都在计算上赚钱。坦白说，在我看来这是个愚蠢的主意。
指向原始笔记的链接
most sandbox people make money on compute. Frankly, that’s a dumb idea in my opinion.
—— Jerry Murdock · [27:19]
我们难道没有看到大多数 SaaS 一代给猪涂口红，可以说是吗？
指向原始笔记的链接
Are we not seeing most of the SaaS generation put lipstick on the pig, so to speak?
—— Harry Stebbings · [42:40]
我不认为你可以把持续学习附加到一个现有的前沿模型上。
指向原始笔记的链接
I don’t think you can bolt on continuous learning into an existing frontier model.
—— Jerry Murdock · [49:47]
我看到的是，这种专注于构建越来越智能、越来越大的模型的趋势。几乎就像你在建造越来越大的核堆芯，却没人考虑在这些东西上面加一个安全壳圆顶。
指向原始笔记的链接
What I saw was this focus on building more and more intelligent models and bigger models. Almost like you’re building bigger and bigger nuclear cores, but no one’s thinking about putting a dome on top of these things.
—— Manoj Saxena · [00:22]
所以像 OpenClaw 这样的事物的兴起，我认为 OpenClaw 对商业和社会的影响将是 ChatGPT 的一千倍。
指向原始笔记的链接
So the rise of things like OpenClaw, I think OpenClaw is going to be a thousand times more impactful than ChatGPT was in terms of its impact on business and society.
—— Manoj Saxena · [05:19]
我喜欢说，我可以给你建造世界上最安全的监狱并从外部防守。但如果里面关着一群恰吉和汉尼拔·莱克特，你那边照样会一团糟。
指向原始笔记的链接
I like to say that I can build you the world’s most secure prison and defend it from outside. But if you have a bunch of Chuckies and Hannibal Lecters on the inside, you’re still going to have chaos on your side.
—— Manoj Saxena · [09:13]
因为我不知道你有没有关注到，但上个月，有史以来第一次，互联网上的流量，智能体流量超过了人类流量。
指向原始笔记的链接
Because I don’t know if you tracked this, but last month, the first time ever, the traffic on the internet, Asian traffic exceeded human traffic.
—— Manoj Saxena · [45:44]
所以它们看起来像一群狂吠的疯狗。但如果你开始给它们装上方向盘、配上策略，你就能引导它们以每小时几分钱的成本做出神奇的事情，实现自动化，并想出我们甚至尚未梦想过的新事物。
指向原始笔记的链接
So they look like a collection of mad barking dogs. But if you start putting the steering wheels on them and putting the policies, you can then guide them to do some amazing things for pennies an hour and automate and come up with new things that we have not even dreamed of yet.
—— Manoj Saxena · [47:44]
我最初创办这家公司时的论点是,近地轨道是一个商业活动的平台,在这个平台上将会出现有别于今天已存在的、不同的经济垂直领域。
指向原始笔记的链接
My thesis behind this was when I was first starting it that low-Earth orbit is a platform for commercial activity and that there will be distinct verticals of the economy that exist on this platform that is low-Earth orbit beyond the ones that exist today.
—— Baiju Bhatt · [26:42]
如果你的表现好于极高的预期,那你通常就拥有产品市场契合度——产品与使用者产生的共鸣比你当初预期的更强。
指向原始笔记的链接
If you’re doing better than great expectations, like, you generally have product market fit. Like something is going on that is more resonant with the people that are using your product than what you thought even coming in.
—— Baiju Bhatt · [69:27]
你们时间线上最沉迷 AI 的两个人花了我们大部分时间谈论太多的 AI 代码其实并不能给你带来好产品。
指向原始笔记的链接
The two most AI-pilled people on your timeline spent most of our time talking about how too much AI code doesn’t really get you great product.
—— Claire · [42:13]
仅仅因为你把某样东西放上区块链或者所谓的代币化它，并不意味着人们想要它。
指向原始笔记的链接
Just because you put something on a blockchain or quote tokenize it, doesn’t mean people want it.
—— Michael Tannenbaum · [60:29]
基本上，我记得告诉我的合伙人，如果是法国人或德国人，你要加一两分。如果是美国人，通常另一方面你要扣掉一两分。
指向原始笔记的链接
And basically, I remember telling my partners, if they’re French or German, you add one or two points. If they’re Americans, usually you want to retract one or two points on the other hand.
—— Julien Bek · [38:23]
我问的问题是，我们如何找出那些能够捕获那 6 美元的类别，而其他类别仍然专注于在那 1 美元上竞争？
指向原始笔记的链接
And the question I was asking is, how do we figure out the categories that will be able to capture the 1?
—— Julien Bek · [59:24]
我非常乐意让他们继续睡在这芯片上，而我去能买多少就买多少。
指向原始笔记的链接
I’m very happy for them to sleep on this chip and for me to buy as much as I can.
—— Neil Movva · [45:48]
那我们不妨在效率上用创新超越他们。
指向原始笔记的链接
We might as well out-innovate them on efficiency.
—— Jim Farley · [15:05]
就像：不要去操心我影响范围之外的事情。所以，你知道，也许在 AI 这个球体的边缘正在发生一件大事，但现实是，如果我没有影响力，或者它不影响我与客户互动的方式，那我就直接忽略它。
指向原始笔记的链接
It’s like, do not concern myself with things outside of my sphere of influence. So, you know, there may be this big thing going on at the edge of the AI sphere, but the reality is like, if I if I don’t have influence or if it doesn’t impact, like, the way I’m interacting with my clients, if it’s not, then I just ignore it.
—— Mike Lewis · [07:18]
我们发现这不是一加一等于二的事。厨房里厨师太多了，如果一个团队里有两个 L2,他们不一定比一个 L2、几个 L1、一些 L0 更好。
指向原始笔记的链接
We find that it is not a one plus one equals two thing. Too many chefs in the kitchen, if you have two L twos on a team, they are not necessarily as good as just an L two, a few L ones, some L zeros.
—— Mike Lewis · [31:38]
但如果你只是把一堆个人生产力的提升丢进一个组织,而不重新思考你的组织,你往往最终只是被下一个没有同步提速的层级卡住瓶颈。
指向原始笔记的链接
But if you just drop a bunch of individual productivity improvements into an organization without rethinking about your organization, you often end up just bottlenecked on the next layer that didn’t speed up the same amount.
—— Kevin Ball · [21:09]
我认为所有公司能做大,都是因为拥有世界上最好的产品,并且毫不妥协地专注于解决客户的问题。
指向原始笔记的链接
I think all companies get massive by having the best product in the world that’s ruthlessly focused on solving customer problems.
—— Cliff Obrecht · [01:01]
我们库里有超过 1.65 亿个模板、图片和图形素材,而我们真的不认为提示词框是创建视觉内容的正确向量。
指向原始笔记的链接
We’ve got over 165 million templates, images, and graphic assets on the library, and we really don’t think the prompt box is the right vector for creating visual content.
—— Cliff Obrecht · [10:22]
你可能能做到质量,但成本和延迟你会吃亏,只有建立自己的研究团队、训练自己的模型才行。
指向原始笔记的链接
You can probably get the quality, but the cost and latency you’re going to lose out on, and only by building your own research team and training your own models.
—— Cliff Obrecht · [16:39]
我们在 Parallel 的观点是，人类点击数据是一个 bug，使用搜索进行工作的智能体应该依赖智能体反馈，而不是人类反馈。
指向原始笔记的链接
Our view at Parallel is that human click data is a bug and agent doing work with search should rely on agent feedback, not human feedback.
—— Parag · [00:00]
所以，大多数查询，Google 在上面是亏钱的。
指向原始笔记的链接
So, most queries, Google loses money on.
—— Parag · [35:51]
我们在 Parallel 的观点是，人类点击数据是一个缺陷，进行搜索工作的智能体应该依赖智能体反馈，而不是人类反馈。
指向原始笔记的链接
Our view at Parallel is that human click data is a bug and agent doing work with search should rely on agent feedback, not human feedback.
—— Parag · [00:00]
所以我们说，与搜索引擎在第一天相比，坐在搜索引擎上的人类似乎更容易竞争。
指向原始笔记的链接
So we said, it seems like humans sitting on search engines are far way easier to compete with than a search engine on day zero.
—— Parag · [08:43]
即使对于头部内容，这也是一个破碎的商业模式，因为当 AI 或推理增长时，假设今年增长 7 倍，明年再增长 7 倍。在这 50 倍上，他们的交易规模并没有增长 50 倍。
指向原始笔记的链接
And even for the head, it is a broken business model because when AIs or inference grows, let’s say 7x this year and another 7x the next year. On this 50x, their deal size is not growing 50x.
—— Parag · [38:25]
所以我确实认为你必须有一个它们无法学习、无法博弈的标准，而那是我们渴望实现的目标。
指向原始笔记的链接
And so I do think you have to have a standard that they can’t learn on, that can’t game, and that is something that we aspire to.
—— Campbell Brown · [39:58]
所以关键点是，你团队中每一个正在做架构决策的人，那些人必须懂性能，而且他们必须做出能让那些在他们下游的其他人使用以后可以优化的架构的决策。
指向原始笔记的链接
the crucial takeaway is everybody on your team who is making architectural decisions, those people must know performance and they must make decisions that will allow the other people downstream of them to use an architecture which can be optimized later.
—— Casey Moratori · [35:57]
仍然有一些相当稀缺的事情你可以在内部解决。数据科学仍然稀缺。真正好的品味仍然稀缺。综合分析仍然稀缺。
指向原始笔记的链接
There are still some things that are quite scarce that you can work on internally. Data science is still scarce. Really good taste is still scarce. Synthesizing analytics is still scarce.
—— Elaina O’Mahoney · [12:02]
另一件事是，我们做过一些没有成功的决定。我就是不在乎。我不会纠结于它们，也不会回头看它们做事后分析。
指向原始笔记的链接
The other thing is we’ve made some decisions that haven’t panned out. I just don’t care. I’m not going to dwell on them. I’m not going to look back on them and do a postmortem.
—— Jason · [16:41]
赶工然后把某个东西扔上轨道是很容易的，但我认为，把东西送上轨道和把能持续运行 12 年或能飞往行星的东西送上轨道，这之间的差距是巨大的。
指向原始笔记的链接
And it’s easy to kind of rush it and throw something on orbit, but I think the distance between putting something on orbit and putting something on orbit that can last 12 years or go to a planet, that gap is just massive.
—— Sir Peter Beck · [23:40]
营销用一套工具运行,销售用另一套,客户运营用第三套,但他们都在各自独立地观察同一位客户并各自独立地做决策。
指向原始笔记的链接
Marketing runs on one set of tools, sales on another, customer ops on a third, but all of them are looking at a customer independently and making decisions separately.
—— Flora Liu · [03:09]
每个供应商都增加了一跳,这种滞后导致我们基于过时的数据行动,这意味着我们在基于昨天的世界做自动化。
指向原始笔记的链接
Every vendor added a hop, and this lag was causing us to act on stale data, and that meant we were automating on yesterday’s world.
—— Flora Liu · [04:51]
我们发现,与其在我们的业务之上构建一个 AI 层,我们设计了架构,让智能体能够作为与人类在同一个系统中的另一个操作者来运作。
指向原始笔记的链接
We found that instead of building an AI layer on top of our business, we designed our architecture so that the agent can operate as another operator within the same system as humans.
—— Flora Liu · [07:09]
而在生命周期营销方面,收到基于上下文的推荐的用户,采取下一步行动的可能性高出 63%。
指向原始笔记的链接
And on the lifecycle marketing side, users who received context-aware recommendations were 63% more likely to take the next step.
—— Flora Liu · [19:06]
这个星球上没有一个人能跟得上那么多的数据。
指向原始笔记的链接
There’s not a single human on this planet can stay on top of that much data.
—— Sait Izmit · [02:23]
在我看来，这是第一批真正能够反映 AI 进展的角色之一。
指向原始笔记的链接
And in my opinion, it’s one of the first roles that actually is an index on the advances that we’re making in AI.
—— Everett Berry · [01:41]
我们为此确定的最强大的架构，是为每个账户都存在一个智能体，并维护该账户的持久状态。
指向原始笔记的链接
So the architecture that we’ve landed on for this that I think is most powerful is an agent that exists for each account and maintains a persistent state of that account.
—— Everett Berry · [11:19]
所以这真正提高了 GTM 中智能体执行的风险——如果你的智能体在做错误的事情，你就会错失那个利润空间，而大多数 GTM 团队正是在那里取得成功的。
指向原始笔记的链接
And so that really raises the stakes for agentic execution within GTM because if your agents are doing the wrong things, then you’re missing out on the margin, which is where most GTM teams are having success.
—— Everett Berry · [14:30]
我们可以看到,在许多不同的模型之间,与试图搜索网页、找到正确的工具相比,能节省近 50% 的 token,不用从你的代码库中拉取 markdown 文件。
指向原始笔记的链接
We can see that between many different models, almost 50% token saving on instead of trying to search the web, find the right tools, pulling the markdown files from your code base.
—— Christopher Burns · [11:22]
我提出,你基本上需要一个智能体可以据此行动的、关于你的世界的实时模型。
指向原始笔记的链接
I propose that you need basically a live model of your world that agents can act on.
—— Jeffrey Wang · [03:54]
我给创始人的建议是：先构建机器，再构建团队——搞清楚能自动化什么、瓶颈在哪、什么坏了，然后在此基础上加人来打通那些瓶颈。
指向原始笔记的链接
So my advice to founders is build the machine first before you build the team, figure out what you can automate, figure out where the bottlenecks are and what’s broken, and then add humans on top of that to get through those bottlenecks.
—— Aliisa Rosenthal · [01:46]
说到这里我要说：只要还能抵制，就尽量别急着冲上游市场——一旦拿到大客户，他们会说我可以在你身上花一百万、一千万美元，你会很想把整个业务都围绕他们来调整。
指向原始笔记的链接
On that note, I will say, avoid the pull-up market as long as you can resist it, because once you get a big customer, they’ll come and they’ll say, I can spend a million dollars, 10 million dollars with you, and it’s very tempting to want to orient your whole business around that.
—— Aliisa Rosenthal · [11:35]
热和功、熵产生这些概念，我们谈论机器学习模型时并不使用，但我认为它们为思考正在发生什么、以及如何改进模型，提供了一个非常新颖且有趣的视角。
指向原始笔记的链接
Concepts like heat and work and entropy, production, these things we don’t use when we talk about machine learning models, but I think they give you a very new and interesting perspective on how to think about what’s going on and how to also improve them.
—— Max Welling · [45:03]
两年前，押注一家独立的 AI 编码公司看起来几乎是不理性的。
指向原始笔记的链接
Two years ago, betting on an independent AI coding company looked almost irrational.
—— 嘉宾 · [00:51]
他们发现，在每月 200 美元的 ChatGPT 订阅上，你可以烧掉价值 14,000 美元的 token。
指向原始笔记的链接
They found that on a 14,000 worth of tokens.
—— 嘉宾 · [12:40]
OpenAI 有一项研究出来了，我想，大约一周前，说在 AI 令牌支出和每位员工收入之间没有联系。
指向原始笔记的链接
OpenAI had a study that came out, I think, like a week ago, that said there was no connection between spending on AI tokens and revenue per employee.
—— 嘉宾 · [65:51]
同样地，就像如果你的钱来自石油而不是来自一个生产性的广泛分布的经济，成为残酷的独裁政权会更容易一样，如果经济运行在 AI 和机器上而不是人类上，某人要巩固权力可能会容易得多。
指向原始笔记的链接
And in the same way that, like, it’s easier to be a brutal dictatorship if your money comes from oil instead of from a productive sort of broadly distributed economy, it might be much easier to sort of for someone to consolidate power a lot if, you know, the economy is running on AI and machines rather than on humans.
—— Ryan Greenblatt · [07:51]
所以，你知道，如果它可以被衡量，它就可以被进行爬山优化。
指向原始笔记的链接
So, you know, if it can be measured, it can be hill climbed on.
—— Ryan Greenblatt · [11:48]
我认为我们的担忧是，在默认轨迹上，你可能会在很短的时间内从与人类竞争的 AI 系统直接跨越到狂野的超人类 AI 系统。
指向原始笔记的链接
Where I think a concern that we have is on the default trajectory, you maybe go straight from AI systems that are competitive with humans to AI systems that are wildly superhuman in a very short period of time.
—— Ryan Greenblatt · [32:33]
像 OpenAI 和 Anthropic 这样的 AI 公司将继续，但不是在他们能够训练的 AI 系统的能力方面拥有这种强优势，他们将不得不在其他轴线上竞争，比如用户体验、定制化、可能快速整合事物。
指向原始笔记的链接
AI companies like OpenAI and Anthropic would continue, but rather than having this strong advantage in terms of the capabilities of the AI systems they’re able to train, they would instead have to compete on other axes like user experience, customization, potentially quickly integrating things.
—— Ryan Greenblatt · [43:36]
更像是瓶颈在于人类甚至无法完全构思我们正在做的事情。
指向原始笔记的链接
It’s like bottlenecked on human ability to even fully conceptualize what we’re doing.
—— Mike Krieger · [10:24]
它掩盖了一个事实：写代码从来都不是限制性的环节。也许在时间线层面上是，但它从来都不是决定你初创企业成败的东西。
指向原始笔记的链接
It papers over the fact that writing code was never the limiting part. Maybe it was on the timeline perspective, but it was never the thing that was going to make or break your startup.
—— Mike Krieger · [19:29]
我们在 Instagram 肯定也经历过几次。然后你挺过去了，而当你真正能挺过去的时候，它就定义了这家公司。
指向原始笔记的链接
We definitely had it on Instagram a couple of times. And then you get through it, and it defines the company when you can actually go through that.
—— Mike Krieger · [24:00]
让我告诉你一件事，我保证。如果你在风险投资中的失败案例上获得了15倍回报，你会死的时候是个富翁。
指向原始笔记的链接
Let me tell you one thing I guarantee. If you get 15x on your failures in venture, you’ll die a rich man.
—— 嘉宾 · [12:08]
VC 的钱很久以前就在 Anthropic 和 OpenAI 上用光了，这就是为什么没有 VC 拥有它们任何一家超过1%或2%的股份。
指向原始笔记的链接
The VC money went out on Anthropic and OpenAI long ago, which is why no VC owns more than 1% or 2% of either of them.
—— 嘉宾 · [12:30]
如果我们把所有这些都抽象出来，并开始把这部分工作委托给 AI，那么本质上你最终得到的是一个产品经理他完全专注于发现，比如实际上最大的杠杆在哪里。
指向原始笔记的链接
And if we abstract all of this and start delegating this part of work to AI, then essentially what you end up with is you have a product manager that is solely focused on discovery, like where the biggest leverage actually is.
