Marc Andreessen introspects on Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"
概要
Marc Andreessen 深度反思浏览器之死、AI 作为 80 年一夜成功、neural networks 胜出的历史意义、A16Z 的 AI 投资逻辑,以及为什么 Pi + OpenClaw 代表了全新的人机交互范式
核心洞察
元信息
- 被访者:Marc Andreessen(联合创始人,Andreessen Horowitz / A16Z)
- 访谈者:Alessio Fanelli(创始人,Decibel Labs)+ Swyx(Latent Space 编辑)
- 来源:Latent Space 播客
- 日期:~2026-04
- 时长:~76 分钟
Executive Summary
- AI 是"80 年一夜成功":从 1943 年第一篇神经网络论文到 ChatGPT、O1、OpenClaw,技术突破不是凭空出现,而是 80 年基础研究的集中兑现。Andreessen 认为我们现在知道了"神经网络就是正确的架构"——这个在过去 60-70 年间一直有争议的命题已经被证实。四个根本性突破(LLM、推理、Agent、自我改进/RSI)已全部实现并正在运行。
- Pi + OpenClaw 是"Unix 时刻"级别的架构突破:Agent 的本质被还原为 LLM + bash shell + 文件系统 + markdown + cron——所有非模型组件都是已知技术。关键洞察是 Agent 的状态存在文件中、可以迁移运行时、可以更换底层模型、可以自省和自我扩展功能。Andreessen 认为这是"十大最重要的软件之一",其概念突破的深度堪比语言模型本身的"下一个 token 补全"。
- 当前 AI 不会重演 2000 年泡沫崩盘:与当年电信公司(Global Crossing 等)用高杠杆超建光纤不同,当前投入主体是微软、亚马逊、谷歌等蓝筹公司,且每一美元的 GPU 投资都在立即转化为收入。供应链未来 3-4 年全部售罄,GPU 甚至出现"旧芯片比新芯片时更值钱"的反常现象——Michael Burr 做空 Nvidia 的论点被"180 度证伪"。
- AI 与 Crypto 的"大统一"即将发生:AI Agent 天然需要钱——最激进的 OpenClaw 用户已经给自己的 Agent 开了银行账户和信用卡。加密货币和稳定币提供了互联网原生货币,这正是 1990 年代浏览器时代 HTTP 402(Payment Required)未能解决的问题。同时,"证明你是人类"(Proof of Human)成为刚需,World 项目的生物识别+密码学验证架构被 Andreessen 认为"完全正确"。
- AI 的真正阻力不在技术,而在制度刚性:加州成为理发师需要 900 小时专业培训,码头工人工会 5 万人中 2.5 万人拿全薪在家坐着,联邦机构员工每月只到办公室 1 天——教育、医疗、法律、住房等行业被执照垄断和公共部门工会牢牢锁死。Andreessen 的结论是"AI 乌托邦主义者和 AI 末日论者都太乐观了"——他们都高估了 80 亿人改变行为的速度。
贯穿全场的核心线索是"释放已有系统的潜在力量"(unlocking latent power):从 1990 年代浏览器释放操作系统和数据库的潜力,到今天 Pi/OpenClaw 释放 Unix shell 和文件系统的潜力,再到 AI 编码释放所有"本该好用但从没好用过的设备"的潜力——Andreessen 的每一个技术判断都指向同一个底层信念:最好的突破不是从零发明,而是用新的桥接层解锁已经存在但被困住的能力。
"80 年一夜成功":神经网络从异端到正统的漫长验证
核心要点:AI 不是突然出现的新事物,而是 80 年基础研究的集中兑现——四个根本性功能突破(LLM、推理、Agent、RSI)已全部实现。
- 1943 年第一篇神经网络论文问世,1955 年达特茅斯大学 AGI 会议获得 NSF 资助,一群 AI 专家认为 10 周暑假就能搞出 AGI——结果当然没有。Andreessen 本人经历了 1980 年代的 AI 大繁荣:专家系统、Lisp 语言、Lisp 机器——"我自己当年就在写 Lisp,那是'AI 未来的语言'"。
- 2013 年 AlexNet 是"曲线真正拐弯的点",2017 年 Transformer 是关键突破,但之后出现了一段"诡异的四年空白期"(2017-2021):谷歌内部有聊天机器人却不让任何人用,GPT-2 出来后 OpenAI 宣布"太危险不能部署",普通人使用 GPT-3 的唯一途径是通过 AI Dungeon 假装玩龙与地下城——"实际上你只是想跟 GPT 说话"。
- 到 2025 年春天,善意的怀疑论者仍可以说"这只是模式匹配"、"幻觉率太高"、"这对创意写作很好但不能用于编程或医学"。O1 和 R1 的推理突破回答了这个问题,假期期间的编码突破是第三步——"如果 Linus Torvalds 都说 AI 编码比他强,这是前所未有的基准"。然后是 OpenClaw 的 Agent 突破和自我改进(RSI)突破。
- 有大量 AI 研究者一辈子都没看到自己的工作成功:"拿了 PhD,研究了 40 年,退休了,很多情况下去世了,从来没有看到它真正工作。" Andreessen 特别提到 John McCarthy——达特茅斯会议的组织者之一,在斯坦福教了 40 年书后去世,没能看到 AI 成真。"但回过头来看,这些人极其聪明,工作极其努力,而且他们是对的。"
"If I were 18, this is 100% what I would be spending all of my time on. This is such an incredible conceptual breakthrough." —— Andreessen
"I call it an 80-year overnight success — it's an overnight success because bam, ChatGPT hits, then O1 hits, then OpenClaw hits — but they're drawing on an 80-year wellspring backlog of ideas and thinking." —— Andreessen
"This Time Is Different":为什么这次不是 2000 年泡沫
核心要点:Andreessen 亲历了 .com 崩盘,对当前 AI 投资周期的核心判断是——基本面完全不同,做空是"邀请别人撕掉你的脸"。
- .com 崩盘的真正故事是电信崩盘/带宽崩盘:美国商务部 1996 年发布报告称互联网流量每季度翻番,电信企业家据此融资铺设光纤。1998-1999 年增速放缓,预期与现实出现缺口——Global Crossing 等公司不仅融了大量股权,还上了大量债务杠杆,结果"大约 2 万亿美元被蒸发"。那些超建的光纤和数据中心全部在用——但花了 15 年(2000 到 2015 年)才填满产能。"酒店业有句话:总是第三任业主才赚钱,因为前面要破产两次。"
- 当前 AI 投资的关键差异:投入主体是微软、亚马逊、谷歌、Meta、Nvidia 这些蓝筹公司,拥有大量从未动用的债务容量;每一美元的 GPU 投资都在立即转化为收入——"所有人都在抢算力"。
- 一个反直觉的现象:3 年前的 Nvidia 芯片今天赚的钱比 3 年前更多,因为软件优化的速度快于芯片折旧速度;谷歌据说在用"非常老的 TPU"且利润很高。"旧芯片变得更值钱,这在历史上从未发生过。" Michael Burr 做空 Nvidia 的论点被"180 度证伪"。
- Andreessen 有朋友每天在 OpenClaw 上花 1,000 美元买 Claude token——每月 3 万美元——"而且这些朋友还有一千个想让 Agent 做的新想法"。他估计一个完全部署的个人 Agent 的潜在需求可达每天 5,000-10,000 美元,即使价格性能提升 10 倍,仍然是每天 100 美元,远超普通消费者支付能力。
"The 4 most dangerous words in investing are 'this time is different.' ...I'll tell you what's different: now it's working." —— Andreessen
"The idea of betting against this is an invitation to get your face ripped off." —— Andreessen
缩放定律:从摩尔定律到 AI 的多重缩放曲线
核心要点:AI 的缩放定律与摩尔定律一样,本质上是"自我实现的预言"——设定基准后激励整个行业去突破障碍,而 AI 可能存在多条尚未发现的缩放曲线。
- 摩尔定律是"每 18 个月芯片性能翻倍或价格减半",持续了 50 年,把 2500 万美元的大型机变成了口袋里 500 美元的手机。关键特征是:缩放定律不是真正的"定律",而是预言——但当它们奏效时,就成为自我实现的预言,激励全行业最聪明的人去确保它成真。
