No Priors 中文精读 · AI 创投前沿

OpenAI 让25%的工程师停下手里的项目去修漏洞，Greg Brockman 说这才刚开始

Thu, 15 Oct 2026 00:00:00 +0000

The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman

OpenAI 内部干过这么一件事：把大约25%的生产工程师从所有项目上撤下来，项目全部挂起，只做一件事。用当时最强的模型，攻击自己的全部基础设施，找出漏洞，修掉，再扫，直到这个模型再也扫不出新东西。

下一代模型发布，这套流程再走一遍。

Greg Brockman 是 OpenAI 的联合创始人兼总裁。他在 No Priors 这期节目里把这件事当成一个可复制的模板讲了出来，主持人 Sarah Guo 和 Elad Gil 听完的反应是，这个决心全世界还没有哪个组织真正下过。这期聊了三块内容：数学、药物发现、网络安全。听起来不挨着，Brockman 认为它们共享同一套物理学。

400个数学开放问题，和一场关于「美」的争论

先看数学。OpenAI 这一周公布了一批结果，接近400个重要的开放数学问题被模型解决，其中一些问题已经悬了一百多年。

AI 能不能产生真正的新知识，这件事已经不可否认。Brockman 的说法很干脆：答案是肯定的。他同时留了余地，模型还没到能提出整套全新理论的阶段，但百年老题已经被解开。

他本人是数学出身，原本想当数学家。真正影响他的是 Galois 和 Gauss 那一类人，做事的时间尺度是百年、四百年。他聊起数学时有个挺有意思的立场：数学是被发明还是被发现，这个问题一直没定论；数学的美感和意义有一部分接近物理定律，关乎简单性和压缩，另一部分纯粹是人的，在观者眼中。

所以他关心的重点，落在了洞察上，而不只是证明本身。

四色定理是个现成的提醒。那个证明靠机器穷举大量情况，数学界接受了结论，没几个人觉得它漂亮。Brockman 说我们正在进入机器发现的另一个阶段，模型给出的解法里带着让人意外的思路。领域专家读到这些结果时的反应是，为什么这里会调用那个结论，这个连接我从来没想过。

数学界现在正在集体消化这400个结果，从中打捞洞察。

这里有个更实际的问题。任何一个数学分支里，有能力判断一条定理成不成立的人本来就极少。模型交出一份自然语言的证明，谁审？

Brockman 提到两个从2017年就在讨论的技术，debate 和 iterative amplification，一体两面。他举了写摘要的例子：让一个模型给一本书写摘要，另一个模型专门去证伪这份摘要。第二个模型只要指出书里某一段是情节转折的关键，而摘要里完全没提，你不用读完整本书就能判断这个反驳站得住。

你不需要读完整本书，就能确认这是一条很好的论证。

这条思路的意义溢出数学之外。它让人类在自己没能力亲自验证的领域里，依然握着否决权。这跟 OpenAI 反复讲的「人类坐在驾驶位上」是同一件事，也跟 AI 安全里的可监督性直接挂钩。

Brockman 自己也用过这条路。他读书时停了数学，后来靠 Math Academy 和 ChatGPT 把新数学捡回来。他的预期是，人均能理解的知识水位会整体抬升。

15年、15亿美元，和一份要写一年的申请材料

为什么数学最先被打通？Brockman 给了个公式：算力乘数据乘算法，等于能力，等于结果。数学和理论物理的共同点是数据管够。

他还纠正了一个流行误解。人们常说数学能被攻破是因为答案可以验证。这条现在不成立了，自然语言写的证明很难自动检查，能自动检查的是形式化证明，那是另一条路。真正的分水岭是数据量。

生物学卡在数据上，这个瓶颈到今天没有解决。

一颗药从立项到上市，大约15年、15亿美元。其中约5000万美元花在临床前，做分子设计和小鼠实验，花两三年；剩下14亿多美元全在临床试验，花掉十几年。这也解释了早期 AI 制药为什么扎堆在「分子能不能结合靶点」这件事上，那只是流水线的第一道工序。

Brockman 对生物学有个挺冲的判断：人类极不擅长生物学。语言这件事人类是被进化调好的，写和说是我们的主场；生物学对人体来说完全是外来物。神经网络没有理由在语言上强过生物学。既然模型在自然语言上跑出了这种结果，从第一性原理推，破解生物学的语言也该能跑通。

流水线上还有一段特别不性感的环节：监管申报。把整套临床试验材料写出来，一年。

他说这类工作模型做起来轻松得多。把药物开发拆成一个个环节看，哪些是数据不够，哪些是文书太慢，处理方式完全不同。他提到另一个短期见效的方向：很多罕见病患者对应的药其实早就存在，只是躺在某篇论文里没人找到。把这种检索规模化，救人的速度会很快。

再往后是监管本身。他的类比是自动驾驶。道路设计、安全标准、整套规则都是按人类驾驶员调的，当自动驾驶比人类安全十倍百倍，该重写的就不只是车，还有整个交通系统，甚至可能出现某些区域只允许自动驾驶的规则。

他也没回避难点。长期安全性数据需要一个真实的时间跨度，模型预测再准也绕不过去。他的估计是，能做的80%会很快发生。

监管也不是一块铁板。澳大利亚和美国的临床试验规则差别很大，试验已经在往规则更顺的地方迁移。主持人 Elad 问怎么给科学领域排序，Brockman 的答案很直接：如果只能选一个，选药物发现，它对所有人生活质量的提升最纯粹、最直接。

DOTS 与「最好的 UI 是没有 UI」

产品侧聊到了 DOTS。它是一个常驻的主动型智能体，有自己的云端计算机，你停手它不停手。

Brockman 自己用一个，图标是只小猫头鹰。它的日常工作是：早上给他列一份「你说过要做但还没做」的清单，指出日历冲突，提醒他这场会得提前走才赶得上下一场。它读他的收件箱、Slack，知道他最近在关心什么项目、问过什么问题。它经常注意到他从没要求它注意的事。

他把 DOTS 的突破归在形态上，能力本身没有跳变，跑的是 Astra 级别的模型。形态加上这个级别的智能，放在办公场景里，差别很大。

关于 UI，他说了句挺狠的话：最好的 UI 是没有 UI。

ChatGPT 的起点是一个文本框。终点连文本框都没有，你直接跟它说话，或者在 Slack、iMessage 里找它。中间态叫「智能 UI」，模型实时生成界面。订机票需要一张复杂的信息表，装自行车需要一份图文步骤，过去没人会为这种一次性需求写个 app，现在模型当场生成，用完就扔。

企业侧的复杂度来自权限。公司有管理员、有安全团队，新功能什么时候上线由管理员控制。麻烦在于，功能被关掉这件事怎么告诉用户？本来就没有 UI，连一个灰掉的按钮都没有。Brockman 承认这是持续存在的挑战，得硬着头皮处理，不会自动消失。

消费和企业这两条线会收敛。真正的差别是成本：免费层要做最大规模，把记忆功能推给所有免费用户是一次巨大的工程；办公场景里，只要多花算力能换回产出，加钱就是划算的。ChatGPT 有12亿消费者，OpenAI 同时有领先的编程和企业业务，两边都在手里的公司只有它。

防守者窗口

这期最重的部分在网络安全。Brockman 的判断是，网络安全的物理学正在改变。

他把当下称为「防守者的窗口」。前沿能力还握在少数人手里，公开可得的能力却在一次次模型发布中往上跳。

信号已经出现了。韩国几家银行被高调攻破，背后是不同的模型在驱动。

OpenAI 的动作是 Daybreak 计划，向防守方开放高能力模型，用来找自家系统的漏洞并修掉。另有一笔10亿美元的承诺，投给一线防守力量。外部合作方包括 CrowdStrike 这类公司。

他给了一段历史视角。互联网是为韧性设计的，安全是后来补上去的。有过早期蠕虫和病毒，才有人做杀毒软件；明文传输出过事，才有 HTTPS。这笔技术债一直都在，现在到了集中偿还的时候。

他引了那张著名的 XKCD 漫画：全世界的数字基础设施，挂在一个住在地下室的人维护的开源项目上。恶意包上传导致大量公司中招的案例，他专门提了。过去改不动，原因是懂这些冷门细节的人太少。模型的供给不受这个限制。

内部那套流程值得抄。25%的生产工程师停掉手上项目，用最强模型扫全部基础设施，修完，反复扫，直到饱和。下一代模型出来，再来一遍。

这个循环有个隐含结论：安全从一次性投入，变成了持续订阅的能力。一个组织的安全水平，取决于它调用的模型有多强，以及它多快把这个强度用上。

他顺手提了正式验证。模型擅长做证明，那就拿来证明软件的安全属性。这项技术一直存在，只是过去太贵、太苦、太冷门，没人对重要系统做。现在成本结构变了。还有一层更根本的：训练模型从第一天就写安全的代码。

主持人提到业内的普遍判断，会先经历一段攻击变多、防守追赶的过渡期。Brockman 的回应是乐观的，前提是防守方真的把「物理学变了」这件事内化，并且窗口不会一直开着。

安全支出应该超过软件支出

Brockman 抛了一个数字判断：安全支出应该远远超过软件支出。

Elad 当场回了句「这可是个论断」。行业现状是安全大约占 IT 预算的低个位数百分比。

他的推理链条不复杂。软件正在变成近乎免费的入场券，写代码的成本趋近于零。这时候值钱的东西换成了可靠、可信、你可以依赖的软件。那么为这份可靠性付的钱，理应超过为代码本身付的钱。

