The State of Startups in 2026 · 苏菲拉底
字幕 字幕位置
--:--
点击播放,这里会跟随视频显示当前句的中英字幕。

The State of Startups in 2026

节目发布 2026-09-17 · Y Combinator
加里·谭 贾里德·弗里德曼 胡佳怡
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
每年与数千名创业者打交道,YC 的合伙人们往往比外界更早看到风向的转折。在这期《光锥》(The Light Cone)里,YC 总裁兼 CEO 加里·谭、管理合伙人贾里德·弗里德曼与合伙人胡佳怡坐在一起,用最近两轮批次的内部数据,拆解 2026 年创业公司的真实状态:硬科技占比从 8% 蹿到 20%,中位营收在三个月里翻了一倍多,单人创始人比例翻了近三倍。他们谈国防、金属、光互连、机器人基础模型,也谈 SaaS 到底死没死、记录系统为什么必须变成 harness。本文依据现场录音编译整理。

开场:先看数据

加里: 欢迎回到新一期《光锥》。在 YC,我们每年要和数千位创始人合作,这意味着很多趋势在它变得显而易见之前,我们就已经看到了。所以今天想把其中一部分分享给各位:什么是当下的前沿,接下来会发生什么,作为建造者的你应该知道些什么。我们开始吧。胡佳怡,你手上有几个数据要跟大家讲。

胡佳怡: 我们把过去 12 到 18 个月录取的所有公司做了一次分析,有几个相当惊人的数字。第一个是批次里硬科技公司的数量,占比已经从 8% 涨到了 20%。背后有不少原因,我们待会儿可以往下挖。第二个是公司的增长速度,以及 YC 这三个月对公司的加速效应,本身也在加快。被录取时,YC 公司的中位状态是零营收、没有产品、没有收入;过去到批次结束时,中位数大概能做到 8000 美元月营收,而现在这个中位数是 20000 美元。这是最值得展开的两条。

硬科技从 8% 到 20%

贾里德: 我们先说硬科技。这些硬科技公司具体是什么?是什么在推动它?

加里: 是真的碰原子、而不只是碰比特的东西。

贾里德: 胡佳怡,你那边有硬科技的细分类别拆解,对吧?

胡佳怡: 有。具体来说,机器人是很大的一块,从占批次的 1% 涨到了大约 6% 到 7%。工业制造,也就是把东西重新在美国造出来,是个巨大的趋势,从大约 4% 涨到了 10%。另一个大头是国防,我们所有人最近都在和大量国防创业公司合作,它从大约 1.5% 涨到了 5% 左右。再一个是 AI 带来的全球算力缺口,所以有很多公司在做半导体栈或者光子学,一年前占批次约 1%,现在接近 4%。比算力更下面一层是电力,做电力基础设施的公司也从 1% 涨到了接近 3%。所以在整个物理原子这一侧,各个数字要么翻了三倍,要么翻了五倍。

加里: 我觉得这是机器的时代。我们的座右铭,印在 T 恤上的那句,是「做人们真正想要的东西」。而现在,人们是真的想要这些东西。

博士创始人与代码生成

胡佳怡: 这些往原子深处走的公司还有一个有意思的地方:我们在资助比以往任何时候都更技术、更有专业积累的创始人。对吧,贾里德?关于今年夏季批次,我们有一个很有意思的数字。

贾里德: 在这一期夏季批次里,每六位创始人中就有一位拥有博士学位。这个比例远高于历史水平。原因也很直接:如果你要做的是硅光子这类东西,你多半需要非常扎实的研究背景。所以我们资助了远比过去更多的这类创始人,而且这批创始人的表现,在整体中是明显偏好的。

加里: 我觉得 AGI 在这件事上还有一层特别迷人、也特别爽的叠加效应。过去你会认为硬科技之所以难,是因为要处理供应链,还因为里面往往有极重的软件成分。帕尔默·拉基(Palmer Luckey)就反复讲过这一点。现在有了代码生成,Anduril 在做的那些事情,突然之间可以快得多地完成。哪怕是三四年前,大家谈软件工程时还会说,顶级软件工程师是能不能做出真正一流的全栈硬件的限制性试剂之一。现在这件事越来越不成立了。你当然还是需要一两个这样的人,或者一个小团队,但你不再需要去和谷歌、Meta 或者别的谁抢着招一千名优秀工程师。这从根本上改变了经济账。

贾里德: 这才是硬科技真正的多头逻辑。不只是大家不愿意投软件生意了,而是我们现在拥有的这些超级聪明的模型,实实在在地在加速科研,让创业公司能在更早的阶段做出更大的研究突破,因此这些深科技公司会真的跑得通。

太空与国防的新一代

胡佳怡: 另一个因素是,推动这波「原子热」的大致有三条宏观趋势。第一条,看到 SpaceX 这样的巨头成功 IPO,催生了一整代想在太空里造东西的创始人。他们分布在整个技术栈上。比如 2026 年夏季批次里有一家我们合作的公司叫 Exosat,想做的基本上是一套主权版的 Starlink。2026 年冬季批次里我还带过一家叫 Beyond Reach Labs 的公司,做的是卫星在太空用的太阳能板。你想想 StarCloud 这类想把数据中心搬上太空的公司,它们一定需要电力,那太阳能板就是个显而易见的解法。第二条宏观趋势是,现在这一代创始人是在战争极其贴近日常、在社交媒体上被反复讨论的环境里长大的,他们想为此做点什么。

加里: 我过去几个批次里资助的公司中,最让我兴奋的有两家,一家是去年秋天的 Icarus,另一家是今年春天的 Nine Mothers,两家都是国防。Icarus 在做的是太阳能动力的 U2 式侦察机,能提供空中监视,同时还能做通信中继,而通信其实非常关键:未来无人机战争的核心,是你能不能真的和地面上的无人机保持通信、看到战场上正在发生什么。他们已经拿下了与新设的战争部(Department of War)的七位数合同。同样是无人机战争这条线,特种部队一直在采购 Nine Mothers 的反无人机防御系统,它本质上是一个带计算机视觉的霰弹枪炮塔。但这几乎是唯一能保护深入敌后的特种部队的办法。你要知道,这些人受训多年,属于极精英的部队,美国并没有几千个这样的人,我们的总量非常非常小。所以保护他们不被一架可能只是量产货的无人机袭击,对战争部来说是关乎生死存亡的事。

加里: 所以看到新一届政府用一种完全不同的方式来对待国防,真的很酷。过去很长时间里,说白了就是被几家国防巨头把持,做的是成本加成那一套,他们把自己当成咨询公司。现在能看到新的创业公司真正利用上所有的 AI、所有的技术、所有新的建造方式,去做出那些巨头老实说根本做不出来的东西,这是眼下一股非常强的大趋势。

把金属供应链搬回美国

胡佳怡: 国防这件事还不只是那些卖给政府的整体方案。还有一大类是军民两用的创业公司,同时卖给私营部门和政府,涉及整条供应链上的一切,比如把制造搬回美国、做定制化的部件。贾里德,你带的那家 Nox Metals 就是这样吧?

贾里德: 对,他们在把金属制造带回美国。美国基本上已经失去了自己的金属工业,过去几十年被掏空了,而没有金属你什么都造不出来。Nox Metals 做的就是重建美国的金属供应链,而且他们把这件事放在美国腹地的底特律,那里有大量空置的工厂,基本就那么一直闲置着。

加里: 他们也是另一个趋势的例子:现在不只是有更多人在做硬件公司,硬件公司本身的增长速度也前所未有。我记得看到 PG 发过一条推,说 Nox Metals 的增速是软件公司的增速。你了解他们为什么能长这么快吗?

贾里德: 一个原因是,他们的很多客户就是这一批新冒出来的国防科技创业公司,造东西都需要金属,而现有的供应商是那种昏昏欲睡的老生意,基本由上了年纪的人经营,完全跟不上新国防公司想要的节奏。这让我想起 YC 早年 Web 2.0 热潮的时候:我们有一堆新创业公司,而它们更愿意从另一批新创业公司那里采购,因为对方能跟上自己的速度,也更好合作。比如 Stripe,你当然可以用传统的信用卡供应商,但和 Stripe 合作就是舒服太多。我觉得 Nox Metals 正在成为整个国防科技生态的那个角色。

算力、芯片与光互连

胡佳怡: 第三条趋势是算力。要让这么多数据中心尽快上线,是一个非常重的物理原子工程,因为我们一直在讨论的 AI 需求正在飞涨。这里有个很有意思的数据:英伟达的 GPU,比如 A100,按小时计的价格居然在升值,这很反常,因为放到今天 A100 已经算是老卡了。

加里: 确实挺老了。

胡佳怡: 但价格还在涨,因为需求太多、供给和算力都不够。所以现在有大量创业公司在做让数据中心尽快上线这件事,从场地建设、到规划用的软件、到真正的数据中心部署,再到怎么给它们供电的各种方案,包括能源和电池的组合。这一整类公司都有。再往下到核心的计算硅,也有一批公司在做英伟达之外的新芯片。泰勒带过一家叫 Lamb Labs 的公司,在做新的计算处理器。我这一批还带着一家叫 Bot 的公司,想做的是一种全新的定制硬件架构,用三进制(ternary)表示来跑模型。因为现在有个很好玩的趋势:你去看英伟达从 A100 到 H100 再到现在 B300 的每一代架构,浮点精度其实是一代比一代低的,从 FP32 到 16、到 8,一路往下。事实证明,大模型的架构并不需要完整的浮点精度。

贾里德: FP2 甚至都有点能用。

胡佳怡: 对,这正是 Bot 想做的事。贾里德,你那边还有一家做光子互连的,挺有意思。

贾里德: 对,这一批里有家公司叫 Dipole Labs,他们要替换的是数据中心里的交换机,也就是不同 GPU 之间的路由系统。如果 GPU A 想和 GPU B 通信,它们就是通过这个设备来对话的,这个设备叫交换机。现在的交换机都是电子的,于是出现了一个问题:交换机跟不上 GPU 了。GPU 的速度一直在涨,而在很多数据中心、很多种负载下,交换机反而成了瓶颈。Dipole Labs 在做的是第一台全光交换机,从 GPU A 到 GPU B 全程都是光子。这会比我们现在用的电子交换机快得多。

机器人的 GPT-3 时刻

胡佳怡: 推动这轮向原子迁移的最后一股力量,是机器人这件事一定会发生。所以有大量公司在围绕它搭建整个栈:从特定行业的垂直机器人,到部署机器人的基础设施,再到给新的机器人实验室卖数据。行业里所有人都感觉到,我们即将迎来机器人的 ChatGPT 时刻,只是还差那么一点。我觉得大家正在把它想明白,也在摸索属于它的新的规模化定律。这一块东西很多,前几期我们请来过 Quan,我们都相信这件事会发生。

贾里德: 机器人,来自 Physical Intelligence 那边的路线。

胡佳怡: 对,来自 Physical Intelligence。一半是 AI,一半是硬件。

加里: 我今天早上还读到,连 Astra 都是机器人领域的一次大跃进。我忘了是哪个基准测试,在那个基准上 Fable 大概只能做到 10%,而 Astra 显示它能完成 60% 到 70% 的任务。

加里: 所以就是这样,过几个星期你一觉醒来,又一个突破发生了,我们又近了一步。

硬科技融资与 SaaS 的护城河

贾里德: 看到硬科技的复兴,对我个人来说特别过瘾。因为 YC 从 2014 年就开始资助硬科技公司了,那真的是我们起步的时候。但在这轮复兴之前,这些公司融资是很难的。我记得在那之前,我们会资助一些让我们超级兴奋的东西,火箭、飞机、芯片、数据中心之类,然后 VC 会说,啊,我们只做 B2B SaaS。而这种公司又很难自举,你确实需要有能力承接后续融资、能一路支撑完整资本开支的下游投资人。所以现在看着硅谷回到自己的根上,挺酷的。风险投资历史上本来就是为了资助硬科技而设立的,只是因为投 SaaS 实在太赚钱,它漂移了一二十年。

加里: 说到投资人重新愿意做硬科技,我从来没见过转向发生得这么快。SaaS 股票在今年早些时候下跌、Claude Code 势头正猛的时候,这件事几乎是立刻发生的。我感觉大概就在今年年初冬季批次进行到一半的时候,到 Demo Day 那天,投资人对硬科技公司的兴趣已经明显大了很多,之后就是一路外推。不过另一面也值得知道:从那之后,一批 SaaS 股票其实已经反弹,而且表现比以往任何时候都好。Salesforce 就是最典型的例子。Snowflake 前两天也刚发了一份爆表的财报。

胡佳怡: 也许两边都能成立。

加里: 也许吧,那当然是最理想的情况。但我还是觉得我们看到的是真东西。未来会被造出来的软件、以及什么东西有价值,显然是不一样的,这是必然的。所以 Salesforce 身上正在发生的事,部分看起来是那个经典的记录系统(system of record)论点正在应验。

贾里德: 至少目前护城河还在。

加里: 对。如果你手上的东西是 agent 能用的,那它就真的有价值。甚至可以说,agent 使用软件的强度会远远超过人类,而这似乎正是 Salesforce 增长的动力。这又回到我们之前聊过的另一个趋势:如果你把 agent 当成你的客户,做 agent 想要的东西,把你的软件做成 agent 愿意用的东西,那大概就是对的那类软件。

AI Harness 之争

加里: Salesforce 特别有意思,因为他们开始发布自己的 Slack harness,也就是 Slack 的 AI 外壳。我觉得我们正站在下一轮 AI harness 大战的开端。Codex 想当那个 harness,Claude Code 想当,OpenClaw 可能当,还有 Hermes、OpenCode。看起来会冒出一堆,而且不太会像浏览器大战那样最终只剩一个赢家,不过这谁也说不准。另外贝尼奥夫(Benioff)有个相当大的优势:世界上最 AI 化的一批公司里,仍然有很多在用 Slack。如果 harness 就长在里面,而它又是大家协作的记录系统,那你就握有一条巨大的数据护城河。只要护城河守得住,SaaS 依然可以像过去一样值钱。

贾里德: 你大概一周前发的那条推我觉得特别有意思,说的是软件公司、或者说记录系统公司,未来必须变成 harness。

加里: 那条基本上是在说 Slack。逻辑是:如果你是一个记录系统,你要么被猎食,你发个 MCP,然后围绕数据的护城河就没了,数据流向别处,迁移变得非常轻易;要么你就得成为 harness,你得成为人们不只是读写、而是真正在你的系统里完成工作、榨取最大价值的那层东西。