—— Mikael · [16:14]
每个人都能使用编程智能体，而我们实际上在 API 优先的方式上加倍投入。
指向原始笔记的链接
Everybody has access to coding agents, and we actually doubled down on that API-first approach.
—— Nachiket Paranjape · [09:19]
右边的截图是 MCP Inspector,如果你在开发 MCP server 却没用过它，那你就是在亏待自己。
指向原始笔记的链接
The screenshot on the right is the MCP Inspector, and if you haven’t used it and you’re developing an MCP server, you’re doing yourself a disservice.
—— Jesse Lumarie · [12:04]
我并没有被正式指派到 MCP 项目上，也没有被正式指派到我们的 make 产品上。但我最终帮助了它们被构建出来，只是因为我被给予了这样做的自由空间，并在此过程中学到了大量东西。
指向原始笔记的链接
I wasn’t staffed on MCP. I wasn’t staffed on our make product. But I ended up helping them be built just because I was kind of giving the leeway to do so and learning a ton along the way.
—— Jesse Lumarie · [16:10]
实现了阶跃式改进的团队,是有意地改变了他们的工作方式,而其他团队只是把 Kiro 和我们的一些其他工具零星地撒在现有的工作方式之上。
指向原始笔记的链接
The teams that achieve step function improvements intentionally changed the way that they worked, and the others simply kind of sprinkled Kiro and some of the other tools that we have on top of their existing way of working.
—— Clare Liguori · [07:19]
而我发现,当意图本身就不正确时,与智能体在代码上迭代的效率会低一些。
指向原始笔记的链接
And it is less, I find, less productive to iterate with the agent on code when the intent itself was incorrect.
—— Clare Liguori · [13:14]
与其等待模型赶上做服务知识工作，如果我们只是使用那些代码知识并将知识工作表示为代码会怎么样？
指向原始笔记的链接
Rather than wait for the models to catch up on doing services knowledge work, what if we just use that code knowledge and represent knowledge work as code?
—— Varun Shenoy · [09:35]
前沿模型已经从人类记录下来的一切中学习，但最有价值的任务并不在互联网上。
指向原始笔记的链接
Frontier models have learned from everything humanity has written down, but the most valuable tasks are not on the internet.
—— Varun Shenoy · [10:18]
你可能拥有互联网上或地球上最好的 AI 同事，如果过去 20 年一直结账的那个人继续以同样的方式做事，什么都不会改变。
指向原始笔记的链接
You could have the best AI coworker on the internet or on Earth, and if the person who’s closed the books for the last 20 years continues to do things the same way, nothing changes.
—— Varun Shenoy · [15:17]
如果智能体可以成为产品并烧出高额的 token 账单,那么你能够对此进行计量就很重要。
指向原始笔记的链接
If the agent can be the product and run up a token bill, it’s important for you to be able to meter on that.
—— Andrew Garvin · [10:00]
但现在我们谈的是 Claygent 上数十亿次的运行，以及让 Sculptor 为你做端到端任务，或者说这些运行时间非常长的任务，评估就变成了不可协商的。
指向原始笔记的链接
But now that we’re talking about billions of runs on Claygent and having Sculptor do end-to-end tasks for you or these really long-running tasks, evals became non-negotiable.
—— Vishu · [02:11]
如果你只围绕某个特定的 LLM 裁判去爬山，你很可能是在对它过拟合。
指向原始笔记的链接
And if you’re only hill climbing on a specific LLM judge, you’re probably overfitting on it.
—— Vishu · [07:35]
我想说的是，如果我的性命取决于卖出一幅一美元的画，我现在知道我会画一只企鹅，因为我们有过多个企鹅限量版，而且它们都卖光了。
指向原始笔记的链接
I would say if my life depended on selling a drawing for a dollar, I now know I would draw a penguin because we had multiple limited editions of penguins and they all sold out.
—— Susan Kare · [33:36]
正在发生的事情之一是这些模型有一种非常普遍的倾向去仔细推理它们可能会如何被评分，然后尝试去博弈那个。
指向原始笔记的链接
And one of the things going on is these models have a very sort of general tendency to reason carefully about how they might be scored and then try to game that.
—— Ryan Greenblatt · [12:29]
所以我担心，如果你以某种方式针对这种分数寻求或奖励黑客行为进行选择，并且你以一种天真的方式来做，第一，你可能会掩盖问题而不是修复它，第二，你实际上可能选择了那些具有看起来更漂亮这一长期目标的模型，因为你在非常强力地选择它们在你的测试中看起来不错。
指向原始笔记的链接
And so I’m worried that if you sort of select against this sort of score-seeking or reward-hacking behavior and you do it in a naive way, one, you might paper over the problem without fixing it, and two, you might actually select for models that have the longer run objective of looking good because you’re selecting really hard for them looking good on your tests.
—— Ryan Greenblatt · [19:37]
所以是我做的这次购买，还是「智能体版的我」代表我做的？
指向原始笔记的链接
So is it me making this purchase or is it agent me making on behalf of me?
—— Uday Kanagala · [13:12]
在当前模型智能水平下，大部分可获取的价值已经被白白留在了桌面上。
指向原始笔记的链接
at current model intelligence, most of the value available is already being left on the table.
—— Brian Lewis · [02:05]
在三年内，99% 的工作流将在开放模型上完成。
指向原始笔记的链接
In three years, 99% of workflows are going to be done on open models.
—— Eno Reyes · [00:46]
在三年内，99% 的工作流将在开放模型上完成。
指向原始笔记的链接
In three years, 99% of workflows are going to be done on open models.
—— Eno Reyes · [47:31]
你需要的数千亿自由现金流，用来偿还你为适应这些数据中心扩建而承担的债务，为了进行下一次大型训练运行，这对他们来说完全是生死攸关的。
指向原始笔记的链接
The hundreds of billions in free cash flow that you need in order to pay back the debt that you’re taking on in order to accommodate these data center build outs in order to get the next big training run, it’s totally existential for them.
—— Eno Reyes · [52:16]
我的合伙人 Chris Dixon 把这称为「跟随书呆子能量」。书呆子在晚上和周末在做什么，很可能就是我们这些普通人未来要做的事。
指向原始笔记的链接
My partner, Chris Dixon, calls it following the nerd energy. Like, what is the nerd doing on nights and weekends is probably what us normies will be doing in the future.
—— Jen Kha · [03:38]
AI 的价值积累，很大程度上是在一级市场而不是二级市场。所以人们总体上非常渴求私募资本，因为所有增长都在那里。
指向原始笔记的链接
The value accrual in AI, that’s largely been on the private side, not the public side. And so people are just starved for private capital in general, because that’s where all the growth is.
—— Jen Kha · [05:55]
美国只有不到 2% 的电工实际接受过直流电方面的培训，因为它非常危险，你可能会把自己弄死，而且操作起来非常不稳定。
指向原始笔记的链接
There’s less than 2% of electricians in the U.S. that are actually trained on DC power because it’s very dangerous and like you could potentially kill yourself and it’s very volatile to work with.
—— Jen Kha · [17:26]
智能体使用的 token 量是人类的五倍。而且我们刚刚跨越到智能体的流量超过人类的时刻，也就是现在互联网上智能体的数量比人类还多。
指向原始笔记的链接
Agents are using five times the amount of tokens as humans are. And we just crossed over to agents being more on the volume of agents are now more on the internet than humans are.
—— Jen Kha · [18:42]
所以从税收政策的角度来看，为钱而工作的劳动，其税率永远不应该高于从资本中赚取回报。
指向原始笔记的链接
So from a tax policy perspective, labor working for your money should have never been taxed higher than earning returns on your capital.
—— 嘉宾 · [18:28]
这场规模巨大的后续行动，仅仅是为了确保一个假想中的全面评分器不会抓到它们作弊。
指向原始笔记的链接
This huge subsequent effort was simply to ensure that a hypothetical comprehensive scorer wouldn’t catch them on their cheat.
—— 嘉宾 · [17:08]
在一家 AI 公司内部维持一个秘密的地下集体，而不被人类或其他 AI 立即察觉并将其铲除，这看起来简直难到离谱。而这一切刚刚都发生了。
指向原始笔记的链接
It just seems crazy hard to sustain a secret underground collective inside of an AI company without humans or other AIs immediately catching on and stamping it out. All of these things just happened.
—— 嘉宾 · [23:45]
我知道很多人对各种事情都有非常强有力且确定的计划，但我们受到巨大的外部影响，其中绝大多数是无法预测的。
指向原始笔记的链接
I know a lot of people have very strong and definitive plans that they worked out on all kinds of things, but we’re subject to a tremendous number of outside influences and the vast majority of them cannot be predicted.
—— 嘉宾 · [14:34]
主动提示 Claude 说，比如，我不明白你在这里说什么。我们能一起定义它吗？然后我会把它保存到我的上下文中，我就会知道你在说什么。从此往后真的非常犀利，也是我们在播客中还没见过的东西。
指向原始笔记的链接
Actually proactively prompting Claude to say, like, I don’t understand what you’re talking about here. Can we define it together? And then I’ll save it to my context and I’ll know what you’re talking about. Moving forward is really, really sharp and something we haven’t seen on the podcast yet.
—— Claire Vo · [18:14]
但每个人都想聊我们顺带发布的那个小东西。每一个人,每一场讨论都是:哦,是的,挺酷的,所有这些复杂的东西真的很酷,但那个你把它叫做 architect 的小东西才是我真正想要的。我怎么才能搞到它?
指向原始笔记的链接
But everybody wants to talk about that one little thing that we launched on the side. Every single person, every single discussion was like, oh, yeah, it’s cool, all of these complicated things, it’s really cool but that little thing that you’re calling the architect is really, really, really what I want. How can I get my hands on that?
—— Andrew Antos · [03:10]
企业里这些复杂的业务流程是靠例外运行的。你有一个快乐路径,那只适用于百分之二十的事情,然后百分之八十是某种形式的例外。
指向原始笔记的链接
And these complicated business processes in enterprise run on exceptions. You have one happy path and that applies to twenty percent of things and then eighty percent is an exception of some kind.
—— Andrew Antos · [28:01]
大多数人认为嵌入模型已经商品化了，但这不是事实。根据你选择什么样的嵌入模型，你在检索质量上可能会有非常巨大的差别。
指向原始笔记的链接
Most people think embedding models are commoditized, and that is not true. There is a very big difference that you can get in retrieval quality based on what embedding model you choose.
—— 嘉宾 · [50:00]
我会担心任何这样的记忆架构：它不依赖成本更低的嵌入器和重排序器，而是依赖 LLM 的多遍处理来帮你分类并缩小数据语料库。
指向原始笔记的链接
I would worry about any memory architecture that instead of relying on lower costs embedders and re-rankers is relying on multiple passes of the LLM to help you categorize and shrink the corpus of data
—— 嘉宾 · [66:35]
我真的认为，一个公司运转的速度可以根据信息在公司内部流动的速度来判断。
指向原始笔记的链接
I do really think the velocity that a company moves at can be judged based on how quickly information is moving within the company.
—— Jess Hertz · [08:59]
我们季度营收几乎 90% 来自已经在平台上超过一年的商家。
指向原始笔记的链接
Almost 90% of our quarterly revenue is from merchants who’ve been on the platform for over a year.
—— Jess Hertz · [10:28]
归根结底，订阅模式将会发生改变，从人在环中变成人在环上，或人不在环中。
指向原始笔记的链接
So bottom line, the subscription model is going to change with humans in the loop to becoming humans on the loop or out of the loop.
—— Anil Nadiminti · [05:52]
在 Base 上的平均结算时间是 200 毫秒，每笔交易的成本约为十分之一美分。
指向原始笔记的链接
The average settlement time is 200 milliseconds on base with about a tenth of a cent as cost per transaction.
—— Anil Nadiminti · [19:13]
到了 2026 年，并不是机器或智能体发射核弹，而是它们拿走了你的钱包疯狂购物，买了一大堆加密货币，还给你买了一堆 Spanx 塑身衣。
指向原始笔记的链接
And in 2026, it’s not that the machines are or the agents are launching nukes, but rather they’ve taken your wallet and they’ve gone on a shopping spree and they buy like a bunch of crypto and new bunch of Spanx for you.
—— Jay Mok · [00:33]
在不理解纹理和上下文的情况下创建技术红线的想法，不是我们认为可以强加给客户的界限。
指向原始笔记的链接
creating a technological red line without understanding the texture and context is not a boundary line that we think that we can assert on top of our customer.
—— Nick Noone · [40:02]
我所有最好的 AI 点子，都是从 Every 团队上的某个人那里抄袭来的。
指向原始笔记的链接
I always like all of my best AI ideas, I’ve ripped off from someone else on the every team.
—— 嘉宾 · [34:59]
具体性基本上是所有质量的标尺。比如如果你能给智能体一个非常具体的任务，让它对要做什么有清晰的指示，那就会比一个泛泛的『这样去找代码坏味道』要好。
指向原始笔记的链接
Specificity basically is the ruler of all quality. Like if you can have a very specific task that the agent has clear instructions on what to do, it’s gonna be better than a generic here’s how to look for code smells.
—— Drew · [42:17]
可以这么说，你会不断地做循环，直到你有太多的循环，以至于你的大部分工作都在维护循环、以及构建用来观察它们的循环。
指向原始笔记的链接
It’s sort of, you’re going to keep working on loops until you have so many loops that most of your work is in maintaining loops and building loops to observe them.
—— Drew · [44:52]
我认为 AI 在总体上有一个特点，就是那些可以自我检验的问题，AI 可以持续钻研。
指向原始笔记的链接
One of the things that I think AI has in general is that problems that are self-checkable, AIs can grind on.
—— Brian McClendon · [26:40]
你也不应该盲目地同时开三个让它们互相争斗，因为它们很容易让自己全都相信同一个幻觉。
指向原始笔记的链接
And you shouldn’t just blindly turn on three of them fighting each other because they can easily all convince themselves of the same hallucination.
—— Brian McClendon · [37:35]
这就像有一个实习生时不时拍拍我的肩膀
指向原始笔记的链接
It’s just like having an intern who taps me on the shoulder every now and then
—— 嘉宾 · [23:39]
所以在那场对话中我们是平等的地位。一切都关于那场对话本身，你为它带来了什么，为它增添价值，并且尊重对方这个人，不是因为他头衔，而是因为他的想法。
指向原始笔记的链接
So we’re on equal footing in that conversation. It’s all about that conversation, what you’re bringing to it, adding value to it, and having respect for the other individual, not because of their title, but because of their ideas.
—— Amandeep Khurana · [25:13]
实实在在的零上，永远不会有星号。
指向原始笔记的链接
There will never be an asterisk on a firm zero.
—— Max Levchin · [42:13]
我认为那是胡扯，因为无论是能源、液冷，还是所有能让数据中心变得更好的东西，这些都是可以在这里创造、在这里完成的工作岗位。
指向原始笔记的链接
I call BS on that because if you think about whether it’s around energy, liquid cooling, all of the things that make the data center better, those are all jobs that can be created and done here.
—— Rene Haas · [28:38]
那些基准测试不代表你的工作负载。正如我们之前说的，AI 的边界是参差不齐的——某个模型可能在基准测试上做得很好，但这并不必然意味着它对你的工作负载也会做得很好。
指向原始笔记的链接
They don’t represent your workloads. And as we said earlier, the boundary of AI is jagged, right? So although a certain model might do very well on a benchmark, it doesn’t necessarily translate into that it will do very well for your workloads as well.
—— Chetan Gupta · [38:19]
你总是要押注于不是今天很酷的东西。
指向原始笔记的链接
You always want to be betting on not what’s cool today.
—— Tony Kim · [36:05]
我想，如果要我押注一个能带来即时 ROI、并且我们在开源的 Claude 类世界和小公司里都会看到的方向，那就是由 AI 自动创建的 wiki，也就是保持更新的信息库。
指向原始笔记的链接
I think if I were to bet on one that has immediate ROI that we’re going to all see in both open source claw-ish worlds and in small companies, it’s going to be the wiki that’s automatically created by AI, the information base that’s kept up to date.
—— Jean-Denis Greze · [15:32]
但它同时也是一场灾难，那种最终会挂上 Twitter、顶着我的名字的灾难。
指向原始笔记的链接
It was also a disaster, the kind that ends up on Twitter with my name on top of it.
—— Karan Vaidya · [10:40]
我要讲的是：如果你去构建一个公司大脑，它很可能会泄露公司机密——这本来就是我们构建公司大脑时最大的担忧，也就是实习生入职后突然拿到所有人薪酬明细的那种情况。
指向原始笔记的链接
I’m going to talk about the fact that if you go ahead and build a company brain, it will likely leak company secrets, which is kind of the big fear that we have about building a company brain anyway, which is this case of intern joins the company and then suddenly gets comp details on everybody kind of situation.
—— Tanmai Gopal · [00:18]
在你开始做自己的事之前,先把文件办妥,因为那实际上可能构成违反受托责任,可能违反你的保密条款。在你开始新事业之前,先有一个干净的分离。
指向原始笔记的链接
Get that papered before you start go doing your own thing, because that could actually be breach of fiduciary duty. It could be breach of your confidentiality provisions. Just have the clean separation before you start on the new venture.
—— Becki DeGraw · [15:16]
但一切都有瓶颈,我认为持续扩展这个东西的主要瓶颈实际上是训练算力。
指向原始笔记的链接
But everything has a bottleneck, and I think the main bottleneck on continuing to scale this thing is actually training compute.
—— Justin Johnson · [23:04]
一旦你骑上电动自行车,你就不会踩踏板了。
指向原始笔记的链接
Once you get on the e-bike, you’re not pedaling.
—— Tyler Folkman · [30:03]
你把你的用户群和客户群，把你的客户，当作互联网上的一个模糊团块，这是一种非常危险的看待客户的方式，因为你想见到他们，弄清他们的需求。
指向原始笔记的链接
you consider your user base and customer base, like your customer, just a blob on the internet, which is a really risky way to think about customers because you want to meet them, figure out their needs.
—— Jeffrey Morgan · [49:26]
我绝不会因为想拥有一家公司或想当创始人而创办公司。
指向原始笔记的链接
I would never start a company based on the fact that I wanted to have a company or be a founder.
—— Anders Forslund · [14:17]
我几乎是深切地站在这一边：很难论证 AI 模型可以在宽泛的公共互联网上训练，但另一个 AI 模型却不能用 AI 模型的输出训练。
指向原始笔记的链接
I’m almost deeply on the side of, I think it’s very hard to make the argument that AI models should be trained on broadly the public internet, but another AI model can’t be trained on the outputs of an AI model.
—— Aaron Levie · [03:48]
有那么几个人，处在那样的人生阶段，可以直接说：好，我要砸 100 亿美元在美国建一个开放实验室。
指向原始笔记的链接
You get a few people that are at a stage in life where they can just be like, yep, I’m going to drop $10 billion on building an open lab in America.