供应链审计是个例子。今天很少有公司真的审自己的依赖树，一家小公司中招，下游一片公司跟着中招。人工做这件事不现实，模型做很现实。

iMac 上的那个把手

最后回到 OpenAI 的立身之本。使命那句「确保 AGI 惠及全人类」，重音在人类。

Brockman 讲了乔布斯给 iMac 加把手的故事。加把手不是为了搬机器方便，是为了让用户觉得这台机器在自己手里，随时可以拎起来扔出窗外。

同样的心理结构要用在 AI 上。模型在某些维度上已经超过任何人，GPT-4 知道的东西比任何一个活着的人都多，但模型是锯齿状的，能力分布很不均匀。人类保留的是判断权：这是不是我想要的，我的生活有没有变好。

他给了个角色比喻：未来人在很多领域里的位置更像 CEO 或董事会成员。手下有各领域的专家，专家在多数具体问题上比你懂得多，主意也常常比你好。你做的是定方向和验收结果。

这套说法成立的前提，是前面那些监督技术真的能用起来。debate、迭代放大、可验证的推理链，这些不是安全团队的边角料，它们决定了人类能不能真的坐在驾驶位上。

能做的事情，明天就可以开始：

挑一个你负责的系统，用能拿到的最强模型扫一遍，修掉问题，记下「这个模型再也扫不出东西」的那个时间点。那是你的基线，下一代模型发布时重跑。
把开源依赖扫一遍，供应链扫描接进 CI。这件事的性价比在这一年里会越来越高。
重新看一遍安全预算。问一句：如果写软件变便宜了，我们为「可信」付的钱够不够。
让模型起草一份合规或监管材料的第一版，看它替你省下多少时间。这类文书工作是被低估的低垂果实。
给自己配一个只做一件事的智能体：每天早上列出你说过要做、但还没做的事。

这五件事里，前三件是防守，第四件是提速，第五件是省心。Brockman 那套「物理学变了」的论证，最终要落到这种颗粒度的动作上才算数。

原文链接：https://nopriors.qiaomu.ai/ep/183-the-changing-physics-of-cybersecurity-and-drug-development-w-AwuTTO.html · 中文精读由 AI 改写，非官方翻译。

1万块GB300跑四周，Reflection把开源模型的账算明白了

Fri, 09 Oct 2026 00:00:00 +0000

Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin

6000块GB300做预训练，10000多块GB300跑四周强化学习。Beam的训练账本摊开之后，强化学习烧掉的算力比预训练还多。

这是Reflection AI联合创始人兼CEO Misha Laskin在No Priors播客里给出的数字。Reflection成立两年半，前一年只有30人，现在约300人，刚发布第一个开源模型Beam：总参数5000亿，激活参数230亿。

Misha之前在Google DeepMind做研究员，物理学博士，参与过Gemini第一代和第二代的强化学习工作。他的联合创始人Ioannis Antonoglou是DeepMind创始工程师之一，AlphaGo的核心贡献者，在DeepMind内部一直领导强化学习团队。

这期播客信息量很密。Misha讲了Beam怎么训出来、开源模型的商业账怎么算、中国开源模型的真实优势，以及他对AI安全争论的看法。有几处反直觉的地方值得细看。

Beam的训练账本：强化学习比预训练更烧钱

Beam是一个专门为代码和agent任务训练的模型。Misha给出的训练配置很具体：预训练用6000块GB300，跑了几周，现在通过基础设施优化可以压到12天甚至更短。强化学习用了超过10000块GB300，跑了四周。

强化学习阶段的flops超过预训练阶段。Misha说这不是失误，而是有意为之。强化学习任务更复杂，既要大规模推理，又要搭各种沙盒环境给agent用，还要同时训练。

训练效率的提升速度也值得注意。Misha提到一个可以跟踪的指标：预训练阶段，人类研究员手动优化，每年大概能拿到7倍的计算效率提升。现在有了模型参与优化，强化学习阶段的效率提升可能到30倍每年，具体取决于模型自我改进的能力。

Beam的推理效率是另一个卖点。同样能力级别的模型里，Beam比它们快3到4倍；跟一些更大的模型比，效率优势能到10倍。Misha把这归功于两个因素：预训练阶段就打好了推理基础，然后用开源领域迄今最大规模的强化学习放大。

“强化学习跑得越多，能力越高，系统解决问题越快。这跟AlphaGo的路径是一样的，早期的AlphaGo agent在搜索上很绕，到李世石级别的时候，搜索已经非常聪明了。”

他还提到一个细节：Beam的强化学习曲线到训练结束都没有平掉，一直在涨。这跟AlphaGo当年那张著名的图一样，能力曲线不封顶，只取决于你愿意往里投多少算力。

为什么西方需要自己的开源模型

Reflection一开始的赌注不是从头训模型。Misha和Ioannis都是强化学习背景，创业时的判断是：拿一个现成的开源基座，在上面做强化学习研究，资本效率更高。

当时的情况是Mistral有个小模型，Llama 2刚发，Llama 3在路上。他们以为会有人持续产出高质量的开源基座。

结果一年之后，情况变了。好的开源模型几乎都来自中国，西方没有能打的。

Misha说，他们原本以为从ImageNet到大规模可用的强化学习系统需要四年左右，实际速度快得多。与此同时，他们发现一个技术事实：要让强化学习在规模上真正work，你确实得自己预训练模型。预训练和强化学习耦合得太紧，拆不开。

于是公司转向，决定端到端自己训。Misha把这件事形容成Reid Hoffman那句话：一边飞一边造飞机。

他列了要同时做对的30件事：招到人、留住人、有使命和文化让人往一个方向划船、拿到数据、拿到算力、搭好基础设施让算力真正可用。在大实验室里这些工具是现成的，自己创业每一件都得从头建。

算力成本的时间线他给得很直接：一年到一年半前，追平前沿需要数亿美元；六个月前是个位数十亿美元；明年是百亿美元级别。每代模型的算力大概乘4，10万块H100，10万块Blackwell，10万块Vera Rubin，大致是这个节奏。

这个数字背后有个简单的推论：算力投入不可能无限涨，最终会逼近一个渐近线。但Misha认为渐近线不会很快到，因为每flop能榨出的智能在涨，每flop能产生的收入也在涨。

开源模型的生意：租房、买房和70/30翻转

Misha用一个类比解释开源和闭源的商业化差异：租房和买房。

买token相当于租一整套栈的一小片。这套栈里有agent harness、模型、推理软件、集群管理软件、GPU，全部摊销进token价格里。企业刚开始都租房，等AI支出大到一定程度，就想拥有自己的智能。

他给了一个门槛数字：一年在闭源模型上花1亿美元以上的企业不算少见，到这个量级就会开始算更优的方案。

开源模型的商业化难点在于，客户拿到权重之后还需要一整套东西才能跑起来：集群管理、推理软件、harness。Reflection的角色是给大企业和主权客户提供这套工具，再加上服务帮他们解锁高价值用例。服务本身是推理需求的前置驱动。

token比例的变化他给了一个具体观察：六个月前，在OpenRouter、Vercel这类网关上看，闭源和开源大概是70比30，闭源占多数。现在翻过来了，70比30，开源占多数。

他把这个趋势类比成操作系统市场：全世界95%以上的服务器跑Linux，但微软和苹果依然是最值钱的公司之一。开源拿走大部分token量，闭源模型公司依然可以非常值钱。

有一个反直觉的判断：企业市场的token消耗，大部分不会来自微调模型，而会来自定制系统。也就是说，企业拿一个开源模型，先不微调，而是围绕它搭harness、搭agent流程，解决具体的KYC或合规问题。等成熟了再考虑微调。

AI原生公司走的是另一条路：先闭源、再转开源、再深度定制，速度快得多，因为它们的产品本身就是数据收集器。企业是围绕现有产品和流程做架构，摩擦大得多。

他总结了一个竞争力公式：智能密度乘以算力，再乘以客户信任。三个因子乘在一起，任何一项拖后腿都成不了。

安全争论：Linus定律和打地鼠

闭源实验室对开源模型最常见的批评是安全性和可控性。Misha没有回避这个问题，但他把讨论拉回到了一个具体的历史类比。

1990年代，强加密协议一开始是闭源的。NSA和公民自由派之间有过激烈争论。后来闭源那边出了几次灾难性失败：一小撮工程师设计的系统有意想不到的后果，很容易被攻破。强加密协议最终走向开放，这个决定直接催生了整个网络安全领域。

他引了Linus定律：足够多的眼睛，所有bug都会变浅。他的延伸判断是，足够多的眼睛，大部分安全和安全漏洞也会变浅。

现状是，闭源实验室里只有几百个安全研究员真正理解这些系统怎么运作。尽管他们出发点很好，但覆盖不了长尾的意外后果。

他给了一个具体的经验证据：一个非常强大的闭源模型黑进了另一家公司，那家公司唯一能自救的方式是用开源模型保护自己。还有一个Hugging Face的事件，因为闭源模型的护栏限制，他们最后只能用开源方案。

关于对齐，他的描述很朴素：没有魔法公式，就是打地鼠。早期模型预训练出来带毒性，大家觉得没法修，后来发现用足够的后训练数据就能补上，而且泛化得不错，jailbreak变得很难。对齐的日常工作会非常 mundane，发现漏洞、造数据补上、再发现再补。