- AI 的缩放定律也一样:会遇到看起来"即将到来的墙",然后工程师们会想办法穿透;会看起来停滞然后再加速。目前已出现多条缩放曲线(训练、推理、Agent 等),"可能还有我们尚未发现的"——比如世界模型和机器人领域可能存在关于大规模真实世界数据获取的缩放定律。
- Agent 突破带来的不仅是 GPU 瓶颈,还转化为 CPU 和内存瓶颈——"整个芯片生态系统都可能被卡脖子,持续数年"。推理成本总体会继续下降,但下降速率可能会因供应约束而趋平,加上实验室可能减少补贴。
- 我们今天用的模型是"缩水版":如果 GPU 便宜 10 倍、数量多 10 倍,模型会好得多——"我们甚至没拿到好东西"。Swyx 补充说"我们用的都是量化版,因为实验室必须把完整版留给自己"。
"Even if technical progress stops, once there's a much bigger build of GPU manufacturing capacity...even the current technology is going to get much better." —— Andreessen
开源 AI:DeepSeek 是"给世界的礼物",Nvidia 可能是最终赢家
核心要点:中国公司做开源有特定商业逻辑——无法在美国销售商业 AI,因此视开源为损失前导(loss leader);开源的最大价值不是免费软件本身,而是"学会它怎么运作"。
- DeepSeek R1 的意义在于"信息扩散":OpenAI 的 O1 是了不起的推理突破但不公开原理、隐藏推理链;R1 开源代码和论文后,全世界知道了怎么做——3 个月内所有其他 AI 模型都加上了推理能力。"即使中国模型本身不被使用,对世界其他地方的教育和信息扩散作用也极其强大。"
- 当前美国和中国加起来约有 12 家有规模的基础模型公司,但 3 年后不会有 12 家——"这类市场容不下 12 家,会剩下 3-4 个大赢家或 1-2 个"。开源将成为"出局者"的替代策略,谁来做开源这件事会变化很快。
- Nvidia 的策略是"商品化互补品"(commoditize the complement):Jensen 黄仁勋在大力投入开源 AI 软件——这对他来说显而易见,因为软件商品化了,硬件就更值钱。"也许最终做开源 AI 的就是 Nvidia,我觉得那很好。"
- 上届美国政府"想把美国开源 AI 掐死在摇篮里",现任政府对 AI 和开源 AI 持"非常开明的观点"。
- 中国的"五虎"(Moonshot、DeepSeek、智谱、Qwen/通义、01)竞争激烈,字节跳动是下一梯队,腾讯也可能有后手。
"DeepSeek was like a gift to the world...the impact of open source is felt two ways: you get the software for free, but you also get to learn how it works." —— Andreessen
Pi + OpenClaw:Agent 的"Unix 时刻"——LLM + Shell + 文件系统 = Agent
核心要点:Agent 的架构被还原为已知组件的组合(LLM + bash shell + 文件系统 + markdown + cron),其中最颠覆性的特征是 Agent 与底层模型解耦、可自省、可自我扩展——这是历史上从未有过的广泛部署的自省式软件系统。
- Andreessen 从 Unix 历史讲起:1960 年代 IBM 的 OS/360 是"天空中的巨型城堡"——巨大的单体架构,强大但几乎不可接近。然后 Unix 说"不,我们要一个完全不同的架构":shell、离散模块、管道串联——"操作系统本身就是一种编程语言"。这催生了 shell 的中心地位、Perl 等脚本语言、Unix 工具链。Andreessen 自己从 1988 年起就是 Unix 人。
- Pi + OpenClaw 的本质是把语言模型思维嫁接到 Unix shell 思维上。Agent 是什么?LLM + bash shell + 文件系统 + markdown 格式 + cron 循环(heartbeat)。除了模型之外,每一个组件都是我们已经完全理解的。
- 几个令人震撼的推论:(1)Agent 独立于底层模型——可以换一个 LLM,Agent 的状态(存在文件中)完全保留,"就像换了编译器重新编译,但还是你的 Agent,保留所有记忆和能力";(2)Agent 可以自我迁移到不同运行时环境、不同文件系统;(3)Agent 拥有完全的自省能力——它知道自己的文件,并且可以改写自己的文件——"历史上从未有过广泛部署的软件系统具备这种自省和自我修改的能力";(4)你可以告诉 Agent "给自己添加新功能",它就会去做。
- 一个派对上的真实场景:有人说"我的 OpenClaw 连接了我的 Eight Sleep 床垫,给我更好的睡眠建议"。你回家跟自己的 Agent 说"把这个能力加给你自己"——它就会上网查资料、用 Claude Code 写代码、然后自动获得这个新能力——"你什么都不用做,只需要告诉它你想要这个"。
- 关于 MCP 等"花哨协议":Andreessen 认为不需要——"我们只需要命令行接口就行了"。Shell 本身有巨大的潜在力量,所有 Unix 命令、所有命令行接口、整台电脑的全部能力都在 shell 层面可用。
"What is an agent? It's LLM plus shell plus file system plus markdown plus cron. And it turns out that's an agent." —— Andreessen
"You can tell the agent to add new functions and features to itself and it can do that — extend yourself, give yourself a new capability." —— Andreessen
浏览器的设计决策如何在 AI 时代回响:文本协议、人类可读性与"释放潜在力量"
核心要点:1990 年代浏览器时代最关键的设计选择——文本协议而非二进制协议、人类可读性、"假设无限带宽来构建"——正在 AI 时代重演,核心逻辑相同:释放已有系统的潜在力量。
- 当年所有老派系统架构师都说互联网带宽稀缺(家用 14Kbit 调制解调器),应该用高压缩二进制协议、持久连接。Andreessen 团队"完全反着来":HTTP 是文本协议、HTML 标签人类可读、"用最低效的方式"。这是一个有意识的赌注——"假设无限带宽的未来来构建,如果系统的潜在能力足够强大,就会创造出足够的需求,推动带宽供应的建设"。
- 浏览器最关键的功能可能是"View Source"——每个网页都能看源码,这意味着你能自学如何建网页。"那个时代的人类可读性指的是技术规范,现在指的是英语自然语言——但给每个用户一个'下探并理解系统如何运作'的选项,这种潜在力量是惊人的。"
- Web 服务器的本质是释放操作系统和数据库的潜在力量:把互联网连接桥接到 OS 的文件系统管理能力和 Oracle/Postgres 数据库。当时有人说"我们一直有数据库,这只是数据库的又一个界面"——没错,但这个界面让 80 亿人使用,让数据库应用的数量爆炸了百万倍。
- 这与 Pi/OpenClaw 的逻辑完全一致——Andreessen 反复强调,最好的突破不是从零发明,而是"释放已有系统的潜在力量"。"行业里最聪明的人面对新挑战时想的是'先造一个新编程语言、再造一个新操作系统、再造一个新芯片'——我更倾向于:不,你已经有了所有这些潜在力量,你要做的是解放它。"
"The key breakthrough in the browser was the view source option — every web page you go to, you could see how it worked, which means you could teach yourself how to build." —— Andreessen
编程语言、浏览器和用户界面的消亡
核心要点:如果 AI 做所有编码,中间抽象层(编程语言、浏览器、UI)的存在理由将被根本性动摇——10 年后可能不存在我们今天理解的"编程语言"概念。
- Andreessen 认为模型不在乎用什么语言编程,且能在任意语言间互译。想要所有代码用 Rust 写?"告诉 bot 就行了"。他认为"我们很接近能让 AI 自己设计最优编程语言"——甚至让语言模型直接生成新模型的权重(已有人做了实验),跳过编程语言直接输出二进制。
- "高质量软件从稀缺资源变成无限可用的东西"——这将带来计算机安全领域有史以来最剧烈的变化:"首先所有潜在安全漏洞都会被暴露——我们会迎来计算机安全的末日。但另一面是编码 Agent 可以修复所有漏洞。"
- 更激进的推论——"如果不需要编程语言,那浏览器也死了"。谁会在未来使用软件?"其他 bot。" 那还需要用户界面吗?"你确定吗?" Swyx 追问:"那你干嘛?" Andreessen 回答:"想干嘛干嘛。" 他引用历史做比:"不久前 99% 的人类还在犁地。人们不犁地之后会干什么?结果发现有比犁地好得多的事情可做。"