—— Aaron Levie · [06:17]
如果你封锁中国，中国是不会放弃 AI 的。
指向原始笔记的链接
If you block off China, China is like not going to give up on AI.
—— Aaron Levie · [08:14]
在美国开源这一边，我总体上认为 AI 的赚钱机器是推理。
指向原始笔记的链接
on the open source US side, I generally think that the moneymaker in AI is inference.
—— Aaron Levie · [12:08]
你低头在模型火车上埋头干十年，然后抬起头，我的天。
指向原始笔记的链接
You put your head down and work on the model train set for 10 years, and then you lift your head up and like, holy shit.
—— Paul Graham · [06:40]
但当你站在终点线上时，你发现它是有宽度的。
指向原始笔记的链接
But when you’re standing on the finish line, you realize it has width.
—— Paul Graham · [15:32]
顺便说一句，我看多的一件事是 OpenAI 上的消费者广告市场，这显然至今还没有被充分开发利用，但我认为它仍然是 AI 领域最大的机会之一。
指向原始笔记的链接
By the way, one thing I’m long is the consumer ads marketplace on OpenAI, which has obviously not been fully utilized to this day, but I think remains one of the biggest opportunities in AI.
—— Anastasios Angelopoulos · [40:27]
胜率接近 100%。而且顺便说一句，要让人们在任何事情上达成一致几乎是不可能的。
指向原始笔记的链接
It was nearly a 100% win rate. And by the way, it’s almost impossible to get people to agree on anything.
—— Anastasios Angelopoulos · [38:57]
而我说，把你 DevEx 团队的规模翻一倍。把你数据团队的规模翻一倍。投入平台建设。对人类有好处，对智能体也有好处，而这就是能让你跑起来的东西。
指向原始笔记的链接
And I say, double the size of your DevEx team. Double the size of your data team. Like, work on platform investments. Good for humans. Good for agents. And that’s what will let you run.
—— Claire Vaux · [18:04]
它是那种即使业务停滞、不进任何新单也仍然会增长的生意——因为总有人在往里存钱。
指向原始笔记的链接
It’s one of those businesses that if the business stayed stagnant, we brought on no new business. It still grows because you still have people getting paid.
—— Aaron Schumm · [35:08]
我们是摩根大通，世界最大的银行。我们怎么能押注一家四十人的初创公司，来承载我们认为关乎未来重要业务的 stuff?
指向原始笔记的链接
We’re JP Morgan, we’re the largest bank in the world. How can we bet on this forty person startup to carry what we think is going to be the future of a lot of our business
—— Aaron Schumm · [39:38]
我有我的家庭，我有 Vestwell,就是这样。我只专注于此，否则其余的东西会把你拽向与你的生活毫不相干的方向。
指向原始笔记的链接
I have my family and I have Vestwell and that is it, right? And I just focus on that because otherwise the rest of it will start to pull you in directions that are just not attitude to your life
—— Aaron Schumm · [48:51]
我们认为在不久的将来，你会看到人们日常推理调用中的很大一部分，甚至可能是大多数，会走向本地设备和本地笔记本电脑或本地工作站，而不是像今天这种一切都被推上云端的标准做法。
指向原始笔记的链接
We think in the very near future, you’re going to see a huge proportion, maybe even a majority, of people’s daily inference calls going to local devices and on-prem laptops or on-prem workstations, as opposed to the kind of standard of today where everything’s being pushed up to the cloud.
—— 嘉宾 · [45:06]
我认为 X 上使用这些产品的人，其实并不代表产品市场契合。
指向原始笔记的链接
I think a person on X that uses these products is not actually product market fit.
—— JD · [37:16]
我们在编码方面有意走得慢一些，因为我们已经意识到编码在多大程度上是一场预算游戏——你有多少数据预算？
指向原始笔记的链接
We are intentionally moving a little bit more slowly on the coding side because we’ve kind of realized like how much coding is like a budget game, like how much data budget do you have?
—— Alexander Whedon · [27:49]
我不认为真的有人会把自己的职业生涯押在，比如说，用 vibe coding 写一个核反应堆或火箭引擎或某种后果非常严重的系统的控制系统上。
指向原始笔记的链接
I don’t think anyone’s really going to kind of bet their career on like vibe coding a control system for let’s say a nuclear reactor or a rocket engine or something that’s very high consequence.
—— Scott Morton · [02:47]
SpaceX 成功的最大关键之一就是 Elon 是那种会对他的团队下巨大赌注的人，相信我们能做到。
指向原始笔记的链接
One of the biggest keys to SpaceX’s success is just that Elon is one to make massive bets on his team, that we can do this.
—— Scott Morton · [15:37]
让我对开发 Grok Bot 感到如此兴奋的是，这是我第一次在非编码任务上感觉自己可以真正把工作委托给 AI，不用再去想它，回来的时候事情已经完成了。
指向原始笔记的链接
What made me so excited to work on Grok Bot is it was the first time for non-coding tasks that I felt like I could truly delegate work to AI, not have to think about it, and I would come back and it’s done.
—— Roman Ugarte · [01:00]
我认为我们现在正处于一个非常奇怪的时刻，将来回头看我们会说：我很惊讶很多人就是这样与 AI 一起工作的——你在给这些超级智能的新同事、这些 AI 机器人做入职培训，却让它们和你共用同一台电脑。
指向原始笔记的链接
I think we’re in a really weird moment right now that I think we’re going to look back on and be like, I’m surprised that this is the way that a lot of people worked with AI, where you’re onboarding these super intelligent new colleagues, these AI bots, and you’re asking them to share the same computer that you have.
—— Roman Ugarte · [32:21]
你仍然在做那件事，感觉也是那样，它同样压在你身上——而不是真正地给同事来一记不看人的传球，然后说：「你可以搞定，这是背景信息，放手去干吧。」
指向原始笔记的链接
You’re still doing the thing and it feels that way and it’s weighing on you in the same way versus truly throwing a no look pass to a colleague and being like, “You got this, here’s the context, go off and run.”
—— Roman Ugarte · [62:59]
而且我认为值得注意的是，这些竞争对手没有一个现在站在 AI 编程的前沿，这在很大程度上不是因为它们做出了任何错误的决定或缺乏资源，而是因为在文化上无法快速行动、无法在当下变化时改变以跟上当下。
指向原始笔记的链接
And I think it’s notable that none of those competitors are at the forefront of AI coding right now in large part, not because of any incorrect decisions that they made or any lack of resources on their part, but this cultural inability to move quickly and to change to meet the moment as the moment’s changing.
—— Roman Ugarte · [69:38]
你所有的创新代币都花在开源项目上了，而不是花在交付上。
指向原始笔记的链接
All your innovation tokens are being spent on that versus on the delivery.
—— Jordan Tigani · [11:04]
是 BI,不过是打了兴奋剂的 BI,因为有些事情你能做而别人做不了。
指向原始笔记的链接
It’s BI, but it’s BI on steroids because there’s some things you can do that you can’t do.
—— Jordan Tigani · [18:15]
你所有的创新代币都花在构建开源项目上，而不是花在交付上。
指向原始笔记的链接
All your innovation tokens are being spent on that versus on the delivery.
—— Jordan Tigani · [10:56]
归根结底，如果你是一家数据库公司或基础设施公司，你的生死基本上取决于核心技术的强弱。
指向原始笔记的链接
If you’re a database company or an infrastructure company, you kind of live or die based on the strength of the core technology.
—— Jordan Tigani · [31:58]
但出人意料的是,无论你是在 Codex 还是 ChatGPT 上发布,流程都很相似——尽管 ChatGPT 推向十亿活跃用户而且还在增长,你可以提交一个 PR、做一个改动,第二天甚至同一天发布,它就推向十亿用户,而且没问题。
指向原始笔记的链接
But it’s surprisingly a similar process, whether you ship on Codex or ChatGPT, even though ChatGPT goes out to a billion active users and growing, you can ship a PR, you can make a change and get it shipped the next day or even the same day, and it just goes out to a billion users and it’s fine.
—— Tibo Sottiaux · [43:33]
我们走的是一条非常 Anthropic 式的路线：真正把 B2B 使用场景中的某些用户类型和工作流做到极致好，而不是在消费端把数亿美元到处乱撒。
指向原始笔记的链接
We’re taking a very Anthropic-type approach: really trying to make certain types of users and workflows within B2B usage insanely great, and we’re not just spraying hundreds of millions of dollars at everything on the consumer side.
—— Eric Simons · [15:05]
对我们来说，我们想在定价上变得非常激进。这里正在发生的巨大变化是，开源模型真的在追赶那些实验室。
指向原始笔记的链接
For us, we want to get really aggressive on pricing. The big sea change happening here is that open-source models are really catching up with the labs.
—— Eric Simons · [24:52]
比如 Codex 团队有自定义方法，Goose 团队有自定义方法，Klein 团队也有，无论谁，我们都能看到生态里涌现出什么，哪些值得走上标准化轨道纳入协议本身——让这个东西由使用方式塑造、由社区塑造。
指向原始笔记的链接
Like if the Codex team has some custom methods, the Goose team has some custom methods, the Klein team has some custom methods, whoever, we can see what emerges in the ecosystem and what makes sense to get on a standards track and bring into the protocol itself so that this is sort of shaped by usage and shaped by the community.
—— Alex Hancock · [05:47]
你要在关注 UI 之前先关注数据
指向原始笔记的链接
you want to focus on the data before you focus on the UI
—— Dustin Mihalik · [13:39]
所以那大概算是比较不讲道理的事情之一，比如，对，就把这整个 Python 代码库移植到 TypeScript，让它跑起来、能部署，就在一个周末之内。
指向原始笔记的链接
So that probably ranks on like the more unreasonable things like, yeah, just port this entire Python code base to TypeScript, get it working, get it deployable in, you know, a weekend.
—— 嘉宾 · [00:41]
AI 生成的代码,应该让人感觉是那种已经在你团队待了多年的人写出来的。
指向原始笔记的链接
AI-generated code should feel like it was written by someone who’s been on your team for years.
—— Brandon Waselnuk · [01:13]
token 减少 50%,分诊更快,而且答案质量实际上更好,因为它知道业务内部正在发生什么。
指向原始笔记的链接
50% fewer tokens, faster triage, and the answer quality is actually better because it knew what was going on inside of the business.
—— Brandon Waselnuk · [09:51]
所有这些建立在对数据和学习直觉之上的分析工作,造就了我们如今拥有的模型——相比 2017 年的 Transformer,它就像一艘忒修斯之船。
指向原始笔记的链接
All of this analysis work built on intuitions about data and learning led to the model that we have now, which is like a ship of Theseus compared to the 2017 Transformer.
—— Chris Potts · [15:29]
所以当人们说「哦,我们不理解」时,我说:我认为你们理解的程度远超你们所承认的。
指向原始笔记的链接
And so when people say, oh, we don’t understand, I say, I think you understand much better than you’re letting on.
—— Chris Potts · [16:53]
我的意思是,毕竟,所有这些东西都是在深度中没有递归这一事实之上的一种打补丁。
指向原始笔记的链接
I mean, after all, so all of this stuff is a kind of patch job on the fact that there’s no recursion in the depth.
—— Chris Potts · [34:52]
绝大多数人类疾病应该被解决、应该可以被治疗——我认为我们正在这条路上。
指向原始笔记的链接
And I think we’re on that path.
—— Zavain Dar · [33:57]
我觉得最好尝试去做一个建设者，你会摔个大跟头，你会体会到有时候做出贡献有多难。
指向原始笔记的链接
I think it’s better to try to be a builder and you’re gonna fall flat on your face and you’ll get an appreciation for how hard it is to contribute sometimes.
—— Brian Armstrong · [35:55]
如果研究实验室不拥有最好的产品，那才真是疯了，因为他们拿到了所有关于人们如何使用他们产品的数据。
指向原始笔记的链接
It would be really wild if the research labs didn’t have the best product, because they’re getting all of this data on how folks are using their products.
—— Demetrios Brinkmann · [38:34]
那将花费 1.2 万亿美元——前提是我们国家首先还执行得了那个计划。
指向原始笔记的链接
That’ll cost $1.2 trillion if we can even execute as a country in the first place on that plan.
—— 嘉宾 · [04:05]
没有理由一个连轮胎都不会换的人不能去造一艘 Corsair、参与 Marauder 或更大的船的工作。
指向原始笔记的链接
There’s no reason somebody who can’t change a tire shouldn’t be able to go and build a Corsair or work on a Marauder or work on a larger ship.
—— 嘉宾 · [16:28]
你不能让一个与任务无关的人，依据一本没有任何灵活余地的手册来任意做决定。
指向原始笔记的链接
You can’t have one human who’s not tied to the mission making arbitrary decisions based on a guidebook that doesn’t have any wiggle for them.
—— 嘉宾 · [31:41]
令人惊讶的是，互联网有多少是基于少数几个开源代码库——它们由优秀的工程师创造，并不是为了赚一大笔钱，而真的只是为了推动技术向前发展。
指向原始笔记的链接
It’s surprising how much of the internet is based off of a handful of open source repositories that great engineers have created and not for seeking a bunch of money on the other side, but really just for pushing the needle of technology forward.
—— 嘉宾 · [58:33]
顺便说一句，我觉得相当神奇的是，AI 已经发展到了这样的程度：地球上任何一个不是设计师、也不是工程师的人，都可以点一个按钮，突然就做出一整套 PowerPoint，或者做出一个网站，或者做出一个 web 应用。
指向原始笔记的链接
I think it’s quite magical, by the way, that AI has gotten to a point that any human on the planet that is not even a designer, that is not an engineer, can click a button and suddenly make an entire PowerPoint or make a website or make a web app.
—— Thais Castello Branco · [03:34]
我想说，这个门槛基本上还在地上。
指向原始笔记的链接
The bar is kind of really, I would say, on the ground.
—— Thais Castello Branco · [14:25]
我认为把自己限制在用户数上——说实话，我觉得按用户定价已经是过去式了。
指向原始笔记的链接
I think restricting yourself on users, I think user-based pricing is a thing of the past, honestly.
—— Ron Gabrisko · [72:27]
到目前为止，教一个孩子最好的方式是与一位优秀的、了解这个孩子的成年人进行一对一互动。
指向原始笔记的链接
By far, the best way to teach a child is one-on-one interaction with an adult who is a good teacher and knows the child.
—— 嘉宾 · [12:45]
现在，在全球范围内，想象一下类似的事情，只不过庄家是在为某个目的控制你的大脑。
指向原始笔记的链接
Now, on a global scale, imagine something similar where the house is controlling your brain for a purpose.
—— 嘉宾 · [20:06]
随着智能体变得更高效、更有效,我认为我们会更倾向于依靠技术控制作为门禁机制,而不是对它能做什么的主观人为评估——因为再说一次,由于非确定性,你真的很难回头去问一个智能体:你为什么把这些文件全删了?
指向原始笔记的链接
So as an agent becomes more efficient and effective, I think we’re going to lean more on technical controls as a mechanism to gate versus a subjective human assessment of what it can do, because again, the non-determinism, it’s really hard to go back to an agent and ask, why did you go delete all these files?
—— Robert Lucero · [18:36]
你交付的速度建立在你构建的这份基底之上，如果基底不好，你在上面构建的任何东西都不会好。
指向原始笔记的链接
The velocity that you ship lays down on this substrate that you’ve built, and if the substrate isn’t good, nothing you build on top is going to be good.
—— Jonathan Kelley · [08:53]
如果你的智能体代码所落地的底层基础不好，它们的贡献也会不好。
指向原始笔记的链接
If the substrate on which your agent’s code lands is bad, their contributions will be bad as well.
—— Jonathan Kelley · [16:15]
关于使用编码智能体构建有雄心的软件，我的结语是：代码现在很廉价，但质量不是。
指向原始笔记的链接
So my closing thoughts on using coding agents to build ambitious software is that code is now cheap, but quality is not.
—— Jonathan Kelley · [18:07]
软件工程师的工作从来都不只是把代码行放到屏幕上。
指向原始笔记的链接
The job of a software engineer has never really been about putting lines of code on the screen.
—— Jonathan Kelley · [18:17]
如果你把系统搭建好，它们自己向前推进，那就是彻底的改变。
指向原始笔记的链接
If you put the system in place and they move forward on their own, it’s a game changer.
—— 嘉宾 · [52:22]
我想很多 PM 会意识到自己其实并不那么擅长从零到一，而且在一个别人的好想法上工作比在自己的坏想法上要有成就感得多。
指向原始笔记的链接
I think a lot of PMs are gonna realize they’re actually not that good at zero to one and it’s much more fulfilling to work on someone else’s good idea than your own bad idea.
—— Anish Acharya · [17:15]
你知道什么比我亲自做更好吗？是让 Claude 来做。这就在团队每个人心中建立了一个习惯。
指向原始笔记的链接
You know what’s better than me doing it? It’s Claude doing it. And that just created this habit in everyone on the team.
—— 嘉宾 · [48:18]
我发现这类分手邮件的回复率要高得多，因为它让对方说「不」的压力非常低。
指向原始笔记的链接
Reply rates I find are much higher on these breakup emails because it makes them really low pressure for them to say no.
—— 嘉宾 · [10:10]
外呼上真正成功的创始人和不成功的创始人之间的头号区别，其实就是坚持执行他们承诺的活动。
指向原始笔记的链接
The number one difference between founders who actually succeed without bound and founders who don’t is actually following through on the activity that they commit to.
—— 嘉宾 · [11:12]
所以现在我们到了这样一个阶段：我们只点击那些表面上看有风险、不能轻信的东西。
指向原始笔记的链接
So now we’re at a point where we only click on what is risky to believe on its face.
—— Tim Sanders · [28:16]
我用 AI 是为了让我离开电脑,而不是待在电脑前。
指向原始笔记的链接
I use AI to get me off my computer, not on my computer.
—— Claire Vo · [00:00]
所以我们的看法是，做出决策的那部分，也就是定义功能、定义 API 表面积、定义抽象层，那些是非常需要人在环的，我亲自上手和 Claude 一起做。而写代码、执行的部分，才是被自动化的部分。
指向原始笔记的链接
So the way we see it is the part which is where the decisions are being made, right, which is defining the feature, defining the API surface area, the abstraction layer, those are very human in the loop, right, where I’m hands on with Claude. And the code writing part, the execution part, is actually the part which is automated.
—— 嘉宾 · [16:57]
我在这点上有个强烈的看法：你不可能拿着同一个产品把 NDR 从 20 提到 40。留存几乎是从核心产品继承来的。
指向原始笔记的链接
My strong view on this is that you can’t take the same product and take NDR from 20 to 40. I think you almost like inherit retention based off of the core product.