贾里德: 这有点像模型公司那边的情况。是哪个基准来着,ARC-AGI……

胡佳怡: ARC-AGI v3。

贾里德: 对。Astra 之前那一代 ChatGPT 模型在上面表现不太好,然后他们说,那只是因为它被接到了错的 harness 上。所以是模型加上 harness,才给出最终的产出。

加里: 是啊。用定制的 harness,他们声称 Astra 在 ARC-AGI 3 上拿到了 90% 以上。

胡佳怡: 对,而几个月前这个数字还停在十几,这个跃升相当惊人。

端到端替人干活的 agent

胡佳怡: 你关于软件的判断我觉得很对。并不是像互联网上很多人讲的那样,软件和 SaaS 已经死了,而是它变形了。我们在批次里就实打实看到了这一点:我们录取的公司中,做全栈、端到端完成某项工作或任务的比例,已经从 10% 涨到超过 25%。它的本质是真的把活干了,是 agent 把活干了,而不只是一个点状解决方案。五年前、八年前的老 SaaS 就是一个点状方案,你还得雇个人去操作那套软件。现在它自己就跑起来了。而在批次里,营收增长恰恰就出现在这一类公司身上。我前面给过那个数据:公司进 YC 时中位营收是零,过去到批次结束时中位数大约是 8000 美元月经常性收入,现在是 20000 美元左右。

贾里德: 这是个巨大的跳跃,对中位数来说是很不平凡的涨幅,平均数还要更高。

胡佳怡: 而这和真正端到端把整份工作做完直接相关。比如真的去做保险经纪的活,真的去做临床接诊,真的把医疗账单的全流程跑完等等,这些就是长得特别快的公司。还有另一个因素,我们前几期也聊过:到现在,从 Opus 4.5 让 agentic coding 真正可用算起,差不多过去了一年,我们看到这些工作流彻底开花结果。结果就是,人们想要的是自己的活被干完,并且愿意为「把活干完」的软件付钱。

贾里德: 当人们听到营收数字涨得这么快,一个很容易的质疑是:这会不会只是 AI 泡沫,这些公司只是因为赶时髦才在 AI 产品上撒钱。公平地说,其中确实有一部分是这样。但多头逻辑其实是我们两年前某一期节目里就说过的:在 agent 刚刚开始成立的时候我们就讲,产品本身会变得更有价值。如果它把整份工作自动化了,它就会比一个只记录工作、却不干活的记录系统更有价值,因此公司自然会为它付更多钱。在我合作的公司里我确实看到了这一点:他们找上一家公司,价值主张强到大型企业在非常非常早期就愿意开出大额支票。

加里: 我看到一家产生这种效应的公司是 Juicebox,一个 AI 招聘工具。过去这两年它一直保持惊人的增长率。但它最初,我会说本质上是一个 LLM 驱动的人才搜索:你输入你想招的人的画像,它能相当好地把你可能想联系的人的资料捞出来。可你还是得自己去联系这些人。最近他们上线了一个 agent 产品,起飞得非常快。这个 agent 不只是找人,它会去联系人,能安排面试,能做一堆事。

贾里德: 这太棒了。

加里: 而且他们看到,单就每个客户账户来说,这会让他们从单个客户身上赚到的收入翻一倍甚至两倍,因为客户想要越来越多这样的 agent。我不认为可以说它是在把招聘人员的工作整个自动化掉,它只是改变了这份工作。招聘人员本来也不想干那种给 500 个人群发触达的机械活。让招聘这份工作更有技术含量、更有意思的,是文化契合度这类判断,让 AI 打一通电话去评估某个人在文化上合不合适,这会非常难,还有其中的人情成分。所以他们发现,招聘人员自己其实非常乐意用这些 agent,因为它把人解放出来,去做那些他们觉得独特且有趣的工作。

三个月冲到七位数

胡佳怡: 还有一个数据很惊人:那些在批次期间真正加速、真正起飞的公司。今年我们第一次经历了过去从未经历过的事,有公司在批次期间从零做到七位数营收,而那只是三个月。这太震撼了。放在过去,一家公司要走到那个位置,需要 18 个月甚至更久,而他们在三个月里做到了。一部分原因是他们解决的是真问题;另一部分是因为 agentic coding,他们做出来的产品成熟度也高得多。这些彻底吃透 AI 的创始人,可能同时开着 20 个编码 agent 的会话,才把产品做到那种成熟度。

卖数据与 RL 环境的生意

贾里德: 还有一类公司最近也长得极快,就是向实验室出售数据或者强化学习环境(RL environments)的公司。

贾里德: 这一类值得聊聊,因为这些公司大多相当隐蔽。跟大多数喜欢宣扬自己做得多好的公司相反,它们有动机不去讲自己做得多好。所以外面的人可能没有意识到这个类别已经变得多大。2016 年 YC 资助 Scale 的时候,这还是个极小的利基,根本算不上一个类别。一开始基本上只有 Scale 在做,然后 Mercor 开始做,接着又有几家。但最近这两年它变成了一个大类别。我们最近拉了数据:仅仅在过去两年里,YC 资助的公司中就有十几家,每一家靠向实验室卖数据或 RL 环境,年收入都超过 1000 万美元。

加里: 而且不少情况下是几亿美元。

贾里德: 对,一年做几亿美元,而这些公司才成立两三年。

胡佳怡: 老实说,这个营收速度相当快。

贾里德: 是啊,简直离谱。加里,你想聊聊其中哪几家吗?

加里: 大的那几家,Afterquery 和 Datacurve 都非常非常棒。可能还有太多家没法一一点名,而且老实说它们大概也不想被提到,毕竟当你手上有个跑得通的东西时,你几乎不希望别人知道。不过这件事本身很好理解。数据是规模化定律的支柱之一,大家把算力讲得很多,可没有数据,你要怎么把这些模型做得更好?RL 环境这件事挺有意思,里面东西很多。有大量针对具体用例的纯定制,比如金融的 RL 环境,有人可以在这上面无限深挖下去,它需要一点专业知识、一堆计算机科学、一些系统工程。RL 看起来是主要的引擎之一,虽然大家可能有点过度地在刷榜,但这件事是要花钱的,而且从大模型公司的做法来看,它似乎会长期留下来。

胡佳怡: 据说大实验室在这上面的花费大约是十亿美元量级。这不是什么广为人知的事实,但围绕它确实能建起一门真正的生意,RL 环境是当下的流行形态,还有长周期任务这类东西正在被搭建起来。我觉得这个趋势也开始出现在机器人领域了。实验室同样想解决让 AI 在物理世界里работать的问题,所以他们需要大量真实世界里的环境。于是第一人称视角数据、灵巧操作任务开始成为一个大的数据类别,实验室在和这些公司签八位数、九位数的合同。我们批次里有好几家公司在做这个,并且已经在这个领域拿下了营收。比如 2026 年夏季批次里的 Practis Robotics;我还带着一家公司,在全世界各地有一张做工业生产的场地网络,从中采集数据;还有一家布拉德带的公司叫 DeepReach,数据来自世界各地的本地创业者。

贾里德: 还有 2026 年冬季批次的 Human Archive。对,最近确实冒出来一批这样的公司。

胡佳怡: 没错。

记录系统也要自己训模型

加里: 如果让我做个预测,回到刚才那句「所有记录系统都得变成 AI harness」,它们可能还得开始训练自己的模型。而这正是 River AI 或者 Tinker 这类东西变得非常有意思的地方。你现在可以坐在 Claude Code 里,甚至我用 OpenClaw 来训练自己的模型,这非常好玩,它会自己做数据清洗,什么都干。到目前为止这还不是一个很大的因素,但我能看到它会变成一个大得多的因素。当你拥有专有数据、又能自己训练的时候,开放权重模型其实已经几乎逼近前沿了,如果你能把它们专门化训练到在某件事上做得比前沿模型还好,那会非常非常强大。

机器人模型必须微调

胡佳怡: 我觉得这件事在机器人领域会更重要。这还只是个假设,尚未被证明,但我认为机器人基础模型和大语言模型的特性很不一样。LLM 的整套逻辑是把现实建模成语言,而机器人是把现实建模在物理的三维空间里,自由度要高得多。要让机器人在某个特定垂直领域里干活,比如在数据中心里做操作的机器人,我带着一家叫 Boost Robotics 的公司,做的就是在数据中心里接线的机器人,要让这类机器人真的能用,更好的办法是拿到一个在定制数据上微调、训练过的模型,专门适配那个环境。机器人还有一个难点是必须实时响应,对刺激的反应要非常快,还要有行动规划,这和 LLM 不一样。LLM 你可以让它自己跑着,过一会儿再回来看。机器人不行,比如你正在给数据中心接一根线,这时有人进来把机器人撞了一下,线就可能插到错误的插口上。

贾里德: 据我所知,所有用 Physical Intelligence 的模型来部署机器人的 YC 公司,都在对 π 模型做微调。我不认为有哪一家能直接拿它开箱即用。它是一个很好的起点,但你必须针对你的具体场景,比如数据中心的线缆,做微调才能真正跑通。

胡佳怡: 嗯。Ultra 这家公司是你在带吧?

贾里德: 对。你从 π 模型起步,但他们有几千小时的「把东西放进箱子」的影像资料,这让模型在往箱子里放东西这件事上特别厉害。

加里: 我听过这样一个说法:你看 Claude Code,它可以用自己的代码记录去判断谁是最顶尖的程序员,然后把这个反过来用,训练出更好的下一代编码模型。如果你恰好拥有 TikTok,你就恰好拥有全部关于人们看什么、点什么、什么内容有吸引力的数据,你可以用它在 Seedance 上做出吸引力强得多的视频。所以这件事已经在发生了。我只是觉得,这个趋势从现在起还会以相当壮观的方式继续下去。

单人创始人翻了三倍

加里: 我们注意到的另一件事,我想大家都感受到了:说实话,最近我们见到的一些最强悍、最狠的创始人,年纪可能在三十多岁后半、四十多岁,甚至五十多岁。资深创始人正在回潮。还有,很多人似乎都想成为「单人创始人的 YC」,但结果是,YC 自己就是单人创始人的 YC。胡佳怡,你有几个让你自己都意外的数据。

胡佳怡: 把一年前的录取数据拉出来分析,过去我们录取的公司里只有大约 5% 是单人创始人,现在超过了 18%、19%。这是我们见过的最高峰值。

加里: 接近批次的五分之一。而且看起来还会继续往上走。在过去,你必须同时具备很多种能力。你得是个很能张罗的人,你得能把事情讲清楚。我们以前常说,创始人里得有人是好的表达者,得有人能站出来说服别人。然后如果你再配上一位世界级的技术专家,那就是最理想的组合。所以传统上你需要联合创始人才能凑齐这一套。倒不是说非有不可,但它会把你的胜率抬高非常非常多。我觉得现在这一点正在改变,因为知道该提示什么、知道该造什么,比 CTO 能不能把东西写出来要难得多、也值钱得多。

贾里德: 我觉得实际情况是,我们其实一直都有极其成功的单人创始人,人们对 YC 这一点没有意识。这里有不同的定义口径,但从各种实际意义上说,Instacart 的阿普尔瓦·梅塔(Apoorva Mehta)、Coinbase 的布莱恩·阿姆斯特朗(Brian Armstrong),至少在批次开始的时候,都是单人创始人。

加里: 对,帕克·康拉德(Parker Conrad)也是以单人创始人身份进的 YC,后来成为他 CTO 的那位,还是我面试的。只不过在过去,要能自己一个人想出点子、能把它卖出去、还能把它做出来,这个门槛真的非常非常高。

胡佳怡: 而现在这件事完全做得到了。

贾里德: 对,我觉得现在的情况正是这样。刚才那三位是极其罕见的人,有能力做到那种程度的人非常非常少。而现在你真的可以先跑起来,所以至少在「把东西做出来」这个维度上,你不需要那么例外,也能开始。但总的算下来,有联合创始人依然是有价值的。它本身也是一种衡量:

加里: 如果你的联合创始人非常顶尖,那说明你大概也非常顶尖,这会把成功概率提高很多。

加里: 而且刚才那几位后来都招了联合创始人。我觉得在每一个案例里,都是先自己跑起来,跑出了牵引力,然后在某个节点加进联合创始人。所以可能股权结构不一样,或者整个动态和传统的那种「一开始你们俩关在一间屋子里,完全五五开」略有不同。贾里德,你要不要补充点什么?

贾里德: 我没有具体数据,但这个趋势我确实明显看到了:人们在公司生命周期更靠后的阶段,在东西已经跑起来之后,才加入联合创始人。

贾里德: 我觉得这会成为常态。我们会在批次里看到更多单人创始人,现在批次开始时就已经能看到了;但在那些最终成功的公司里,我仍然预期他们会随着公司推进而补上联合创始人。

资深创业者的回归

贾里德: 加里,你也讲讲更资深的人出来创业这个趋势吧?

加里: 看起来那些已经在江湖上闯过几轮的人,做得明显好得多。我想到彼得·施泰因贝格(Peter Steinberger)是个标准样本。他,我记得是四十出头,做过开发经理,之前也做过创业公司,然后他很早就彻底皈依了 AI 和这些「铁皮人」,接着他尝试了大量东西,于是他知道该造什么。这件事我觉得特别鼓舞人。基本上,如果你已经在江湖上走过一圈,你知道龙在哪里,你有品味,而这类人现在的战斗力高得反常。

加里: 过去有太多经典的、被设了门槛的说法:哦,你必须有联合创始人;你需要一批够酷的投资人看上你。现在这些越来越不成立了。真正要紧的是你得知道该造什么,这是现在最高位的那一比特。如果你活过一些年头、去过一些地方、有非常鲜明的观点,那你现在可能已经没有借口了。你的借口是什么?你在互联网上大声嚷嚷了这么久,为什么还不去造点东西?打开 OpenCode 直接干就完了。拿出真本事来。

贾里德: 我也在想,管理编码 agent 在某些方面其实和管理人没那么不同。

加里: 太对了。

贾里德: 所以像彼得、像你、像博里斯·切尔尼(Boris Cherny)、像 Shopify 的托比这样,整个职业生涯都在管理工程师团队的人,反而极其适应这件事,能拉起庞大的编码 agent 团队并管好它们,甚至比一个没有这些年经验的聪明的 19 岁小孩管得更有效。

加里: 是啊,我们对 agent 可以稍微不那么刻薄一点,试着理解它们的处境。只是你需要照顾它们情绪的次数少了 99.9%。所以确实挺有帮助的。

现在就打开终端

胡佳怡: 我想问,对一个现在就想开始、想在这个时代做一家公司的人,具体的建议是什么?