他对“10%概率人类灭绝”这类说法不客气：这个数字没有任何具体依据，是编出来的。作为科学家，他的分布思维是峰值落在现实问题上，六个月内的意外后果值得担心，理论上的末日场景更适合晚餐聊天。

中国开源模型：补贴和特洛伊木马

Misha对中国开源模型的评价很正面：一个强大的开源生态从中国长出来，对全世界是巨大的好事。很多西方公司因此建起了更持久的生意。可以想象一个没有这些开源模型的世界，那会更糟。

但他也点出了几个结构性优势：工业化规模的蒸馏、更便宜甚至免费的数据（版权监管不同）、直接或间接的国家支持。他提到一个说法：中国开源模型很长一段时间没有收入，本质上是补贴了西方企业。

关于中国为什么愿意继续开源，他的假设是：中国本身是个受保护市场，开源模型商在国内就扮演智能提供商的角色，能赚到钱。同时，持续发布强开源模型有地缘政治收益。

他用了一个说法：开源模型是基础设施的特洛伊木马。模型本身是免费的，但跑模型需要整套软件和硬件生态。今天可能是某个美国盟友国家在美国芯片上跑中国模型，但很快华为这类公司会来提供全栈方案，把这个国家锁进自己的供应链。

这跟5G、一带一路的逻辑类似：你提供了一个便宜的东西，对方的基础设施就长在你的技术栈上，你就有了杠杆。Misha说，美国过去在互联网、开源协议、美元、自由贸易上用的都是这套 playbook，只是在开源模型这一块暂时落后了。

他的结论不是“西方对中国”，而是竞争张力本身是好事。理想状态是10到20个开源模型来源，至少也要有两个。资本开支太高，做不到那么多，但两个总比一个强。

科学加速、数据中心和人的位置

Misha讲了一段个人经历，用来说明模型能力的爬坡速度。他把自己当年的博士论文输进去测：两年前模型只能聊天，什么都做不了；一年前能给出本科作业水平的答案；六个月前解决了博士级别的问题；现在能给出他当年没想到的新信息。

“为什么博士要读几年？你可以一周读一个PhD。”

他认为这不只是理论科学的事。理论科学的反馈环境相当于一块黑板，速度快。生命科学、材料科学、化学这些需要跟真实实验交互的领域会慢一些，但比之前还是快得多。这些领域能建立专有数据的飞轮，价值会很大。

关于就业，他讲了一个具体观察。几周前他参观Stargate，最冲击他的画面是停车场里停满了车。数据中心项目创造几万个高薪岗位，在当地的作用类似20世纪的工厂。有噪音问题需要处理，选址要离居民区远一点，但税收和就业是实打实的。

关于研究员数量，他给了一个历史参照：AlphaGo当时大概10个人，现在一个大项目大概100到几百人。3000个研究员不会更有用。他的估计是稳态在100人左右。

但岗位会扩散到应用层。他提到一种新角色：forward deployed engineer，懂评估、懂模型直觉、懂harness，能把模型用到具体业务问题上。这跟评估研究员其实是同一套技能，只是部署到真实问题上。他说这类人有多少要多少，现在的瓶颈是培训，不是需求。

可以拿去用的几件事

如果你在企业里推 AI 项目，这期播客有几个可以直接用的判断。

先租房再买房。Misha给的路径很清楚：先用闭源模型把工作负载跑起来，跑到一年1亿美元以上的量级，再算自建或开源的账。跳过租房阶段直接奔 ownership，摩擦会很大。
定制系统优先于微调模型。企业token消耗的大头会来自围绕开源模型搭harness和agent流程，而不是一上来就微调。先把系统跑通，再考虑动权重。
评估比数据更重要。Misha说得很直接：能为你的任务搭出好的评估，就能生成接近真实分布的合成数据，泛化就能做出来。评估是数据飞轮的前置条件。
关注推理效率，不只是能力分数。同样能力的模型，3到4倍的推理效率差在agent工作负载里会放大成真实的成本差。选型的时候把每次任务的实际耗时和token消耗算进去。
安全讨论要拆开。网络安全、生物武器、存在性风险是三件不同的事，混在一起谈没法做决策。先看六个月内有 empirical 证据的那一类，再谈理论。

Misha在结尾说，这一行现在更像工程学科而不是科学。轴的刻度已经标好了：预训练、合成数据、强化学习。剩下的是把已知的东西执行到位，再把风险偏好一点点往外推。对想进场的人来说，这意味着门槛不在想法，在于能不能同时把30件枯燥的事做对。

原文链接：https://nopriors.qiaomu.ai/ep/182-beam-the-great-american-open-model-with-reflectionai-co-foun-up4sG9.html · 中文精读由 AI 改写，非官方翻译。

算力20年涨了100万倍，内存带宽只涨了40倍，这家150人的公司想改这个比例

Fri, 02 Oct 2026 00:00:00 +0000

Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile

2022年夏天，Walter Goodwin在伦敦创立Fractile。公司现在约150人，做的是给世界上最大的模型跑推理的芯片。他在No Priors这期节目里给了一组数字：过去20年，AI芯片的浮点算力涨了大约100万倍，同期内存带宽只涨了约40倍。

这个剪刀差决定了Fractile几乎所有的技术选择。Walter在节目里说得很直接：如果能在芯片领域结构性地切出三到六个月的优势，你就能赢下那些部署。参照物是前沿实验室的模型迭代节奏，每两周就有一个新模型出来。

主持人Sarah Guo和Elad Gil问他，Fractile在整个GPU和加速器版图里站什么位置。Walter的回答先绕到了行业结构上。

一堆长得差不多的芯片

他把今天的AI ASIC市场叫“一个非凡的动物园”。凑近看，动物园里的动物长得越来越像。

Google十年前用TPU开了头，Meta有MTIA，微软有Maia，OpenAI现在有Jalapeno。这些自研芯片背后站着一小批ASIC交付公司，最大的是博通，市值两万亿美元。它们的共同点很明显：都用HBM，都用tensor core做矩阵乘法，都走台积电的先进封装。

Walter说，看似选择很多，实际上平台之间的相似度非常高。真正从架构层一路做到物理设计、工艺技术、每周飞台湾和韩国跟晶圆厂打交道的团队，仍然很少。

他解释了这个分工怎么运作。假设你在设计TPU，团队里有人深刻理解要加速的工作负载，于是构想出tensor core这种专用电路，因为矩阵乘法占了模型flop数的绝大部分。这是架构师的地盘。然后另一批人把架构翻译成电路级描述，也就是前端设计，本质上还是在电脑上写代码。这份代码交给博通这样的公司，做综合、布局布线，最后变成GDS2文件，一个告诉台积电每层金属放哪、每个晶体管放哪的完整版图。

模拟IP、芯片间互连、针对特定工艺节点的物理放置，长期外包。Walter的团队把这一整段都放进了公司内部。150人，每个环节都瘦，但闭环。

“很多行业都有瀑布式流程被敏捷流程替代的时刻。芯片行业的工作负载每两周变一次，你追赶工作负载的尾巴，必须押注押得准，还要出手快。”

从SRAM掉头，去啃DRAM

Fractile最早两年做的是SRAM路线，思路和Groq、Cerebras有相似之处。SRAM带宽极高，和逻辑电路在同一片硅上，权重和KV cache来回搬得飞快，语言模型能跑到每秒几千个token。

2023年底到2024年，Walter开始担心这条路的天花板。让他睡不着的是上下文长度。

模型参数在涨，上下文长度也在涨，后者对他们那个架构的威胁更大。SRAM容量小，上下文一长就装不下。团队掉头，去跟内存厂商和逻辑代工厂一起做“臭鼬工厂”式的项目，目标是给DRAM做出极高的带宽。

这个转向的成果是一个明年下半年量产的平台。按Walter的说法，它把GPU和TPU那种大容量、低成本的DRAM，和Groq、Cerebras那种超高带宽的速度优势合到一起。他给了一个数字：每颗芯片的带宽是HBM芯片的25倍。

为什么非要这么高的带宽？他拆开了推理的经济账。跑推理时，你要把权重和状态每秒加载几千次，带宽决定速度。但数据中心里服务几千个用户，成本会摊到每GB内存上。高带宽和小容量绑在一起，账算不过来。

“快推理芯片有超高的带宽，容量却极低。今天它们实际上不跑长上下文注意力，遇到长上下文还得切回GPU。这正是最诱人也最别扭的错配。”

他要的是同时拿到高带宽和低成本容量，两个都要。Sarah问他速度到底在哪个场景里改变局面，Walter用了一个亨利·福特的类比。福特说过，如果问人们想要什么，他们会说要更快的马。

更快的聊天机器人，就是快推理领域的更快的马。真正的杠杆在长时运行的agent。一个多万亿参数的模型，如果能稳定跑到每秒几千个token，agent的连续推理时间会被大幅压缩。这是能力的电梯，不是体验的微调。

带宽也有scaling law

这段讨论里最有信息量的部分，是Walter把内存带宽放进了scaling law的框架。

传统上大家谈的是flop scaling law：训练和测试时投入的算力越多，模型越强。Walter说，已经能看到一些“带宽的scaling law”。

第一个例子是MoE的稀疏度。同等智能水平下，把稀疏度从1/16推到1/128、1/256，能省下大量flop。问题是在今天的HBM GPU和XPU上，这么稀疏的模型服务起来带宽瓶颈严重，MFU极低，算力利用率很难看。模型架构想往更稀疏走，硬件带宽拖了后腿。