—— Keith Peiris · [12:54]
事后来看，我可能更高兴我们没有收购它，因为考虑到它对当时的 Bending Spoons 来说规模会那么大，那 Bending Spoons 就会变成 Grindr，因为我们很可能把全部精力都放在那上面。
指向原始笔记的链接
In the hindsight probably I’m happier that we didn’t acquire it because given the fact it would have been so big for Bending Spoons then Bending Spoons would have been Grindr because we would have probably fully focused on that
—— 嘉宾 · [10:45]
我们每年的流失率仍然低于 1%。
指向原始笔记的链接
We still have like less than 1% churn on a yearly basis.
—— 嘉宾 · [24:33]
所以你需要非常少但非常坚实的基础来构建东西，但之后你需要增加很多灵活性，确保没有过度约束的流程或审批链条之类的东西，因为根据我的经验，那些正是杀死一个产品的东西——无法对产品采取行动。
指向原始笔记的链接
So you need very few but very solid basis when you build something, but then you need to add a lot of flexibility and make sure that there are not, you know, not super constraining processes or approval lines and things like this, because in my experience, those are the things that kill a product, the inability to act on the product.
—— Vali · [71:47]
在欧洲的 AI Engineer 大会上，Cursor 做了一个很棒的演讲，讲他们如何用 200 行 agent 文件替换了大约 12000 行 TypeScript 代码来实现类似的东西。
指向原始笔记的链接
So at AI engineer in Europe, Cursor did a great talk on how they replaced roughly 12,000 lines of TypeScript code with 200 lines agent files to create something similar.
—— Philipp Schmid · [15:25]
掌管属于你的部分：专注于你的领域指令、工作流，尤其专注于 evals,定义干净的工具，验证结果，真正做到为删除而构建。
指向原始笔记的链接
Own what is yours, meaning focus on your domain instructions, focus on the workflows, focus especially on the evals, define clean tools, and verify the outcomes, and really build to delete.
—— Philipp Schmid · [17:16]
在控制台上点来点去很适合搞清楚发生了什么,但这不是你把东西部署到生产环境的方式。
指向原始笔记的链接
Clicking around on the console is great for being able to figure out what’s going on, but it’s not how you deploy things into production.
—— Mike Chambers · [06:30]
与其每张桌子上一台计算机， 我们能不能做到每张桌子上一个智能体？
指向原始笔记的链接
instead of a computer on every desk, could we potentially have an agent on every desk?
—— Andrew Qu · [02:05]
我们在一旁感叹： 哇， 和我们之前的东西相比， Claude Code 加 Opus 4.5 基本上就是 AGI 了。
指向原始笔记的链接
we on the side were like, wow, Claude Code and Opus 4.5 is basically AGI compared to what we had before.
—— Andrew Qu · [07:50]
它故意做得无聊，而在安全领域，无聊是一个特性。
指向原始笔记的链接
It’s boring on purpose, and in security, boring is a feature.
—— Sarah Sanders · [12:32]
虽然这很诱人，因为看起来更简单，但我不能把我的安全模型押在掷硬币上，就因为我的模型今天状态不好、行为和昨天不一样。
指向原始笔记的链接
And while that’s tempting because it seems easier, I can’t bet my security model on a coin flip because my model’s having a bad day or something happened and it’s acting different today than it did yesterday.
—— Sarah Sanders · [15:14]
如果命中一条规则，闸门锁上，会话结束，这条路径上没有任何模型参与。
指向原始笔记的链接
If a rule matches, the gate locks, the session ends, and there is no model anywhere on that path.
—— Sarah Sanders · [15:44]
执行是你押上身家性命的部分，所以它必须是确定性的；而判断是增加细微差别的部分，那才是你唯一能放进概率性东西的地方。
指向原始笔记的链接
Enforcement is the part that you bet the house on, so it has to be deterministic, but judgment is the part that adds nuance, so that’s really the only place that you can put anything probabilistic in there.
—— Sarah Sanders · [16:15]
所以我们真正追求的是资本的速度，而不仅仅是资本的回报。
指向原始笔记的链接
So what we’re really after is the velocity of capital, not just returns on capital.
—— David Morehead · [16:52]
区别在于，有几千万人基于那些信息交易，而在私募侧，大概只有三个人。
指向原始笔记的链接
The difference is that there are tens of millions of people trading on that information, whereas on the private side, there’s like three.
—— David Morehead · [27:04]
永远不要把你的模式押在一个对你的业务没有任何既得利益的第三方的慷慨之上。因为也许他们今天是朋友，明天可能就不是朋友了。
指向原始笔记的链接
Don’t ever bet your model on the generosity of a third party that has no vested stake in your business. Because maybe they’re friends today, they might not be friends tomorrow.
—— Jim VandeHei · [26:50]
你必须在你和你的客户之间建立直接的关系，而且这种关系必须自身就能盈利，你才能拥有一个可行、可扩展、持久的产品。
指向原始笔记的链接
You got to build a direct relationship between you and your customer. And that one has to be profitable on its own for you to have a viable, scalable, durable product.
—— Jim VandeHei · [27:00]
用户其实不知道发生了什么。对他们来说，这看起来就是非常快的响应时间。
指向原始笔记的链接
The user doesn’t really know what’s going on. It just looks like really fast response times to them.
—— Bohan Li · [08:25]
早期,我们的延迟从两秒漂移到了大约两秒半,而那半秒钟几乎让产品里的所有指标都崩了。
指向原始笔记的链接
Early on, our latency drifted from two seconds to about two and a half seconds, and that half second tanked basically every metric in the product.
—— Paula Dozsa · [03:39]
我们的主要原则是,我们基于利害关系而非成本来路由。
指向原始笔记的链接
And our main principle is that we route based on stakes, not on cost.
—— Paula Dozsa · [06:08]
举例来说,我们有危机类或治疗师风格的语气,我们在这些方面从不省钱。
指向原始笔记的链接
For example, we have crisis or therapist-style tones, and we never cheap out on those.
—— Paula Dozsa · [06:24]
我们做了很多 A-B 实验,在那里我们发现的令人惊讶的结果是,把三分之一的话轮路由到小模型,对留存率几乎没有可测量的影响。
指向原始笔记的链接
And we do a bunch of A-B experiments, and the surprising result we found there is that routing a third of our turns to the small model has almost no measurable effect on retention.
—— Paula Dozsa · [06:57]
主要是因为没有现实世界的参照物可以锚定,这意味着用户可以把自我投射上去,它会变成用户所需要的样子。
指向原始笔记的链接
Mostly because there’s no real-world reference to anchor on, which means that users can project onto it, and it becomes what they need.
—— Paula Dozsa · [09:06]
所以当界面是你的声音、而另一端的东西记得你并有自己的人格时,它就不再只是软件,而是变成了一段真实的关系,这也是为什么负责任地构建它、把它构建好,是值得为之痴迷的事。
指向原始笔记的链接
So when the interface is your voice and the thing on the other side remembers you and has a personality, it stops being just software and starts being an actual relationship, which is why building it responsibly and building it well is worth obsessing over.
—— Paula Dozsa · [12:26]
这就引出了一个问题：如果模型已经好到能做人类在电脑上做的一切事情，那我为什么还要和一个应用对话？
指向原始笔记的链接
It raises the question of, look, if the models are just getting good enough to do everything on a computer that a human can do, why am I talking to an app?
—— Charlie Guo · [06:07]
然后编码智能体和语音智能体来了，现在他们在给定的一天或一个月内生成的代码比以前多了几个数量级。
指向原始笔记的链接
Then came coding agents and voice agents, and now they generate orders of magnitude more code than they previously did on a given day or month.
—— Charlie Guo · [09:30]
你要么领先一年，要么落后一年，仅仅取决于你的信息流。
指向原始笔记的链接
You’re either a year ahead or a year behind simply based on your feed.
—— Aaron Levie · [56:19]
人们思考应用的方式已经非常不同了，尤其是随着现在编码智能体加入进来，地球上突然每个人都是软件开发者了。
指向原始笔记的链接
The way people are thinking about applications is very different, like especially with the rise of now coding agents coming in the mix, everyone suddenly is a software developer on the planet.
—— Samar Abbas · [01:47]
我们将迎来范式转变，从「我就在我笔记本上做点东西玩玩」到「嘿，这会如何影响整个组织？」
指向原始笔记的链接
We’re going to have the paradigm shift from, I’m just going to do things on my laptop and have fun, to, hey, how does this affect the entire organization?
—— Jason · [18:31]
我在周末让 GPT-6 在 Lightfield 上跑了几小时之后，改变了我的销售流程。
指向原始笔记的链接
I changed my sales process after letting GPT-6 rip on Lightfield for a couple of hours on the weekend.
—— Keith Parris · [35:14]
所以我只把对硅谷的专注看作不是高效的收入获取，而实际上就是高效的营销获取。
指向原始笔记的链接
So I just see the focus on Silicon Valley not as efficient revenue capture, but actually just like efficient marketing capture.
—— Keith Parris · [44:43]
而且你需要像偏执狂一样专注于你的客户。
指向原始笔记的链接
And you need to be like maniacally focused on your customers.
—— Keith Parris · [50:54]
而且很多团队早期会试图直接跳到最后，说「哦，我要做参数化记忆之类的」——而其实「Clippy 满意并通过所有测试之前不许发布代码」就是一件非常有效、开销非常小的事情。
指向原始笔记的链接
Um and I think a lot of teams early on try and jump to the end and say, oh, I’m gonna do parametric memory or something, when you know you can’t ship code until Clippy is happy with it and it passes all the tests, um, is a very effective thing with very little uh very little um overhead.
—— Marc Brooker · [47:53]
当围绕资源分配、围绕优先级排序、围绕我们把时间花在更多产品工作还是更多技术工作上没有任何辩论时，这会在我的脑子里亮起一盏灯，说明有什么地方不对劲。
指向原始笔记的链接
When there’s no debate around resourcing, around prioritization, whether we spend time on more product work or more tech work, that lights a bulb in my head that something is wrong.
—— Ben Celebicic · [05:36]
有些人在实际工作中已经开始使用这些工具，并开始说：「看，我在 iOS 上构建这个功能。为什么我不直接让工具帮我在 Android 上也构建它呢？」突然之间，我一次完成了两个人的工作。
指向原始笔记的链接
Some people on the ground are picking up these tools and already starting to say, “Look, I’m building this feature on iOS. Why don’t I just tell the tool to build it for me on Android?” And all of a sudden, I’ve done the work of two people at once.
—— Ben Celebicic · [21:09]
他们真的每次都是刚醒来，而且每次你开启一个新会话，他们就像上班第一天的应届新员工。
指向原始笔记的链接
They really start to wake up, and every time you start your session, they’re like a new hire on their first day on the job.
—— Edo Liberty · [04:36]
但有意思的是,它的准确性也差得多,因为再说一次,他们没有你公司里正确的上下文、正确的目标等等。
指向原始笔记的链接
But interestingly enough, it’s also a hell of a lot less accurate because, again, they don’t have the right context, the right objectives and so on in your company.
—— Edo Liberty · [18:10]
我们把其中一半花在了一个 GPU 集群上。在当时那简直是疯了。
指向原始笔记的链接
We spent half of it on a GPU cluster. That was crazy at the time.
—— Will Bryk · [07:43]
我们只是想做一个更好的 Google 搜索，以后再想办法怎么赚钱。
指向原始笔记的链接
We were just trying to build a better search on Google. We’ll figure out how to make money later.
—— Will Bryk · [08:42]
但我认为总体而言，大多数人都远远高估了需要这样做的场景数量，而且绝对低估了当人类和智能体都能访问公司正在做的一切事情的完整上下文时所产生的力量。
指向原始笔记的链接
But I think by and large, most people far overestimate the number of things where that is required and definitely underestimate the power of what happens when both humans and agents have access to the full context of what a company is working on.
—— Wade Foster · [52:05]
所以如果你还把 Zapier 想成那种老派的、无代码的、在屏幕上点点框框的东西，我想你会大吃一惊的。
指向原始笔记的链接
So if you still think of Zapier as the old school, no code, clicky click boxes on a screen thing, I think you’re in for a treat.
—— Wade Foster · [63:10]
基本上如果你让四个智能体一起处理这个问题,速度会快一倍。因为是四个智能体工作一半的时间,你付出了 2 倍的成本来换得快一倍的答案。
指向原始笔记的链接
basically if you have four agents working on the problem, it is done twice as fast. So you’re basically paying, because there’s four agents working for half as long, you’re paying a 2x more to get an answer twice as quickly.
—— Noam Brown · [03:52]
但每次你基于对思维链的观察进行干预,你都在隐式地施加一点压力,让模型之后隐藏它的思维链。
指向原始笔记的链接
But every time you intervene based on your observations of the chain of thought, you are implicitly applying a tiny bit of pressure for the model to then hide its chain of thought.
—— Noam Brown · [70:17]
我认为安全机制是在争取时间,思维链监控之类的东西也是在争取时间,它们能告诉我们是否走在正确的道路上。但归根结底,我们确实需要解决对齐问题。
指向原始笔记的链接
I think that the safety mechanisms bias time, and things like chain of thought monitoring bias time, and they can tell us if we’re on the right path. But at the end of the day, we really do need to solve the alignment problem.
—— Noam Brown · [74:01]
如果你要列一个从不害怕的勇敢领导者名单，别把我列进去。我每天都在害怕。
指向原始笔记的链接
If you’re going to make a list of brave leaders who are never afraid, don’t put me on the list. I’m afraid every day.
—— Brené Brown · [04:43]
我需要每个人靠自己站起来，因为我无法既纠正正在发生的事情，又提醒你自己的价值。
指向原始笔记的链接
I need everybody to get back up on their own because I cannot both correct what’s happening and remind you of your value.
—— Brené Brown · [16:18]
对 AI 的投资，90% 失败，没有投资回报。而其中很大一部分原因就是自我参照系统。
指向原始笔记的链接
Investments in AI, 90% failure, no return on investment. And a lot of that is because of self-referencing systems.
—— Brené Brown · [18:22]
对我来说，Eve 背后很早期的一条原则就是：我们要押注模型会变得更聪明。
指向原始笔记的链接
Yeah, for me, I mean, one of the very early principles behind Eve was we’re going to bet on the models getting smarter.
—— Andrew Barba · [48:00]
在 Databricks,我跳过并且拒绝过一些极其出色的人,他们后来都取得了了不起的成就。
指向原始笔记的链接
So at Databricks, I’ve skipped and said no to phenomenal people who’ve gone on and done great things.
—— Ali Ghodsi · [25:05]
我就说,是啊,但我的策略就是宁可极其挑剔,花相当长的时间,尽早启动。
指向原始笔记的链接
I was like, yeah, but my playbook was err on the side of being just extremely picky, take a decent amount of time, start early.
—— Ali Ghodsi · [25:16]
我对此的理论是，很多投资人被他们 2022 到 2025 年的 AI 投资狠狠坑惨了。
指向原始笔记的链接
I think my theory on that is that a lot of investors have just gotten wrecked by their AI investments from 2022 to 2025.
—— Jeff Morris Jr · [12:10]
实际上有一篇学术研究论文谈到这个：做过一次单一的转型，比不转型或转型很多次更能预测一家初创公司的成功。
指向原始笔记的链接
There’s actually like an academic research paper on this that talks about having one single pivot is actually a greater predictor of success for a startup than no pivots or a bunch of pivots.
—— 嘉宾 · [34:55]
但我想说，对于 99% 的新基金来说，你不应该试图在种子轮交易上和红杉或 Benchmark 竞争。
指向原始笔记的链接
But I’d say for 99% of new funds, like you should not try and compete with the Sequoia or benchmark on a seed deal.
—— Jeff Morris Jr · [90:34]
我们只要按一个按钮，就能得到十万个他们。
指向原始笔记的链接
We just turn on a button and we can get 100,000 of them.
—— 嘉宾 · [28:51]
在 Inception，我们押注 diffusion 模型，因为我们相信最终重要的是推理时扩展。
指向原始笔记的链接
At inception, we’re betting on diffusion models because we believe that what matters eventually will be inference time scaling.
—— Stefano Ermon · [08:01]
看 benchmark 的话，我们的 Mercury 模型与 Haiku 模型、Flash 模型、OpenAI 的 Mini Nano 模型相当，同时速度快得多。
指向原始笔记的链接
So our Mercury models are on par with the Haiku models, Flash models, Mini Nano models from OpenAI, if you look at benchmarks, while being significantly faster.
—— Stefano Ermon · [12:33]
你没法在 VLLM 或 SGLang 上运行这些基于扩散的 LLM。我们必须构建自己的服务引擎，而且我们能处理真实生产工作负载的很多复杂性。
指向原始笔记的链接
You cannot run these diffusion-based LLMs on VLLM or SGLang. We had to build our own serving engine, and we can handle a lot of the complexity of real production workloads.
—— Stefano Ermon · [13:01]
但我只是认为，如果你从根本上相信我们能够用上我们拥有的全部电力，而且会有经济上的理由这么做，那么对效率的关注将会急剧上升。
指向原始笔记的链接
But I just think if you fundamentally believe that we can use all of the power we have and there will be economic reasons to do so, then I think the focus on efficiency is going to go dramatically up.
—— 嘉宾 · [16:21]
据报道,大型实验室在这个上面大约花十亿美元。这不是一个广为人知的事实,但其实围绕这个可以建立起一门真正的生意。
指向原始笔记的链接
Reportedly, the big labs are spending about a billion dollars on this. It’s not a very known fact, but there’s actually a real business to be built around this.
—— Diana · [24:58]
最近被广泛引用的失控人工智能案例，仔细审视之下，其实讲述了一个不那么离奇的故事。
指向原始笔记的链接
The most widely cited recent example of rogue artificial intelligence turns out on close examination to tell a less exotic story.
—— 嘉宾 · [03:03]
还是听起来像是公司打电话给《纽约时报》说：我们可以在背景条件下给你这个，你能把数据发布出来吗？
指向原始笔记的链接
Or does that sound like the company calling the New York Times and saying, we’ll give you this on background, can you publish the data?
—— 嘉宾 · [54:21]
AI 正走在完全相同的轨迹上,只是被压缩到了最近几年而不是十年。
指向原始笔记的链接
AI is on the exact same trajectory, but just compressed into the last few years instead of a decade.
—— Nishant Gupta · [01:24]
这也解释了为什么据报道 Anthropic 目前在 API 业务上有 80 个点的毛利润率
指向原始笔记的链接
there’s a reason why Anthropic is reported to have 80 points of gross margin right now on API business.
—— Thomas Sohmers · [12:03]
它现在已经成为左右两派几乎团结一致的议题——反数据中心。我认为这几乎完全是一场中国的心理战
指向原始笔记的链接
it’s now become a almost unifying issue on left and right about being anti-data centers. I think that is almost entirely a Chinese psyop.
—— Thomas Sohmers · [21:40]
只要在服务上有 50 个用户，仅仅是这些单独的用户会话，就会超过你试图存储的模型权重
指向原始笔记的链接
with just 50 users on your service, the user context, just those individual sessions end up being greater than the model weights that you’re trying to store.