加里: 就去写提示词。今天打开 GPT-6 是相当震撼的体验。就是那个瞬间,你的 agent 明显变聪明了,一堆你一直很恼火、但自己没时间深挖的 bug,你可以直接说:要不你回去看看那些你之前没搞定的事情?把我们之前所有的对话都翻一遍,凡是看起来你当初没弄明白的,现在再试着弄明白。它每一次都能做到。我们身处一个多奇怪的历史时刻:你早上醒来,接上一个新模型,然后那些哪怕一个月前你还在挠头「这为什么就是不行」的东西,突然就开始行了。想到这可能是我们接下来 18 个月、24 个月、36 个月都能一直做的事,我也不知道它什么时候结束,但这大概就是 AGI 时代的编程。

加里: 今天的时间就到这里。不过如果你还没听出来,我们所有人都对眼下正在发生的一切感到非常兴奋,你也应该如此。我们迫不及待想看到你造出什么。

本期讲者
加里·谭Y Combinator 总裁兼 CEO,早期投资机构 Initialized Capital 联合创始人,曾任 Palantir 工程师与 Posterous 联合创始人。本期主持开场并主导国防、harness 与创始人画像的讨论。
贾里德·弗里德曼Y Combinator 管理合伙人,负责硬科技方向,文档分享平台 Scribd 联合创始人兼前 CTO。本期提供硬科技融资史、数据与 RL 环境赛道的判断。
胡佳怡Y Combinator 合伙人,增强现实公司 Escher Reality 联合创始人(被 Niantic 收购后任 AR 平台工程负责人)。本期负责公布批次统计数据与算力、机器人趋势分析。
章节 · 点击跳转视频
0:00 冷开场:最猛的创始人已四十岁 ▶ 正在看
0:41 批次数据:硬科技占比与收入中位数 ▶ 正在看
2:24 国防、制造、算力、电力的回潮 ▶ 正在看
5:45 太空与国防:主权星链与反无人机 ▶ 正在看
9:05 数据中心瓶颈与新芯片路线 ▶ 正在看
12:28 机器人的 ChatGPT 时刻尚未到来 ▶ 正在看
13:30 风投重新回到硬科技老本行 ▶ 正在看
15:03 SaaS 没死:记录系统变身 harness ▶ 正在看
17:34 端到端 agent 把收入中位数拉高 ▶ 正在看
22:22 卖数据与 RL 环境的隐形赢家 ▶ 正在看
29:07 单人创始人从 5% 涨到 19% ▶ 正在看
32:34 经验型创始人与管理 agent 的手艺 ▶ 正在看
本期论点
本期回应
4:40
代码生成让顶级软件工程师不再是做硬件的限制性因素,硬件公司的用人经济账已被改写 硬科技变热,是模型加速了科研,还是投资人躲开了软件?加里·谭
5:07
硬科技看涨的真正理由不是资本回避软件,而是强模型正在加速科研、让突破来得更早 硬科技变热,是模型加速了科研,还是投资人躲开了软件?胡佳怡
14:20
风险投资本来就是为投硬科技而生的,只因投 SaaS 太赚钱才偏离了一二十年 硬科技变热,是模型加速了科研,还是投资人躲开了软件?加里·谭
7:46
新创业公司靠 AI 和新的制造方式,能做出传统国防巨头做不出来的东西 已经领先的人会被换掉吗?贾里德·弗里德曼
21:42
电话初筛、判断文化契合这类带人际因素的招聘工作,AI 很难做 AI 会怎样改变人的工作?胡佳怡
11:37
大模型架构不需要完整精度的浮点运算,更低精度甚至三进制表示就足够 算力提升还能从哪里来?胡佳怡
30:47
知道该怎么提示、该做什么产品,比知道怎么让 CTO 写出代码更难也更有价值 面对技术冲击,人该学什么?加里·谭
34:23
带过工程师团队的资深从业者驾驭编程 agent,比聪明的年轻人更高效 为什么说四五十岁的创业者,现在反而比年轻人更有优势?贾里德·弗里德曼
其他论点
13:07
机器人领域还没有找到属于自己的 scaling law 胡佳怡
15:53
当下对路的软件是把 agent 当成客户、做成 agent 愿意用的东西 贾里德·弗里德曼
16:48
只要数据护城河守得住,SaaS 依然可以像 AI 之前一样值钱 企业软件的护城河,在于囤积数据还是在于承接工作?加里·谭
17:04
记录系统公司要么护城河被 MCP 吃掉,要么变成人们真正在里面干活的 harness 企业软件的护城河,在于囤积数据还是在于承接工作?加里·谭
26:48
有专有数据时,为专门用途训练的开放权重模型能胜过前沿模型 加里·谭
27:12
机器人基础模型与 LLM 特性截然不同,必须用定制数据微调才能真正跑起来 胡佳怡
31:59
更可行的路径是先独自做出成绩,到某个阶段再引入联合创始人 加里·谭
01冷开场:最猛的创始人已四十岁
0:00
Frankly, some of the most powerful and badass founders that we've been seeing lately, there might be in their late 30s, 40s, even 50s. I mean, there's a sort of resurgence of the experienced founder. A lot of people seem to say that they want to be uh YC for solo founders, but it turns out YC is the YC for solo founders. What a weird moment we are in history where you wake up in the morning, you like wire up a new model and then these things that even a month ago you're just like why isn't it working? It just starts working.
坦白讲,我们最近见到的一些最强、最厉害的创始人,可能都已经三十好几、四十多,甚至五十多岁了。我是说,有经验的创始人正在某种程度上重新崛起。很多人似乎都在说他们想做单人创始人版的 YC,但结果是,YC 本身就是单人创始人的 YC。我们正处在一个多么奇特的历史时刻啊,你早上醒来,接上一个新模型,然后那些哪怕一个月前你还在想「这怎么就是跑不通」的事情,突然就跑通了。
便签引用
02批次数据:硬科技占比与收入中位数
0:41
Welcome back to another episode of the light cone. At YC, we work with thousands of founders per year, which means we start to see things before they're obvious. So, we wanted to share some of that with you today. What's the state of the art and what's coming next? What should you, the builder, know? Let's get started. Diana, you have a few things to share with us. So we did a bit of an analysis for all the companies we accepted in the last 18 12 months and we have some pretty shocking stats to share with everyone. So one of the big ones is the number of heart tech companies that are in the bat. It has gone from 8% to 20%. There's a lot of uh underlying reasons why that has happened. We we will go deeper into that. The other one is uh the rate of growth of companies and what YC does to the companies has accelerated. So the median YC company when it gets accepted is at zero in revenue is pre- revenue pre-product and by the end of the batch in the past companies would get to about 8K median
欢迎回到 Light Cone 的又一期节目。在 YC,我们每年和数千位创始人合作,这意味着很多事情在变得显而易见之前,我们就已经看到苗头了。所以今天我们想和你分享其中的一些内容。技术的前沿现状是什么?接下来会发生什么?作为建设者的你,应该知道些什么?我们开始吧。Diana,你有几件事要和我们分享。我们对过去 18 个月、12 个月里录取的所有公司做了一些分析,我们有一些相当惊人的数据要和大家分享。其中一个重磅数据是这一批里硬科技公司的数量。它已经从 8% 涨到了 20%。这背后有很多原因,我们稍后会更深入地聊。另一个是公司的增长速度,以及 YC 对公司产生的作用都加快了。以前,YC 公司在被录取时,收入的中位数是零,处在收入前、产品前的阶段,而到了这一批结束时,公司通常能做到大约 8000 美元的中位数
便签引用
1:52
revenue and now the companies in median are getting to 20,000 monthly revenue as opposed to 8K. So those are the top two that we can dive deeper into. >> Yeah. Let's dig into hard tech first. Like what are these hard tech companies and what's what's driving this >> things that actually touch atoms and not just bits. >> Yeah. And I I I think you have the the the category breakdown of the heart tech companies, right Diana? >> Yeah. So specifically robotics has been a big one. It has gone from 1% of the batch to about six 7% of the batch.
收入;现在这些公司的中位数能达到每月 2 万美元的收入,而不是 8000。所以这就是我们可以展开聊的两个最主要的点深入探讨。>> 是啊。我们先来聊聊硬科技。比如这些硬科技公司都是做什么的,背后的驱动力是什么 >> 就是那些真正跟原子打交道的东西,而不只是>> 比特。>> 对。而且我我我记得你手上有硬科技公司的品类拆分数据,对吧,Diana?>> 对。具体来说,机器人一直是个大热门。它从占整批公司的 1% 涨到了大约百分之六七。
便签引用
03国防、制造、算力、电力的回潮
2:24
Industrial manufacturing building things back in the US has been a huge trend. It has it has gone from about 4% to 10% of the batch. The other one is defense is a big one. We all have been working with a lot of uh defense startup. It has gone from about 1.5% to about 5% of the batch. The other big one is there's this uh compute need that the world is getting into with AI. So there's a lot of companies building the semiconductor stack or photonix. It has gone from about 1% of the batch from a year ago to about close to 4% of the batch. And the other one even below the stack of compute is power. So there's a lot of power infrastructure as well has gone from also 1% to about close to 3% of the batch. So all these numbers across the physical atom stacks have somewhere triple or quintupled. Yeah, this is the uh age of the machine, I think. And that's I mean, >> as the world goes, you know, our our motto um the t-shirt says, "Make something people want." And people sure do want those things right now.
工业制造、把生产制造搬回美国,一直是个巨大的趋势。它从大约 4% 涨到了整批公司的 10%。另一个大类是国防,这块很重要。我们大家都在跟很多国防领域的创业公司合作。它从大约 1.5% 涨到了整批公司的大约 5%。还有一个大类是,随着 AI 的发展,全世界对算力的需求越来越大。所以有很多公司在做半导体技术栈或者光子学(photonics)。它从一年前占整批公司的大约 1% 涨到了接近 4%。还有一个比算力更底层的,是电力。所以也有很多电力基础设施方面的公司,同样从 1% 涨到了接近整批公司的 3%。所以在这些跟实体原子相关的技术栈里,各项数字基本都翻了三倍甚至五倍。对,我觉得这就是机器的时代。而且这就是说 >> 随着世界的发展,你知道,我们的口号,T恤上印的那句话是:“做人们想要的东西。”而现在人们确实非常想要这些东西。
便签引用
3:28
>> And the other interesting factor about all these companies that are going deep into Adams is that we've been funding more technical founders and with more expertise than ever. Right Jared? We have this fun stat about the current summer batch with >> in the current summer batch, one in six of the founders actually has a PhD. It's way more than that's been historically and it's because yeah, if you're doing, you know, something with like silicon photonics, you're probably going to need a pretty strong research background in that. And so we've been funding a lot more of those founders and those founders, I think, have disproportionately been doing like especially well. I think AGI compounds this in a really fascinating and awesome way in that like you might think in the past you actually like hard techch was hard because you had supply chains you had uh an incredible software component of often uh I think Palmer Lucky talked about this a lot when it came to and it's like having codegen means that
>> 另一个关于这些深耕"原子"(硬件)领域公司的有趣现象是,我们资助的创始人比以往任何时候都更技术型、更有专业深度。对吧 Jared?我们有个很有意思的数据,关于这一期夏季批次—— >> 在当前这期夏季批次里,六个创始人中就有一个拥有博士学位。这比历史上的比例高多了,原因就在于,你要是在做像硅光子学这类东西,那你多半需要在那个领域有相当扎实的研究背景。所以我们资助了更多这样的创始人,而我觉得这批创始人做得格外出色,比例远高于平均。我觉得 AGI 让这件事以一种特别迷人、特别酷的方式叠加放大,就是说,过去你可能会觉得硬科技之所以难,是因为有供应链问题,而且往往还有一个极其庞大的软件部分——我记得 Palmer Luckey 就谈过很多这方面的事——而现在有了代码生成,就意味着突然之间,连 Anduril 在做的那些事情都能快得多、
便签引用
4:26
suddenly even all the things that they do at Anderil uh can happen much much faster right even three or four years ago you would talk about software engineering and like the top tier software engineers as one of one of the limiting reagents to being able to do really really top tier full stack uh hardware and that's less and less true. I mean you still need one or two of them or you need like a small team but you don't need to hire a thousand great engineers versus Google or Meta or whoever else and that really changes the economics.
快得多地完成,对吧?哪怕是三四年前,大家聊软件工程时,都会把顶级软件工程师视为做出真正顶尖的全栈硬件的"限制性试剂"之一,而现在这一点越来越不成立了。当然你还是需要一两个这样的人,或者需要一个小团队,但你不需要去跟 Google、Meta 或者别的谁抢着招一千名优秀工程师,这就彻底改变了整个经济账。
便签引用
5:01
>> I mean that's the that's the true bullcase for heart. It's that it's not just that people are shying away from funding software businesses, but it's actually that the super smart models that we have now are actually accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier and that therefore these deep tech companies will actually work better. I think the other factor is um there's basically three macro trends that are also driving all this all this growth on atoms and is seeing huge companies like SpaceX have such a successful IPO has created a generation of founders wanting to build in space. There lots of these companies that are building across the whole stack. So there's been companies uh in in the current batch in summer 26.
>> 我是说,这才是硬科技真正的看涨逻辑。不只是大家在回避投软件生意,而是我们现在拥有的这些超聪明的模型,确实在加速科学研究,让创业公司能更早地取得更大的研究突破,因此这些深科技公司真的会跑得更好。我觉得另一个因素是,大概有三个宏观趋势也在推动"原子"领域的这一轮增长。看到像 SpaceX 这样的巨头 IPO 如此成功,催生了一代想在太空创业的创始人。有很多公司在整个技术栈的各个环节上做事。所以在 2026 年夏季这一期批次里就有一些公司。
便签引用
04太空与国防:主权星链与反无人机
5:45
This is company that we work with called Exosat that's trying to build basically a sovereign Starlink solution. There's other company that I work with uh in winter 26 called Beyond Reach Labs that's building solar panels for satellites in space. If you imagine companies like StarCloud wanting to have all these data center in space, they will need to have power. So this is a obvious solution. Now the other macro trend is um I think we have a generation of current founders right now that have uh grown with the war that's been very front and center and spoken a lot in social media and they want to do something >> like two two of the companies I'm most excited about that I funded the last couple batches one was Icarus last fall and then nine mothers this last spring and both of them were defense Icorus is doing like a solarp powered U2 spy plane that gives overwatch and can also do comms, which is actually really important. The future of drone war is being able to actually communicate with your drones on the ground and see what's
有一家我们合作的公司叫 Exosat,基本上是想做一套主权版的星链方案。还有另一家我在 2026 年冬季批次合作的公司叫 Beyond Reach Labs,在做太空卫星用的太阳能板。你想想看,像 StarCloud 这样想把数据中心搬到太空的公司,它们都需要电力。所以这就是一个很自然的解法。另一个宏观趋势是,我觉得现在这一代创始人是伴随着战争长大的——战争一直非常显眼,社交媒体上也讨论得很多——他们想为此做点什么。>> 比如说,我过去几期批次里投过的公司中最让我兴奋的有两家,一家是去年秋天的 Icarus,另一家是今年春天的 Nine Mothers,两家都是国防方向。Icarus 在做一种太阳能驱动的 U2 侦察机,既能提供战场监视,也能做通信中继,这其实非常重要。无人机战争的未来,就在于你能真正和地面上的无人机保持通信,看清战场态势。他们已经拿到了新成立的"战争部"的
便签引用
6:44
going on. They've been able to get to seven figure contracts with the new department of war. Um, and then likewise with drone war, um, special forces has been buying, uh, nine mothers anti- drone defense. So, it's basically a shotgun turret with CV. Uh, but it's actually almost the only way that you could protect special forces uh, deep behind en enemy lines. I mean these are people who have been training for years and years uh you know in a very elite special force that um you know America doesn't have uh you know thousands of these people you know we we have a very very small set and so protecting them from uh what could be like a commodity drone attack is actually really existential for the department of war.
七位数合同。嗯,同样在无人机战争这块,特种部队一直在采购 Nine Mothers 的反无人机防御系统。它基本上就是一个带计算机视觉的霰弹枪炮塔。但这几乎是保护深入敌后的特种部队的唯一办法。我是说,这些人是训练了很多很多年的,属于极其精英的特种部队,美国并没有成千上万这样的人,我们的人数非常非常少,所以保护他们不被那种可能只是廉价量产的无人机攻击干掉,对战争部来说真的是生死攸关的事。
便签引用
7:27
So just really cool to see this new administration actually uh approach defense in a very different way. like it, you know, classically there was just a lot of frankly capture from the big defense primes that are just doing sort of um cost plus. They think of themselves as consultants and you know to be able to see new startups that can actually take advantage of all of the AI, all of the tech, all of the new ways of building things to build things that frankly uh the defense primes can't build. You know, that's a really powerful mega trend right now. Now the thing about defense is not just those full solutions that get sold to the government. There's also a lot of category of startups that are dual use that they sell both to the private sector and to the government and that have to do with everything down the supply chain. So things like manufacturing things back in America, building custom um I think you had this company Nox Metal. Yeah, they're bringing metal manufacturing back to America. America has like largely lost
所以看到这届政府用一种非常不一样的方式对待国防,真的挺酷的。你知道,传统上说白了就是被几家国防巨头垄断了,它们做的基本是成本加成那一套。它们把自己当成咨询公司。而现在能看到新的创业公司真正用上所有这些 AI、所有这些技术、所有这些新的造东西的方式,去做出说实话国防巨头做不出来的东西。这真的是眼下一股非常强劲的大趋势。而国防这块不只是那些直接卖给政府的整体方案。还有一大类创业公司是军民两用的,既卖给私营部门也卖给政府,涉及整条供应链的方方面面。比如说把制造搬回美国,做定制化的……我记得你投了一家叫 Nox Metal 的公司。对,他们在把金属制造带回美国。美国基本上已经失去了自己的金属工业,过去几十年被掏空了,而没有金属
便签引用
8:33