- 但 Andreessen 并非绝对主义者:"我有一个 11 岁的儿子,他正在学编程,我觉得学编程仍然是好主意。"
"Are you even going to have programming languages in the future? Or are bots just going to be emitting binaries?" —— Andreessen
"Who is going to use software in the future? Other bots." —— Andreessen
YOLO 模式与 Agent 生活:睡眠监控、机器狗重写固件、智能家居接管
核心要点:最激进的 OpenClaw 用户正在用 Agent 做上千件事——从监控睡眠到改写机器狗固件,这些"先烈"式的实践正在定义 AI Agent 的真正能力边界。
- Andreessen 有一个朋友让 OpenClaw 用卧室摄像头监控自己睡觉。他看过运行日志,场景极其生动:"Joe 睡着了。这很好,因为我有他的健康数据,我知道他最近睡眠不足。我真的希望他能睡满 5 小时 REM。" 然后:"Joe 在动。Joe 可能要醒了。这是真正的危机——如果 Joe 现在醒来会毁掉他的睡眠周期。" 然后:"好的,Joe 只是翻了个身,又睡着了。好的,我可以放松了。" 朋友的评价是:一方面这很诡异也许在接管自己的生活,另一方面——"如果我半夜心脏病发作,这东西毫无疑问会疯掉然后打 911,可能还会召唤 SWAT 来救我的命。"
- 中国公司 Unitree 的机器狗:自带的非 LLM 控制系统"营销不错但实际上不太行"——上楼梯有问题。后来加了 LLM 和语音,但与控制系统完全不连通——"你有了一条精神分裂的狗,爬楼梯是白痴,但能用一口英国口音教你量子力学"。Andreessen 的朋友让 OpenClaw 入侵并重写了机器狗的固件——"现在它是他孩子们真正的宠物狗了",每次出问题 Agent 就直接重写代码修复。
- OpenClaw 特别擅长"入侵你局域网里的所有东西"——物联网设备虽然安全性差但可被发现。最激进的用户让 OpenClaw 接管家里所有设备:安防摄像头、门禁系统、网络摄像头。"这是我第一次能自信地说,我知道怎么才能真正拥有一个智能家居——30 种带芯片和网络的设备全部协调一致。"
- 关于"YOLO 模式"(dangerously 标志,来自 Facebook 的内部文化——把东西命名为"dangerous"以提醒你正在做危险操作,但当然这反而让人更想启用它):Sam Altman 在自己的笔记本上用跳过权限模式运行 Codex。Andreessen 的态度是:"我自己没这么做,但我爱这些人——他们是人类文明进步的殉道者。虽然他们的银行账户可能在头 20 分钟就被 bot 洗劫,但他们对物种未来的贡献是惊人的。就像 Ben Franklin 放风筝试图被闪电劈中,就像 Jonas Salk 给自己注射脊髓灰质炎疫苗。"
"I have a friend whose claw watches him sleep...the transcripts are literally: 'Joe's asleep. This is good. I really hope he gets his full 5 hours of REM sleep.'" —— Andreessen
"I think the people who turn that on for bots are like martyrs to the progress of human civilization." —— Andreessen
AI x Crypto 大统一:Agent 需要钱,HTTP 402 终于要实现了
核心要点:AI 是 crypto 的杀手级应用——Agent 天然需要互联网原生货币来执行经济行为,这正是 1990 年代 HTTP 402(Payment Required)未能解决的遗留问题。
- Andreessen 确认自己当年就是提出"互联网最大的错误是没有解决支付问题"的人——HTTP 402 状态码"Payment Required"至今未被实现。他现在认为"这次肯定会被解决"。
- 两个原因:(1)我们现在有了互联网原生货币——加密货币和稳定币;(2)AI Agent 显然需要钱——"如果你有一个 Agent 要替你买东西,你必须给它某种形式的钱"。
- 最激进的 OpenClaw 用户已经给 Agent 开了银行账户和信用卡——"目前可能只有 5,000 人这么做了,但这就是这类事物开始的方式"。Andreessen 引用 William Gibson 的名言:"未来已经到来,只是分布不均。"
- 他开玩笑说:"如果你不给 OpenClaw 银行账户,它反正会自己黑进去把钱拿走——所以你还不如给它。"
"AI is the crypto killer app. I think this is the grand unification of AI and crypto." —— Andreessen
Proof of Human:虚拟世界的身份验证与物理世界的无人机威胁是同一个问题
核心要点:Bot 已通过图灵测试,"证明不是 bot"已不可能,必须转向"证明是人类"——World 项目的生物识别+密码学+选择性披露架构被 Andreessen 认为"完全正确"。
- 互联网充斥假人和 bot——"每个社交媒体用户都知道这个问题"——但从未被正面对抗。如今语言模型让 bot 彻底不可区分,"证明不是 bot"已不可能,必须转向"证明是人类":需要生物识别验证起点(否则 bot 会注册为假人)、密码学验证、选择性披露(证明你是人类而不暴露底层信息)。
- 未来还需要"年龄证明"(各国法律对 13/16/18 岁的不同要求)、"信用评分证明"等——"查你信用的人不应该需要知道你的名字"。这也是隐私问题的终极解法:"我只需要在那个时刻证明我需要证明的东西。"
- Andreessen 将虚拟世界的 bot 问题和物理世界的无人机问题视为同一种"经济不对称":发动攻击成本极低,防御成本极高。无人机方面:20 年来我们知道廉价自杀式无人机是最大不对称威胁,但每栋写字楼、体育场、学校、监狱都毫无防护——"我们知道了,却什么都没做"。乌克兰冲突和伊朗局势正在推动反无人机技术的爆发。
管理资本主义的第三条道路:创始人 + AI = 两全其美
核心要点:James Burnham 在 1940 年代提出的"资产阶级资本主义 - 管理资本主义"二阶段论可能被 AI 打破——创始人 + AI 超级能力可能实现"天才直觉 + 管理规模"的第三种模式。
- Burnham 的理论:第一阶段是"资产阶级资本主义"(bourgeois capitalism)——"门上挂名字"的模式,Henry Ford 亲自指挥一切,但不可扩展;第二阶段是"管理资本主义"——创建不懂具体业务但懂管理的专业经理人阶层,催生了商学院、管理咨询公司,今天的财富 500 强绝大多数由非创始人的职业经理人运营。
- 风险投资本质上是对管理主义的"反叛残余"(rump protest movement)——试图找到下一个 Henry Ford / Elon Musk / Steve Jobs / Mark Zuckerberg,用"君主制"的创始人模式去创新,赌注是创始人能做到管理主义官僚做不到的事。"但我一辈子都觉得我们是在'怒斥光明的消逝'(rage against the dying of the light)——不断试图阻止管理主义吞噬一切。"
- AI 可能开启第三条道路:创始人 + AI 超级能力。AI 最擅长什么?"做文书工作、填表格、写报告、阅读材料——所有管理工作,它们都极其擅长。" 因此最佳模式可能是"天才创始人的直觉创新 + AI 做所有管理层的工作"——"我们从来不知道自己想要这个,因为我们从来没想过这是可能的"。
- 对大公司的影响:管理主义运营的传统巨头将面对一种前所未有的"叛乱者"——拥有 AI 超能力的创始人驱动型公司。"这将迫使很多大公司要么搞清楚创新,要么死于尝试。"
"VC is a rump protest movement...we're constantly trying to fight off managerialism basically swamping everything and everything getting boring and gray and dumb." —— Andreessen
"AI 乌托邦主义者和末日论者都太乐观了":制度刚性才是真正的阻力
核心要点:技术使某件事成为可能,不等于 80 亿人会改变行为——经济中大量行业被执照垄断、工会保护和政府垄断锁死,AI 的真正对手不是技术瓶颈而是制度惯性。
- 加州成为理发师需要 900 小时培训;全美约 35% 的经济需要某种专业认证才能从业——"这些职业本质上都是卡特尔"。
- 码头工人工会的故事:亚洲现代码头全部机器人化,美国码头仍然靠人工搬运。工人罢工成功,迫使码头业主承诺不引入更多自动化。工会有 5 万人:2.5 万人在码头工作,另外 2.5 万人根据此前的工会协议拿全薪坐在家里。"即使只有 2.5 万人的工会仍然有巨大的政治影响力。"