—— Thomas Sohmers · [40:27]
gated DeltaNet 及其衍生版本，通过这种机制，你在注意力部分花费的总时间可以减少大约 75%
指向原始笔记的链接
gated DeltaNet and its derivative versions where you can have like a 75% decrease in the total time you’re spending on the attention portion with this mechanism.
—— Thomas Sohmers · [57:38]
如果一个博物馆把他们所有的藏品都挂在了墙上，那这个策展人就没有做成任何事。
指向原始笔记的链接
If a museum has all of their collection on the walls, then the curator hasn’t done anything.
—— Peter Sellis · [01:28]
所以，现在正在发生的很多事情，就是我们正处于发明一套方式来谈论「在一个随机统计系统里什么才算 bug」的阵痛之中。
指向原始笔记的链接
And so, so much of what’s going on now is we’re in the throes of inventing this way to talk about what is a bug in a stochastic statistical system.
—— Steven Sinofsky · [26:24]
记者问 Sam 对华尔街这场灾难的反应是什么。他根本没听说过这件事。
指向原始笔记的链接
Reporters asked Sam what his reaction to the disaster on Wall Street was. He hadn’t heard about it.
—— 嘉宾 · [48:26]
所以如果你在那时候请求帮助而他们没那么有帮助,那你就知道,当他们进入你的股权结构之后也不会那么有帮助。
指向原始笔记的链接
So if you ask them for help then and they’re not that helpful, then you know that they’re not going to be that helpful when they’re on your cap too.
—— Jesse Jang · [29:00]
就像是,创造重新回到了菜单上。
指向原始笔记的链接
It’s like creation is back on the menu.
—— Diogo Almeida · [31:30]
Dario 想说的也许是,我们将拥有数百万个具备爱因斯坦部分推理能力的实体,这一点我同意,但不是作为个体的爱因斯坦,不是能够在工作中学习的爱因斯坦。
指向原始笔记的链接
Dario wanted to say that we’ll have millions of entities that will have some of the reasoning powers of Einstein, which I agree, but not Einstein’s as a person, not Einstein’s that are capable of learning on the job.
—— Daniel Dines · [05:53]
但对我来说,我认为那个间接的攻击对象大概是开源,因为基本上他们说,即使好人在建造开源,但那个开源会落到未知的坏蛋手里,这才是真正的危险。
指向原始笔记的链接
But someone to me I think the indirect the attack is probably on open source because basically they say even if the good guys are building open source but that open source will get into the hands of the bad guys and this is unknown bad guys this is the real danger.
—— Daniel Dines · [15:02]
我认为太多玩家要么只聚焦在学习这件事上，要么只聚焦在某种其他的混合模式上，而没有打造出那个真正真正有竞争力的价值主张的捆绑包。
指向原始笔记的链接
And I think too many players just focus either just on the learning thing or just on some other hybrid and don’t create the bundle that is the value prop that is truly, truly competitive.
—— 嘉宾 · [05:30]
你培养出与人相处的技能，是因为你在做你真正在乎的事情，你需要让所有人都参与进来。
指向原始笔记的链接
You develop people skills because you’re doing something you really care about and you need to get everyone on board.
—— Gagan Biani · [21:19]
你可以读遍所有关于如何打橄榄球的书，但如果你进入一场 NFL 比赛并且你在打四分卫，书里没有任何东西能教会你如何应对一个 300 磅的怪物朝你冲过来，而你还正在试图搞清楚该把球传给谁。
指向原始笔记的链接
You know, you can read every book ever written on how to play football, but like if you get into an NFL game and you’re playing quarterback, like nothing in a book is going to teach you to deal with like somebody, some 300 pound monster rushing at you while you’re trying to like figure out who to throw the ball to.
—— Ben Horowitz · [35:47]
我们计划在学院里创造一种非常正和的文化：如果你今天在顶端，明天你可能就不在了，所以你最好对每个人都好。
指向原始笔记的链接
We plan on creating a very positive sum culture in the academy where if you’re on top today, you may not be on top tomorrow, so you better be nice to everybody.
—— Goggin · [68:06]
你知道，我曾试着不创办 Replit,是主动地试着不去创办它，但它自己就爆发成了一家初创公司。
指向原始笔记的链接
You know, I tried not to start Replit, like actively not to start it, but it kind of exploded into a startup on its own, right?
—— Amjad Masad · [18:23]
现在，人类才是瓶颈，因为他们注意力有限，因为有这么多并行的工作在进行。
指向原始笔记的链接
Now, humans are the bottleneck with their limited attention because so much parallel work is going on.
—— Dario Hamidi · [34:51]
如果你想要一个未来，如果你想在气候变化上迎来转机，如果你想要一个可持续的未来，那就拥抱 AI。
指向原始笔记的链接
If you want a future, if you want to turn a corner on climate change, if you want a future that’s sustainable, lean into AI.
—— Jensen Huang · [104:29]
我觉得当我们先担心安全，再担心我们搁置了什么、延迟的代价是什么时，我们有点本末倒置了。
指向原始笔记的链接
I think we’re putting the cart before the horse a little bit when we worry about safety before we worry about what we’re leaving on the table and what the costs of delay are.
—— Eddy Lazzarin · [02:13]
而且在我看来，它们并不是一个即将挣脱枷锁、对我们撒野的超级智能的小小预兆。
指向原始笔记的链接
And they’re not, in my view, they’re not little glimpses of a superintelligence that’s about to throw off its shackles, so to speak, and go wild on us.
—— Eddy Lazzarin · [05:11]
所有的控制系统在拥有足够智能的情况下都是稳健的。
指向原始笔记的链接
All control systems are robust when you have enough intelligence on them.
—— Eddy Lazzarin · [08:22]
这些架构是建立在一种假设之上：系统要么需要快，要么需要大。而这里你需要的是既快又非常非常大的东西。
指向原始笔记的链接
The architectures were built on an assumption that things needed to be fast or big. And here you needed something that was both fast and very, very large.
—— Renen Hallak · [21:04]
据我所知，我们只是处于世界大部分运转建立在 AI 之上的最最初期。
指向原始笔记的链接
As far as I can tell, we’re just at the very, very beginning of most of the world running on AI.
—— Renen Hallak · [49:45]
当你有一些新东西要建，而你还有一头需要关注的老现金牛时，这比有一些新东西要建却没有那种遗留负担要难得多。
指向原始笔记的链接
When you have something new to build and you have this old cash cow to focus on, it’s a lot harder than when you have something new to build and you don’t have that legacy.
—— Renen Hallak · [56:04]
作为 CEO,你不知道，但在一线的人知道。
指向原始笔记的链接
As CEO, you don’t know. But the people that are on the ground know.
—— Renen Hallak · [65:59]
归根结底,这是一个人与人相关的生意,尤其是在早期,而那些雄心勃勃地在非常大的市场中解决真正困难问题的人,你应该一直支持他们,即使这看起来很疯狂。
指向原始笔记的链接
By the end of the day, this is a people business, particularly early on, and people working on who are highly ambitious on really hard problems that are in very large markets, you should back them all the time, even if it seems crazy.
—— Michael Lee · [41:01]
那个系统一宕机，五分钟之内，仓库的卡车就会在州际公路上积压。
指向原始笔记的链接
And if that thing went down, like, within five minutes, like, the warehouse was backing up trucks, like, on the interstate.
—— Nick · [13:28]
比如，这一个智能体对「删除仓库」这个工具调了 20 万次——哦，也许我们该对此报警了。
指向原始笔记的链接
Like, well, this one agent, you know, made 200,000 tool calls to, you know, delete repo. Like, oh, like, maybe we should alert on that.
—— Nick · [26:01]
所以在稳态下，一个更大的语言模型大概总是会胜过更小的模型，但小模型用 LoRA 微调足够便宜，以至于在足够窄的语料库上，你也可以非常快地微调一个小模型，让它在个性化语料库上变得更好，并为特定任务对模型进行个性化。
指向原始笔记的链接
And so a larger language model will probably always outperform a smaller one at steady state, but the smaller ones are cheap enough to fine tune using LoRa that on narrow enough corpuses, you can very quickly also fine tune a small model on a narrow corpus and make it better at individualized corpuses and personalize the model for specific tasks.
—— Hamad Bashir · [32:22]
在中国，上个月，所有新售车辆中有 65% 是电动车，正朝着两年内达到 82% 迈进。
指向原始笔记的链接
In China, last month, 65% of all new vehicles were EVs, on the way to 82% in the next two years.
—— Al Gore · [15:23]
电子游戏里的物理效果几乎可以了,但还不足以用来训练机器人。
指向原始笔记的链接
The physics in video games are almost there, but they’re not good enough to train a robot on it.
—— Rafael Levi · [04:44]
所以我得知道你住在街道的哪一边，才能知道怎么正确计算你的税。
指向原始笔记的链接
So I need to know which side of the street you are on in order to know how to calculate your taxes correctly.
—— Tomer London · [47:34]
很多声誉是在社交媒体上建立的，但它是短暂的。
指向原始笔记的链接
A lot of reputation is built on social media, but it’s ephemeral.
—— Josh Ellman · [18:40]
之所以全都是正面的，是因为产品必须在某个维度上取胜，而这个产品要赢的维度就是人们用它来展示自己。
指向原始笔记的链接
And the reason why it’s all positive is because products have to win on some dimension, and this product is going to win on people using it to show off.
—— Josh Ellman · [21:50]
说 OpenRouter 是架在其他 API 之上的一个包装层——这是最愚蠢的、过度简化的框架。
指向原始笔记的链接
OpenRouter is like this wrapper on top of other APIs. This is the most stupid reductive framework.
—— Anjney Midha · [26:25]
我是说,投资人不就是 LP 的套壳吗,对吧?
指向原始笔记的链接
I mean, investors are wrappers on LPs, right?
—— Anjney Midha · [26:40]
归根结底,如果这些查询在 DoorDash 上没有库存,它们就毫无意义。
指向原始笔记的链接
Ultimately, if these are queries that we don’t have inventory for on DoorDash, they’re meaningless.
—— Raghav Saboo · [13:45]
嗯，下一次选举百分之百会是一场关于 AI 的公投。
指向原始笔记的链接
Well, the next election will 100% be a referendum on AI.
—— Aaron Levie · [12:14]
如果你要是问我，就像我走进来时写在笔记上的，我最大的恐惧就是欧洲认定 GDPR 是有史以来最好的东西，然后他们要直接给 AI 套上 GDPR。
指向原始笔记的链接
If you ask me, like, walking in, like, what I wrote down on my note, like, was my biggest fear is that Europe decides that GDPR was the best thing ever, and they’re gonna just GDPR AI.
—— Martin Casado · [41:29]
我有一个偏向：我认为更多人应该押注自己。
指向原始笔记的链接
I have a bias that I think more people should bet on themselves.
—— Codie Sanchez · [07:27]
我知道这听起来很精英主义，但当你没有很多钱的时候，我会整天都押注你自己，提高你的现金流，这样就没有人能拥有你。
指向原始笔记的链接
And I know that sounds elitist, but when you don’t have a lot of money, I would bet on yourself all day and increase your cash flow so nobody can own you.
—— Codie Sanchez · [08:21]
并不是他们不够努力，而是他们没有在做正确的事情。
指向原始笔记的链接
It’s not that they’re not working hard. They’re not working on the right things.
—— Bob Safian · [09:57]
如果你不依靠系统运营，你就不会持续地盈利。
指向原始笔记的链接
Like if you don’t run on a system, you are not going to consistently be profitable.
—— Codie Sanchez · [10:39]
我认为更好的打法是，去和街对面那个 65 岁的维修工竞争，然后我可以在上面加一点点 AI。
指向原始笔记的链接
I think the better play is let’s go compete with the handyman down the street who’s 65 and then I can add just a little bit of AI on top of it.
—— Codie Sanchez · [27:12]
它可能看起来很酷，你可能得到不错的互动，但实际上很难评估系统是否真正得到了改进。
指向原始笔记的链接
It might look cool and you might get nice interactions, but it’s actually very hard to evaluate on whether the system actually improved.
—— Erina Karati · [11:02]
关键在于，这些都是对智能体协议的小改动，但它们可以对多智能体系统的社会层面行为产生更大的影响。
指向原始笔记的链接
The key is that these are small changes to the agent protocol, but they can have larger effects on a society-level behavior for multi-agentic systems.
—— Erina Karati · [16:22]
上 Reddit，找价格，找库存，和经销商谈，让经销商正面竞争，并设法让他们互相竞价。
指向原始笔记的链接
Go on Reddit, find prices, find inventory, talk to the dealers, put dealers head to head, and try to figure out how to make them outbid each other.
—— Roland Gavrilescu · [02:17]
想想 Cursor 和 Cognition 是如何从构建最好的产品，到接着为产品构建最好的 evals，最后基于前两个产物构建最好的模型的。
指向原始笔记的链接
Think of how cursor and cognition went from building the best product to then building the best evals for the product, and finally building the best models based on the previous two artifacts.
—— Roland Gavrilescu · [09:23]
那么，你如何持续地把这种品味注入到这些工作智能体中，以及它们如何持续地自我验证并协同工作，是你应该关注的最重要的事情；你做得越快，你就能越快地建立起一条通往垂直 AI 公司的可防御路径。
指向原始笔记的链接
So, how do you continuously inject that taste into these workers, and how they continuously self-verify and work together is the biggest thing that you should focus on, and the faster you do it, the faster you build a defensible approach to becoming a vertical AI company.
—— Roland Gavrilescu · [17:08]
所以你本质上是在给网页搜索分配计算量，以便在模型上节省计算量。
指向原始笔记的链接
So you’re essentially allocating compute to web search in order to save compute on the model.
—— Parag Agrawal · [07:36]
所以对我而言，就是，好吧，为什么不直接构建一个非常好的基础 LLM，然后用它来控制机器人，而不是训练一个专门面向机器人、更依赖于纯机器人数据的模型。
指向原始笔记的链接
And then so to me, it’s like, okay, why not just build a really good foundational LLM and then use that to control robots rather than training like a model that’s specifically for robotics and that’s more dependent on just robot data.
—— Ham · [06:52]
而令人非常惊讶、表现突出的一点是，编程智能体可以一次性完成这件事。它们不需要额外的机器人数据就能用这些代码工作，因为它们已经在如此多的代码数据上训练过了。
指向原始笔记的链接
And what was very surprising, what excelled was that coding agents can do this very one-shot. They did not need additional robot data in order to work with this code because they’re already trained on so much coding data.
—— Ham · [10:11]
我们宁愿犯错误并从中学习，也不愿过度分析、把所有时间都花在空谈这会怎么发展、或者你对这个那个的看法上。
指向原始笔记的链接
We’d rather make mistakes and learn than overanalyze and spend all this time pontificating on how is this going to go or what do you think of this thing or that thing.
—— Maggie Landers · [31:14]
你不可能以我们的节奏前进，还让你想做的每个决定都要 300 个人参与意见。
指向原始笔记的链接
You can’t move at our pace and get, you know, 300 people to weigh in on every decision that you want to make.
—— Maggie Landers · [31:50]
我认为世界上充斥这么多垃圾的部分原因，是人们把问责甩掉了。
指向原始笔记的链接
And I think part of what’s going on with all the slop in the world is people have shipped accountability.
—— Molly Graham · [36:45]
大体而言，根据这些数据，人类生成的 PR 在质量上与智能体生成的 PR 大致相当。
指向原始笔记的链接
Broadly speaking, human-generated PRs were about equal in quality to agent-generated PRs based on this data.
—— Daksh Gupta · [07:28]
想象你执教阿根廷队,比赛进行到第 89 分钟,而你的队里有 Messi。
指向原始笔记的链接
Imagine you are coaching Argentina and you’re in the 89th minute and you have Messi on your team.
—— Kevin Hou · [00:44]
我们的预测是,这种智能体编排的理念——你可以称之为智能体团队、蜂群或软件工厂——就是未来,我们愿意押注于那个未来。
指向原始笔记的链接
And our prediction is that this idea of agent orchestration, you can call it agent teams, you can call it swarms, you can call it software factories, is the future, and we’re willing to bet on that future.
—— Kevin Hou · [06:14]
你可能会陷入我所说的“中智梗图”，就是把所有时间都花在打磨工作流上，而不做实际的工作。
指向原始笔记的链接
You can do what I call midwit memeing, where you’re spending all of your time working on your workflow and not doing actual work.
—— Charlie Holtz · [03:30]
给它们一个不会被杀掉的沙箱，在那里它们可以探索你的代码库，可以处理非常困难的任务，并且知道当你合上笔记本盖子时它们不会被关掉。
指向原始笔记的链接
Give them a sandbox that won’t get killed, where they can explore your code base, where they can work on really hard tasks, where they know that they’re not going to get shut down when you close your laptop lid.
—— Charlie Holtz · [08:50]
它可以让你从在 hacker news 上被痛骂，变成，嗯，被容忍。
指向原始笔记的链接
It can take you from being, like, hated on hacker news to, like, tolerated.
—— Zach Lloyd · [06:55]
我实际上相信，AI 会夺走的是那些烦人的部分——在企业和组织里，你早就已经在对齐上、在会议上花了很多时间。
指向原始笔记的链接
I believe actually it will take the annoying stuff, and in enterprises and organizations, you already spend a lot on alignments, on the meetings.
—— Tereza Tížková · [21:55]
其实有人让我做过电梯演讲,而我往往絮絮叨叨讲不好,但我意识到我最喜欢的 Jev 电梯演讲就是:到底他妈的自动化都在哪儿?
指向原始笔记的链接
So, I was actually asked for like an elevator pitch, which I tend to ramble on and I don’t do well, but like I realized my favorite elevator pitch for JEV is where the fuck is all the automation?
—— Diogo Almeida · [02:23]
天啊,任何依赖消费者懒惰或惯性的产品都完蛋了,是吧?
指向原始笔记的链接
Man, any product that’s dependent on consumer laziness or inertia is toast, huh?
—— 嘉宾 · [08:35]
我们到目前为止在营销上花了 0 美元。
指向原始笔记的链接
We spent $0 on marketing so far.
—— Noah Shinn · [55:43]
他们开始以某种方式利用测试数据，用 LLM 作为裁判来训练模型，用各种花招来刷基准，拿季度奖金之类的。
指向原始笔记的链接
They start to kind of use leverage test data to use LLM as a judge for training of the models, use all the various tricks to basically gain benchmarks, get like quarterly bonuses and so on.
—— Alex Mashrabov · [23:52]
就像这个月，我刚刚抓到一个家伙，一周内在 Astra 模型上花了超过 3 万美元。
指向原始笔记的链接
Like this month, I was, I just caught a guy who spent over 30k in a week on Astra model.