its metal industry. It got hollowed out over the last few decades and can't build stuff without metal. And so Knox Metals is like rebuilding America's metal supply chain. And they're doing it in in the heartland of America in in in Detroit where there's all these like empty factories that have basically just been like sitting there. >> And they're an example of like the trend where the it's not just people are doing hardware companies, but the hardware companies themselves are growing faster than ever. Like I think I saw a PG tweet that Nox Metals is growing like at software growth rates. Do do you understand that? Like kind of what how are they growing so fast?
你什么都造不出来。所以 Knox Metals 相当于在重建美国的金属供应链。而且他们是在美国的腹地做这件事,在底特律,那里有一大堆空置的工厂,基本上就一直那么荒着。>> 而且他们也体现了另一个趋势:不只是有人在做硬件公司了,而是这些硬件公司本身的增长速度比以往任何时候都快。我好像看到 PG 发过一条推,说 Nox Metals 的增长速度堪比软件公司。你了解这个吗?他们到底为什么能长这么快?
便签引用
05数据中心瓶颈与新芯片路线
9:05
>> So one reason is that a lot of their customers are these new defense tech startups that have sprung up and need metal to build all their all their stuff and the existing suppliers that are these sort of like sleepy old businesses mostly run by old people just like can't keep up with the pace that the new defense tech startups want to build at. And it reminds me a bit of like when the web 2.0 boom happened early in in the YC days. We would have these new startups, but then they would prefer to buy from new startups that could sort of like move at their speed and like work well with them. Like Stripe, for example, you could use a legacy credit card vendor, but like it's just like way better to work with Stripe. And so I feel like they're sort of becoming that for the whole defense tech ecosystem. Now the third trend is um basically compute is a very heavy physical atoms process to get all these data centers live very quickly because a lot of the demand for AI that we've been talking has been skyrocketing and there's a very
>> 一个原因是,他们很多客户就是这些新冒出来的国防科技创业公司,它们需要金属来造各种东西,而现有的供应商都是那种老气横秋、大多由上了年纪的人经营的公司,根本跟不上新国防科技创业公司想要的节奏。这让我有点想起 YC 早期 Web 2.0 热潮的时候。那会儿我们有一批新创业公司,而它们更愿意从别的新创业公司那里采购,因为对方能跟上它们的速度、配合得好。比如 Stripe,你当然可以用传统的信用卡供应商,但跟 Stripe 合作就是爽得多。所以我觉得他们正在成为整个国防科技生态里的那个角色。那么第三个趋势就是,算力本质上是一个非常重的物理"原子"工程,要让这些数据中心尽快上线,因为我们一直在聊的 AI 需求一直在暴涨。有个非常
便签引用
10:05
interesting stat where GPUs from Nvidia let's say like a A100 GPU per hour is actually appreciating in cost which is unusual in the past when you get a A100 by now sort of old. >> They're pretty old. Yeah. >> The price is going up because there's just too much demand and not enough supply and compute. So there's a lot of uh startups that are now working on bringing data centers live and you have everything from the construction of the sites to the software to planet to actually doing the data center buildout to interesting solutions that have to do with how to power them and combination of energy, battery. So there's all these category of startups and even to the point of uh going down to the core compute silicons. There's a number of uh startups that are building new new silicon for an alternative to to to Nvidia. There's this company that Tyler worked with called Lamb Labs that's building new processors for compute.
有意思的数据:英伟达的 GPU,比如说 A100 的每小时价格居然在涨,这很反常,因为按以往,到现在 A100 已经算是老卡了。>> 是挺老的了,对。>> 价格还在涨,就是因为需求太大、供给和算力都不够。所以现在有很多创业公司在做把数据中心跑起来这件事,从场地建设到规划软件,到真正执行数据中心的建设,再到各种有意思的供电方案、能源和电池的组合。所以这一整类创业公司都有,甚至一路往下做到核心的算力芯片。有不少创业公司在做新的芯片,作为英伟达的替代方案。有一家 Tyler 合作的公司叫 Lamb Labs,在做新的算力处理器。
便签引用
11:05
There's another one that I'm working on this batch called bot that is trying to build basically new custom hardware architecture that's using turnary representation for models because what it turns out which is a funny trend right now if you look at all the Nvidia architectures from A100s to H100s and now the B B300s each of these generations they're actually going down in floating point precision in terms of uh what they were they're going from FP3 32 168 etc. And it turns out that the LLM architecture doesn't need the full precision floating.
还有一家是我这期在带的,叫 bot,基本上是想做一种全新的定制硬件架构,用三进制表示来跑模型。因为事实证明——这是现在一个挺好玩的趋势——如果你看英伟达从 A100 到 H100 再到现在 B300 的各代架构,每一代其实都在降低浮点精度,从 FP32 到 16、到 8,等等。而事实证明,LLM 架构根本不需要完整精度的浮点。
便签引用
11:42
>> FP2 is even somewhat usable. >> Right. So this is what bot is trying to do. And I think you have an interesting one that's doing um the interconnect with uh with photonics. >> Yeah, there's a company called Dipole Labs in the current batch that is replacing the switches that are in data centers which are essentially the like routing systems between different GPUs. If like GPU want a wants to talk to GPU B, they talk to each other through this device. It's called a switch. And these switches right now are electronic. And so there's actually an issue which is like the switches are not keeping up with the GPUs. The the speed of the GPUs keeps going up. And the switches are actually the bottleneck for many data centers in many different workloads. And so Dipole Labs is building the first fully optical switch where it's like all photons from GPU A all the way to GPU B.
>> FP2 甚至都有点能用。>> 对。这就是 bot 想做的事。而我记得你那边也有一家很有意思的,在做互连,用光子学的方式。>> 对,这期批次里有家公司叫 Dipole Labs,他们在替换数据中心里的交换机,交换机本质上就是不同 GPU 之间的路由系统。如果 GPU A 想跟 GPU B 通信,它们就通过这个设备互相通信,这就叫交换机。而现在这些交换机都是电子的。所以就出现了一个问题:交换机跟不上GPU 了。GPU 的速度一直在往上涨,而交换机在很多数据中心、很多不同的工作负载里反而成了瓶颈。所以 Dipole Labs 在做第一款全光交换机,从 GPU A 一路到 GPU B全程都是光子。
便签引用
06机器人的 ChatGPT 时刻尚未到来
12:28
And so it will actually be much faster than the electronic switches that we use now. And >> now the last one that's driving all this move to atoms is this uh aspect where robotics is going to happen. So there's a lot of companies building the stack around that and everything from vertical robotics in specific industries to the infrastructure to deploy robots to uh data selling to the new robotics labs because there's this moment that everyone in in the industry is feeling that we're going to get to the chat GBT moment. just not quite there yet. And I think we're figuring it out and new a scaling law around it. So there's a lot of that and we had Quan here uh a couple episodes ago and we're believers that's that's going to happen. I mean robotics >> from Pi >> from Pi, right? Half is AI and half is uh hardware. I was hearing this reading this morning that even Astra is like a big leap forward for for robotics like on I forget the benchmark but there's a benchmark where like Fable was maybe at
所以它会比我们现在用的电子交换机快得多。>> 那么推动这一轮向"原子"转移的最后一个因素,就是机器人时代要来了。所以有很多公司在围绕这个搭建技术栈,从特定行业的垂直机器人,到部署机器人的基础设施,再到给新的机器人实验室卖数据。因为业内所有人都感觉到,我们快要迎来机器人领域的 ChatGPT时刻了,只是还差一点。我觉得我们正在摸索,也在找它的新的 scaling law。所以这块的事情很多。我们几期之前请过 Quan 来聊,我们是相信这件事一定会发生的。我是说机器人——>> 来自 Pi 的。>> 对,来自 Pi。一半是 AI,一半是硬件。我今天早上读到,连 Astra 都是机器人领域的一次大跃进,在某个基准上——我忘了是哪个基准了——Fable 大概只能做到
便签引用
07风投重新回到硬科技老本行
13:30
10% and um Astra is showing like you can do like 60 to 70% of the tasks. >> So data to just you know wake up in you know another couple weeks and another breakthrough happens and we're a little bit closer. >> It's been really cool for me to see the resurgence of heart because you know YC we've been funding hard companies since 2014. That's really when we we started. But it was like pretty hard to get these companies funded before. Like I remember like pre pre this recent resurgence, we would like fund awesome stuff that we were super excited like rockets and planes and chips and data centers and stuff like that. And then like VCs would just be like ah we only do B2B SAS. Um and it's hard to bootstrap a company like this. So you really do need like downstream investors who can fund lots of who who who can who can fund you know a full a full capital buildout. And so it's cool that like it seems like Silicon Valley which historically like venture was set up to fund hard techch but it like drifted away from it for a
10%,而 Astra 能完成其中 60% 到 70% 的任务。>> 所以说,可能你再过几周一觉醒来,又一个突破出现了,我们就又近了一步。>> 看到硬科技的复兴对我来说真的很酷,因为你知道,YC 从 2014 年就开始投硬科技公司了。那时候真的是我们的起点。但以前这类公司要拿到融资相当难。我记得在这一波复兴之前,我们会投一些我们特别兴奋的酷东西,像火箭、飞机、芯片、数据中心这类。然后那些 VC 就会说,啊,我们只投 B2B SaaS。而且这类公司很难靠自筹资金做起来。所以你真的需要下游有能出大钱的投资人,能支撑你完成一整轮资本性建设。所以挺酷的是,硅谷——历史上风险投资本来就是为投硬科技而生的,但有一二十年它偏离了这条路,因为只投 SaaS 公司实在太赚钱了——
便签引用
14:28
decade or two because it was so profitable to just fund SAS companies and so it's cool to have it coming back to its roots. >> Yeah. On the point about like the the investors wanting to do hard tech again it does seem like that I've never seen that happen so quickly. I mean it seemed like it happened pretty immediately when like SAS stocks were down earlier this year. Claude code was surging and that just became like the the I mean I feel like even at demo day it literally happened I feel like that happened probably mid the winter batch at the start of this year and it seemed by even demo day that investors were starting to be a lot more interested in hard tech companies and that's just extrapolated.
现在能看到它回归本源,真的挺酷的。>> 是啊。说到投资人重新愿意投硬科技这一点,确实是这样,我从没见过这种转变发生得这么快。今年早些时候 SaaS 股票下跌的时候,感觉几乎是立刻就发生了。那会儿 Claude Code 正火,然后那就成了……我感觉甚至在 Demo Day 上就已经真真切切发生了。我觉得大概是今年初冬季批次进行到一半的时候开始的,到 Demo Day 的时候就能明显看出投资人对硬科技公司的兴趣大了很多,之后就一路延续下来了。
便签引用
08SaaS 没死:记录系统变身 harness
15:03
I mean it is worth knowing though on the other side like since that a bunch of the SAS stocks have actually recovered and are doing better than ever. Like Salesforce is like the prime example of that. I think Snowflake recently had like like two days ago had these like blowout earnings and so it's possible that >> it all hits. >> Yeah, maybe. I mean that would be the dream case. Um I mean I still think we're seeing real stuff though. Like it clearly the software that gets built in the future and what's valuable is different. like it just has to be. And so partly it seems like what we're seeing with Salesforce is the classic the system of record argument is actually playing out.
不过另一面也值得注意:从那之后,不少 SaaS 股票其实已经反弹了,而且表现比以往更好。Salesforce 就是最典型的例子。我记得 Snowflake 最近——好像就两天前——发布了非常炸裂的财报。所以也有可能 >> 两边都成立。>> 对,也许吧。那会是最理想的情况。不过我还是觉得我们看到的是实打实的变化。很明显,未来被造出来的软件、以及什么东西有价值,都会不一样。这是必然的。所以我们在 Salesforce 身上看到的,一部分像是经典的"记录系统"(system of record)论点真的在应验。
便签引用
15:37
>> The Moes are intact for now. >> Yeah. Like if you have a thing that agents can use um that is actually valuable and if anything you'll just like agents will use software a lot more than humans will and that seems to be driving Salesforce growth. And so kind of takes us back to the other trend that we've talked a little bit about is if you think of agents as your customers and you make things that agents want and your software as something that agents want to use, then that seems like the right type of software.
>> 护城河目前还在。>> 对。如果你有个东西是 agent 能用的,而且是真的有价值的,那反过来说,agent 用软件的频率会远高于人类,而这看起来正在推动 Salesforce 的增长。这就又回到我们之前稍微聊过的那个趋势:如果你把 agent 当成你的客户,做出 agent 想要的东西,把你的软件做成 agent 愿意用的东西,那这看起来就是对路的软件。
便签引用
16:04
>> Yeah, Salesforce is super interesting because uh I think they started releasing their own Slack harness, Slack AI harness. And so I think we're right at the beginning of like the next AI harness wars. It's like Codex wants to be it, Cloud Code wants to be it. Um Open Cloud could be it. Hermes, open code. It seems like there are going to be a bunch of them and it's not going to be quite like the browser wars and that the browser wars tend toward like one winner, but you know, I guess it's anyone's guess. And then um yeah, Ben off has a pretty big advantage in that you a lot of the most AI pill and companies in the world still use Slack.
>> 对,Salesforce 特别有意思,因为我记得他们开始推出自己的 Slack harness,Slack 的AI harness。所以我觉得我们正处在下一轮 AI harness 大战的开端。Codex 想当那个赢家,Claude Code 想当那个赢家。OpenClaude 也可能是。还有 Hermes、open code。看起来会有会有一大批,而且不会完全像当年的浏览器大战那样,浏览器大战往往最后只剩一个赢家,但你知道,我想这谁也说不准。然后嗯,是的,Benioff 有一个相当大的优势,就是世界上很多最「AI 化」的公司现在仍然在用 Slack。
便签引用
16:39
And you know, if the harness is in there and it's your system of record for how people collaborate, then you have like this mega data mode. And then you know SAS can still be as valuable as it's ever been valued if those modes hold. >> Yeah. I thought you had a really interesting tweet maybe a week or so ago about how the software or system of record companies will have to become like harnesses. >> I mean and that's bas that was about slack I would say. It's like basically if you are a system of record you either will be prayed upon like you'll release an MCP and then maybe like the data you know you lose your moat around the data the data goes elsewhere like becomes very trivial to switch or you kind of have to be a harness you have to be the way people not just read and write but actually do their work inside you know your system of record >> and get the most value out of it I mean it's a it's a little bit like the model companies like was the um was it RKGI was the benchmark Mark where there was a
你知道,如果 harness 就在里面,而它又是你们协作方式的记录系统,那你就拥有了这个超级数据护城河。然后你知道,只要这些护城河守得住,SaaS 依然可以像以前一样值钱。>> 是啊。我觉得你大概一周前发的那条推特特别有意思,讲的是软件公司或者说记录系统公司将不得不变成 harness。>> 我是说,那基本上……我会说那条是在讲 Slack。基本上就是,如果你是一个记录系统,你要么会被人「吃掉」,比如你发布一个 MCP,然后你围绕数据的护城河可能就没了,数据流到别处去,切换变得非常容易;要么你就得成为一个 harness,你得成为人们不只是读写、而是真正在里面完成工作的地方,也就是在你的记录系统里 >> 并从中获得最大价值。我觉得这有点像模型公司的情况,就像那个……是 ARC-AGI 吗?就是那个基准测试,有一个
便签引用
09端到端 agent 把收入中位数拉高
17:34