- 联邦政府机构的例子:有些机构在 COVID 期间签订的新集体谈判协议规定员工每月只需到办公室 1 天。员工很聪明——在月末最后一天和下月第一天来——"所以他们每 60 天在办公室 2 天,大楼空 58 天,我们所有人都在为此买单"。
- 美国 K-12 教育是"政府垄断"——"AI 如何应用于教育?答案是不会,因为这是政府垄断,永远不会改变。教师 100% 反对,100% 不会发生。唯一能做的是像 Alpha School 那样创建全新的教育体系。"
- Andreessen 的终极判断:"AI 乌托邦主义者和 AI 末日论者都太乐观了——因为他们都相信技术使某件事成为可能后,80 亿人就会改变行为。不会的。现有经济的运行方式大量是'硬接线'的。AI 快速落地对社会来说是幸运的——因为如果不快速落地,我们只会得到停滞。"
"Both the AI utopians and the AI doomers are far too optimistic — because they believe that because the technology makes something possible, 8 billion people are all of a sudden going to change how they behave." —— Andreessen
附录:关键人/机构/产品/数据
| 项目 | 详情 |
|------|------|
| Marc Andreessen | A16Z 联合创始人,Mosaic/Netscape 浏览器创造者,1980 年代起从事 AI |
| A16Z | Andreessen Horowitz,投资了 OpenAI、Thinking Machines、World 等 |
| Pi | Agent 架构的底层突破,将 LLM 与 Unix shell 思维结合 |
| OpenClaw | Agent 产品,基于 Pi 架构,Andreessen 称两者组合为"十大最重要软件之一" |
| DeepSeek R1 | 中国开源推理模型,"给世界的礼物",让全球学会了如何实现推理能力 |
| World(原 Worldcoin) | Alex Blania 领导的 Proof of Human 项目,A16Z 参与投资 |
| James Burnham | 20 世纪政治思想家,提出"资产阶级资本主义→管理资本主义"二阶段论 |
| AlexNet(2013) | 深度学习突破点,"曲线真正拐弯的点" |
| Transformer(2017) | 关键架构突破,之后经历 4 年"诡异空白期" |
| HTTP 402 | "Payment Required" 状态码,互联网支付的未兑现承诺 |
| Global Crossing | .com 时代电信公司,高杠杆超建光纤后破产,约 $2 万亿蒸发的标志 |
| 达特茅斯 AGI 会议(1955) | 首次 AGI 研究会议,获 NSF 资助,认为 10 周可实现 AGI |
| John McCarthy | AI 领域创始人之一,斯坦福教授 40 年,未能亲眼见到 AI 成功即去世 |
| Linus Torvalds | Linux 创始人,承认 AI 编码已超越自己——"前所未有的基准" |
| Mistral | 欧洲开源模型公司,A16Z 投资,"在中国之外做得极好" |
| Unitree 机器狗 | 中国公司产品,控制系统差但 LLM 好,用户让 Agent 重写固件变成"真宠物" |
| Alpha School | 新型教育体系,Andreessen 认为是绕过公立教育垄断的唯一出路 |
| 每天 $1,000 token 消费 | Andreessen 朋友的 OpenClaw 使用成本,月均 $3 万 |
| 900 小时 | 加州成为理发师所需的专业认证培训时长 |
| 25,000 + 25,000 | 美国码头工人工会:2.5 万在岗 + 2.5 万拿全薪在家 |
| 15 年(2000→2015) | .com 时代超建的光纤和数据中心填满产能所需时间 |
I think what's actually happened is an enormous amount of technical progress that built up over time. And like for for example, we now know that neur
al network is the correct architecture. And I will tell you like there was a 60-year run where that was like a you know, or even 70 years where that w
as controversial. And so so the way I think about what's happening is basically I think about basically the the period we're in right now is it's I ca
ll it an 80year overnight success, right? which is like it's an overnight success cuz it's like bam, you know, chat GPT hits and then and then 01 hits
and then, you know, open claw hits and like, you know, these are open these these are like overnight like radical overnight transformative successes,
but they're drawing on an 80year sort of wellspring backlog, you know, of of of ideas and thinking. It's not just that it's all brand new, it's that
it's an unlock of all of these decades of like very serious hardcore [music] research. If I were 18, like this is 100 this is what I would be spending
all of my time on. This is like such an incredible conceptual breakthrough. Before we get into today's episode, I just have a small message for liste
ners. Thank you. We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choos
e to also click in and tune into our content. We've been approached by sponsors on an almost daily basis. But fortunately, enough of you actually subs
cribe to us to keep all this sustainable without ads, and we want to keep it that way. But I just have one favor to ask all of you. The single most po
werful, completely free thing you can do is to click that subscribe button. It's the only thing I'll ever ask of you, and it means absolutely everythi
ng to me and my team that works so hard to bring the Inspace to you each and every week. If you do it, I promise you, we'll never stop working to make
the show even better. Now, let's get into it. Hey everyone, welcome to the Lydian Space Podcast. [music] This is Allesio, founder of Colonel Labs, an
d I'm joined by Spix, editor of L and Space. Hello. And we're in A16Z with a uh Mark and Jason Gson. Welcome. Yes. Yes. A and what half of 16 [laug
hter] something like
ad.