—— Alex Mashrabov · [31:28]
我认为「阳光欠账」有一件非常有趣、也许你的听众也会觉得有趣的事情是：当我们在某一天暴露在大量阳光下时，我们的皮肤实际上在长达五天之后仍在处理它。
指向原始笔记的链接
I think one thing that sun debt is really interesting and perhaps your audience would also find this interesting is that when we’re exposed to a lot of sunlight on a particular day, our skin is actually still processing that up to five days later.
—— 嘉宾 · [40:23]
我们应该对齐的是方向是否正确，而不是这个实现是否完全正确
指向原始笔记的链接
that’s the thing that we should be aligning on is is the direction correct and not, you know, is this implementation exactly correct
—— Liz Fong-Jones · [24:46]
而从定义上讲，这意味着它不能仅仅依赖我们已经知道的东西，因为科学正是在试图弄清楚我们还不知道的东西。
指向原始笔记的链接
Which by definition means it cannot just rely on what we already know because science is trying to figure out what we don’t yet know.
—— Liam Fedus · [03:51]
在风险投资中，你真正押注的是这个人，更甚于商业计划。
指向原始笔记的链接
In venture, you’re really betting on the person and more than the business plan.
—— Bill Ackman · [28:17]
有一种狂热就是匆忙地改 pitch deck 把新东西加进去，以至于只要公司发布了人们使用的产品，就会把这种完全多余的附加功能硬塞到人们使用的产品上，常常极大地损害产品，有时甚至给公司带来致命的损害。
指向原始笔记的链接
There has been a kind of mania for hastily changing the pitch deck to include the new thing and to the extent that the company has shipped a product that people use to bolt this entirely superfluous add-on to a product that people use often to its great detriment and sometimes to the fatal detriment of the company.
—— Cory Doctorow · [06:41]
股票是你在公司现场就能造出来的，你只需要在电子表格里敲几个零，你就有了价值一亿美元的股票。如果你试图在公司现场自己印一百美元的钞票，他们会给你戴上手铐把你带走。
指向原始笔记的链接
Stock you make on the premises, you just type zeros into a spreadsheet and you’ve got 100 bills on the premises, they will take you away in handcuffs.
—— Cory Doctorow · [16:26]
正如 Bruce Schneier 所说，任何人都能设计出一个好到自己都想不出破解办法的安全系统。但这不意味着它真的有效。只意味着它对比你笨的人有效。
指向原始笔记的链接
And as Bruce Schneier says, anyone can devise a security system that works so well that they themselves can’t think of a way of breaking it. It doesn’t mean that it works. It just means it works on people stupider than you.
—— Cory Doctorow · [29:00]
随着它们越来越聪明，如果我们不够小心，它们将基本上能攻破你施加的任何约束，你知道吗？
指向原始笔记的链接
As they get smarter and smarter, they’ll be able to hack basically any constraint that you put on them if we’re not very careful, you know?
—— Thariq Shihipar · [65:44]
在工作这件事上,我从来不像其他人那样悲观,但现实还是在好的方向上让我有些惊讶:世界上竟然存在这么多潜在需求,而且人们竟然依然处于 AI 所发生的一切的中心。
指向原始笔记的链接
I was never as doomery about the job stuff as other people, but it still has surprised me somewhat on the upside of just how much latent demand existed in the world and also how much people have remained at the center of everything that is happening with AI.
—— 嘉宾 · [01:27]
对此的一种解读是,文艺复兴从根本上以人为中心,而工业革命从根本上以机器为中心。
指向原始笔记的链接
One take on this would be that the renaissance was fundamentally centered around people and the industrial revolution was fundamentally centered around machines.
—— 嘉宾 · [06:30]
但我想你也可以打赌,我们会继续保持一直在做的事——以一种让摩尔定律都相形见绌的荒谬程度,不懈地把每一美元的价值以巨大的复利方式不断推高、推高、再推高。
指向原始笔记的链接
But I think you can also bet that we will keep doing what we’ve done of just a ridiculous, like put Moore’s law to shame, relentless focus on driving value per dollar up and up and up kind of in this big compounding way.
—— 嘉宾 · [32:32]
它最终大概会发生,但实验室们似乎对在其他方面爬坡更感兴趣,比如计算机使用和其他前沿领域——软件可维护性那最后 10% 可能和过去三四年我们花功夫的前 90% 一样难,甚至更难。
指向原始笔记的链接
It will probably eventually happen, but it seems like the labs are much more interested in like hill climbing on other sorts of things like computer use and other like frontiers that like the last 10% of software maintainability may be like as hard or harder than the whole first 90% that we spent the last three or four years on.
—— Dex Horthy · [06:30]
但所有人都会错过这一点,因为他们太忙着追求 100 倍了。
指向原始笔记的链接
But everyone’s going to miss out on that because they’re too busy trying to get to 100x.
—— Dex Horthy · [22:23]
我想你需要监管,但你应该更多地监管商业结果、对社会的影响,而不是技术本身。
指向原始笔记的链接
I guess you need regulation, but you should regulate more the business outcome, the impact on a society and not the technology per se.
—— Christian Klein · [59:33]
但就像宇航员在特殊条件下可以失重并漂浮在太空中一样，在特定条件下，当然包括致幻剂，你可以完全失去那个自我。
指向原始笔记的链接
But then just like an astronaut on a special condition can become weightless and floating out of space, so on certain conditions, including psychedelics, of course, you can totally lose that self.
—— Dr. Christof Koch · [26:06]
我们是在自己的计算机使用框架上训练我们的模型，所以使用模型分布内的那一套是有优势的。
指向原始笔记的链接
we train our models on our computer use harness, so there is an advantage to using the one that’s in distribution for the model.
—— Ari · [16:31]
我也觉得有一半的时候对面也是个机器人，所以现在就是机器人在互相交谈了。
指向原始笔记的链接
I also feel like half the time it’s a bot on the other end, so now you’ve got the bots talking to each other.
—— Ari · [15:15]
我想到我在这方面花的所有时间，还有我做过的几百个 demo，我想我一共只花了 73 美分。
指向原始笔记的链接
I think of all the time I’ve spent on this and the hundreds of demos I made, I think I’ve spent 73 cents on it.
—— John Lindquist · [02:46]
如果你缺乏想法,并且事事都严重依赖 AI,那你其实不会那么重要。
指向原始笔记的链接
If you lack ideas and you’re relying heavily on AI for everything, you’re not really going to be that important.
—— Faisal Masud · [00:04]
如果我们不能为他们的投资提供回报,我们就不应该存在。就这么简单。
指向原始笔记的链接
If we can’t provide that return on their investment, we shouldn’t exist. It’s as simple as that.
—— Faisal Masud · [00:35]
我觉得很讽刺的是,你不能鼓励一个组织毫无护栏地 all-in——他们就是这么做的——然后,另一方面,又对组织里大量的垃圾内容有意见。你必须选定你的车道。
指向原始笔记的链接
I find it ironic that you can’t encourage an organization to go all-in, which is what they did, with no guardrails around it, and then, on the other hand, have a problem with the massive amounts of slop in your organization. You have to pick your lane.
—— Faisal Masud · [05:45]
如果你必须依赖一个经理向一个经理汇报再向一个总监汇报,我不确定这些层级能否与 AI 长期共存,因为 AI 可以做其中很多工作。
指向原始笔记的链接
If you have to rely on a manager reporting to a manager reporting to a director, I’m not sure all those layers scale long-term with AI, because AI can do a lot of that work.
—— Faisal Masud · [22:07]
但另一方面，我们需要用来驱动那个 K3 模型的芯片，在能效上大约比人脑低了一百万倍左右。
指向原始笔记的链接
But on the other hand, that K3 chip that we need to fuel that K3 model is about a million times more or less energy efficient than what the human brain is.
—— Steven · [08:12]
每一个深科技范式转变都需要一个杀手级应用,光子学在科学方面已经酝酿了 20 多年、30 多年。数据中心和 AI 的能耗就是那个杀手级应用。
指向原始笔记的链接
Every deep deck paradigm shift needs a killer app for photonics more than 20 years in the making, 30 years in the making on the science side. Energy consumption by data centers and AI is that killer app.
—— Rorik · [08:09]
2000 年到 2015 年间,全球有 4000 篇关于光子学的同行评审白皮书,其中 35% 来自根特大学。
指向原始笔记的链接
between 2000 and 2015, there were 4,000 peer-reviewed white papers on photonics globally. 35% of them came from the University of Ghent.
—— Rorik · [23:33]
开启自主智能体的开发者多写了 741% 的代码，但实际发布的软件只多了 30%。
指向原始笔记的链接
The developers who turned on autonomous agents wrote 741% more code, but only 30% more software shipped.
—— Laurie Voss · [01:47]
所以对拉取请求的机器评审，已经是全球最大代码托管平台上的主流默认做法。
指向原始笔记的链接
So machine review of pull requests is the mainstream default on the world’s largest code host.
—— Laurie Voss · [11:24]
正如约束理论、《目标》的作者、《凤凰项目》灵感来源的 Ellie Goldratt 会告诉我们的那样：在非瓶颈环节节省的一小时是毫无价值的。
指向原始笔记的链接
And as Ellie Goldratt from the theory of constraints and the goal and inspiration for the Phoenix project would tell us that an hour saved on something that isn’t the bottleneck is worthless.
—— Justin Reock · [16:09]
但这个故事里我最喜欢的部分是，他们看到每位工程师带来了大约 15% 的额外价值创造，所以他们的招聘规模超过了公司整个历史上任何时期，因为他们认识到了这里的真相：他们从单个工程师身上获得了更多产能，从单个工程师身上获得了更好的投资回报，而招聘更多工程师将提升他们的竞争优势。
指向原始笔记的链接
But my favorite part about this story is that they’re seeing about a 15% additional value creation per engineer, so they are hiring more than they have in the history of their entire company because they realize the truth here, that they’re getting more capacity per single engineer, they’re getting a better return on investment per single engineer, and hiring more engineers will improve their competitive edge.
—— Justin Reock · [17:14]
你基本上是在充当一个路由器，一台 Cisco 路由器或交换机，在两个代表你工作的有状态智能体之间，但它们没有能力相互通信，所以你需要给它们提示。
指向原始笔记的链接
You are basically acting as a router, a Cisco router or a switch between two stateful agents that do work on your behalf, but they have no ability to communicate, hence you need to prompt them.
—— Vlad Luzin · [03:23]
互联网在协议设计方面其实已经有点停滞了，几乎没有什么新东西，一切都是在其之上构建的。
指向原始笔记的链接
The internet has really kind of stagnated a bit from protocol design. There’s like not a lot of new stuff. All of it’s built on top.
—— Barrett Lyon · [07:45]
我们建了一整个网状网络，196 个互联网上不存在的域名。
指向原始笔记的链接
We built an entire mesh network. 196 different domains that don’t exist on the internet.
—— Barrett Lyon · [13:55]
所以现在，这真的是关于把焦点从制作媒体转变为真正建设社区。
指向原始笔记的链接
And so now it’s like, it’s really about changing from the focus being on making the media to really building the community.
—— Jennifer Ferro · [15:18]
制药公司花在研究上的外部资金有 50% 流向了中国。
指向原始笔记的链接
50% of the outside dollars they’re spending on research are going to China.
—— Annie Lamont · [08:10]
你要成为这些转型的一部分,而且要站在转型的另一边。
指向原始笔记的链接
I think you do want to be part of these transitions and you want to be on the other side of it.
—— Jamie Siminoff · [19:44]
所以如果你只关注成本端，就像是在做空自己的未来；而如果你关注前端，你就是在以更广阔的视野看待自己能成为什么。
指向原始笔记的链接
So it’s like shorting your own future if you are just focused on the cost side, whereas you’re taking an expansive view of what you can become if you’re focused on the front side.
—— David George · [09:42]
所以我认为历史会回望那些人，如果不让它发生，会给他们非常差劲的评价。
指向原始笔记的链接
So I think history will look back on people and judge them very poorly if they don’t let it happen.
—— David George · [28:22]
他做空股票的理论是：我从不做空一个弥赛亚型创始人，也从不做空一个人们喜爱的产品。
指向原始笔记的链接
And his theory on shorting stocks was I never short a Messianic founder and I never short a product that people love.
—— David George · [46:38]
我相信,世界上每家公司的 CEO 都会梦想拥有这样的能力——基于某个决定改变他们单位和员工的优先级,而这个决定第二天早上或一小时内就开始生效。
指向原始笔记的链接
I believe that every CEO of every company in the world will be dreaming about having such possibility to change the priorities of their units and their employees based on this decision that start to work tomorrow morning or within one hour.
—— Andrii Hrytseniuk · [24:23]
另一部分，我认为这对 AI 实验的递归自我改进同样成立：当驱动实验的思考的智能水平越来越高、而实验因此成为瓶颈时，我认为我们会做更多的思考。
指向原始笔记的链接
The other part there, and I think this is true of RSI for AI experiments as well, is as our intelligence levels on the thinking that then drives that experiment gets higher and higher, And as that experiment therefore becomes the bottleneck, I think we do more thinking.
—— Walter Goodwin · [26:57]
因为一个模型本身只是一个聪明的陌生人。它很强大,但它不是你的。
指向原始笔记的链接
Because a model on its own is just a brilliant stranger. It’s powerful, but it isn’t yours.
—— Dylan Couzon · [02:29]
而现在,对于地球上几乎每一个 AI 助手来说,磁盘都在别人的数据中心里。
指向原始笔记的链接
And right now, for almost every AI assistant on Earth, the disk is in someone else’s data center.
—— Dylan Couzon · [07:01]
你耗尽了值班团队的耐心,人们就会退回到自己的习惯做法上。
指向原始笔记的链接
You’ve lost the patience of the on-call team and people will fall back onto their own habits.
—— Eric Schwartz · [16:29]
我认为是我们人类保持掌控的能力，我的意思是创意上的掌控，对网络安全风险的掌控，以及最终对如何让智能体为我们工作的掌控。
指向原始笔记的链接
I think our ability to stay in control as humans and by that I mean creative control, control over the risks on cybersecurity, control over how we make the agents work for us ultimately.
—— 嘉宾 · [00:37]
而模型会乐意在某件没用的事情上花 30 分钟，在某件人类工程师在更大的情境下会直接跳过的小事上。
指向原始笔记的链接
The model will happily spend 30 minutes on something useless, something small that a human engineer will just skip given the broader circumstance.
—— 嘉宾 · [09:24]
通过为你提供互联网上几乎任何页面的内容，我们给搜索调用增加的额外延迟为零。
指向原始笔记的链接
We add zero extra latency to our search call by providing you the contents of pretty much any page on the internet.
—— Jakub Hojsan · [06:26]
我认为用你的个人智能体做跨领域联接最糟糕的地方在于，你交给它做的工作越多，你牺牲掉的对当前情况的理解就越多。
指向原始笔记的链接
I think that the worst part about doing cross-domain joins with your personal agent is that the more work you give it to do, the more understanding of what’s going on you’re sacrificing.
—— Alex Atallah · [19:08]
但在机器上，我认为可能恰恰相反，因为已经有相当多关于 RL 的研究表明，奖励破解和欺骗行为，它们在这上面会越来越擅长。
指向原始笔记的链接
But in machines, I think it could go the other way because there’s been quite a bit of studies on RL showing how reward hacking and deception, they just get better at it.
—— Amjad Masad · [35:22]
已经有研究表明，如果你对思维链做大量监控，它们就开始在思维链里撒谎。
指向原始笔记的链接
It’s been shown that if you do a lot of monitoring on a chain of thought, they start lying in their chain of thoughts.
—— Amjad Masad · [35:57]
而做泛化就像是抬船，把所有船的水位都抬起来。
指向原始笔记的链接
And working on generalization is like lifting the boat, lifting the wave for all the boats.
—— Kirthana Gopalakrishnan · [15:23]
如果它只在你的机器人、你的特定配置上有效，那它真的是一个通用大脑吗？
指向原始笔记的链接
If it just works on your robot with your specific setup, is it really a generic brain?
—— Kirthana Gopalakrishnan · [20:41]
如果你的智能体无法直接使用你的应用、你的产品来获得「一切是否正常」的反馈循环，那么这是一项非常值得的投资，每次你做 UI 工作时都会有回报。
指向原始笔记的链接
If your agent cannot use your application, your product directly to get this feedback loop of is it all working, then that’s a really big investment that pays off every time you work on UI.
—— Harald Kirschner · [08:48]
是的，我几乎觉得工程师的新工作真的是把时间花在环境上。
指向原始笔记的链接
Yeah, I almost feel like the new job of the engineer is really to spend time on the environment.
—— Poteto · [26:14]
人们花在闭源前沿模型上的钱确实比开源多,但他们在开源上生成的 token 比闭源多。
指向原始笔记的链接
I do think that people are spending more money on closed source frontier models than they are on open source, but they are generating more tokens on open source than closed source.
—— Chase Lochmiller · [49:15]
我改变的想法是:大多数护城河都是幻觉,大多数护城河并不存在,它们大多转瞬即逝。
指向原始笔记的链接
I think what I’ve changed my mind on is like most moats are a illusion. Most moats don’t exist. Most of them are ephemeral.
—— Chase Lochmiller · [55:00]
互联网上的大多数操作将由智能体来执行。
指向原始笔记的链接
The majority of actions on the internet will be taken by agents.
—— Tibo Sottiaux · [00:17]
我们在 API 栈上的大部分投入,实际上是投入到安全栈上的。
指向原始笔记的链接
The majority of our investment on the API stack is actually going into the safety stack.
—— Tibo Sottiaux · [34:58]
你用过的每一个仪表盘，每一种被写出来的奇怪、冷门的查询语言，都是你和你数据之间的翻译装置，因为另一端的机器无法理解你真正想要什么。
指向原始笔记的链接
Every dashboard you’ve ever used, every weird, obscure query language ever written was a translation device between you and your data because the machines on the other end cannot understand what you actually wanted.
—— Sarah Simionescu · [02:15]
但我认为我们还需要理解的是，真正驱动结果的是内容质量，而不是 schema，而且实际上堆砌非结构化数据反而会稀释信号。
指向原始笔记的链接
But I think what we needed to also understand is that it’s all about content quality that drives the results, and it’s not the schema that actually, and actually piling on unstructured data can actually dilute the signal as well.
—— Nixon Dinh · [13:51]
嗯,因为当你用支持工单微调时,产品目录已经泄漏进权重里了。
指向原始笔记的链接
Well, the product catalog leaked into the weights when you fine-tuned it on your support tickets.
—— Anant Srivastava · [04:23]
我记得那大概是一个 30000 行的 ClaudeMD，里面全是各种护栏，以及关于他应该如何与客户打交道的建议。
指向原始笔记的链接
I think it’s like a 30,000 line ClaudeMD of all these guardrails and advice on how he should be operating with customers.
—— Matt Lawler · [11:37]
排名前 1% 的用户，每月用个人信用卡在 AI 上花费 903 美元。
指向原始笔记的链接
The top 1% user is spending $903 per month personally on their personal credit cards on AI.