benchmark where of the >> RPGI V3. >> Yeah. Where the pre or pre-Astra ChachiBT model didn't do as well, but then they said, "Well, that's just cuz it was plugged into the wrong harness." So, it's like the model plus the harness gets you the output. >> Oh, yeah. I mean, uh, with a custom harness, they claim Astra got to n north of 90% on RKGI 3. >> Yes. I think a couple months ago, this was in the low two digits, right? Which is an impressive leap. >> And I think you have a very good point around software. It's not that software and SAS is dead what people claim on the internet. It's just that is has transformed. We actually seen this in the batch. The percentage of companies we accepted that do sort of full stack end to end work or a task has gone from just 10% to over 25% of the batch. This has to do with actually doing the job.
基准测试 >> ARC-AGI v3。>> 对。就是那个 Astra 之前的 ChatGPT 模型表现没那么好,但后来他们说:「其实那只是因为它被接到了错误的 harness 上。」所以就是说,模型加上 harness 才决定了最终输出。>> 哦对。我是说,呃,用定制的 harness,他们声称 Astra 在 ARC-AGI 3 上做到了 90% 以上。>> 是的。我记得几个月前这个分数还只有十几分,对吧?这是一个很惊人的飞跃。>> 而且我觉得你关于软件的观点非常对。并不是像互联网上很多人说的那样,软件和 SaaS 已经死了。只是它发生了转变。我们其实在这一批里就看到了。我们录取的公司中,做那种全栈式、端到端完成工作或任务的比例,已经从只有 10% 涨到了超过 25%。这跟真正把活干完有关。
便签引用
18:26
the agent does the job, not just like a point solution, which old SAS in five, eight years ago was just like a point solution and you needed a you needed someone to operate the SAS software. Right now, it just runs by itself and actually in the batch, this is where we're seeing a lot of the growth in revenue. I think I gave that stat of the median uh startup when it gets into YC is at zero in revenue and it has gone from by the end of the batch it was about 8K in MR now it's like about 20K in MR and a lot of these >> that's a huge jump that's like non-trivially big jump for the median right the average is even higher >> and it has to do with doing the full end toend job with for example doing insurance broker actually doing the clinical intake doing the full end to end workflow of I don't know medical billing etc and these are the ones that are growing a lot and I think there's another factor where that's happened I think we talked about this in couple episode ago we right now are about
是 agent 把活干了,而不只是一个点状解决方案。五年、八年前的老式 SaaS 就只是一个点状解决方案,你还需要有人去操作这个 SaaS 软件。而现在,它自己就跑起来了。实际上在这一批里,我们看到收入增长很多都来自这里。我之前给过一个数据,就是创业公司进 YC 时收入的中位数是零,而到批次结束时,这个数字已经从大约 8K MRR,涨到现在大约 20K MRR,而且很多这些 >> 这是个巨大的跃升,对中位数来说是相当可观的跃升,对,平均数还更高 >> 这跟端到端把整件事做完有关,比如做保险经纪,真正去做临床接诊,做完整的端到端工作流,比如医疗账单之类的,这些公司增长特别快。我觉得还有另一个因素促成了这件事,我想我们在前几期节目里聊过,我们现在差不多是在 agentic coding 开始真正可用、也就是 Opus 4.5 之后将近一年,我们看到这些工作流彻底
便签引用
19:32
almost a year since agentic coding started to work since opus 4.5 that we're seeing these workflows fully blossom and the result are basically people want their job just be done and are willing to buy software that just gets the job done. >> I think when people hear these revenue numbers growing so fast, an easy knock on it is like maybe it's just AI hype and these companies are just like shelling out money for AI products because it's like the cool thing to do. And to be fair, that's probably some of that. But I think like the bullcase is actually something that we said in an episode like 2 years ago when agents were really just beginning to be a thing where we were like actually the products are just going to be more valuable. Like if they automate the whole job, they will actually just be more valuable than some like system of record that tracks the job but doesn't do the job and therefore companies will just like pay more money for the for for the product.
开花结果,结果基本上就是:人们想要的是活被干完,并且愿意为那种真正把活干完的软件付钱。>> 我觉得当人们听到这些收入数字涨得这么快时,一个很容易的质疑是:也许这只是 AI 的炒作,这些公司只是因为这事很潮就把钱砸在 AI 产品上。公平地说,确实有一部分是这样。但我觉得看多的理由,其实是我们大概两年前在某一期节目里说过的事情,那时候 agent 才刚刚开始成型,我们当时就说,其实这些产品会变得更有价值。如果它们把整件工作都自动化了,那它们确实会比那种只记录工作、但不干活的记录系统更有价值,因此公司自然愿意为这个产品付更多的钱。
便签引用
20:27
I definitely see that in companies that I work with where yeah, they just go to a company and like the value proposition is so great that like large enterprises are willing to write big checks very very early. >> Yeah. A company that I'm seeing um having this effect is Juicebox, an AI recruiting tool. Like it's been on incredible growth rate for like the last couple of years now, but they started out as I mean I would say it was essentially sort of LLM powered people search. Like the thing that they did was you could type in sort of the spec of the type of person you wanted to hire and it did a really good job of pulling the um good profiles of people that you may want to contact. Um but then you still have to go and contact the people.
我在合作的公司里绝对看到了这一点,就是他们去找一家公司,价值主张实在太强了,以至于大型企业非常非常早期就愿意开出大额支票。>> 是的。我看到一家有这种效应的公司是 Juicebox,一个 AI 招聘工具。它在过去这几年一直保持着惊人的增长速度,但他们最开始,我会说,本质上算是一种 LLM 驱动的人才搜索。他们做的事情是你可以输入你想招的人的大致画像,然后它能非常好地找出你可能想联系的那些优质候选人档案。但接下来你还是得自己去联系这些人。
便签引用
21:03
And recently they've launched an agent product which is really taking off and the agent like doesn't just search for the people. It like contacts the people. It'll be able to schedule the interview, do a bunch of things. That's awesome. >> Yeah. and they're seeing that that's going to just on a like per account basis I think is going to double or triple like the revenue they make from a single customer cuz customers want more and more of these agents. I don't think it's fair to say that it's like it's not like it's like automating the job of the recruiter at all. It's just that it's just changing it like it like the recruiters didn't necessarily want to be doing like that sort of wrote reach out to like 500 people anyway. Like the thing that makes the recruiter job I would say like more skilled and interesting is like there's like culture fit. that's just going to be really hard for like an AI to do a phone screen that assesses like how well someone's going to be like a culture fit and um and the
而最近他们推出了一个 agent 产品,现在增长得非常猛,这个 agent 不只是帮你搜人,它还会去联系人。它能安排面试,做很多事情。这太棒了。>> 是的,而且他们发现,单看每个账户的话,我觉得这会让他们从单个客户身上赚到的收入翻一倍甚至三倍,因为客户想要越来越多这样的 agent。我不觉得可以说这是在完全替代招聘人员的工作。它只是改变了这份工作,因为招聘人员本来也不太想做那种机械式的、去联系 500 个人的事。让招聘这份工作更有技术含量、更有意思的部分,我会说,是文化契合度这类东西。让 AI 去做电话初筛、评估一个人在文化上契合度如何,这会非常难,还有其中的人际因素。所以我觉得他们发现,招聘人员
便签引用
21:50
human element of it. And so I think they're finding that the recruiters themselves are actually really excited to use the agents because it frees them up to do the work that they feel is like unique and interesting. >> The other shocking stat is that the companies that really accelerate during the batch, they really start taking off. One of the things that we started experiencing this year that we never experienced in the past is we have companies breaking from zero to seven figures in revenue during the batch. And that is in a span of 3 months.
自己其实非常愿意用这些 agent,因为这把他们解放出来,去做那些他们觉得独特而有趣的工作。>> 另一个令人震惊的数据是,那些在批次期间真正提速的公司,是真的起飞了。今年我们开始遇到一件以前从来没遇到过的事:有公司在批次期间就从零做到了七位数的收入。而那只是三个月的时间跨度。
便签引用
10卖数据与 RL 环境的隐形赢家
22:22
And that's shocking. In the past, that would have taken for a company to get to that about 18 months or more. And they're doing it in that amount of time. Part of it is they're solving real problems. And because of agentic coding, they're actually building products that are a lot more mature as well. And they're able these founders that are super AID run, I don't know, 20 coding agent sessions to get to that product maturity. >> And there's another category of companies that's also been growing super fast recently, which is companies that sell data or RL environments to the labs.
这很震撼。过去一家公司要做到那个程度,大概需要 18 个月甚至更久。而他们在这么短的时间里就做到了。部分原因是他们在解决真实的问题。而且因为有了 agentic coding,他们做出来的产品也成熟得多。而这些非常「AI 化」的创始人能同时跑,比如说,20 个 coding agent 会话,把产品做到那种成熟度。>> 还有另一类公司最近也在超高速增长,就是那些向各大实验室出售数据或 RL 环境的公司。
便签引用
22:59
>> This one might be interesting to talk about because a lot of these companies are pretty stealthy. they tend to have a disincentive to talk about how well they're doing instead, you know, compared to most companies that like to talk about how well they're doing. And so I think people out there might not realize how big a category this has become. When YC funded scale back in 2016, this was like a tiny little niche thing. It wasn't even a category. There was initially it was basically just scale who was doing it and then Meror began to do it and then like a couple other companies, but the last couple years has become a big category. We pulled the data recently and um just in the last two years, YC has funded more than a dozen companies that are each making more than $10 million a year selling data or RL environments to the labs >> and in many cases hundreds of millions of dollars.
>> 这个可能挺值得聊的,因为这些公司里很多都相当低调。它们往往有某种动机不去谈自己做得有多好,跟大多数喜欢宣传自己做得多好的公司正相反。所以我觉得外面的人可能没有意识到这已经变成多大的一个品类。2016 年 YC 投 Scale 的时候,这还只是个很小的细分领域,甚至都算不上一个品类。最开始基本上只有 Scale 在做,然后 Mercor也开始做,接着又有几家公司,但过去这两年它已经变成一个很大的品类。我们最近拉了数据,嗯,仅仅在过去两年里,YC 就投了十几家每年靠向实验室出售数据或 RL 环境赚超过 1000 万美元的公司 >> 而且很多情况下是好几亿美元。
便签引用
23:43
>> It makes hundreds of millions of dollars and these are companies that were just just a couple of years old. >> That's pretty fast to revenue honestly. >> Yeah, it's like pretty bananas. Do you want to talk about any of them Gary? >> Uh I mean the big ones I mean I think After Corey and data curve both really really great. I mean there probably too many to name that are honestly like maybe don't even want to be mentioned because well you know once you have something that's working you almost don't want people to know. I think that it's kind of natural to understand this though. I mean data is one of the legs of the scaling law and you know much has been made of compute but without the data how are you going to make these models that much better? Um the RL environment thing is interesting. I mean there's a lot there. I mean there's a lot of like pure customization that's happening for specific use cases like you you'll have like RL environments for finance for instance and someone can go
>> 赚好几亿美元,而这些公司也就成立了短短几年。>> 说实话,这个收入速度相当快。>> 是啊,挺疯狂的。Gary,你想聊聊其中哪几家吗?>> 呃,我是说那些大的,我觉得 After Corey 和 Data Curve 都非常非常棒。可能多到数不过来,而且说实话有些公司大概根本不想被提到,因为你知道,一旦你手里有个奏效的东西,你几乎不想让别人知道。不过我觉得这件事其实挺容易理解的。数据是 scaling law 的一条腿,你知道,大家谈算力谈得很多,但没有数据,你怎么把这些模型做得好那么多?嗯,RL 环境这块很有意思。里面东西很多。有大量针对特定用例的纯定制化工作,比如你会有专门做金融的 RL 环境,而有人可以在这上面无限深挖下去,这既需要一点领域专长,也需要
便签引用
24:36
very very infinitely deep with that and it's like a little bit of expertise it's a bunch of computer science it's some systems work but RL seems to be I mean one of the big engines for how I mean people are maybe bench benchmark maxing a little bit more than they should but uh it costs money to do it and it's seemingly here to stay in terms of how big model companies are going to approach it. >> Reportedly, the big labs are spending about a billion dollars on this. It's not a very known fact, but there's there's actually a real business to be built around this and a real environment is the current flavor of it and there's things with long-term horizon tasks that are getting built up and I think that is starting to also emerge in robotics. The labs also want to solve the problem of getting AI to work on the physical world. So they need a lot of the environments in the real world. So things with egocentric data, tea optic task starting to emerge as a big data category where labs are spending eight
一堆计算机科学的东西,还有一些系统工程,但 RL 看起来是,我是说,是最大的引擎之一,虽然人们可能有点过度地在刷榜,但呃,这是要花钱的,而且从大模型公司打算怎么做来看,它看起来是会长期存在的。>> 据说,大实验室在这上面花了大约十亿美元。这不是一个广为人知的事实,但围绕这件事确实能做出一门真正的生意,RL 环境是当下的形态,还有一些长周期任务的东西正在被搭建起来,我觉得这在机器人领域也开始出现。这些实验室也想解决让 AI 在物理世界里工作的问题。所以他们需要大量现实世界中的环境。所以第一视角数据、遥操作任务这类东西,正在成为一个很大的数据品类,实验室在跟这些公司
便签引用
25:38
nine figure deals with these companies. We had a number of companies in the batch that work on that and been able to close close revenues in that in that space. companies like um in the current bachelor of summer 26 there's practis robotics there's one that I'm working with that has like a network of places across the world where industrial moduction gets done they collect data from that there's this other company that Brad work with called deep reach that also has data that local entrepreneurs in across the world do >> and human archive and winter 26 yeah there have been a bunch of these companies recently >> right >> I think like if I were going prognosticate like one of the things going back to the uh you know all systems of record need to be AI harnesses they might also need to start training their own models and that's where things like river AI or tinker start becoming really interesting out of the box like you can sit there in cloud code or even open I use openclaw to train my own models which is very fun
签八位数、九位数的合同。我们这一批里有好几家公司在做这个,并且已经在这个领域拿下了不少收入。比如说,在当前 2026 夏季批次里有 Practis Robotics,还有一家我在带的公司,他们在全球有一个做工业生产的场地网络,从中收集数据;还有另外一家Brad 在带的公司叫 Deep Reach,他们的数据来自世界各地的本地创业者 >> 还有 Human Archive,2026 冬季批次,是的,最近确实出现了一批这样的公司 >> 对 >> 我想如果让我做点预测的话,其中一件事是,回到刚才说的,所有记录系统都需要变成 AI harness,它们可能还需要开始训练自己的模型,而这正是River AI 或者 Tinker 这类东西开始变得非常有意思的地方,开箱即用,你可以就坐在 ClaudeCode 里,甚至我用 OpenClaw 来训练自己的模型,这很好玩,它会自己做数据清洗之类的一切,但到目前为止这还没有
便签引用
26:39
it'll do its own data cleaning and everything but to date like that hasn't been a huge factor but I can see that becoming a much much bigger factor I And when you have proprietary data and you can train I mean the open open weight models are uh really nearly frontier if you can like sort of special purpose train these things to do um even better than what the frontier can do like that that's going to be really really powerful. >> I think this is actually going to be even bigger in robotics. I mean this is a hypothesis is not proven yet but robotic foundation models in robotics I think uh have a very different characteristics versus LLMs. LM is the whole thing is you model reality as language and for robotics you model reality in the physical 3D space which has way more degrees of freedom and perhaps in order to get robots to work in a specific vertical like let's say robots that do operations in data centers. I have this company called Boost Robotic that build robots for data centers like doing the cabling. It is