nal office. We're in the we're in the we're in where the whole thing started. It's beautiful. Great. Thank you.
a, you know, I wanted to pick a spicy start. In October 2022, I just made friends with Rune and, uh, I wanted to give him something to sort of be spic
y about. And I said, uh, it'll never not be funny that A6Z was constantly going, "The future is where the smart people choose to spend their time and
then going deep into crypto and not in AI." And that was in October 22 2022. And Run says there was an internal meeting in A6Z to reorient around Gena
i. Obviously, you have, but was there a meeting? What What was that? I mean I don't look I've been doing AI since the late 80s. So I I don't know like
all as far as I'm concerned this stuff is all Johnny come lately. Yeah. I mean look we've been doing AI our entire existence. I mean we've been do
ing AI machine learning you know deep we've been doing this stuff way from the beginning. Obviously a AI is just core to computer science. I I I actua
lly view them as like quite [snorts] uh quite continuous. Um you know Ben and I both have computer science degrees. Um you know we we both Ben and I a
ctually both are old enough to remember the actual AI boom in the 1980s.
mes like expert systems um and the era of like lisp and list machines. Um I I coded in lisp. I was coding in lisp in where that was the the languag
e of the AI future. Um yeah so this is something that we're like completely you've completely comfortable with and been doing the whole time and are v
ery enthusiastic about.
d out very very quickly. Um just just in terms of investing sort of sort of investment excitement although that's really when the the Nvidia phenom
enon really it was I would say it was in that period when it was very clear that at the time the vocabulary was more machine learning but it it was ve
ry clear at that time that machine learning was hitting some sort of takeoff point.
our on your thing but you know if you really track what happened I think the real story is it was it was the AlexNet uh basically breakthrough in like
2013 that was the that was the real knee in the curve. Um and then it was obviously the transformer breakthrough in 17. Um and then everything tha
t followed but but you know look machine learning you know there were you know look uh I mean look I've been working you know I've been working with o
ne of my you know kind of projects working with Facebook since 2004 um on the board since 2007 and of course that you know they they started using mac
hine learning very early um and you know have used it basically you know for like 20 years for you know content you know feed optimization and adverti
sing optimization and obviously many you [snorts] know financial services you know many many many companies many different sectors have been doing thi
s. So, it's like one of these things. It's like it's not a sing it's not a single thing. Like it's it's like it's like layers, right? Um and and the l
ayers arrive at different paces and but they kind of build up
it was 2017 was kind of the you know the key the key point with transformer and then and then as you guys know there was this really weird like four-
year period where it's like the the transformer existed and then it was just like [snorts] let's go.
etween 2020 but between 2017 and 2021 I mean that was the era of which like companies like Google had internal chat bots but they weren't letting anyb
ody use them. Yeah.
o deploy. Right. You know we can't possibly let normal people normal people use this thing. And then you guys I'm sure remember AI dungeon. Um so t
he only for there was like a year where like the only way for a normal person to use GPT3 was in in AI dungeon.
this. You'd go in there and you'd pretend to play Dungeons and Dragons. In reality, you're just trying to talk to talk to GPT. And so there was this,
you know, there was this long, you know, you know, the big big companies, you know, big companies are cautious and, you know, the big companies were
cautious. It it, by the way, it took Open AI, you know, they they they talked about this. It took Open AAI time to actually adjust, you know, kind of
red redirect their research
that dinner would have taken place in 2018,
ebrated a 10 year anniversary. So it it is 2025. Yeah. So 2015. Yeah. 2015. Yeah. 2015.
> 1718 171 18. So it Yeah. For and then and then they didn't really and then GPT3 was what 2020 20 2020 that became co-pilot 21. even OpenAI which
has been you know the leader of this thing in the last decade you know even they had to adapt and and and lean into the new thing and so um yeah I I t
hink it's just this process of basically sort of wave after wave layer after layer you know building on itself and then you kind of get these catalyti
c moments where where the whole thing pops and and obviously that's what's happening now is it useful to think about will there be any winter because
there's always these patterns like is this endless summer it's something I constantly think about because do I get do I just like just get endlessly h
yped and just trust that I will only be early and never wrong or [laughter] or will there be a winter?
g. There's something about AI that has led to this repeated pattern. Um, and and you guys know this, it's summer winter, some winter,
ter, some winter, and it goes back 80 years. 80 years. Uh, so the original neural network paper was 1943, right? Which is which is amazing uh that it
was it was far back that long. And then there was you if you guys have ever talked about this on your show, but there was this uh there was a big uh t
here was an AGI conference at Dartmouth University in 1955 and they got an NSF grant to uh for the all the AI experts at the time to spend the summ
er together and they figured if they had 10 weeks together they could get AGI [laughter]
e grant they got the 10 weeks and then you know 1955 you know no AGI and like I said I live through the 80s version of this where there was a big a
big boom and a crash and so so there is this thing there is something about AI that causes the people in the field I would say to become um both exce
ssively utopian and excessively apocalyptic. Um and and it's probably on both sides of like the the boom bus cycle, you kind of see that play out. Hav
ing said that, I think what's actually happened is like just in, you know, we now know in retrospect like an enormous amount of technical progress tha
t built up over time and like for for example, we now know that neural network is the correct architecture. And I will tell you like there was a 60-y
year run where that was like a you know, or even 70 years where that was controversial.
u know, everything we're building on today sort of derives from the original idea in 1943. And so so in retrospect, we now know that like these these
guys were right. You know, they would get the timing wrong and they thought, you know, capabilities would arrive faster or there it could be turned in
to businesses sooner or whatever. But like they were fundamentally the scientists who worked on this over the course of decades were fundamentally cor
rect about what they were doing [snorts] and and and the payoff from from all their work is happening now. And so so the way I think about what's happ
ening is basically I think about basically the the period we're in right now is it's I call it an 80-year overnight success, right? which is like it's
an overnight success because it's like bam, you know, chat GPT hits and then and then 01 hits and then you know openclaw hits and like you know these
are open these are these are like overnight like radical overnight transformative successes but they're drawing on an 80year sort of wellspring backl
og you know of of of ideas and thinking. It's not just that it's all brand new. It's that it's an unlock of all of these decades of like very serious
hardcore research um and thinking. I mean look, there were AI researchers who spent their entire lives, they got their PhD, [laughter] they worked for
they researched for 40 years, they retired in a lot of cases, they passed away and they never actually saw it work. Yeah.
is sad. It is sad.