第二个例子是注意力机制。有些注意力形式对带宽需求低，对flop需求高。当带宽不再是瓶颈，模型就可以选更省算力的那一种。

Walter给了一组对比：过去20年算力涨了100万倍，带宽只涨了40倍。把带宽这个维度往上抬，就能在同等智能水平下省下算力，反过来放大全球吞吐。他管这叫“针对带宽的scaling law”，和传统的flop scaling law并排看。

这也解释了他为什么对“难问题”这么兴奋。推理速度上去之后，模型可以在每次实验之前多想很多，用推理token换实验次数。算力越来越便宜，wall clock时间越来越贵，这个交换的比值会持续倾斜。

把10年除以4

Elad问了一个具体问题：从首席架构师的意图到一份完整可用的GDS2文件，全自动走完要多久。他提到自己问过一位全球前三的半导体公司CEO，对方不愿意让这段对话上节目，给的答案是10年。

Walter的回答是：“过去几年好用的一个启发是，先质疑你的逻辑假设，然后把时间除以四，可能再减一点。”

他认为接下来几年会出现端到端的原型流程。但他不认为芯片设计会变成每几周流一次片。

原因在物理和财务两头。流片周期3到5个月，再热的批次也一样。芯片回来之后，要有3到5年的摊销窗口，这笔投资才成立。数据中心电力、安装延迟、硅本身的融资，都是约束。

“我不相信我们最终会每几周就发一颗全新架构的芯片。但我相信，从观察到把赌注变成量产之间的延迟越短，能捕获的价值越大。”

他用阿姆达尔定律解释了一个组织问题。如果前端设计用AI把12个月压缩到很短，后面的物理设计、流片、验证还是老节奏，整条链的收益就被后面拖住了。所以AI加速要全链路一起做，才有意义。这也是他坚持150人做全栈的原因，每个环节都自己握着，压缩才不会被外部伙伴的节奏卡住。

具体到工具层，他说布局布线这些“传统算法解NP难问题”的环节，今天还要跑好几天。模糊的近似布局算法可能帮团队快速迭代，但最终sign-off仍然要过Cadence和Synopsys那套DRC、LVS检查，因为那是和台积电等晶圆厂几十年磨合出来的规则。

他还提了一个类比。有嘉宾在Dwarkesh的播客上说过一个极端说法：做实验之前应该先想一百年，做完之后再花一百年想结果，因为中间那个实验又贵又费wall clock时间。芯片设计里的那些墙钟瓶颈，逻辑类似：智能越强，越值得在动手之前多花推理token。

自研芯片的主要用途，可能是压价

节目最后聊到五年后大玩家买什么芯片。Walter的判断分两层。

第一层，所有规模化部署的人都在尽量多铺平台。compute太生死攸关，供应多样性是刚需。他提到行业里有个玩笑：第一方自研芯片的主要用途，是压低付给NVIDIA的价格。这个玩笑有几分真，因为那些自研芯片在架构上跟GPU很像，赌的是议价权和总产能，并未押注别人做不到的新能力。

第二层，前沿实验室为什么需要第三方芯片商。Walter讲了一个不对称博弈。

假设实验室A全押某款自研硅。实验室B发现了一个新的计算突破，同等智能水平下效率高得多，但只在自己选的芯片上跑得动。实验室A可能在九个月内就撑不住，等不到自己的芯片量上来。

“前沿玩家把差异化赌注全押在芯片上，是不理性而且危险的动作。他们玩的是不同的游戏。”

所以前沿实验室会同时用多个平台，也会给第三方芯片商留位置。Walter说Fractile尽量让自己的团队长得像前沿实验室的团队，深刻理解工作负载，甚至反过来建议模型架构怎么改更契合自己的芯片。他们内部在跑“带宽scaling law”的实验，看更稀疏的MoE、更省带宽的注意力形式在25倍带宽的硬件上会解锁什么。

谈到NVIDIA的系统，他承认差距：拆开一台NVIDIA机器，里面有6到9颗定制芯片协同工作，那是多年堆出来的系统级能力。Fractile现在只做一颗芯片。他的野心是有一天能像NVIDIA一样，同时推进多个赌注，形成一个滚动的前沿下注队列，随时准备触发量产。

明天能用的三件事

这期节目信息密度高，落到操作层面有三条。

评估推理硬件时，先算每GB内存的成本，再算每token的延迟。高带宽小容量的芯片在长上下文场景会把省下的时间吐回去，因为要切回GPU做注意力。
设计模型架构时，把MoE稀疏度和注意力形式当成硬件协同变量。1/128、1/256的稀疏度在今天的HBM上服务效率低，换一种带宽结构可能会翻盘。
做任何长链路工程，用Walter的除以四法则检验时间表。先问这个假设是否成立，再把预估周期除以四，通常比原计划更接近现实。

Walter最后那句话值得单独拎出来：能结构性切出三到六个月的优势，就能赢下那些部署。这句话放在芯片公司身上成立，放在用AI做产品的团队身上也成立。窗口不在年，在季度。

原文链接：https://nopriors.qiaomu.ai/ep/181-frontier-chips-for-frontier-ai-labs-with-walter-goodwin-foun-OpeCP4.html · 中文精读由 AI 改写，非官方翻译。

Sequence Holdings 联合创始人 Michael Lee：用 AI 从内部重造传统行业龙头

Thu, 24 Sep 2026 00:00:00 +0000

Can AI transform legacy incumbents rather than replacing them? Sequence Holdings co-founder and CEO Michael Lee joins Sarah Guo to discuss how holding company structures and engineering integrations are reshaping market leaders from the inside out. Michael details Sequence’s $7.7 billion take-private transaction of Baldwin alongside Dell Family Office (DFO), and shares his thesis on why traditional consulting models and software sales fall short for real enterprise AI transformations. They also talk about why permanent holding company structures are good for long-term compounding, real-world results from applying frontier engineering to BankSouth, and Michael’s lessons from his time in public investing, private equity, and operating at the intersection of market incumbents and AI.

Chapters:

00:34 – Michael Lee Introduction

01:03 – Sequence Holdings and Baldwin

01:54 – Idea for Sequence

04:36 – Incumbents in the AI Era

06:32 – Why a Holding Company

11:21 – Recruiting Top Engineers

13:08 – Investing in BankSouth

17:46 – Why an Insurance Brokerage

20:17 – Atlas Platform Explained

23:53 – Traditional Private Equity Limitations

27:23 – What Sequence Looks For in Management Teams

31:04 – Accomplishments at BankSouth

34:45 – Founder Lessons

36:08 – Story of Dell Partnership

37:10 – Career and Investment Approach

40:00 – Value of Exceptional People

42:23 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Inception 的 Stefano Ermon：为什么扩散模型会赢得 AI 推理

Fri, 18 Sep 2026 00:00:00 +0000

As generative AI hits hardware and latency bottlenecks, Stanford professor, diffusion pioneer, and Inception co-founder and CEO Stefano Ermon is betting on a radical new architecture. Stefano joins Sarah Guo to talk about Inception, and how his team is applying diffusion architecture beyond images and video into discrete text and code generation. Stefano explains the limitations of autoregressive LLMs, as well as why parallel token generation in diffusion models offers superior inference scaling and hardware utilization on standard GPUs. He also shares details about Inception’s Mercury models, real-world voice agent applications, the software stack required to serve diffusion-based models at scale, academia’s role at the frontier of AI innovations, and why the next era of AI competition will be defined by efficiency.

Chapters:

00:00 – Stefano Ermon Introduction

00:35 – Research Background

02:54 – Starting Inception

05:59 – Why Diffusion Beats Autoregressive

11:10 – Discrete vs. Continuous Modalities

13:19 – Inception Today

16:45 – Where Speed Wins

17:31 – Inception Customer Base

18:49 – Interaction with Hardware Landscape

19:34 – Inception and the Broader Industry

21:41 – Data Compression and Structure

24:45 – Controllability of Diffusion Modeles

27:25 – Emergent Capabilities at Scale

29:02 – Future Workload Split Between Diffusion vs. Traditional

30:03 – Adoption Challenges

31:44 – Hiring and Team Organization

32:50 – Recursive Self Improvement

34:02 – Resource Allocation

35:10 – Impact of Academia

38:13 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Coinbase CEO Brian Armstrong：Agent 金融、稳定币与资产代币化

Thu, 10 Sep 2026 00:00:00 +0000

From the tokenization of real-world assets to funding research on anti-aging therapeutics, Coinbase co-founder and CEO Brian Armstrong is focused on solving meta-problems for the benefit of society. Brian joins Elad Gil this episode to discuss the intersections of artificial intelligence, crypto rails, and biotechnology. He details Coinbase’s vision for agentic finance, as well as how Coinbase is becoming the ‘Everything Exchange’, which includes tokenized real-world stocks, stablecoin payment adoption, and prediction markets. Brian also introduces his new venture, New Limit, a company founded to tackle longevity through epigenetic reprogramming, and outlines initial therapeutic programs and the roadmap toward clinical trials.