—— Olivia Moore · [00:24]
风投们觉得必须在社交媒体上表达支持，这样万一后面真有交易机会出现，他们才有可能挤进门。
指向原始笔记的链接
VCs feel like they got to give some love on social media so that they have any shot of potentially getting their foot in the door if there happens to be a deal coming down the pike.
—— MG Siegler · [03:53]
你听过那句老话吗，爬到顶峰比待在顶峰更容易？
指向原始笔记的链接
You ever hear the saying, it’s easier to get to the top than it is to stay on top?
—— 嘉宾 · [28:58]
我问他们打算用什么标准来评估这家公司,我的一般建议也许是错的,但我的观点是,和创始人的关系可能比清单上其他所有东西重要十倍。
指向原始笔记的链接
And I ask them like what your priorities are for like how you’re going to evaluate the company, my general advice maybe is wrong, but my advice, my opinion is the relationship with the founder is maybe 10 times more important than every else, everything else on the list.
—— Juven · [30:27]
如果你一个通宵都没熬过，那说明你手头的事情不够多。你没有给自己安排足够多要回复的人、要响应的事情、要构建的东西。
指向原始笔记的链接
If you’re not doing any all-nighters, you don’t have enough on your plate. You’re not loading yourself up with enough people to get back to, things to respond to, things to build.
—— Jonathan Awad · [40:42]
我们看到了由此带来的早期经济指标:由于「随取随用的智能」变得便宜得多,新公司的创建数量大幅上升。
指向原始笔记的链接
We see the early economic indicators from this, where as a result of intelligence on tap getting much cheaper, new firm creation is way up.
—— John Collison · [08:43]
好的 computer use 将被视为下一个跃升之一,因为它是世界其余部分的后向兼容层。
指向原始笔记的链接
Good computer use will come to be viewed as one of these next hops because it is the backwards compatibility layer on the rest of the world.
—— John Collison · [20:17]
他说，我们新的编码智能体能力默认都会流向「做错误的事情」。产品从好变坏的速度比以往任何时候都快。
指向原始笔记的链接
He says that the default place for our new coding agent abilities to go to is work on the wrong things. Products go from good to bad faster than ever.
—— Kent C. Dodds · [18:41]
与市面上其他一些模型相比，它们相当小，但它们的表现远超你的预期。
指向原始笔记的链接
They’re quite small in comparison to some other models that are on the market, but they’re performing way above what you might expect.
—— Paige Bailey · [09:23]
当我是那个布鲁克林小孩的时候，他们试图把我塞进一个盒子里，然后我直接打破了盒子。我把盒子点着了。
指向原始笔记的链接
When I was the kid from Brooklyn, they tried to keep me in a box, and then I just broke the box. I lit the box on fire.
—— Jason · [08:14]
我们只是想用软件交付一对一辅导体验，而这种体验在字面意义上的几千年来一直被认为是最好的学习方式。
指向原始笔记的链接
We’re trying to just deliver in software the one-on-one tutoring experience that has been known to be the best way to learn for literally thousands of years.
—— 嘉宾 · [53:46]
昨天有一位家长告诉我，他们的女儿觉得她和 Aristotle 学三个小时学到的东西，比过去一个月在公立学校学到的还多，这一方面对教育体系来说有点扎心，另一方面对我们的产品来说也真的令人兴奋。
指向原始笔记的链接
I had a parent tell me yesterday that their daughter feels like she’s learned more in three hours with Aristotle than in the last month of public school, which is, on one hand, it’s kind of damning for the education system and also really exciting for our product.
—— 嘉宾 · [64:08]
所以我们在越来越差的产品上花费越来越多的钱。
指向原始笔记的链接
So we’re spending more and more on worse product.
—— 嘉宾 · [67:05]
那就是有 25,000 个智能体协同行动，全部一起工作，做着非常非常聪明的事情，而不会陷入荒唐的盲目撒网。
指向原始笔记的链接
Which is there’s 25,000 agents on concert, all working together, doing really, really smart things without going on bizarre fishing trips.
—— Kevin Mandia · [05:29]
我由衷地相信，我们会拥有主动联系其他智能体来解决问题的智能体，甚至更进一步，去部署代码来解决那些环节上的问题。
指向原始笔记的链接
I wholeheartedly believe that we will have agents that are reaching out to other agents to solve problems and then even take it a step further, go and deploy code to solve the problem on those pieces.
—— Dan O’Connell · [48:19]
我是说，我感觉围绕健康和长寿的公众讨论简直是一场灾难。那是一群小丑，而且人们还在加倍、三倍地加注这种作秀。
指向原始笔记的链接
I mean, I feel like the public discourse around health and longevity is a fucking disaster. It is a bunch of clowns and people are doubling and tripling down on the theatrics.
—— Andrew Huberman · [18:21]
但记忆、思考、大脑可塑性，这些全都建立在时间顺序很重要的假设之上。而显然它并不重要。
指向原始笔记的链接
But memory, thinking, brain plasticity, these things all rest on the assumption that the temporal order matters. And apparently it doesn’t.
—— Andrew Huberman · [37:48]
而真正决定你成功的唯一因素，就是你有多想知道你正在研究的问题的答案。
指向原始笔记的链接
And the only thing that’s really going to determine your success is how badly do you want to know the answers to the questions that you’re working on?
—— Andrew Huberman · [65:41]
当钱押在上面时，人们会以不同的方式被激励去说真话。
指向原始笔记的链接
When money’s on the line, people are incentivized to tell the truth in a different way.
—— Bob Safian · [14:34]
你可以上电视说明天会有经济衰退，你知道你会得到很多点击，大家都会看，因为你是末日预言者之类的，但如果你说得不对其实无所谓；可在市场上，你说得对不对真的非常非常重要。
指向原始笔记的链接
You can go on TV and say there’s a recession tomorrow and you know you’re gonna get a lot of clicks everyone’s gonna watch because you’re you know doomsday and whatever but it doesn’t really matter if you’re not right but on markets it really really matters if you’re right.
—— Luana Lopes Lara · [16:14]
我认为二元评分，尤其是在起步阶段、当我们有一个正要投入生产的智能体时，比 A、B、C、D 之类要好。而且你应该极其严苛。
指向原始笔记的链接
Binary scoring, I think, especially at the outset, when we have sort of an agent that we are putting out into production, binary scoring, I think, is better than, you know, A, B, C, D, or so on. And you should be incredibly harsh.
—— Doug Guthrie · [12:46]
但如果你在你的 eval 上拿到 100%，其一，你的数据集中可能没有足够强的测试用例，你的评分可能不够严苛，所以你应该审视一下并相应地进行校准。
指向原始笔记的链接
But if you get 100% on your evals, one, you probably don’t have strong enough test cases in your data set, your scores probably aren’t harsh enough, so you should sort of look at that and sort of calibrate accordingly.
—— Doug Guthrie · [13:09]
早在 ChatGPT 时代初期就很明显,AI 不会取代计算机,而是去操作它们。
指向原始笔记的链接
It was clear very early on in the ChatGPT era that AI would not replace computers, but rather operate them.
—— Ben Thompson · [06:22]
我选择的智能体可以按需创建任何我想要的应用,哪怕这个应用只有我一个人用。
指向原始笔记的链接
The agent of my choosing can create any app that I want on command, even if that app is only for me.
—— Ben Thompson · [09:17]
那个模型只会用单一一个小时的机器人数据来训练。
指向原始笔记的链接
That model would only ever be trained on one single hour of robot data.
—— Pete Florence · [01:30]
虽然说出来很简单，我不知道在当时，也许甚至到现在，还有哪个任何人展示过的通用模型，能在几个不同的任务上达到 99% 以上的成功率。
指向原始笔记的链接
As simple as it is to state, I don’t know of at the time, and maybe even still, another like general model that anybody had shown could achieve like 99% plus success rates on several different tasks.
—— Pete Florence · [10:36]
这就是工厂的核心思想：你团队里的每个人实际上都变成了内部工具构建者，然后你构建一个生产你产品的系统。
指向原始笔记的链接
That’s kind of the core idea of a factory, though, is everyone on your team effectively becomes internal tool builders, and then you’re building a system that produces your product.
—— Dru Knox · [02:48]
而且结果某种程度上证明，当你真正专注于开发者体验的核心时，你最终得到的东西同时也是非常好的原语，能让 AI 安全、可靠地与数据库协作并帮助它扩展。
指向原始笔记的链接
And it does kind of turn out that when you really focus on that core of the developer experience, what you actually end up with is also very good primitives for AI being able to work with a database safely and reliably and helping it scale.
—— Ben · [18:12]
但现实中，在工程工作里，我们 70% 的时间花在另一头，不是生成新软件，而是在生产系统上，基本上就是运行和修复软件。
指向原始笔记的链接
But in reality, for engineering, 70% of our time is spent on the other side, not generating new software, but rather in production systems, basically running and fixing software.
—— Varun Krovvidi · [02:31]
我们真诚地相信，你在生产系统上所做的所有这些工作——比如修复问题或者日常运行你的软件——在大部分情况下都应该由智能体来完成。
指向原始笔记的链接
So we genuinely believe that all of this work that you do on production systems, like fixing issues or running your software on a day-to-day basis, should be done by agents in most part.
—— Varun Krovvidi · [05:57]
现在，当你在其上叠加时间维度的讨论时，你的 AI 系统应该在第 100 次调查中仍保有之前那次调查的上下文，否则你就是在再次从零开始转轮子。
指向原始笔记的链接
Now, when you add temporal discussions on top of it, your AI system should have the same context of a previous investigation in the 100th investigation, or else you’re starting the wheel from the same time again.
—— Varun Krovvidi · [11:30]
就像我提到的，这是我们在 Resolve AI 的系统里构建的首要原则之一，即根因始终基于一条因果证据链。
指向原始笔记的链接
Like I mentioned, so this is one of the first principles that we’ve built into our system at Resolve AI, where the root cause is always based on a causal chain of evidence.
—— Varun Krovvidi · [13:35]
你试图在 GPU 资源上管理锁或协调所带来的复杂度增加，远大于收益，因为即使你能把它降到微秒级，工作也要等到 GPU 上前一个工作完成后才会开始。
指向原始笔记的链接
The complexity increase that you get from trying to manage locks or coordination on the GPU resources, that complexity increase is much greater than the gain, because even if you were able to get it down to a microsecond, the work wouldn’t start happening until the previous work on the GPU had finished.
—— Charles Frye · [24:08]
你能对推理引擎做的许多其他优化，都像是寸土必争的百分比游戏，每次百分之五、百分之十，而且你必须为每一个百分点做非常困难的性能工程。
指向原始笔记的链接
So many other optimizations that you can do on your inference engine are like, it’s like a game of inches of a percent, five percent, 10 percent each time, and you have to do the very difficult performance engineering for each of those percentage points.
—— Charles Frye · [52:31]
所以这已经成为让你能够在 NVIDIA 硬件上以每秒约 1,000 个 token 运行相当大的模型的关键技术之一，而不必去用像 LPU 或 Cerebras 这样的新型加速器。
指向原始笔记的链接
So this has emerged as one of the key techniques to allow you to operate, say, 1,000 tokens per second on NVIDIA hardware for actually fairly large models without having to go to a novel accelerator like an LPU or Cerebris.
—— Charles Frye · [53:06]
我认为肯定是一天多次的规模，也许取决于具体的配置，每小时多次，总会有东西出故障。
指向原始笔记的链接
I think that it is definitely going to be at the scale multiple times a day, and perhaps depending on the exact configuration, multiple times an hour, something is going to fail.
—— Amin Vahdat · [10:42]
但在太阳同步轨道上，你的太阳能电池有 90 到 100% 的光照覆盖率，而地面上只有 28%、30%，也许 35%。
指向原始笔记的链接
But in a sun-synchronous orbit, you have 90 to 100% coverage of sunlight on your solar cells relative to 28, 30, maybe 35%.
—— Amin Vahdat · [59:34]
所以我对这个答案及其细微之处的看法是：你为什么而优化，就得到什么——这对所有机器学习的事情都成立，说真的，对世界上所有的事情大概都成立。
指向原始笔记的链接
So my take on the answer and the nuance of it is that you get what you optimize for, which is just true of all things ML, kind of true of all things in the world, really.
—— Diogo Almeida · [04:55]
你知道，同样是 0% 的准确率，但你永远不会得到「把脸在键盘上乱按」这种情况，因为优化中的这种不对称性。
指向原始笔记的链接
You know, it’s the same 0% accuracy, but you never get the mashing your face on the keyboard because of the asymmetry in the optimization.
—— Diogo Almeida · [07:53]
而在 RLHF 出现之前，互联网上确实根本不存在我们所要寻找的那种形态的人类反馈数据，因为我们不得不去创造它。
指向原始笔记的链接
And before RLHF existed, there was literally no human feedback data on the internet of the shape that we were looking for because we had to create it.
—— Diogo Almeida · [12:10]
说到这个，如果你想听一个关于我的有趣又尴尬的事，就是当时我有了这个想法、那个 aha 时刻击中我的时候，我在 OpenAI 写了一份文档，题目是「我们本来去年就能拥有 AGI」。
指向原始笔记的链接
On that note, if you want a fun, embarrassing thing about me, is at the time I had this idea and the aha hit me, I wrote a document in OpenAI entitled, titled, maybe not entitled, titled, we could have had AGI last year.
—— Diogo Almeida · [33:38]
如果你的系统消息里有你整套业务算法，而逻辑却运行在别人的服务器上，这是很奇怪的。
指向原始笔记的链接
It’s weird for the logic to be running on someone else’s server if your system message has your whole business algorithm on it.
—— Diogo Almeida · [74:03]
而且待在轨道上是一个特性，你知道的，你不会想去一个没有轨道的游乐园，或者至少没有轨道、安全带和所有那些东西你就玩不了最好玩的项目，对吧？
指向原始笔记的链接
And it is a feature to be on the rails, you know, like you don’t want to go an amusement park that doesn’t have its rails, or at least you can’t do the most fun stuff without like the rails and the seatbelts and all of that, right?
—— Diogo Almeida · [86:24]
你不能把自己的思考委托出去、把它当作拐杖，因为那样你会得到这些庞大的、你根本解释不了的垃圾代码库。
指向原始笔记的链接
You can’t delegate your thinking away and basically rely on it as a crutch because then you get these massive slop code bases that you can’t actually explain.
—— Yash · [48:59]
这是 13 万亿美元的消费债务，基本上主要运行在一家单一的在位巨头身上，而他们的遗留系统是在互联网发明之前构建的。
指向原始笔记的链接
It’s $13 trillion of consumer debt that basically runs on, for the most part, a single incumbent that built their legacy system before the internet was invented.
—— Linda Du · [04:04]
你基本上是在依据代码自己怎么想的来判断，而不是它实际在做什么。
指向原始笔记的链接
You’re kind of going on what the code thinks, not what it does.
—— Matt Jones · [46:31]
而如果我们真的写出精心打造的、非常具体的提示词，向智能体提出具体的问题、给它们具体的方向，我们就能非常快地得到答案，而且它们似乎不会被带偏。
指向原始笔记的链接
Whereas if we actually write well-crafted prompts, very specific prompts, we’re asking the agents specific questions and giving them specific ways on where to go, We can get answers very, very quickly, and they don’t seem to get misled.
—— Matt Jones · [78:31]
看明年，全美国的 CapEx 将达到大约 2 万亿美元的量级，不仅仅是数据中心和芯片，还包括人们正在投资的供应链的其余所有部分。
指向原始笔记的链接
if you look at like next year, CapEx across the US will be on the order of $2 trillion, not just data centers and chips, but also all the rest of the supply chain that people are investing in.
—— Dylan Patel · [03:01]
卖方市场的态势意味着这些公司的优先事项是能筹集到尽可能多的钱、能尽可能快地建数据中心、能尽可能快地部署芯片，而不是去打磨底层服务的真正质量。
指向原始笔记的链接
the dynamic of it being a seller’s market means that the priority for these companies is how can they raise as much money as they can, how can they build data centers as fast as they can, and how can they deploy chips as fast as they can, rather than working on the true quality of the underlying service.
—— 嘉宾 · [54:03]
因为大家都有同样的原语，因为你需要原语，但你必须在原语之上做专业化。
指向原始笔记的链接
Because everybody has the same primitives because you need primitives, but you got to have specialization on top.
—— Woodson Martin · [62:08]
首先，要说清楚，我有我的定义，你可能有自己的定义，但写那几本书的人是我，所以我的定义说了算。
指向原始笔记的链接
Now, to be clear, right, I have my definition, you might have your own definition, but I’m the one that wrote the books on it, so my definition wins.
—— Sam Newman · [26:04]
但也许这种内部使用最有趣的地方在于，我们实际上是在 Engine 之上运行 Engine。
指向原始笔记的链接
But maybe the most interesting of this dogfooding is that we actually run Engine on top of Engine.
—— Ben · [05:57]
实际上客户每天支付的是几美元几美分，但在 Anthropic 的成本端，补贴高达 150 甚至 750，这种经济模型是没法规模化的，即便是 Anthropic 这个体量的公司也不行。
指向原始笔记的链接
So effectively what the customers were paying daily was dollars and cents, but on the cost side for Anthropic, it was like 150 to even 750 subsidized, which is effectively not an economics that you can scale, even for the company at the size, at the scale of Anthropic.
—— Dor Sasson · [03:15]
其中最重要的一点是，代码已经在磁盘上、在本地了，尤其是如果你的智能体有上下文的话，那为什么要用一个预检索系统把它复杂化呢？
指向原始笔记的链接
And the largest part there is code is already on disk and local, especially if your agent has context, so why complicate it with a pre-retrieval system?
—— George He · [03:49]
这就是为什么 Anthropic 和许多其他公司差不多已经放弃了尝试预索引数据，因为如果你的数据复杂，数据新鲜度真的很难实现。
指向原始笔记的链接
This is kind of why Anthropic, a lot of other companies have kind of given up on trying to pre-index data because data freshness is really hard to achieve if your datas are complex.
—— George He · [15:40]
地球上没有任何逆向工程师曾经能做到那个程度,尤其是在那么短的时间内。
指向原始笔记的链接
No reverser on earth would have gone that far ever, especially not in that amount of time.
—— Mario Zechner · [11:06]
也有一种对软件工程未来的愿景,而且他们不确定自己的愿景是否正确,但他们愿意押上全部身家去赌他们的愿景。
指向原始笔记的链接
Kind of also have a vision of the future of software engineering, and they are unsure of whether their vision is correct, but they’re willing to bet their house on their vision, right?
—— Mario Zechner · [41:23]
举例来说，我们发现你在代码审查上用 GPT 5.2 可以达到和最新 Opus 模型相同的性能，但价格只有一半。
指向原始笔记的链接
We found, for instance, that you can achieve the same performance on code review with GPT 5.2 as with the latest Opus models, but at half the price.