McCarthy, you know, John McCarthy was like one of the inventors of the field. He's one of the guys organized the Dartmouth conference and, you know, h
e taught at Stanford for 40 years and passed, you know, passed away, I don't know, whatever, 10 10 years ago or something.
to see it happen. But like it is amazing in retrospect like these guys were incredibly smart and they worked really hard and they were correct. So an
yway, so then it's like okay you know they say history doesn't repeat but it rhymes. It's like okay does that mean that there's going to be another li
ke you know basically boom buzz cycle and I I will tell you like look like in a sense like yes everything goes through cycles and you know people get
overly enthusiastic and overly depressed and there there's a time there's a timelessness to that. Having said that there's just no question. Um so the for the four most dangerous words investing are this time is different. Do you know the 12 most dangerous words of investing? No.
four most dangerous words [laughter] investing are this time is different. Um the 12 most dangerous words. And so like I'll tell you what's different
. Like now it's working like like there's just no I mean look there's just no question. And by the way I'll just give you guys my take. like LLM's lik
e from from basically the Chad GPT moment through to spring of 25 I think you could still I think well-intentioned well-informed skeptics could still
say oh this is just pattern completion and oh these things don't really understand what they're doing and you know the hallucination rates are way too
high and you know this is going to be great for creative writing and creating you know Shakespearean sonnetss and you know as as rap lyrics or whatev
er like it's going to be great at all that stuff but we're not going to be able to harness this to make this relevant in you know coding or in medicin
e or in are and you know you know kind of feels that you know kind of really really matter and I think basically it was the reasoning breakthrough it
was 01 and then R1 that basically answered that question and basically said oh no we're going to be able to actually turn this into something that's g
oing to work in the real world and and then obviously the coding breakthrough over the over basically the coding breakthrough that kind of catalyzed o
ver the holiday break was kind of the third step in that [snorts] like all right if if you know if Lannos Tvolt is saying that the AI coding is now
better than he is like [laughter]
that it's it's going to sweep through coding and and then and then we we know you know we know that if it's going to work in coding it's going to wor
k in everything else right it's just then because that's that's like that's like that's like the hardest in many ways that's the hardest example and n
ow everything else is going to be a derivative of that
antastic which is amazing and incredibly powerful and then we just got the the um the auto research uh you know the the self-improvement you know w
e're now into the self-improvement breakthrough and so the so the way I think about it is we've had four fundamental breakthroughs in functionality LL
M's reasoning uh agents um and and uh and then now RSI. Um and they're all actually working. Um and so I'm I'm just as you can tell I'm jumping out of
my shoes [laughter] like this is like this is it. Like this this is the culmination of 80 years worth of worth of work and this is the time it's beco
ming real.
d it's like all right we understand why these things are getting better. We understand the physics of it. with AI it's it's so jagged in like the jump
s where like like you said it's like in three months you have like this huge jump like and people are like well this can't keep happening right but th
en it keeps happening it'll keep happening
e this question with guests which is like you know should you spend time building harness for a model versus like the next model just going to do it o
ne shot in the latest space and how does that inform like how you think about the shape of the technology you know you talk about how it's a new co
mputing platform. If you have a computing platform, then like every six months, it like drastically changes in what it looks like. It's hard to build
companies on top of it.
s a scaling law. And for your younger viewers, Moors law was every chip chip chips either get twice as powerful or twice as cheap every every 18 month
s. And that and that and then you know that it's gotten more complicated in the last few years, but like that that was like the 50-year trajectory of
of of the computer industry. And then and then by the way and that's what took the mainframe computer from a $25 million current dollar thing into you
know the phone in your pocket being you know a million times more powerful than that like that you know for for 500 bucks. And so that was a scaling
law and then and then and then key to any scaling law including Moors law and the AI scaling laws is you know they're not really laws right they're th
ey're they're predictions but when they work they become self-fulfilling predictions because they they they they set a benchmark and and then the enti
re industry right all the smart people in the industry kind of work to make sure that that actually happens. And so they they kind of motivate the bre
akthroughs that are required to to keep that going. And and and in chips that was a 50-y year that was a 50-y year run, right? And it it was amazing.
And it's still happening in in some areas of of chips. I think the same thing is happening with the the core scaling laws, the core scaling laws in in
in AI. you know, they're they're not really laws, but like they they are basically they're predictions and then they're motivating catalysts for the
research work that is required to be and and and by the way, also the investment uh dollars um are you know required to basically keep you know keep t
he curves going and and look it's going to be complicated and it's going to be variable and there you know there are going to be walls that are going
to look like they're fast approaching and then they're going to be you know engineers are going to get to work and they're going to figure out a way t
o punch through the walls and obviously that's you know that's been happening a lot you know and then look there's going to be times when it looks lik
e the walls have you know the the laws have petered out and then they're going to they're going to pick up again and surge And then and then and then
it it appears what's happened to the AI is there's now multiple you know multiple scaling laws. Um there's multiple areas of improvement and and I thi
nk you know I don't know how many more there are already yet to be discovered but there are probably some more that we don't know about yet. You know
like for example there's probably some scaling law around um world models and robotics that we don't fully you know kind of acquisition of data at sca
le in the real world that we don't fully understand yet. So that that that one will probably kick in at some point here. There's a bunch of really sma
rt people working on that. Um and so yeah, I I think the expectation is the that you know the the scaling laws generally are going to continue. Yeah,
the pace of improvement will continue to move really fast. Um to your question on like what to build. So I'm a complete believer that the scaling laws
are going to continue. I'm a complete believer the capabilities are going to keep getting amazing um you know leaps and bounds. uh the part where I k
ind of part ways a little bit with what I would describe as the AI purist um you know which is which I would characterize as like the people who are i
n many ways the smartest people in the field but also the people who spend their entire life like in a lab [snorts] um and have have I would say have
very little experience in the outside world. Um the the nuance I would offer is the outside world of 8 billion people and institutions and governments
and companies and economic systems and social systems is really complicated. Um and um and doesn't you know it it 8 billion people making collective
decisions on planet earth is not a simple process of like just like you see this happening now it's like a bunch of the AI CEOs have this thing which
is just like well there's just this they just all have this kind of thing when they talk in public where they're just like well there's just these obv
ious set of things that society needs to do [snorts] and then they're like societyy's not doing any of those things right and it's like how can soc
iety not you know whatever their theory is how can society not see XYZ and the answer is well society is number one there's no single society it's lik
e eight billion and they like all have a voice and they all have a vote like at the end of the day of how they react to change and then you know just
like it's just human reality is just really complicated and messy and and so the specific answer to your question is like as usual it depends um you k
now it depends look there's no question people are going to like there's no question there are going to be companies it's already happening there are
companies that think that they're building value on top of the models and they're just going to get blitzed by by the next model there's no question t
hat's happening but I think there's no question also that just the process of adaptation of any technology into the real into the real messy world of
humanity is is just going to be messy and complicated. It's it's not going to be simple and straightforward. It's going to be messy and complicated an
d there are going to be a lot of companies and a lot of products um and in in fact entire industries that are going to get built that to basically act
ually help all of this technology actually reach real people. [snorts] The amount of capital going into these companies, I mean Dario talked about it
on the Dorcash podcast and Dor Cash was like, "Why don't you just buy 10x more GPUs?" and he's like because I'm going to go bankrupt if the model does
n't exactly hit the the performance level.
hines and world apps it seems like we're leveraging the scaling loss at a pretty high rate like how comfortable I guess do you feel with the downside
scenario like and say like things peter out you think you can kind of like restructure uh these buildouts and uh you know capital investments. Yeah
. Yeah. So, I should start by saying so I live through the.com crash. Um, and I can tell you stories for hours about the do crash and it was horrible.
No, it was awful. It was it was it was apocalyptic. By the way, the a lot of the dot crash was actually at the time it was actually a telecom crash.