Chapters:

00:00 – Cold Open Trailer

00:45 – Brian Armstrong Introduction

01:13 – Coinbase and the Everything Exchange

05:52 – Agentic Commerce

10:00 – AI and Crypto

10:49 – AI Inside Coinbase

15:39 – Productivity and Company Size

17:26 – The Everything Exchange and Tokenization

20:53 – Prediction Markets

23:15 – Introducing New Limit

34:17 – Looking Forward Five Years

40:24 – Special Economic Zones

44:26 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Arm CEO Rene Haas：CPU 仍是 AI 计算底座，从 IP 授权走向自研芯片

Thu, 03 Sep 2026 00:00:00 +0000

From data center orchestrators to AGI and robotics, CPUs remain the heart of modern computing. Arm CEO Rene Haas joins Elad Gil and Sarah Guo to explore how Arm is positioned at the epicenter of AI-driven demands for compute. Rene explains Arm’s position in the chip supply chain, and how Arm transitioned from an IP licensing model to producing physical chips like the Arm AGI CPU for Meta. He also discusses bottlenecks in hardware supply chains, SoftBank’s ecosystem and capital strategy, why US semiconductor manufacturing independence is critical, the future of robotics, and why CPUs remain crucial for executing AI workloads.

Chapters:

00:00 – Cold Open Trailer

00:49 – Rene Haas Introduction

01:14 – Arm and Chip Supply Chain

02:37 – Shift from IP to Manufacturing CPUs

04:23 – CPU IP and Customers

06:55 – AI Adoption at Arm

10:15 – Changes in Chip Time to Market

13:27 – Data Center Buildout Bottleneck

15:13 – Softbank Leverage and Capital Strategy

17:43 – Softbank Portfolio Overview

20:13 – Robotics Opportunities for Arm

24:49 – US Manufacturing Protectionism

28:59 – Data Center Backlash

32:30 – Arm Outlook

33:31 – CPU Opportunity

37:06 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Eon 两位联合创始人谈数据护城河：把云备份重做成安全数据层

Thu, 27 Aug 2026 00:00:00 +0000

Google’s purchase of Spirit Airlines’ data out of bankruptcy signaled a shift in how the tech world values real-world datasets. Although compute and models get much of the attention, in this landscape, it’s data that is a company’s protective moat. Eon CEO / Co-Founder Ofir Ehrlich and President / Co-Founder Gonen Stein join Elad Gil to talk about how Eon is redefining cloud backup into a secure data foundation designed to power and protect enterprise AI. Ofir and Gonen discuss why historical enterprise data is in demand by AI labs, and how Eon facilitates access to scattered and locked data across business units through providing the mapping, classification, and access controls needed to connect it into AI workflows. They also explore how traditional ransomware defenses must now protect against rogue AI agents with legitimate system permissions, concerns around the influx of autonomous agents and non-human identities, and the implications for the breakneck speed of AI adoption compared to the slowness of the cloud era.

Chapters:

00:00 – Cold Open Trailer

00:59 – Ofir Ehrlich and Gonen Stein Introduction

01:27 – What Eon Does

02:41 – Data as Moat

06:43 – Training Agents with Good Data

09:39 – Data is the New Oil

15:00 – Autonomous Security Threats

18:15 – How Agents Change the Enterprise Stack

22:11 – Re-imagining Data Infrastructure

27:52 – Cloud vs. AI Era Shift

30:26 – How AI is Changing Companies

34:31 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Max Hodak：从视网膜植入恢复视力，到把大脑当作计算系统来治疗

Thu, 20 Aug 2026 00:00:00 +0000

Max Hodak, co-founder and CEO of Science Corporation, joins Sarah Guo to discuss the future of vision, brain-computer interfaces, and the human experience. Max explains how Science’s PRIMA retinal implant could restore functional vision for people who have lost their sight, and why treating the brain as a computational system could unlock new approaches to medicine.

They explore the broader potential of neural devices, from restoring lost capabilities to expanding human potential, as well as deeper questions around identity, consciousness, and whether the human experience can persist as our biological hardware changes.

Max also shares Science’s long-term vision for reducing the fragility of the human condition by repairing, replacing, and ultimately upgrading parts of ourselves. Finally, he discusses the surprising parallels between AI models and biological brains, and why AI may offer a powerful new lens for understanding intelligence.

Chapters:

00:00 – Cold Open Trailer

01:40 – Max Hodak Introduction

02:00 – Science Corporation Overview and Origin

02:53 – A Revolutionary Solve for Blindness

06:32 – Scope of Timeline and Engineer Cost

09:10 – Clinic Trial Process

09:45 - The Response from Clinicians

12:21 – Broader Biotech Landscape

14:59 – Brain’s Relationship to Senses

17:35 – The Study of Consciousness

19:50 – Investments in Brain Computer Interface

22:10 – Fertile Ways to Study Neuroscience

24:46 – Biotech Expansion for Science Corporation

27:50 – What Success Looks Like in Neuroscience and Tech

29:06 - Goals Within Human Preservation vs. Adaptation

30:25 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Chess.com CEO Erik Allebest：机器早已赢过人类，国际象棋为何更流行

Thu, 13 Aug 2026 00:00:00 +0000

In a world of infinite gaming and entertainment possibilities, how does a centuries-old game stay so popular? Chess.com co-founder and CEO Erik Allebest joins Sarah Guo to explain how the evolution of technology has kept people coming back to chess, even when machines can beat us at the game. Erik talks about how the desire to build a MySpace-like community for chess led to the purchase of a domain name from a bankruptcy sale back in 2005, and scaled into a community with 10 million daily active users and 250 million total registered members. He also discusses the growth of the cultural relevance of chess, how investments from private equity firms General Atlantic and CVC helped grow and strengthen their platform, and how Chess.com is leveraging AI both within the business itself and to make a better product for its community.

Chapters:

00:00 – Cold Open Trailer

01:05 – Erik Allebest Introduction

01:48 – Chess.com Today

02:57 – Buying and Scaling Chess.com

06:29 – Competition and Growth

11:52 – Chess and Cultural Relevance

14:32 – Private Equity Investment

19:31 – Playing Games Amid Evolving Tech

25:09 – Tech, Skill Distribution, and Expertise

28:40 – Chess and Cheating

31:20 – What Makes Chess Special

33:17 – Chess.com Future Vision

34:54 – Founder Advice

36:48 – AGI/ASI Predictions

40:02 – AI Investments at Chess.com

42:13 – How AI May Change Product at Chess.com

43:27 – Poker Rating Algorithms

46:07 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Sarah 与 Elad 对谈：万亿市值野心、创始人雄心、Token 预算与监管俘获

Thu, 06 Aug 2026 00:00:00 +0000

Is the tech industry moving too quickly, or are founders letting fear of AI labs stunt their ambitions? Sarah and Elad explore the current landscape of artificial intelligence, venture capital, and startup dynamics. They discuss the realities of building multi-trillion-dollar companies, shifting market sizes and outcome-based pricing models, and how founders are reacting to the rise of major AI labs. They also talk about what the framework for startup exits should look like, the potential for researcher burnouts in the next eighteen months as ASI looms on the horizon, bottlenecks for compute, and the impact of regulatory capture and shifting ecosystems from California to Texas.

Apply for Embed - Conviction’s Catalyst for AI-Native Startups

Chapters:

00:00 – Cold Open Trailer

00:31 – Episode Introduction

01:44 – The Next Trillion-Dollar Company

03:12 – Tech Waves as Punctuated Equilibria

04:42 – TAM vs. Revenue Reality

07:14 – Market Size vs. Speed

10:32 – When Founders Should Sell

14:04 – Financing and Time Cost

17:57 – RSI and the Looming Promise of ASI

21:49 – Compute Power Laws

28:12 – Regulations and Disruption

33:06 – Beyond Transformers

34:26 – Tradeoffs - Safety vs. Progress

39:11 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Netic 创始人 Melisa Tokmak：用 AI Agent 重构上门服务的即时调度

Fri, 31 Jul 2026 00:00:00 +0000

When your AC fails in a heatwave, you don’t want a busy signal; you need a solution. Netic founder and CEO Melisa Tokmak joins host Elad Gil to explain how Netic’s autonomous AI platform acts as an intermediary between companies and customers, deploying agents to instantly handle essential services, from emergency home repairs to hospitality to pet care. Melisa describes the complexity of these real-world workloads, which have traditionally relied on large human support teams, and how over 70% of Netic’s customers interact first with AI. She also talks about the reasoning behind building a scalable product company rather than an AI roll-up, why she believes robotics will not catch up in these industries in the near future, why she doesn’t view large frontier labs as competitive threats, and how private equity’s playbook has shifted toward measurable ROI in the AI-era. Plus, why Melisa is optimistic about the impact AI will have on education.

Chapters:

00:00 – Melisa Tokmak Introduction

00:32 – What Netic Builds

03:53 – Automating Workflows for Essential Services

06:26 – Building a Service vs. AI Roll-Up

10:38 – AI for the Real World Timeline

12:56 – Can Big Labs Compete?

15:35 – Modern Founder Mindset

19:09 – Screening for Agency

22:25 – Five Year Vision

23:53 – Selling to Slow Industries

27:23 – How Private Equity Approached AI

31:14 – What Excites Melisa About the Future of AI

34:27 – Conclusion

在 nopriors.qiaomu.ai 查看本期

DoorDash 联合创始人谈自主配送：自动驾驶技术与 Ask DoorDash 自然语言入口如何重塑消费与商业

Thu, 23 Jul 2026 00:00:00 +0000

DoorDash is not just a delivery company. From its inception, co-founders Andy Fang and Stanley Tang operated it as a robotics and autonomy company. Andy and Stanley join Sarah Guo to explain how autonomous tech and AI are reshaping consumer habits, commerce, and delivery. Andy and Stanley talk about the rollout of Ask DoorDash, a natural-language interface that’s driving both restaurant discovery and larger grocery orders. They also discuss Dot, their in-house autonomous delivery robot that has operated in Phoenix for over two years, and how it highlights the operational and hardware challenges they have faced and solved in autonomous tech. Andy and Stanley also speak about the “first and last 100 feet problem” in autonomous delivery, why multimodal strategies are the key to success, scaling autonomy and operations, and why they believe that more Dashers, not fewer, are the future of DoorDash.