成为一个很大的因素,不过我能看到它会变成一个大得多的因素。而当你有专有数据、并且你可以训练的时候——我是说现在的开放权重模型已经非常接近前沿了——如果你能针对专门用途训练这些模型,让它们做得比前沿模型还好,那会是非常非常强大的。>> 我觉得这件事在机器人领域其实会更大。这只是一个假设,还没被验证,但我认为机器人基础模型跟 LLM 有非常不同的特性。LLM 的核心是你把现实建模为语言,而机器人是你把现实建模在物理的三维空间里,那有多得多的自由度。也许要让机器人在某个特定垂直领域里真正可用,比如说在数据中心里做作业的机器人——我有一家叫 Boost Robotics 的公司,做数据中心机器人,比如布线。这些机器人是有可能真正跑起来的,而更好的做法是用一个在定制数据上
便签引用
27:41
possible for these robots to work. It's better to get a model that's fine-tuned and trained on custom data that just works on that environment because the the thing that's also challenging for robotics, they need to be in real time and respond very quickly to to the stimuli and have an action plan which is different than LMS. LM you can have this this feature where you can just let it go and come back. But for Borax, you can't be because if if I don't know, let's say you connect a cable to the data center and then someone comes in and like knocks a robot out and things could get connected to the wrong plug, let's say.
微调和训练过、专门适配那个环境的模型。因为机器人还有一个挑战是,它们必须实时运行,对刺激做出非常快速的反应,并给出行动方案,这跟LLM 不一样。LLM 有那种特性,你可以让它自己跑,过一会儿再回来看。但机器人不行,你不能这样,因为如果……我不知道,比如说你正在给数据中心接一根线,然后有人进来把机器人撞了一下,东西就可能被插到错误的插口上,打个比方。
便签引用
28:17
>> Yeah. My understanding is that all the YC companies that are using physical intelligence as models to deploy robotics, they're all fine-tuning the PI models. I don't think any of them are able to use the PI models out of the box. Even though it's a great like starting point, you have to actually fine-tune it for like your specific case like data center cables in order for it to work. >> Mhm. You work with this company uh Ultra, right? >> Yeah. You start with the PI model, but then they have like thousands of hours of footage of like putting things in boxes that makes it really good at putting things in boxes. I've heard the argument basically that you know you could look at claude code like claude code can use its mo its uh code transcripts to figure out who the top coders are and you can take that and turn it around and you know basically train the next coding model to be even better. If you happen to own Tik Tok you happen to have all of the data on uh what people watch and click on and
>> 是的。据我了解,所有用 Physical Intelligence 的模型来部署机器人的 YC 公司,都在微调 π 模型。我不觉得有哪家能直接开箱即用地使用π 模型。虽然它是一个很好的起点,但你还是得针对你的具体场景去微调才能正常运转,就像数据中心的线缆一样。>> 嗯。你跟这家公司叫 Ultra 的合作,对吧?>> 是的。你先从 PI 模型开始,但他们有好几千小时的视频素材,都是把东西装进箱子里的画面,这让模型特别擅长把东西装箱。我听过这样一种说法,就是说你可以看看 Claude Code,Claude Code 可以用它的代码记录来分析出谁是顶尖程序员,然后你可以反过来利用这一点,基本上就是训练出下一代更强的编程模型。如果你恰好拥有 TikTok,你就恰好拥有全部数据,知道大家看什么、点什么、什么内容有吸引力,然后你可以用这些数据在 Seedance 里做出更有吸引力的视频。
便签引用
11单人创始人从 5% 涨到 19%
29:07
what's compelling and you can use that to make much more compelling videos in seed dance. So, you know, that's already been happening. I just, you know, I think that that that trend is going to continue in a fairly spectacular way from here. So, one of the things that we've been noticing, I think all of us have, is that frankly some of the most powerful and badass founders that we've been seeing lately, they're might be in their late 30s, 40s, even 50s. I mean, there's a sort of resurgence of the experienced founder. A lot of people seem to say that they want to be uh YC for solo founders, but it turns out YC is the YC for solo founders. Diana, you have a few stats that uh you found surprising.
所以你看,这种事其实已经在发生了。我只是觉得,我认为这个趋势从现在开始还会以相当惊人的方式延续下去。所以我们注意到的一件事,我想我们所有人都注意到了,坦白说,我们最近见到的一些最强、最厉害的创始人,可能都已经三十好几、四十多岁,甚至五十多岁了。我是说,经验丰富的创始人出现了某种复兴。很多人似乎都说他们想做单人创始人版的 YC,但结果发现,YC 本身就是单人创始人的 YC。Diana,你有几个数据,你觉得挺意外的。
便签引用
29:49
>> One of the shocking stats from analyzing the septic companies from a year ago, we used to only have about 5% of the companies accepted be solid founders and now we're over 18 19%. Which is a huge This is the highest spike that we've seen >> almost 1/5if of the batch. So, and it seems like it's going to keep going. You know, before you had you had to have like, you know, so many different skills. You had to be a great hustler. You know, you had to be able to explain and, you know, we would say like they have to be good talkers, right? Like you need someone who can, you know, uh be a hot person. You need someone who can actually come in and convince someone of something. Uh and then if you paired that with someone who is a world-class technologist, that's sort of the combo that is so ideal. Um, and so classically you would need co-founders to do that.
>> 分析这批录取公司时,跟一年前比有个很惊人的数据:我们过去被录取的公司里,只有大约 5% 是单人创始人,而现在已经超过 18%、19% 了。这是个巨大的——这是我们见过的最高涨幅——已经见过 >> 差不多五分之一的这一批公司了。所以,看起来这个趋势还会继续。你知道,以前你必须得具备很多种不同的技能。你得是个很能张罗的人。你得能把事情讲清楚,我们会说他们得很会说,对吧?你需要一个能够,呃,能撑场面的人。你需要一个真能坐下来说服别人的人。呃,然后如果你把这样的人和一个世界级的技术人才搭在一起,那就是那种特别理想的组合。嗯,所以按传统来说,你需要联合创始人才能做到这一点。
便签引用
30:37
Like you know any it you didn't necessarily need one but like it would increase your chances by so so much. I feel like a lot of that is like changing to this degree. It's becoming such that like knowing what to prompt and knowing what to build is so much more difficult and valuable than just knowing you know the CTO being able to code the thing. I think what's going on is that we've always actually had um hugely successful single founders. I think people don't realize this about YC. There's different sort of definitions of it, but for all intents and purposes, our Puver with Instacart, Brian Armstrong with Coinbase, at least when the batch started, were single founders.
就是说,你倒也不是非得有一个,但有的话你的成功概率会大得多。我感觉这一点正在发生很大的变化。现在变成了:知道该怎么提示(prompt)、知道该做什么东西,比单纯知道怎么让 CTO 把代码写出来要难得多,也有价值得多。我觉得现在的情况是,我们其实一直都有非常成功的单人创始人。我觉得大家没意识到 YC 的这一点。这个说法有不同的定义,但从各方面来说,Instacart 的 Apoorva、Coinbase 的 Brian Armstrong,至少在那一批开始的时候,都是单人创始人。
便签引用
31:15
>> Yeah, Parker Conrad got into YC as a single founder and then I uh interviewed Lakshiny his who ended up being his CTO. the bar for being able to like have the idea, be able to sell it, and be able to build it all by yourself was just really, really high. Um, and >> that's actually totally doable. >> Yeah, I think that's what's going on like in that case like those three are just like incredibly exceptional people. And so there's like very very few people who are capable of that. And now you can actually like get going and so I think you just don't have to be quite that like exceptional at least on one of those dimensions, the building part to be able to get going. But net net, it's still valuable to have co-founders. It's still a measure of like, >> you know, if your co-founders are super elite, like that means you're probably super elite and it just increases the chance of success by a lot. And >> in each of those cases, they did bring on co-founders. I think in each of those cases, you just get going and they got
>> 是的,Parker Conrad 当初是以单人创始人的身份进的 YC,然后我呃面试了 Lakshmi,他后来成了他的 CTO。那时候要能独自想出这个点子、能把它卖出去、还能自己把它做出来,这个门槛真的非常非常高。嗯,而 >> 现在这件事其实完全做得到了。>> 是啊,我觉得那种情况就是这样,那三位都是极其出众的人。所以能做到这一点的人非常非常少。而现在你其实可以直接上手开干了,所以我觉得至少在其中一个维度上——也就是动手做产品的能力上——你不需要那么出类拔萃,就能把事情启动起来。但总的来说,有联合创始人仍然是有价值的。它仍然是一种衡量标准,你知道的,如果你的联合创始人都是超级顶尖的人,那说明你大概率也是超级顶尖的,这会大大提高成功的概率。而且在这几个案例里,他们后来都招了联合创始人。我觉得每个案例都是这样,你先自己干起来,做出了些成绩,然后到了某个阶段再加入联合创始人。所以也许股权分配不太一样,或者说这种关系跟传统模式略有不同——传统模式是,
便签引用
32:06
traction and they added on co-founders sort of at a certain point. And so maybe like the um the equity ownership is different or maybe the dynamic is just slightly different to the traditional hey like you're you start out and like the two of you in a room and uh and you're completely 50/50. I don't know if you want to put stuff in but >> yeah I don't have the stats but I have definitely seen a greater trend towards that. People adding co-founders later in the company life cycle after the thing has already like gotten off the ground.
嘿,一开始就是你们两个人待在一个房间里,然后股权完全五五开。我不知道你想不想补充点什么,但是>> 是啊,我没有具体数据,但我确实看到这种趋势越来越明显了。人们在公司发展的后期阶段才加入联合创始人,也就是在项目已经跑起来之后。
便签引用
12经验型创始人与管理 agent 的手艺
32:34
>> I think that will be the trend. I think we'll see a like more single founders in the batch which we're already seeing like starting the batch but at least of the things that succeed I still expect that they're going to be adding co-founders um as the company progresses. >> Gary, do you also want to talk about the trend towards like more experienced people starting companies? >> So it does seem like uh people who have been around the block a few times are doing much much better. I think of Peter Steinberger as like sort of the canonical example like you know he's uh I believe in his early 40s and he'd been a dev manager. He'd worked on startups before and then you know he sort of uniquely got extremely AI pill with the clankers early but then he just tried a lot of stuff and then he knows what to build and so that's one thing that I think is actually really encouraging.
>> 我觉得这会成为一种趋势。我觉得我们会在这一批里看到更多的单人创始人,其实我们已经看到了,比如刚进批次的时候就是一个人,但至少在那些最终成功的项目里,我仍然预计随着公司的推进,他们还是会加入联合创始人。>> Gary,你要不要也聊聊那个趋势——就是越来越多有经验的人开始创业?>> 所以看起来,那些已经摸爬滚打过几轮的人,做得要好得多得多。我想到 Peter>> Steinberger 就是那种典型的例子,你知道,他我记得四十出头,之前当过研发经理。他以前做过创业公司,然后你知道,他算是很特别地很早就被 AI 洗脑了,很早就用上了那些 "clanker"(AI 工具),然后他就是尝试了大量的东西,而且他知道该做什么,所以这是我觉得其实特别令人鼓舞的一点。
便签引用
33:25
Like basically if you've been around the block, you know where the you know where the dragons are. You sort of uh have taste and then those people in particular are like unusually powerful right now. Yeah. I mean there just so many uh classic gatekept things that happen. It's like oh you know you have to have a co-founder, you need like a certain set of you know cool investors to be into you. And like now it's just less and less true. It's actually like you need to know what to build. That's like the high order bit now is you need to know what to build. And if you've lived a little bit and you've been in places and you're very opinionated, like actually now you might not have an excuse. Like what's your excuse? Like you've been this loudmouth on the internet for so long. Like you know, why are you not building something? Like just pop open open code and just go do it. You know, like put your money where your mouth is.
基本上就是,如果你摸爬滚打过,你就知道坑在哪儿。你多少有点品味,而这些人现在尤其厉害得离谱。是啊。我是说,以前有太多被把持着的门槛之类的东西。就像是,哦,你知道,你必须得有个联合创始人,你需要有一批比较酷的投资人看好你。而现在这些越来越不成立了。现在其实是,你得知道该做什么。这才是现在的关键位——你得知道该做什么。而如果你已经活过一些日子、待过一些地方、而且你很有主见,那其实现在你可能真没什么借口了。你的借口是什么?比如你在网上大放厥词这么久了,那你知道,你怎么不去做点东西呢?就像直接打开 open code 干就完了。你懂的,别光说不练。
便签引用
34:16
>> I also wonder if managing coding agents is actually like in some ways not that different from managing people. Oh yes. >> And so like people like Peter or you or or Boris Churnney and like Toby from Shopify who who have had whole careers like managing teams of engineers actually like take to this super well and can like spin up huge teams of coding agents and manage them maybe more effectively than even like a really smart 19-year-old who hasn't had those years of experience. >> Yeah. We can be a little bit less abusive to our agents, try to understand where they're coming from. uh you have to catch their emotions like 99.9% less.
>> 我还在想,管理编程 agent 其实在某些方面跟管理人没那么大区别。哦,是的。>> 所以像 Peter、像你、或者 Boris Cherny、还有 Shopify 的 Toby 这些有过完整职业生涯的人比如管理工程师团队——他们其实特别适应这种方式,能同时启动一大批编程 agent 并把它们管好,可能比一个真正聪明但还没积累那些年经验的 19 岁小孩还要更高效。》是啊。我们可以对 agent 少苛责一点,试着理解它们的出发点。呃,你得照顾它们的情绪,不过大概要少操心 99.9%。
便签引用
34:55
>> So yeah, it's pretty helpful. >> I wonder what's the concrete advice for uh someone that wants to get started and want to build a company like right now in the current era. >> I mean just start prompting. I mean opening up GPT6 today was pretty wild. I mean just that moment where your agents are, you know, palpably smarter. they, you know, a bunch of things that you've been annoyed about, like these bugs that, you know, you haven't had time to deep deep dive yourself. You just be like, "Actually, could you just go back to the list of things that you couldn't figure out?" Like, look at your, you know, all of our last chats and, you know, anything that looks like you didn't figure out like try to figure it out now and it'll do it like every single time. Like you know it's what a weird moment we are in history where >> you wake up in the morning you like wire up a new model >> and then these things that even a month ago you're just like why isn't it working it just starts working and like
》所以说,这还挺有帮助的。》我在想,对于那些现在就想上手、想在当下这个时代创业的人,有什么具体建议吗,就在当前这个阶段。》我觉得就是开始写 prompt 吧。说真的,今天打开 GPT-6 的感觉挺震撼的。就是那一刻,你的 agent明显变聪明了。你知道,那些一直让你烦心的事,比如那些 bug——你自己一直没时间深入去查的——你就可以说:“要不你回去看看之前那些你搞不定的问题清单?”比如去翻你的、我们所有过往的聊天记录,凡是看起来你当时没解决的,现在试着解决一下。它每次都能做到。你知道吗,我们身处一个多么奇妙的历史时刻——》你早上起来,接上一个新模型,》然后那些哪怕一个月前还让你抓狂“这怎么就是跑不通”的东西,突然就开始正常运行了。》一想到这可能
便签引用
35:53
>> you know to think that that might be this thing that we get to do for the next 18 24 months 36 months like who you know I don't know where when it ends but um that's coding in the time of AGI I guess. Well, that's all we have time for for today, but um if you can't tell, we're all pretty excited about what's going on right now, and you should be, too. So, we can't wait to see what you build.
就是我们接下来 18 个月、24 个月、36 个月要做的事……我也不知道这会在什么时候结束,但呃,这大概就是 AGI 时代的编程吧。好了,今天的时间就到这里,不过呃,你应该看得出来,我们都对眼下正在发生的事情非常兴奋,你也应该如此。所以,我们迫不及待想看看你会做出什么。
便签引用
视频总结 · 一句话概括与核心要点