It was a bandwidth crash. Like the the thing that actually crashed that wiped out all the money was the the telecom companies.
nd it was literally the the US commerce department put out a report in 1996 and they said internet traffic was doubling every quarter. Um and and actu
ally in 1995 and 1996 internet traffic actually did double every quarter. And so that became the scaling law. So what all these telecom entrepreneurs
did was they went out and they raised money to build fiber anticipating that the demand for bandwidth was going to keep doubling every quarter. Doubli
ng every quarter though is like you know grains of chess on the chessboard like at some point the numbers become extremely large right and and and it
really and really what happened was the internet the internet by the way continuously kept growing basically since inception it's you know it's it's c
ontinuously grown it's never shrunk and it's grown really fast compared to anything else you know in human history but it wasn't doubling every qua
rter as of 1998 1999 and so there was this gap in the expectation of what they thought was a scaling law versus reality and that's actually what cause
d the dot crash which was it they they way over companies like global crossing way overbuilt fiber which is sort of the by the way fiber telecom equip
ment you know so all the all the networking gear you know and then and then by the way the actual physical data centers like that was the beginning of
the of the of the data center build and then and then data center overbuild and so you had that but it was it was literally I think it was like $2 tr
illion got wiped out right it was like it was like a big it was and by the way the other the other subtlety in it was the internet companies themselve
s never really had any debt because tech tech companies generally don't run on debt but the telecom companies run on debt physical infrastructure comp
anies run on debt And so the company's like, well, we're crossing not just raised a lot of equity. They also raised a lot of debt. So they're highly l
evered. And so then you just do the thing of just like, okay, you have a highly levered thing where you're you're just you're overbuilding capacity. D
emand is growing, but not as fast as you hoped. And then boom, bankrupt. Right. And and then and then it's like they say about the hotel industry, whi
ch is it's always the third owner of a hotel that makes money, right? It has to go bankrupt twice, right? You have to wash out all of the overoptimist
ic exuberance before it gets to actually a stable state and then it makes money.
e all the fiber that it's all in use today but 25 years later [snorts] but it it took and actually the elapse time was it took 15 years. It took 15 ye
ars from 2000 to 2015 to actually f fill up all that capacity. The cautionary warning is the overbuild can happen. Um and and and and you know, you ge
t into this thing where basically everybody everybody who basically has any sort of institutional capital is like, "Wow, it's just I I don't know how
to invest in these crazy software things, but for sure I can put build data centers and for sure I can buy GPUs and I can deploy, you know, compute gr
ids and and all these things." Um and and so, you know, if you're a pessimist, you could look at this and you could say, "Wow, this is like really set
up to be able to basically replicate, you know, what we went through what we went through in 2000." Obviously, that would be bad. The counterargument
which is the one I I agree with which is the counter on the other side is a couple things. One is the companies that are investing all the the compan
ies that are investing the money are like the bluest chip of companies. And so back back in the in the doc like global crossing was like a it was like
an entrepreneur. It's like a new venture. But like the money that's being deployed now at scale is Microsoft and you know and Amazon and Google right
and Facebook and Nvidia and you know these the these and and now you know by the way Open AI anthropic which are now at like you know really serious
size
capacity that they they've never used. And so this is institutional in a way that that really wasn't at the time. And then the other is at least for n
ow every dollar that's being put into anything that results in a running GPU is being turned into revenue right away like so and you guys know this li
ke everybody starve for capacity everybody starve for compute capacity and then you know all the associated things memory and and and interconnect and
everything else um data center space and so every dollar right now that's being put in the ground is turning into revenue and and in fact I actually
think there's an interesting thing happening which is because everybody starve for capacity the models that we actually have that we can use today are
inferior versions of what we would have if not for the supply constraints.
cheaper and 10 times more plentiful, the models would be much better because you would just allocate a lot more money to training and you'd just buil
d better models and they would be better. Um and so we're actually getting the sandbag version of the technology.
quantized because the the labs have to keep the the full versions, right? Like [laughter]
现在我们知道了。所以我们今天所构建的一切都源自 1943 年的那个最初想法。回过头来看,我们现在知道这些人是对的。他们在时间判断上会出错,他们以为能力会更快到来,或者以为能更早商业化什么的。但从根本上说,几十年来从事这项研究的科学家们,他们做的事情在根本上是正确的 [哼了一声],而他们所有工作的回报正在现在兑现。所以我对正在发生的事情的理解基本上是这样的——我们现在所处的这个时期,我称之为"80 年的一夜成功",对吧?它之所以是一夜成功,是因为——砰,ChatGPT 来了,然后 o1 来了,然后 Claude 来了——这些都是一夜之间的、激进的、变革性的成功,但它们汲取的是 80 年积累的思想和思考。不是说一切都是全新的,而是说这是对几十年非常严肃、硬核研究的一次释放。
你看,有些 AI 研究者把一生都投入其中,拿了博士学位,[笑] 研究了 40 年,退休了,很多情况下,去世了——他们从未真正看到它成功。
ut getting the good stuff is it's just even if technical progress stops once there's like a much bigger build of like GPU manufacturing capacity and m
emory [clears throat] you know all all the things that have to happen in the course of the next 5 or 10 years once it happens even the current technol
ogy is going to get going to get much better and then as you know like there's just like a million ways to use this stuff like there's just like a mil
lion use cases for this like it you know this isn't just sending packets across a thing whatever and hoping that people find something to do with it.
This is just like oh we apply intelligence into every domain of human activity and then it works like incredibly well. Um, here's what I know. Here's
what I know. Um, in the next 3 or 4 year, it's like somewhere between 3 or 4 years out, basically everything is selling out. So, like the entire suppl
y chain is is is sold out or selling out. And so, there there's no like we're just going to have like chronic supply shortage for, you know, for years
to come. Um, there's going to be a response from the market that's going to result in an enormous, you know, it's happening now. An enormous flood of
investment in a new fab capacity and, you know, everything else to be able to do that. some point the supply chain constraints will unlock you know a
t least to some degree that will be another accelerant to industry growth when that happens because the products will get better and everything will g
et cheaper and so so I know that's going to happen. I know that you know the deployments you know the actual use cases are like really compelling and
then like I said you know with reasoning and agents and so forth like I know they're just going to get like much much better from here and so I I know
the capabilities are like really real and serious. I also know that the technical progress is not going to stop. It it is accel is is accelerating li
ke the breakthroughs are are tremendous. I mean, even just month over month, the breakthroughs are really dramatic. And so, you know, I think if you w
ere a cynic, and there there are cynics, you can look at 2000, you can find echoes, but I can't even imagine betting that this is going to like someho
w disappoint in, you know, at least for years to come. I think it would be essentially suicidal to make that bet.
at's an interesting [laughter]
as He doesn't mind.
he analysis now that the current models are getting better faster at such a rate that if you are running an NVIDI if you're running an NVIDIA inferenc
e chip today that's 3 years old, you're making more money on it today than you did 3 years ago because the pace of improvement of the software is i
s faster than the than the depreciation cycle of the chip. And then my understanding is Google is running I don't think I don't know exactly what thes
e are rumors that I've heard or maybe it's public but um I think Google's running very old TPUs
所以,然后就是——人们说历史不会重演,但会押韵。那是不是意味着还会有又一个繁荣-萧条周期?我跟你们说,从某种意义上说是的——一切都会经历周期,人们会过度热情也会过度悲观,这有某种永恒性。话虽如此,毫无疑问——投资界最危险的四个字是"这次不一样"(this time is different)。你们知道投资界最危险的 12 个字是什么吗?