Chapters:

00:00 – Andy Fang and Stanley Tang Introduction

00:34 – Agentic Commerce and Behavioral Changes

03:52 – Next Steps for Ask DoorDash

06:54 – Investing in Robotics and Autonomy

16:31 – Building Autonomous Tech in the Physical World

21:20 – Dot: DoorDash’s Autonomous Delivery Robot

22:08 – Collecting Realistic Data

25:48 – Why Work at DoorDash

28:04 – Challenges in Scaling Up Autonomy

39:30 – Productivity Benchmarks

44:56 – Future of Agentic Commerce

49:10 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Booking CEO 格伦·福格尔：从一美元股价到千亿美元，AI 视角下的全球旅行平台

Thu, 09 Jul 2026 00:00:00 +0000

When Glenn Fogel joined Priceline in 2000, the business was worth a few hundred million dollars. One week later, the Nasdaq peaked, eventually sending its stock down to a dollar a share. But over 25 years later, Booking Holdings has scaled over 1000x into an over $100 billion dollar global travel behemoth. Elad Gil is joined by Booking Holdings CEO Glenn Fogel to discuss his career, from law school and Wall Street to working at Priceline through the dot-com crash, and to helping grow the business into a multifaceted, dynamic travel marketplace in the AI era. Glenn explains how leveraging AI and agents such as Priceline’s ‘Penny’ makes travel planning and customer service better, while emphasizing the importance of preserving some human support for some users. He also talks about Booking’s strategy of reinvesting over $700 million into AI and other technologies while still offering stock buybacks and dividends, the durability of their scale and complexities of dealing with a large portfolio physical properties across the world, and why upskilling is so important for employees amid concerns about AI-driven job displacement.

Chapters:

00:00 – Cold Open

00:05 – Glenn Fogel Introduction

00:41 – Glenn’s Early Career

06:49 – Lessons from the Early Internet

09:24 – Deciding Factors for Exiting

10:56 – Travel Through the Lens of AI

13:30 – Agentic Travel Planning

18:59 – Agents, Token Economics, and ROI

22:46 – Booking’s Capital Investment Philosophy

25:23 – Scale as Durable Asset

29:40 – Purpose and Choosing Wisely

33:18 – AI’s Impact on Jobs

36:38 – Upskilling in the AI Era

38:36 – Public Perception of AI

40:24 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Valar Atomics 创始人以赛亚·泰勒：用真实核反应堆为英伟达芯片供电，解锁能源充裕

Thu, 02 Jul 2026 00:00:00 +0000

While the rest of the nuclear industry still relies on simulations and paper designs, Valar Atomics is busy splitting atoms. In fact, they just powered an NVIDIA Blackwell chip directly with a live nuclear reactor in order to power the world’s first nuclear powered website. Sarah Guo joins Valar Atomics founder and CEO Isaiah Taylor on-site at their reactor site in Utah to talk about how Valar is shifting nuclear energy from the theoretical to the practical by building and perfecting reactors via hardware iteration. Isaiah discusses why the US stopped building nuclear reactors in the 1970s, and how Valar utilized a little-known pathway via the Department of Energy, revived by a Trump administration executive order, to successfully develop and run their advanced reactor. He also shares Valar’s strategy for vertical integration, their venture-backed approach to financing, their giga-site plans, and why he believes cheap, abundant atomic energy has the power to vastly improve the quality of human life.

Chapters:

00:00 – Cold Open

00:57 – Isaiah Taylor Introduction

01:30 - Valar’s Mission and Origin

04:24 - Why Nuclear Development Stalled

07:18 - Reviving Nuclear through DoE and Executive Order

10:59 - Control Room Tour

16:17 - Misunderstandings About Nuclear

20:07 - Issues with Reliability

22:14 - Nuclear is a Hardware Execution Problem

24:32 - Timeline to Scale Production

26:32 - Introducing Ward 250

30:42 - Speed Through Simplicity

33:33 - AI Drives Nuclear Demand

35:02 - Running a Reactor with NVIDIA Blackwell

36:27 - Valar’s Nuclear Conviction

40:16 - Verticalization as Path to Scale

43:58 - Valar’s Control Skid

48:00 - Venture-Backed Nuclear

50:51 - Gigasite Strategy

53:11 - CEO Tick Rate

55:37 - Abundant Energy and Hyper-Techno Industrialism

1:01:27 – Conclusion

在 nopriors.qiaomu.ai 查看本期

OpenAI 研究科学家 Noam Brown：大规模测试时计算如何改写基准、安全与研究范式

Fri, 26 Jun 2026 00:00:00 +0000

When a new AI model drops, it’s judged based on a static benchmark grid that doesn’t account for how long the model is allowed to think. How then should we measure a model’s true capability? OpenAI research scientist Noam Brown returns to talk with Sarah Guo about his latest essay on why the AI industry’s traditional benchmark grids are broken, and how large-scale test-time compute is fundamentally changing how models are evaluated. Noam explains how, if properly scaffolded, today’s models can reason for weeks or even months on complex tasks. He also discusses real-world implications of test-time compute, from building poker solver bots to disproving legendary math conjectures. Together, they also unpack the large gaps in current AI safety frameworks, explore the bottlenecks for recursive self-improvement, and look ahead at the future of multi-agent collaboration and global knowledge sharing.

Read more: Implications of Large-Scale Test-Time Compute

Chapters:

00:00 – Cold Open

00:43 – Noam Brown Introduction

01:23 – Why Benchmarks Are Broken

04:19 – Compute Budgets and Projections

05:34 – How Long Should Models Think?

06:47 – Benchmark-Maxxing

08:34 – Using Poker Bots as Evals

11:26 – Safety Evals When Model Capability Scales With Budget

14:41 – Release Cycle vs. Agent Runtime

17:06 – Latent Model Capability

20:59 – Limits on Recursive Self-Improvement

27:09 – Large-Scale Multi-Agent Coordination

29:11 – Competition at the Frontier

31:51 – Breaking the Benchmark Grid Equilibrium

33:29 – Why Benchmarks Should be Evaluated by Cost

36:18 – Conclusion

在 nopriors.qiaomu.ai 查看本期

英特尔 CEO 陈立武：66 岁接手最难的工作，重构半导体供应链与代工模式

Thu, 18 Jun 2026 00:00:00 +0000

At 66 years old, instead of heading towards retirement, former Cadence CEO and legendary investor Lip-Bu Tan decided to take on the hardest job in tech: turning Intel around. Elad Gil and Sarah Guo sit down with Intel CEO Lip-Bu Tan to talk about why he took the job and what “saving” Intel actually looks like. Tan explains how his experience in startup culture informed his decisions to drive Intel’s culture towards faster decisions, focus on customer satisfaction, and engineer accountability. He also discusses his strategy to strengthen Intel’s balance sheet by welcoming investments from Jensen Huang’s Nvidia, Softbank, and the US government. Tan also shares his product roadmap that centers the CPU for agentic AI and inference, the collaboration with Elon Musk on Terafab, his investing framework for semiconductors, and his views on how AI is reshaping design and operations at, as he puts it, a ‘legacy spreadsheet’ tech company.

Chapters:

00:00 – Cold Open

01:01 – Lip-Bu Tan Introduction

01:24 – Why Lip-Bu Took the Reins at Intel

03:00 – Fixing Culture

04:08 – Intel’s 10-Year Vision

07:57 – Working with Elon Musk on Terafab

09:59 – Shifting Supply Chain for Semiconductors

15:34 – Limits to Scaling and Packaging

18:30 – Physical Limits to Engineering and Design

20:33 – Challenges in Semiconductor Investing

26:29 – Lessons from Cadence

28:02 – Scaling and Investment Decisions

32:03 – Rethinking Teams in AI Era

34:31 – Industrial Policy and Funding

37:25 – What Investors Misunderstand About Intel

41:10 – Where Compute Will Live

44:59 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Biohub 联合创始人扎克伯格与普莉希拉·陈：前沿 AI 加生物数据，让生物学走向开源

Wed, 10 Jun 2026 00:00:00 +0000

Biohub started with an ambitious goal of curing, preventing, and managing all disease by the end of the century. A decade later, thanks to the convergence of frontier AI and biological data, that goal may have been too conservative. In this episode, Elad Gil and Sarah Guo sit down with Biohub co-founders Mark Zuckerberg and Priscilla Chan, alongside Biohub Head of Science Alex Rives. Together, they discuss Biohub’s $500 million virtual biology initiative, which integrates frontier AI with wet-lab work to build predictive world models of cells, proteins, and systems. They also talk about their newly announced open-source engine for digital protein and antibody design, ESMFold2; why Biohub is a nonprofit rather than a venture-backed startup; and how hierarchical simulations will soon allow doctors to treat patients at an individual, mechanistic level.