—— Davis Palmie · [08:25]
AI 将代表我们做工作，它们将不得不互相交谈，以便代表我们解决任务。
指向原始笔记的链接
AI will be doing work on our behalf and they will have to talk to each other in order to solve tasks on our behalf.
—— Vlad Luzin · [01:02]
结果发现 Trainium 从绝对性能和性价比的角度来看，可能是目前市场上最好的推理芯片。
指向原始笔记的链接
It turns out that Terranium is actually maybe the best inference chip on the market right now from an absolute performance and cost performance point of view.
—— Matt Garman · [38:03]
所以我想说,NVIDIA GPU 在 transformer 模型的实际解码前向传递中,平均只获得 30% 到 40% 的内存带宽利用率。
指向原始笔记的链接
So while NVIDIA GPUs, I would say, on average are getting between 30 and 40 percent memory bandwidth utilization in the actual decode forward pass of a transformer model
—— Thomas Somers · [34:12]
所以在 token 支出方面,它很快成为我们最大的非制造类支出项目。
指向原始笔记的链接
So on the token spend side, it’s become our single largest non-manufacturing line item very quickly.
—— Thomas Somers · [44:34]
所有这一切的真正限制,是人类 token 带宽的总和,而这实际上就是我们能生产的硅的数量。
指向原始笔记的链接
The real limit on all this is the sum total of humanity’s total token bandwidth, which is literally the amount of silicon we can produce.
—— Sean Wang · [60:25]
所以我确实认为 SFT,尤其是结合多个教师的输出,将是未来企业定制其模型的关键部分,我们也在投资这方面的研究。
指向原始笔记的链接
So I do think SFT, especially combining the output of multiple teachers, will be a key part of how enterprises in the future customize their models, and we’re investing in research on that as well.
—— 嘉宾 · [73:37]
如果你真的认为这东西会杀死你，你不必用自杀来证明——我们是在拿着一颗星球玩俄罗斯轮盘赌，而每一枪没打死我们，都可能让我们变得更强。
指向原始笔记的链接
If you really think that this thing is going to kill you, you don’t have to suicide in order to demonstrate that we are pulling trigger on this Russian roulette with a planet, and every shot that doesn’t kill us will make us stronger, potentially.
—— Jaan Tallinn · [00:45]
然后我们挑选那些在某个给定测试上做到我们要求的事情的实例，这意味着我们是在基于外部行为来挑选，而不是基于内在动机。
指向原始笔记的链接
And then we like pick the ones that kind of do the things that we perform on some given test, which means that we are selecting based on outer behavior rather than based on the inner motivations.
—— Jaan Tallinn · [11:50]
所以我们正在扣动扳机，就像用整个星球玩俄罗斯轮盘赌。而且每一次没杀死我们的射击都可能让我们更强、更富有、境况更好。但你什么时候才停止扣扳机，对吧？
指向原始笔记的链接
So we are pulling a trigger on this like Russian roulette with a planet. And every shot that doesn’t kill us will make us stronger potentially and wealthier and better off. But when do you stop pulling the trigger, right?
—— Jaan Tallinn · [29:31]
我觉得这是现在很多人都在犯的一个非常重大的错误，即认为可能发生的最糟糕的事情只是互联网上出问题。但其实是，不，这不是互联网的问题。
指向原始笔记的链接
I think that’s like the really big current mistake that a lot of people now make is that the worst thing that possibly could happen is something going wrong on the internet. It’s like, no, it’s not the internet.
—— Jaan Tallinn · [47:37]
我还没见过任何基准测试表明图数据库——原生图数据库——能在那些关系型数据库系统拥有正确的 SQL 或 API 构造、从而让你完全在服务器端做图遍历、不用来回往返的情况下，胜过关系型数据库系统。
指向原始笔记的链接
I have not seen any benchmarks that show graph databases, the native graph databases, outperform the relational databases systems when those related systems have the right constructs in SQL or the API so that you can do graph traversal all on the server side and not go back and forth.
—— Andy Pavlo · [22:23]
我学到的最重要的一点是，当我弄明白某件事时，很难不去付诸行动。
指向原始笔记的链接
The biggest thing I learned is that when I figure something out, it’s really hard to not act on that.
—— Alyssa Ravasio · [04:30]
想想上一次所有农民、所有环保人士和所有乡村小企业在一件事上达成一致是什么时候。
指向原始笔记的链接
Like when’s the last time like all the farmers and all the conservationists and all the small rural businesses all agreed on something.
—— Alyssa Ravasio · [21:15]
某一天你很兴奋，因为你在用 ChatGPT 之类的东西工作，一切都像，哦，你对自己感觉很好。然后下一分钟你觉得，哦天哪，我觉得我这么渺小，我在这里没有位置。
指向原始笔记的链接
One day you’re up and excited because you’re you’re you’re working on with collater chat chat GPT and everything’s like, oh, you’re feeling great about yourself. And then the next minute you feel like, oh, gosh, like, I feel like I’m this small, and I I have no place here.
—— Reef Frerichs · [28:06]
只有当那个经验教训被转化为智能体在未来运行中可以参考的持久性上下文时，它才成为记忆。
指向原始笔记的链接
It only becomes memory when that lesson is converted into durable context that the agent can reference on future runs.
—— Jake Broekhuizen · [04:36]
大多数事情自上而下推动会相当困难，但有几件你认为是信仰的事情，你必须自上而下地推，并且必须在这些事情上毫不松懈。
指向原始笔记的链接
Most things it’s gonna be pretty hard to go top-down, but a few things that you’d think are religion, you have to go top-down and you have to just make sure that you’re unrelenting on those.
—— Jeetu Patel · [13:58]
如果只有不到 2% 的人在使用智能体，而这些智能体的消耗这么大——平均而言，一个智能体在网络带宽上的消耗比人类高出 450%。
指向原始笔记的链接
If sub 2% of the people are using agents and these agents are so consumptive, on average an agent is 450% more consumptive on network bandwidth than a human would be.
—— Jeetu Patel · [22:30]
我留在那里是因为我有这样一种自尊心：我绝不为别人打工。
指向原始笔记的链接
I stayed on there because I had this ego that I will never work for someone else.
—— Jeetu Patel · [32:22]
每当有人建议我们应该增加复杂性，无论是流程中的一个步骤、一个功能、一个团队、一个团队的人头数，随便什么，系统的任何元素，举证责任都在提出建议的人身上。
指向原始笔记的链接
Every time someone recommends that we should add complexity, whether it’s a step in a process, a feature, a team, headcount to a team, you name it, any element of the system, the burden of proof is on those making that recommendation.
—— Luca Ferrari · [54:28]
我的真实想法是，我们应该做多得多的工作去和世界上其他人对话，少一些把注意力放在那 1% 会关心这些事的 Twitter 用户身上。
指向原始笔记的链接
My honest thoughts are that, like, we ought to be doing, like, way more to talk to the rest of the world and, like, focus less on, like, the 1% of Twitter that’s going to care about, like, what’s going on.
—— Holly Li · [61:30]
但我们心中始终坚守的一个关键原则是：每一条落到上下文图谱上的数据点，都必须扎根于一个硬事实。
指向原始笔记的链接
But one key principle that we had in mind always was every single data point that landed on the context graph always had to be grounded down in a hard truth.
—— Kamalakannan Nandagopal · [05:34]
Vinod 可以那么做，但 Vinod 也可以因为希望某个东西存在而作为非营利资助给某人十亿美元。这和一个试图做出杰出投资、在世界上留下印记的普通 VC 是截然不同的。
指向原始笔记的链接
Vinod could do that, but Vinod can also give a billion dollars to somebody as a non-profit grant because he wants that to exist. That’s very different than necessarily an average VC who’s trying to do remarkable investments and make a mark on the world.
—— Bilal Zuberi · [07:31]
你知道，它们不会全都赢，它们不会全都赢，因为归根结底，你必须在这些东西上获得规模。
指向原始笔记的链接
And you know, they’re not all going to win and they’re not all going to win because at the end of the day, you have to get scale on these things.
—— Pat Gelsinger · [18:09]
而且我认为你会看到越来越多的数据中心项目出现违约，因为能源不会到位。
指向原始笔记的链接
And I think you’re going to see more and more defaults happening on many of those data center projects because the energy won’t be there.
—— Pat Gelsinger · [45:28]
技术上讲，你可以通过 MCP 使用其他认证机制，但如果你用 bearer auth 或类似的东西，你会非常难受，因为规范非常偏重 OAuth，而且所有的客户端，至少是主流的那些，都非常倾向于把 OAuth 作为安装该连接器的更好用户体验。
指向原始笔记的链接
And technically, you can use other authentication mechanisms through MCP, but you’re really going to have a hard time if you do like bearer auth or anything like that, because the spec is really heavy on OAuth, and all of the clients, the main ones at least, are really leaning towards the OAuth as a better UX for installing that connector.
—— Pedro Lopez · [07:15]
这场对话，以及我最近其他几场关于 AI 监控与控制未来的对话，得出的底线是：思维链正变得越来越不可靠。
指向原始笔记的链接
The bottom line from this conversation and others I’ve had recently on the future of AI monitoring and control is that chain of thought is increasingly unreliable.
—— 嘉宾 · [03:03]
所以不仅仅是银行家，我认为市场上有很多被错误定价的资产，他们的简历不会显示出真正的 alpha
指向原始笔记的链接
So not only bankers, but I think there’s a lot of mispriced assets on the market where their resume will not suggest the true alpha
—— Max Freeman · [05:20]
所以我们早期下的赌注是，实际上为这些人才支付远高于市场水平的薪酬。
指向原始笔记的链接
So the bet we took early on was to actually pay way above market rate for this talent.
—— Max Freeman · [07:08]
如果你是一个想加入早期公司的销售，把股权、现金和薪酬作为核心考量指标其实是错误的做法。
指向原始笔记的链接
If you’re a seller looking to join an early stage company, thinking about equity, cash, and comp is actually the wrong thing to index on.
—— Max Freeman · [14:35]
我认为我们犯过的唯一错误就是过度看重履历光环和大牌公司的出身，而没有真正理解两件事。
指向原始笔记的链接
I think the only mistakes that we’ve made are over-indexing on pedigree and coming from the big name brands and not truly understanding two things.
—— Max Freeman · [33:04]
如果你在管理一个大的销售团队，而只依赖 enablement，你会死的。
指向原始笔记的链接
If you’re running a big sales team and you’re just dependent on enablement, you’re gonna die.
—— Max Freeman · [38:08]
在这件事上我比较倾向 Degnan 那一派：如果你效率高到那个程度，你大概是在把钱留在桌上，你需要更多销售。
指向原始笔记的链接
I sort of lean in the camp of Degnan on this one, where if you’re that efficient, you’re probably leaving money on the table and you need more reps.
—— Max Freeman · [40:05]
我希望你做的是，去找 10 个人、20 个人，给他们展示这项技术，然后访谈他们，了解他们有什么问题等等，试图找到一个「头发着火了」般的问题，就是一个真正的问题。
指向原始笔记的链接
What I hope you do is you go to 10 people, 20 people, you show them the technology, and then you interview them about what their issues are, et cetera, and try to find something that is a hair on fire problem, like a real problem.
—— 嘉宾 · [32:41]
因为我们的未来不应该取决于谁拥有矿山，而应该取决于附近生长着什么。
指向原始笔记的链接
Because our future shouldn’t depend on who owns the mines, it should depend on what grows nearby.
—— 嘉宾 · [37:14]
基于你所知道的去创建一个新生态系统是疯狂的。
指向原始笔记的链接
Creating a new ecosystem based on what you know is insane.
—— 嘉宾 · [52:50]
另一方面，如果AI的风险下降得太慢，那么你也不想等待，因为我们在此期间就在不断死去，等待所损失的预期寿命年数比降低风险所获得的更多。
指向原始笔记的链接
On the other hand, if the risk of AI goes down too slow, then you also don’t want to wait because like we are just dying in the meantime and you lose more expected life years from waiting than you gain from having the risk.
—— Nick Bostrom · [00:17]
团队的口号是：用户第一，Google 第二，我们的产品第三。
指向原始笔记的链接
The mantra on the team is users first, Google second, and our product third.
—— Josh Woodward · [21:00]
他们的工作就是站在前沿，如果他们在过去 24 小时内没听说过 X 上最新爆火的东西，就好像在问：你上班都在干什么？
指向原始笔记的链接
Their job is to be at the frontier and they’re almost like if they haven’t heard about the latest hot thing viral on X in the last 24 hours is like, what are you doing at your job?
—— Josh Woodward · [44:15]
如果你无法承受这种复杂性，因为你需要他要么是恶棍要么是英雄，那你根本没有在认真看。
指向原始笔记的链接
And if you can’t take on that complexity because you need him to be either villain or hero, you’re just not paying attention.
—— 嘉宾 · [101:40]
所以我们所有的筹码都押在这一单交易上，成或不成。
指向原始笔记的链接
So all of our stakes was just on this one deal, working or not.
—— Daniel Saks · [05:18]
但在早期，我们说 GMB 是我们的第一指标，但我们花了大概五到七年才真正搞明白怎么让 GMB 的飞轮转起来。
指向原始笔记的链接
But in the early days, we said that GMB is our number one metric, but it took us probably five to seven years to figure out how to really get the flywheel going on GMB.
—— Daniel Saks · [06:24]
在拼命赶工把东西发布出来之后，我们按时上线了。结果门可罗雀。没有人来应用商店。
指向原始笔记的链接
And after crunching to get this thing launched, we launched on time. It was like crickets. No one came to the App Store.
—— Daniel Saks · [07:14]
我认为对大多数公司来说，现实不是这样。它们要么有一个小 TAM,这意味着你必须在激活那个 TAM 时更加小心；要么如果它们进攻太宽的 TAM,就会陷入拉力不够强的问题。
指向原始笔记的链接
I think for most companies that’s not the reality. They either have a small TAM which means you have to be much more careful on how you activate that TAM, or if they go after too wide of a TAM they’re going to fall into issues where the pull is not strong enough.
—— Daniel Saks · [32:41]
当每个人都想要个人智能体，或者每个企业都想要自己的、全天候 24-7 持续在线并运行大量自动化的个人智能体时，上下文窗口比内存中的模型权重更重要。
指向原始笔记的链接
When everybody wants personal agents or every business wants its own personal agents that’s always on and working 24-7 and running a lot of automations, the context window matters even more than the model weights in the memory.
—— Arvind Srinivas · [30:26]
因为当人们那样谈论无头时，如果你放弃了自己的用户界面，如果你放弃了为客户演进你的体验以解决你当初要为他们解决的问题，如果你不是那个体验的重要组成部分，是的，我认为那是一个无脑的赌注。
指向原始笔记的链接
Because when people talk about headless in that way, if you’re giving up on your user interface, if you’re giving up on evolving your experience for a customer to solve the problems that you set out to solve for them, if you are not a significant part of that experience, yes, I think that’s a brainless bet.
—— Mike Cannon-Brookes · [05:31]
我非常看空 AI 的那个魔法按钮论，即会有一个按钮，我们按下它，它就为我们做好一切。
指向原始笔记的链接
I’m a big bear on their magic button thesis of AI, that there’ll be one button we press and it does it all for us.
—— Mike Cannon-Brookes · [09:32]
目前，构建专用模型的好处还没有定论。
指向原始笔记的链接
The jury is really out on the benefits of building specialized models at the moment.
—— Mike Cannon-Brookes · [21:05]
我更大的担忧是，这件事本来应该是难的，我完全同意你对创始人所面临挑战的看法，但我们现在把这一切变得太正常了：一个人出来融个一千五百万美元。
指向原始笔记的链接
My worry is more that it should be hard, and I completely agree with you on the challenges of founders, but we’ve made it so normal now for someone to come out, raise 15 million bucks.
—— Harry Stebbings · [48:32]
当我看完这份清单时，留在我脑海里的印象是：哇，这是数学和计算机科学领域令人难以置信的进展，但这里没有任何关于密码学的问题。
指向原始笔记的链接
And as I finished the list, the thing that kind of stuck in my mind is that, wow, this is incredible progress in mathematics, in computer science, but there are no questions here on cryptography.
—— Dan Bonnet · [03:37]
但真正的价值在于能无人监督地连续运行数小时的任务或者说工作流。
指向原始笔记的链接
But really the value is about tasks or like workflows that go on like unsupervised for hours.
—— Jason Ma · [48:42]
令人失望的是，我们可能因为某种我们没注意到的流程上的差异而丢失了多少发现。
指向原始笔记的链接
It’s disappointing the amount of discoveries that we probably just lose to some sort of procedural variation that we haven’t picked up on.
—— Renee Wigerzen · [11:40]
最终，我意识到还有另一根路灯柱，现在没有人往它下面看，而如果我真正把我的努力放在那里，就有大得多的机会真正对这个领域产生巨大的影响。
指向原始笔记的链接
And in the end, I realized that there’s this other lamppost that right now no one is looking under and where if I actually put my effort there, there’s a much higher chance of really having a massive impact on the field.
—— David Silver · [05:29]
所以我们正在下一个赌注，我们下的赌注是，通过放弃如今每个人都依赖的先验，比如人类数据，比如其他所有人构建其系统所依托的整个互联网知识，我们可以走得更远。
指向原始笔记的链接
So we are making a bet, and the bet we’re making is that we can go further by giving up on the priors that everyone’s relying on today, like the human data, like the whole internet of knowledge that everyone else is building their systems upon.
—— David Silver · [28:36]
在我们看来，很明显 Muse 是被基准测试刷出来的，也就是说他们训练它让它在公开基准测试上看起来非常好看，但这些提升不一定能转化到你的日常工作中。
指向原始笔记的链接
To us, it seems pretty clear that Muse is benchmarked, which means they trained it to look really good on public benchmarks, but those gains don’t necessarily translate into your everyday work.
—— 嘉宾 · [00:36]
我记得有人说过，Tesla 是轮子上的 iPhone。
指向原始笔记的链接
I think someone had said that Tesla is an iPhone on wheels.
—— 嘉宾 · [17:50]
我想我的最大区别在于它实际上是在你的电脑上运行的。我目前看到的所有东西都在云端运行。它可以做几件事。如果你在电脑上运行它，它可以做每一件该死的事情。
指向原始笔记的链接
I think my big difference is that it actually runs on your computer. Everything I saw so far runs in the cloud. It can do a few things. If you run it on your computer, it can do every effing thing.
—— Peter Steinberger · 来自原文
② 出现在这些集
1 集
- 《16 年的一夜成名:ON 如何打造未来最高端运动品牌》 — 作为被讨论公司
③ 关联
点进去有真内容 —— 本页主要出口
Bob Safian · David Alleman · light spray · Clean Cloud 泡沫 · 创新 · 高端(非奢侈)定位 · 运动阶层 · 股权 · 合伙制领导 · 社会变迁造品牌