一句话概括

YC 合伙人复盘 2026 年创业格局:AI 编程能力的跃迁正把资本和人才重新推回"硬科技"与端到端 Agent 产品,同时把创业的入门门槛(联合创始人、团队规模、融资背书)大幅拉低,让单人创始人和 40+ 岁老将成为新的优势群体。

核心要点

  • 硬科技在 YC 批次中的占比 12~18 个月内从 8% 飙到 20%,细分几乎全线翻三到五倍:机器人 1%→6~7%,工业制造回流美国 4%→10%,国防 1.5%→5%,半导体/光子 1%→近 4%,电力基础设施 1%→近 3%。
  • 驱动硬科技复兴的是"代码不再是瓶颈",而非软件失宠。过去做硬件公司需要几十上百名顶级软件工程师才能撑起全栈,现在 codegen 让一两个人加小团队就够(Palmer Luckey 就此讲过 Anduril 的提速),彻底改写了单位经济模型;同时更聪明的模型直接加速科研本身,让深科技初创更早拿到突破。
  • YC 里读博的创始人比例升至六分之一,与硅光子、新型芯片这类需要研究背景的方向直接相关,而这批人的表现明显优于平均。
  • 国防创业出现结构性窗口:老牌防务巨头靠 cost-plus 合同做"顾问",做不出新一代产品。Icarus(太阳能 U-2 式侦察兼通信中继机)已拿下七位数国防部合同,Nine Mothers(配 CV 的霰弹枪炮塔反无人机)卖给特种部队——保护极稀缺的特种兵免遭廉价商用无人机袭击,是刚需。
  • 硬件公司开始跑出软件式增长曲线:Knox Metals 在底特律空厂房里重建美国金属供应链,客户正是这批国防初创——老供应商"由老人经营、节奏太慢"跟不上,重演了 Web 2.0 时代新创业公司偏爱 Stripe 而非传统支付商的格局。
  • 算力紧缺已经反常到 A100 这种旧卡单位小时价格在上涨,供给远不足需求。由此催生从数据中心选址施工、软件规划、供电与储能,到底层芯片的整条创业链:Lamb Labs 做新处理器,Bot 押注三值(ternary)表示的定制架构(A100→H100→B300 精度一路从 FP32 降到 FP8,说明 LLM 根本不需要高精度浮点),Dipole Labs 做全光交换机——因为电子交换机已成 GPU 间通信的瓶颈。
  • SaaS 没死,但形态变了:批次中做端到端完整工作流的公司占比从 10% 涨到 25%+,反映在收入上就是入营时中位数为零收入,出营时中位数 MRR 从 8K 跳到 20K,且开始出现 3 个月内从 0 做到七位数年收入的案例(过去要 18 个月以上)。逻辑很直白:替你把活干完的产品,比只记录活的系统值更多钱。案例如 Juicebox,从 LLM 人才搜索升级到会主动联系候选人、安排面试的 Agent 后,单客户收入预计翻倍到三倍。
  • "数据/RL 环境"已从 2016 年 Scale 时期的小众品类变成隐形富矿:近两年 YC 投出十几家年收入超 1000 万美元的此类公司,部分达数亿美元,而它们往往刻意保持低调。大实验室在这上面据传每年花约 10 亿美元;下一波是机器人相关的第一视角数据与灵巧操作数据,已出现八九位数量级的合作。
  • 系统记录型软件的出路是变成"harness":只开放 MCP 等于把数据护城河让出去、迁移成本归零,必须成为用户和 Agent 真正干活的容器。Salesforce 推出自家 Slack AI harness 就是这个逻辑,而"Agent 比人更频繁地使用软件"反而可能支撑其估值——Codex、Claude Code、open code 等多个 harness 正在混战,且未必像浏览器大战那样只剩一个赢家。
  • 单人创始人比例从约 5% 跳到 18~19%,且 40 岁上下的老将异常能打。过去要一个"能说会道的 hustler + 世界级技术合伙人"组合,现在难点从"会写代码"转移到"知道该造什么";有阅历、有主见的人反而占优。另有一条被反复提及的机制:管理编程 Agent 与管理工程团队高度同构,所以带过团队的人(Peter Steinberger、Toby 等)能同时驾驭远比 19 岁天才更多的 Agent 会话。