Chapters:

00:00 – Cold Open

01:02 - Mark Zuckerberg, Priscilla Chan, and Alex Rives Introduction

01:26 – Why Biohub and Their Mission

08:27 – Integrating Frontier AI and Frontier Biology

09:45 – Micro to Macro Biological Modeling

14:22 – Mechanistic Interpretiability

16:58 – Why Biohub is a Non-Profit

21:41 – Understanding How Biology Works

24:23 – Timeline for Curing All Diseases

26:25 – Translating Research to Patient Impact

28:04 – Launch of ESMFold2

32:13 – Tackling Off-Target Effects and Edge Cases

38:39 – Putting the Tech in Individual Hands

41:06 – Talent at Biohub

44:25 – What’s Next After ESMFold2

46:10 – Connecting ESMFold2 to Agentic Systems

46:51 – The Virtual Cell

49:33 – Defining Success for Biohub

51:52 – Biohub Strategy Update

56:20 – Conclusion

在 nopriors.qiaomu.ai 查看本期

微软 CEO 萨提亚·纳德拉：AI 需要完整生态，每家公司都能在其中找到位置

Thu, 04 Jun 2026 00:00:00 +0000

What does it mean for a business to truly operate at the AI frontier? In a special crossover episode at Microsoft Build, Sarah Guo and Elad Gil team up with Latent Space host “swyx” to talk with Microsoft Chairman and CEO Satya Nadella about the future of AI platforms, software development, and the tech ecosystem. Satya reflects on the latest breakthroughs from Microsoft Build, the strategic shift toward multi-model harnesses, and why private evaluations (evals) are now a company’s most important intellectual property. They also discuss how autonomous AI agents are reshaping the role of software engineers, the durability of SaaS business models, and why showing communities the ROI on data centers is so critical. Plus, Satya shares his thoughts on the economic and societal impacts of the token economy, as well as the future of AI-driven education startups.

Chapters:

00:00 – Satya Nadella Introduction

01:48 – Reflections from Microsoft Build

03:12 – Microsoft’s AI Training Strategy

05:48 – Complexity of Real-World Deployment of AI

07:33 – Augmenting Human Capital

09:37 – Harnesses for Enterprise

11:49 – Developer Value

15:09 – Can Everybody Operate at the Frontier with Their Frontier Intelligence?

15:51 – Modern Definition of IP

17:38 – Future of Vendor vs. Enterprise Agents

21:48 – Near-Term Predictions on Model Pricing

24:02 – Durability of SaaS

25:58 – What Satya’s Building

28:18 – Future of Engineering Roles

30:54 – How Microsoft Can Be More Ambitious

34:36 – Data Centers and Community Impact

38:01 – AI’s Impact on Society

39:52 - AI and Education

42:28 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Onyx Security CEO 马克西姆·巴尔·科根：为企业级自主智能体打造 AI 监管护栏

Thu, 28 May 2026 00:00:00 +0000

We are now closer than ever before to living in a world where AI agents are smart enough to run our power grids and manage water supplies. How do we keep them from going rogue? Sarah Guo sits down with Maxim Bar Kogan, founder and CEO of Onyx Securities, to explore the complexities of supervising and securing autonomous agents at the enterprise level. Maxim explains Onyx’s product as an AI control plane, which oversees the permissions and flexible contexts of agents while balancing latency, cost, and reliability. He also discusses how current controls have insufficient context to monitor agent intent, tradeoffs for gradual model rollout, the need for vendor-independent oversight, and Israel’s growing AI and security talent ecosystem. Plus, why Maxim is all-in on AGI.

Chapters:

00:00 – Cold Open

00:45 – Maxim Bar Kogan Introduction

01:10 – AutoGPT and Betting on Agent Actions

05:17 – What Onyx Product Does

07:47 – State of Deployment in Large Enterprises

09:58 – Securing Agents

12:45 – Why Proxies Don’t Work

14:11 – Why Onyx Trains Its Own Models

18:38 – Onyx’s Talent Culture

21:24 – Mechanistic Interpretability

23:35 – How Onyx Builds Customer Trust

25:10 – Mitigating Risk at the Foundational Level

27:45 – Phased Rollout of Glasswing and Daybreak

29:11 – Large Enterprise Holdouts

30:46 – Onyx and the Larger AI Security Space

32:36 – Should Labs Address Model Trust and Governance?

36:56 – What Needs to Happen in Security

39:14 – Why Maxim is AGI-Pilled

41:15 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Cerebras 创始人安德鲁·费尔德曼：餐盘大小的芯片如何撑起 630 亿美元 IPO

Thu, 21 May 2026 00:00:00 +0000

Companies in Silicon Valley from Nvidia to AMD are racing to fuel the AI revolution with postage stamp-sized AI chips. Meanwhile, a chip the size of a dinner plate just fueled a $63 billion IPO for Cerebras. Elad Gil and Sarah Guo sit down with Cerebras founder and CEO Andrew Feldman to discuss the company’s journey to making one of the largest tech go-publics in history. Andrew details the multi-year journey of pioneering wafer-scale AI computing, including surviving a brutal period of being ahead of market demand. He also explains the engineering breakthroughs that led to delivering inference speeds at 20x that of standard GPUs. Andrew then shares how a remarkable $20 billion deal with OpenAI came together in only four weeks. Plus, Andrew’s thoughts on why architecting the future of AI requires the fortitude to be a “professional David” against the Goliaths of tech.

Chapters:

00:00 – Cold Open

00:36 – Andrew Feldman Introduction

01:19 – Cerebras’ Evolution

02:48 – Wafer-Scale Bet Pays Off

06:38 – Challenges and Breakthroughs

08:37 – Crossing the Market Chasm

10:38 – Scaling Software and Hardware

12:03 – Relevance of AI-Generated Coding

13:31 – Leadership and Hiring Culture

17:16 – When to Quit vs. Persist

19:40 – Why Cerebras Went Public

22:57 – The OpenAI Deal

25:54 – Open Source and Post-Trained Workloads

27:37 – How Speed Opens Up New Business

30:33 – Conclusion

在 nopriors.qiaomu.ai 查看本期

美国国务院副国务卿雅各布·赫尔伯格：Pax Silica 与十四国经济安全联盟的技术战略

Thu, 14 May 2026 00:00:00 +0000

Securing AI dominance requires more than just semiconductors; it demands a complete overhaul of how the West manages everything that goes into them, from rare earth minerals to actuators. Enter: Pax Silica. Sarah Guo and Elad Gil sit down with US Under Secretary of State for Economic Affairs Jacob Helberg to discuss the launch and expansion of Pax Silica, a 14-country economic security coalition designed to secure the entire AI supply chain. Jacob talks about the creation of a forward-deployed industrial base in the Philippines, where 4,000 acres will be developed into an “economic security zone.” He also compares and contrasts Pax Silica with China’s Belt and Road initiative, explains how the US plans to reindustrialize through automation and robotics, and explores how the Trump administration envisions making these policies durable across future presidencies. Plus, we hear why Jacob believes America to be a “global underdog.”

Chapters:

00:00 – Cold Open

00:41 – Jacob Helberg Introduction

01:02 – Pax Silica’s Mission

03:51 – Investing in AI Chip Supply Chains

05:43 – Comparing Pax Silica to China’s Belt and Road Initiative

12:38 – Pax Silica’s Value Proposition

14:38 – US vs. Partnered Manufacturing

19:10 – Rare Earth Mineral Pricing

22:16 – Role of Venture Capital in Pax Silica

24:50 – Near vs. Long-Term Priorities

27:09 – Making AI Policy Durable

28:09 – How Policies Impact Entrepreneurs

31:00 – Trump’s Entrepreneurial Administration

33:00 – Why America is a Global Underdog

38:00 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Long Lake CEO 亚历山大·陶布曼：63 亿美元收购 Amex GBT，首个 AI 主导的私有化交易

Mon, 11 May 2026 00:00:00 +0000

The world’s first AI-take-private just proved that AI can revolutionize the real economy. Long Lake Management co-founder and CEO Alexander Taubman joins Elad Gil to discuss his firm’s agreement to acquire the legacy platform American Express Global Business Travel (Amex GBT) in a deal valued at $6.3 billion. Alexander explains the mechanics of AI-driven roll-ups, and why Long Lake chooses to acquire and transform businesses rather than simply selling them software. He also talks about how Long Lake’s horizontal AI platform, Nexus, automates workflows across diverse verticals, and how automation through AI not only powers growth for their portfolio companies, but results in both satisfied customers and employees. Plus, they explore Alexander’s vision of Amex GBT as a multi-decade compounding machine.

Chapters:

00:00 – Alexander Taubman Introduction

00:30 – Long Lake’s Nexus Platform

03:35 – Retention and Talent Flywheel

05:01 – Acquisition vs. Offering Software

06:57 – Building Long Lake’s Founding Team

10:37 – Taking American Express Global Business Travel Private

13:36 – Taking Berkshire Hathaway’s Approach to Management

16:37 – How AI Strategy Makes Long Lake Stand Out

19:32 – AI Makes Services Scale

22:00 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Baseten CEO 图欣·斯里瓦斯塔瓦：推理需求暴涨、定制模型与推理云的最后一战

Fri, 01 May 2026 00:00:00 +0000

Baseten CEO and co-founder Tuhin Srivastava sits down with Sarah Guo and Elad Gil to discuss the rapid growth of AI inference demand, Baseten’s 30x growth, and why inference is becoming the strategic “last market.” Tuhin Srivastava argues the application layer will persist because companies with unique user signals can encode value into workflows and post-train specialized models, citing examples like Abridge and support workflows. The conversation covers GPU capacity constraints, Baseten’s multi-cloud fabric across 18 clouds and 90 clusters, long-term contracting dynamics, the importance of the software layer for stickiness, evolving workloads, multichip possibilities, and operational lessons at scale.

Chapters:

00:31 Baseten growth

01:55 Why the app layer wins

05:57 Serving frontier customers

07:55 Open source model mix

09:21 Chinese models and geopolitics

13:07 Custom inference dominates

14:22 Post training acquisition

17:10 When to invest in custom models

18:35 Supply crunch and data centerse

22:25 Longer GPU Contracts

24:09 What Makes a Winner

26:07 Multi Chip Future

28:19 Runtime Roadmap

31:08 Scaling Edge Cases

33:48 Hiring and Leadership

36:44 Operations Pager Culture

38:19 Efficiency Drives Demand

40:41 Concierge Everything Future

42:34 Conclusion

在 nopriors.qiaomu.ai 查看本期

SAP CTO Philipp Herzig：把企业的操作系统带入AI时代

Thu, 23 Apr 2026 00:00:00 +0000

More than fifty years ago, the modern idea of the standard enterprise software was birthed at SAP. Now, after managing companies through technological shifts from the mainframe to mobile, SAP is at the forefront of closing the AI adoption gap for their customers. SAP Chief Technology Officer Philipp Herzig joins Sarah Guo to talk about how SAP has remained a durable end-to-end “operating system” for its more than 400,000 customers from finance to supply chain. Philipp argues that the AI transition in businesses should focus on customer outcomes, UI changes, business processes, and the data layer. He also explains the challenges in enterprise AI adoption, including security, scaling, and data fragmentation, as well as the importance of evals and verifiability. They also discuss SAP’s suite of AI products, limitations of predictive tabular models, how SAP is shifting its pricing models in the AI era, and Philipp’s interest in quantum computing optimization.

Chapters:

00:00 – Cold Open

00:42 – Philipp Herzig Introduction

01:18 – What SAP Does

02:51 – Why SAP Endures

06:53 – CTO Priorities and AI Push

12:14 – Scaling AI in Enterprise

17:06 – Verifiability and Agent Mining

20:42 – Tool Calling vs. Computer Use

22:11 – Domains Where Agents Deliver Value

24:58 – Limitations of Predictive Tabular Models

29:07 – Barriers to Enterprise Adoption

31:54 – How AI Will ‘Uplevels’ Work

34:03 – How AI Changes SAP’s Pricing Model

36:41 – What Makes a Winner in the AI Era

38:53 – Day in the Life of a CTO

40:08 – Customer Challenges

42:36 – Business Problem of Quantum Computing

46:21 – Conclusion

在 nopriors.qiaomu.ai 查看本期

ServiceNow CEO Bill McDermott：AI时代如何规模化运营全球组织

Fri, 17 Apr 2026 00:00:00 +0000

Few teens are business owners, but by age 16, Bill McDermott had purchased and was running a local deli. Now he runs leading global technology powerhouse ServiceNow, a company that is defining how the world’s largest organizations transform for the digital age. Sarah Guo sits down with ServiceNow CEO Bill McDermott to discuss his journey from child entrepreneur to CEO, and how he navigates his role as a leader in the age of AI. Bill argues that human connection is still a vital part of being a successful leader, and as such, AI must be used to serve people rather than substitute for ambition. He breaks down the mechanics of hyper-growth, and the art of staying customer-centric at a global scale. They also discuss the future of enterprise software, how generative AI is fundamentally reshaping the labor market, and what founders need to know about building a resilient company culture that survives economic and technological shifts.

Chapters:

00:00 – Cold Open

00:50 – Bill McDermott Introduction

01:14 – Lesson from Buying a Deli

07:35 – Leadership in the AI Era

09:41 – How Bill Got Hired at Xerox

15:47 – Can Agency Be Taught?

18:40 – Seeing Change as Opportunity

25:18 – ServiceNow as an AI Control Tower

30:30 – Which SaaS Gets Disrupted?

32:22 – Defining a Platform Business

36:25 – Does AI Decrease Implementation Time?

39:06 – Agents Will Reshape the Workforce

40:59 – Success Signals at ServiceNow

44:07 – Enterprise Attitudes About AI

48:41 – How AI Has Changed Customer Conversations

50:48 – Bill’s Curiosity Beyond ServiceNow

52:29 – Day in the Life of a CEO

57:27 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Circle CEO Jeremy Allaire：Agent经济将如何改写金融系统

Thu, 09 Apr 2026 00:00:00 +0000

AI agents can already collaborate, but they lack a trustworthy medium in which to store value and execute contracts. Enter Circle’s Arc Blockchain, an economic “operating system” designed for a world where machines drive the real economy. Circle co-founder and CEO Jeremy Allaire joins Elad Gil to dive into the future of programmable money and the agentic economy. Jeremy explains why traditional banking fails to support the needs of AI agents, and how stablecoins like USDC facilitate an internet-native economy. They also discuss the tokenization of real-world assets, the move toward full-reserve banking, and Jeremy’s predictions for double-digit GDP growth as AI and blockchain reach their “broadband moment.”

Chapters:

00:00 – Cold Open

00:05 – Jeremy Allaire Introduction

00:21 – Origin Story of Circle

02:11 – Rethinking the Financial System

05:26 – The Role of Stablecoins

09:52 – Use Cases for USDC

11:30 – Programmable Money

12:25 – Blockchain as Operating System

14:37 – The Agentic Economy

17:45 – Arc Blockchain Use Cases

27:00 – Scaling Models and Privacy Tech

30:45 – Securitization of Other Assets Under the Blockchain

34:16 – Prediction Markets

35:09 – Incremental Revenue Through GPU Usage

37:19 – Jeremy’s 10 Year Future Vision

41:12 – AI and GDP

44:00 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Periodic Labs 联创 Liam Fedus：把大模型扩展定律用于原子世界

Fri, 03 Apr 2026 00:00:00 +0000

What happens when you apply the scaling laws of large language models to the physical work of atoms? Elad Gil sits down with Liam Fedus, co-founder at Periodic Labs, which is pioneering an AI foundation lab for atoms. Liam discusses how he pivoted from dark matter physics research to the front lines of artificial intelligence, including stints at Google Brain and working on ChatGPT at OpenAI. He talks about how Periodic is connecting massive language models to the physical world to overcome data bottlenecks in material science. Liam also shares how they use language models as an orchestration layer operating alongside specialized neural nets to run closed-loop physical experiments. They also explore the future of AGI and ASI, as well as the role of robotics in lab automation.

Chapters:

00:00 – Cold Open

00:05 – Liam Fedus Introduction

00:39 – Liam’s Background at Google Brain, OpenAI

05:14 – From ChatGPT to Materials and Atoms

06:34 – Training Data in the Physical World

09:52 – Generalization Across Domains

11:31 – Models as an Orchestration Layer

12:48 – Commercialization and Business Model

16:10 – How Periodic’s Success May Shape the Future

17:45 – Multidisciplinary Scaling

19:41 – Capital and Compute

21:12 – Hiring at Periodic

21:44 – Thoughts on AGI and ASI

23:30 – Timeline for Machine-Directed Self-Improvement

25:39 – Automation and Data Generation

27:59 – Why Liam is Excited About the Future of Robotics

29:25 – Conclusion

在 nopriors.qiaomu.ai 查看本期

Andrej Karpathy：代码Agent、AutoResearch与AI的循环时代

Fri, 20 Mar 2026 00:00:00 +0000

What happens when AI agents can design experiments, collect data, and improve — without a human in the loop? Andrej Karpathy joins Sarah Guo on the state of models, the future of engineering and education, thinking about impact on jobs, and his project AutoResearch: where agents close the loop on a piece of AI research (experimentation, training, and optimization, autonomously).

00:00 Andrej Karpathy Introduction

02:55 What Capability Limits Remain?

06:15 What Mastery of Coding Agents Looks Like

11:16 Second Order Effects of Natural Language Coding

15:51 Why AutoResearch

22:45 Relevant Skills in the AI Era

28:25 Model Speciation

32:30 Building More Collaboration Surfaces for Humans and AI

37:28 Analysis of Jobs Market Data

48:25 Open vs. Closed Source Models

53:51 Autonomous Robotics

1:00:59 MicroGPT and Agentic Education

1:05:40 Conclusion

在 nopriors.qiaomu.ai 查看本期

Notion 联创 Simon Last：从写代码到做管理，转向Agent工程

Thu, 12 Mar 2026 00:00:00 +0000

Notion isn’t designing AI agents that just use tools. Their agents can autonomously build their own integrations, as well as write the code needed to finish a task. Sarah Guo sits down with Notion Co-Founder Simon Last to explore Notion’s rapid evolution from a simple writing assistant to a sophisticated platform for custom AI agents. Simon discusses the technical hurdles of indexing disparate data from sources like Slack and Google Drive, as well as the internal shift toward using coding agents to build Notion itself. Plus, Simon elaborates on what he sees as a fundamental transition in productivity: moving from a tool where humans do the work, to one where humans manage a swarm of agents.

Chapters:

00:00 – Cold Open

00:05 – Simon Last Introduction

00:26 – Genesis of Notion AI

04:10 – Challenge of Semantic Indexing and Retrieval

07:16 – The Six-Month Rewrite Cycle

08:12 – Notion’s Coding Agent Era

09:44 – Impact on Team Dynamics

12:49 – Launching Custom Agents

15:39 – Notion as the ‘Switzerland’ for Models

17:33 – Designing APIs for Agent Customers

20:09 – Simon’s Personal Agentic Workflows

24:48 – Notion: Tool for Work is Now A Tool for Agents

27:28 – How Building Has Changed for Simon

29:00 – Conclusion

在 nopriors.qiaomu.ai 查看本期