结论与值得注意的细节

  • 机器人被明确定位为"还没到 ChatGPT 时刻但很近":Physical Intelligence 的 π 模型是好起点,但 YC 系公司几乎全都要微调才能用——Ultra 用数千小时装箱录像把模型调到擅长装箱,Boost Robotics 做数据中心布线机器人。机器人比 LLM 更吃垂直定制的根本原因是物理 3D 空间自由度远高于语言,且必须实时响应(缆线被人碰掉就得立刻纠正),没有"让它跑完再回来看"的余地。
  • harness 的价值被一个数据点具象化:同一模型在 ARC-AGI v3 上,换成定制 harness 后据称从低两位数提升到 90%+,说明"模型 + harness"才是产出单位,单独比模型分数越来越失真。视频同时提到 Astra 在某机器人基准上把完成率从约 10% 推到 60~70%。
  • 一个值得警惕的反例平衡:主持人自己指出,收入暴涨可能掺杂"AI 热潮下企业乱花钱"的成分;同时 SaaS 股(Salesforce、Snowflake 财报)已从年初低点强势反弹,说明"硬科技起飞 = 软件衰落"的叙事并不成立,更可能是两者同时成立。
  • 趋势的外推方向:拥有专有数据的系统记录公司下一步会开始自训模型(提到 River AI、Tinker 这类工具让训练变得像在 Claude Code 里对话一样简单),因为开源权重模型已接近前沿,专用微调有机会在垂直场景上超过通用前沿模型。Claude Code 可从代码轨迹反推谁是顶级程序员再去训下一代模型、TikTok 用消费数据训 Seedance,都是同一逻辑的既成事例。
  • 单人创始人的成功案例并非新事(Instacart 的 Apoorva、Coinbase 的 Brian Armstrong、Rippling 的 Parker Conrad 进 YC 时都是单人),变化在于过去需要"三项全能的极稀有人",现在建造这一维度被 AI 补齐了。但共识仍是:最终还是要加联合创始人,只是时点后移到有了牵引力之后,股权和关系结构因此不再是默认 50/50。
  • 给当下想开始的人的建议只有一句:现在就去 prompt。原话描述的状态是"早上醒来接上新模型,一个月前怎么都跑不通的东西突然就能跑了",并把这个窗口估计为还有 18~36 个月。
核心句型 · 10
1. It's not that …, it's just that …
“It's not that software and SAS is dead what people claim on the internet. It's just that is has transformed.”
先否定一个流行结论,再给出修正版本。用于纠偏时非常有力:不推翻现象,只重新定性。仿写:It's not that X is over, it's just that X has moved.
2. It has gone from X to Y
“It has gone from about 4% to 10% of the batch.”
汇报变化的标准句式,主语用 it 回指前文指标,避免重复长名词。数据播报连用多次即可形成节奏,比 increased by 更口语也更清楚起止点。
3. … but it turns out …
“A lot of people seem to say that they want to be YC for solo founders, but it turns out YC is the YC for solo founders.”
先摆出普遍说法,再用 it turns out 抖出反直觉事实。写作中常用来制造转折,比 however 更有揭示感,后接完整从句。
4. X as opposed to Y
“The companies in median are getting to 20,000 monthly revenue as opposed to 8K”
用于并置新旧数字或两个对立选项,重点落在前项。比 instead of 更中性、更书面,适合报告与数据对比。
5. one of the limiting reagents to doing …
“The top tier software engineers as one of the limiting reagents to being able to do really top tier full stack hardware”
借化学术语表达「决定上限的稀缺要素」。这类跨学科比喻在英语科技写作中很常见,仿写时注意 to 后接动名词。
6. What a weird moment we are in history where …
“What a weird moment we are in history where you wake up in the morning you like wire up a new model”
感叹句后接 where 引导的定语从句,把抽象感受落到具体场景。适合演讲收束,注意感叹句中主谓不倒装。
7. if anything, …
“If anything you'll just like agents will use software a lot more than humans will”
表示「要说有什么不同的话,反而是……」,用来加强而非削弱前面的判断。是英语里常被误当作让步的强调结构。
8. you either will …, or you kind of have to …
“If you are a system of record you either will be prayed upon … or you kind of have to be a harness”
把复杂局面压成非此即彼的二选一,是英语论证中制造压迫感的常用手法。kind of 在此软化语气,避免结论过硬。
9. net net, …
“But net, it's still valuable to have co-founders.”
商务口语,意为「把正反都算上,最终结论是」。用于承认复杂性之后给出净判断,位置在句首并带逗号。
10. that's less and less true
“And that's less and less true.”
用比较级叠加表达渐进变化,比 that's no longer true 更准确——承认旧规律仍有残余。仿写:that's becoming more and more common.
词汇精讲 · 136 · 按出现顺序
badass /ˈbædæs/ adj. 0:00
极猛的、狠角色的(口语,褒义)
resurgence /rɪˈsɜːrdʒəns/ n. 0:00
复兴、重新崛起(a resurgence of …)
wire up phr. v. 0:00
接上、连线接入(此处指把新模型接进工作流)
state of the art phr. n. 0:41
技术前沿现状;作形容词时写作 state-of-the-art
batch /bætʃ/ n. 0:41
一批;YC 语境特指一期加速器批次
median /ˈmiːdiən/ n. / adj. 0:41
中位数(区别于 average 平均数)
dive deeper phr. 1:52
深入展开、细谈(dive deeper into sth)
breakdown /ˈbreɪkdaʊn/ n. 1:52
拆分明细、分项数据(category breakdown 品类拆解)
photonix /foʊˈtɑːnɪks/ n. 2:24
光子学(规范拼写 photonics),用光传输与处理信号
quintupled /kwɪnˈtuːpəld/ v. 2:24
变成五倍(对应 tripled 三倍)
motto /ˈmɑːtoʊ/ n. 2:24
格言、口号
disproportionately /ˌdɪsprəˈpɔːrʃənətli/ adv. 3:28
不成比例地、远高于平均地
compounds /kəmˈpaʊndz/ v. 3:28
使复合叠加、加剧(金融义为复利增长)
supply chains n. 3:28
供应链
codegen /ˈkoʊdʒen/ n. 3:28
代码生成(code generation 的行业缩略)
silicon photonics n. 3:28
硅光子学,在硅基芯片上做光信号传输
limiting reagents n. 4:26
限制性试剂(化学术语),引申为决定上限的稀缺要素
top tier adj. 4:26
顶级的、第一梯队的
full stack adj. 4:26
全栈的,从底层到应用层通吃
economics /ˌiːkəˈnɑːmɪks/ n. 4:26
此处指一门生意的成本收益结构,即「经济账」
bullcase /ˈbʊlkeɪs/ n. 5:01
看涨理由(规范写法 bull case,反义 bear case)
shying away from phr. v. 5:01
回避、不敢碰(shy away from sth)
macro trends n. 5:01
宏观趋势
sovereign /ˈsɑːvrɪn/ adj. 5:45
主权的;科技语境指由本国自建自控的系统
front and center phr. 5:45
处在最显眼位置、备受关注
overwatch /ˈoʊvərwɑːtʃ/ n. 5:45
军事术语:居高警戒、战场监视掩护
comms /kɑːmz/ n. 5:45
通信(communications 的军用简称)
seven figure adj. 6:44
七位数的(百万美元级)
turret /ˈtɜːrɪt/ n. 6:44
炮塔、旋转枪座
elite /ɪˈliːt/ adj. / n. 6:44
精锐的;精英
existential /ˌeɡzɪˈstenʃəl/ adj. 6:44
关乎存亡的(existential risk 生死攸关的风险)
commodity /kəˈmɑːdəti/ adj. / n. 6:44
大路货的、无差别量产的;大宗商品
capture /ˈkæptʃər/ n. 7:27
此处指「俘获」:少数厂商对政府采购的把持
primes /praɪmz/ n. 7:27
主承包商(prime contractors 的简称)
cost plus adj. 7:27
成本加成(按实际成本加固定利润结算的合同)
dual use adj. 7:27
军民两用的
mega trend n. 7:27
大趋势、超级趋势
hollowed out phr. v. 8:33
被掏空、空心化(常用于产业衰退)
heartland /ˈhɑːrtlænd/ n. 8:33
腹地;美国语境特指中西部传统工业与农业区
sprung up phr. v. 9:05
迅速冒出、涌现(spring up 的过去分词)
sleepy /ˈsliːpi/ adj. 9:05
死气沉沉的、不思进取的(形容老旧企业)
keep up with phr. v. 9:05
跟上……的节奏
legacy /ˈleɡəsi/ adj. 9:05
遗留的、老旧的(legacy system 传统系统)
skyrocketing /ˈskaɪrɑːkɪtɪŋ/ v. 9:05
暴涨、直线飙升
appreciating /əˈpriːʃieɪtɪŋ/ v. 10:05
升值(反义 depreciate 贬值、折旧)
buildout /ˈbɪldaʊt/ n. 10:05
建设扩容、成规模的基础设施铺设
turnary /ˈtɜːrnəri/ adj. 11:05
三进制的(规范拼写 ternary),权重只取 −1/0/1
floating point precision n. 11:05
浮点精度(FP32/FP16/FP8 等)
interconnect /ˌɪntərkəˈnekt/ n. / v. 11:42
互连(芯片或设备间的数据通路)
switches /ˈswɪtʃɪz/ n. 11:42
交换机(网络设备,非「开关」)
bottleneck /ˈbɑːtlnek/ n. 11:42
瓶颈,限制整体速度的环节
workloads /ˈwɜːrkloʊdz/ n. 11:42
工作负载(计算任务类型)
optical /ˈɑːptɪkəl/ adj. 11:42
光学的、光路的(fully optical 全光)
vertical robotics n. 12:28
垂直行业机器人(vertical 指细分行业)
scaling law n. 12:28
缩放定律:模型性能随算力、数据、规模变化的经验规律
benchmark /ˈbentʃmɑːrk/ n. 12:28
基准测试、评测集
bootstrap /ˈbuːtstræp/ v. 13:30
不靠外部融资、用自有收入把公司做起来
downstream /ˈdaʊnstriːm/ adj. 13:30
下游的(投资语境指后续轮次的投资人)
drifted away phr. v. 13:30
逐渐偏离、漂离(drift away from sth)
extrapolated /ɪkˈstræpəleɪtɪd/ v. 14:28
外推、按既有趋势延续下去
surging /ˈsɜːrdʒɪŋ/ v. 14:28
激增、势头猛涨
blowout earnings n. 15:03
远超预期的财报(blowout 意为压倒性的)
system of record n. 15:03
记录系统:企业中存放权威数据的系统,如 CRM、ERP
intact /ɪnˈtækt/ adj. 15:37
完好无损的、未被破坏的
harness /ˈhɑːrnəs/ n. 16:04
原义马具;AI 语境指包裹模型、提供工具与上下文的外壳
browser wars n. 16:04
浏览器大战,1990 年代末网景与 IE 的市场争夺
anyone's guess phr. 16:04
谁也说不准(It's anyone's guess.)
prayed upon phr. v. 16:39
被捕食、被吞噬(规范拼写 preyed upon)
moat /moʊt/ n. 16:39
护城河,指难以被复制的竞争壁垒
trivial /ˈtrɪviəl/ adj. 16:39
轻而易举的、微不足道的
leap /liːp/ n. 17:34
跃升、飞跃
end to end adj. 17:34
端到端的,从头到尾整段完成
point solution n. 18:26
点状解决方案:只解决流程中单一环节的产品
intake /ˈɪnteɪk/ n. 18:26
接收登记;医疗语境指初诊接诊流程
non-trivially /nɑːnˈtrɪviəli/ adv. 18:26
相当可观地、绝非微不足道地
knock on it phr. n. 19:32
对某事的质疑或贬低(a knock on sth)
shelling out phr. v. 19:32
(不情愿地)大把掏钱
to be fair phr. 19:32
平心而论、公平地说(让步语)
value proposition n. 20:27
价值主张:产品给客户带来的可量化收益
write big checks phr. 20:27
开出大额支票,指爽快地大笔付款或投资
taking off phr. v. 21:03
起飞、爆发式增长
culture fit n. 21:03
文化契合度(招聘评估维度)
phone screen n. 21:03
电话初筛面试
frees them up phr. v. 21:03
把某人腾出来去做别的事(free sb up to do sth)
span /spæn/ n. 21:50
跨度、一段时间(in a span of three months)
maturity /məˈtʃʊrəti/ n. 22:22
成熟度(产品完善程度)
stealthy /ˈstelθi/ adj. 22:59
隐秘低调的(创业语境 in stealth 指未公开运营)
disincentive /ˌdɪsɪnˈsentɪv/ n. 22:59
反向激励、阻碍去做某事的动因
niche /niːʃ/ n. 22:59
细分市场、小众领域
pulled the data phr. 22:59
调取、拉取数据(pull data 为常用搭配)
bananas /bəˈnænəz/ adj. 23:43
疯狂的、离谱的(俚语,= crazy)
much has been made of phr. 23:43
关于……已被大书特书、谈得很多
customization /ˌkʌstəməˈzeɪʃən/ n. 23:43
定制化
benchmark maxing phr. 24:36
刷榜:为提高评测分数而针对性优化
here to stay phr. 24:36
会长期存在、不会消失
egocentric /ˌiːɡoʊˈsentrɪk/ adj. 24:36
第一人称视角的(egocentric data 头戴设备记录的视角数据)
long-term horizon tasks n. 24:36
长周期任务,需要跨越很长时间跨度才能完成
prognosticate /prɑːɡˈnɑːstɪkeɪt/ v. 25:38
预测、预言(正式用词,略带自嘲色彩)
out of the box phr. 25:38
开箱即用,无需额外配置
proprietary /prəˈpraɪəteri/ adj. 26:39
专有的、独家掌握的(proprietary data 专有数据)
open weight adj. 26:39
开放权重的,模型参数可公开下载
frontier /frʌnˈtɪr/ n. / adj. 26:39
前沿;AI 语境指能力最强的一线模型
degrees of freedom n. 26:39
自由度,系统中可独立变化的维度数
hypothesis /haɪˈpɑːθəsɪs/ n. 26:39
假设(复数 hypotheses)
fine-tuned /ˌfaɪnˈtuːnd/ adj. 27:41
经过微调的(在基座模型上用专有数据再训练)
stimuli /ˈstɪmjəlaɪ/ n. 27:41
刺激、外界输入(单数 stimulus)
footage /ˈfʊtɪdʒ/ n. 28:17
影像素材(不可数)
transcripts /ˈtrænskrɪpts/ n. 28:17
记录文本、会话记录
compelling /kəmˈpelɪŋ/ adj. 29:07
极有吸引力的、令人信服的
spectacular /spekˈtækjələr/ adj. 29:07
惊人的、蔚为壮观的
spike /spaɪk/ n. 29:49
陡增、尖峰(数据曲线上的急剧上升)
hustler /ˈhʌslər/ n. 29:49
能张罗、能跑业务的人(创业语境为褒义)
combo /ˈkɑːmboʊ/ n. 29:49
组合、搭配(combination 的口语缩略)
for all intents and purposes phr. 30:37
实质上、从各方面来说(固定搭配,不可改词)
the bar n. 31:15
门槛、标准(raise/lower the bar 提高或降低标准)
exceptional /ɪkˈsepʃənəl/ adj. 31:15
出类拔萃的、罕见优秀的
net net phr. 31:15
总的算下来、净结果是(商务口语)
traction /ˈtrækʃən/ n. 32:06
牵引力;创业语境指用户或收入上的实际进展
equity ownership n. 32:06
股权所有权、持股结构
gotten off the ground phr. v. 32:06
(项目)成功启动、跑起来了
been around the block phr. 32:34
见过世面、摸爬滚打过(完整说法 been around the block a few times)
canonical /kəˈnɑːnɪkəl/ adj. 32:34
典范的、最标准的(canonical example 典型例子)
AI pill phr. 32:34
被 AI 彻底说服(源自 red-pilled 的造词,常写作 AI-pilled)
clankers /ˈklæŋkərz/ n. 32:34
对机器人/AI 的戏谑称呼,源出《星球大战》
gatekept /ˈɡeɪtkept/ adj. / v. 33:25
被把持、设门槛不让人进入(gatekeep 的过去分词)
high order bit phr. n. 33:25
最高位;引申为影响结果最大的那个因素
opinionated /əˈpɪnjəneɪtɪd/ adj. 33:25
有强烈主见的(可褒可贬)
loudmouth /ˈlaʊdmaʊθ/ n. 33:25
大放厥词的人、话多的人
taste /teɪst/ n. 33:25
品味、判断力(知道什么值得做)
put your money where your mouth is phr. 33:25
别光说不练,拿真金白银证明你的主张
take to this phr. v. 34:16
很快上手、适应得好(take to sth)
spin up phr. v. 34:16
迅速启动、拉起(一批进程或团队)
abusive /əˈbjuːsɪv/ adj. 34:16
辱骂的、苛待的
where they're coming from phr. 34:16
某人的立场与出发点(understand where sb is coming from)
palpably /ˈpælpəbli/ adv. 34:55
明显地、可切身感知地
deep dive n. / v. 34:55
深入钻研、深挖(do a deep dive into sth)
精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.178Shopify Founder, Tobi Lütke: Why Taste Matters More Than Technical Skill 下一期 · NO.180 →Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview
苏菲周报 · THE WEEKLY 每周一封,
追问一个大问题。
苏菲拉底的每周来信,写这一周在追问的问题和看到的回应。
苏菲拉底
ASK THE BIG QUESTIONS · THINK DEEPLY · SEE THE WORLD DIFFERENTLY
苏菲拉底微信公众号二维码 微信公众号
© 2026 苏菲拉底 · 内容仅供学习 [email protected]