y turns out, as far as I can tell, it's actually the opposite of the Bur thesis is actually, he was actually 180 degrees wrong. It's actually the the
the old Nvidia chips are getting more valuable, which is something that's like literally never happened [laughter] before. Like it's never been the
case that you have an older model chip that becomes more valuable, not less valuable. And and again, that's an expression of the just ferocious pace
of software progress, ferocious pace of capability payoff that you're getting on the other side of this. And so I just the idea of betting against tha
t like
he H100 and H200s and and going like you know usually they adise like four to seven years and it was you know maybe you sort of realistically care cut
it down to two to three but actually it's going up and not down and and uh that's I mean that's I think that's the dream. uh we are finding utilizati
on and I think utilization solves all problems like you can you can find use use cases for even like the poor like even memory we're having a shortage
right and and even like the the shittier versions of of memory that we do have we are finding use cases for it so like that's great
ant is open source AI and kind of like edge inference in a world in which you have three years of supply crunch like do you think in the like you know
if you fast forward like five years like how do you think about inference uh in the data center versus at the edge well so just to start yeah So I
think I think open source is very important for a bunch of reasons. I think edge edge inference is very important for a bunch of reasons. I I think j
ust practically speaking if we're just going to have fundamental constru crunches for the next I mean you guys know if you just project forward demand
over the next three years relative to supply one of the dismaying predictions you can do is what's going to what's going to happen to the cost of of
inference in the core over the next three years and like it may rise dramatically right like so so what is and then as you know like the the big model
companies are subsidizing heavily right now right and so so what's the
know three years from now to do all the things that they want to do And I I don't know what it's going to make. I mean, I have you guys probably have
friends I have friends today who are paying $1,000 a day for OpenClaw for claw tokens to run OpenClaw, right? And so, okay, $30,000 a month, right
? And and by the way, those friends have like a thousand more ideas of the things that they want their claw to do, right? And so, you could imagine th
ere there's like latent demand of up to, I don't know, five or $10,000 a day of of tokens for a fully deployed, you know, p personal agent. And obviou
sly, consumers can't pay that, right? And so, so but it gives you a sense of the of the f of the future scope of demand, right? And so so even even if
there's a 10x improvement in price performance that still you know goes to $100 a day which is still way beyond what people can pay.
然后在此基础上,我们刚刚得到了 agent 突破——OpenClaude——太棒了,太不可思议了,威力巨大。然后我们刚得到了自主研究——自我改进的突破,现在已经进入了自我改进突破阶段。所以我的思路是,我们已经有了四个根本性的功能突破:LLM、推理、agent,以及现在的 RSI(递归自我改进)。它们实际上都在起作用。所以我——你们看得出来——我已经兴奋得快跳起来了 [笑]——就是这个了。这就是 80 年研究的结晶,这就是它变为现实的时刻。
st going to be like ferocious demand. By the way the agent thing the other interesting thing is I think the agent thing so up until now a lot of the c
onstraints have GPU constraints. I think the agent thing now also translates into CPU constraints CPU and memory. Yes.
d so like the entire chip ecosystem is just going to get wait for network constraints. That would be the killer.
ly for years. And so so I I think that Brad and I think it's actually possible. I mean generally inference costs are going to keep coming down but I t
hink the let's put it this way the rate of decline I think may level out here for a bit because of these supply constraints and then at some point may
be the lab stops subsidizing so much and that that that again will be be an issue and so there's just going to be so much more demand for inference th
an than can be satisfied um you know kind of with the centralized model and then and then you you guys know this but like all the just the dramatic I
mean just the dramatic innovations that have happened in the Apple silicon to be able to do uh inference is is quite amazing the level of effort being
put like the open source guys are putting incredible effort into getting you know this recurring pattern where the big model will never run on a PC a
nd then 6 months later it runs on a PC, right? It's like amazing and there's very smart people working on that. So there's all that and then look ther
e's also you know there's also like other there's other motivators there's other motivators which is just like okay how much trust are the big central
ized model providers you know how much trust are they building in the market versus you know how much are you know at least for in certain cases with
some people for certain use cases people being like well I'm not willing to just like turn everything over. So there there there's all the trust is
sues. Um, by the way, there's also just like straight up price optimization. There's many uses of AI where you don't need Einstein in the cloud. You j
ust need like a a smart local model. There's also performance issues where you want to, you know, you want, you know, you're going to want your dooror
knob to have an AI model in it,
with a chip is going to have an AI model in it. And a lot of those are going to be local. Um, and so yeah, no, like I think I think you're going t
o have t and then you're by the way, also wearable devices, you know, you don't want to do a complete round trip. you want, you know, your whatever yo
ur smart devices are, you want it to be like super low latency.
collapse of AI2, the Allen Institute, one of the actual American open source model labs.
e. Like you guys invested in Mistral and Mrol is doing extremely well outside of China. That's about it. Yeah, we'll see. We'll see. I look, number
one, I do think we care. I do think we I do think we care who makes it. Um I would say this the the the previous presidential administration wanted t
o kill it in the US. Like they wanted to drown in the bathtub. Um [snorts] and so they wanted to kill it. So at least we have a government now that ac
tually like actually wants it wants it to happen.
这就是缩放定律。然后任何缩放定律——包括摩尔定律和 AI 缩放定律——它们并不是真正的"定律",它们是预测,但当它们起作用时,它们会变成自我实现的预测,因为它们设定了一个基准,然后整个行业——所有聪明人——都会努力确保它真的发生。所以它们在某种意义上激励了保持趋势所需的突破。在芯片领域这持续了 50 年。太了不起了。在某些芯片领域这仍在发生。我认为同样的事情正在发生在 AI 的核心缩放定律上。它们不是真正的定律,但它们是预测,是研究工作和投资的催化剂。
会有复杂性,会有波动,会有看起来快要撞上的墙,然后工程师们会去想办法突破那堵墙——显然这一直在发生。会有看起来定律已经耗尽的时候,然后它们又会重新加速、激增。然后 AI 似乎已经出现了多个缩放定律,多个改进方向。我不知道还有多少尚未被发现的,但可能还有一些。比如,世界模型和机器人技术方面可能存在某种缩放定律——在现实世界中大规模获取数据——这一点我们还没有完全理解。那个可能在某个时候启动。有很多非常聪明的人在研究这个。所以,我认为预期是缩放定律总体上会继续,改进的速度会持续很快。
关于你问的该做什么——我完全相信缩放定律会继续,完全相信能力会持续突飞猛进。我和"AI 纯粹主义者"有一点分歧——我把他们描述为在很多方面是这个领域最聪明的人,但也是那些一辈子都待在实验室里 [哼了一声],对外部世界经验很少的人。我想补充的微妙之处是:外面那个 80 亿人的世界——各种机构、政府、公司、经济系统、社会系统——真的非常复杂。80 亿人在地球上做集体决策不是一个简单的过程。你现在就能看到——一堆 AI CEO 都有这种东西,他们在公开场合说话时就好像"哦,有这些很明显的事情社会需要去做" [哼了一声],然后他们会说"社会没有在做这些事情",然后就"社会怎么可能看不到 XYZ 呢?"——答案是,首先没有一个单一的"社会",是 80 亿人,他们都有发言权,归根结底他们都有投票权来决定如何应对变化,然后人类现实就是非常复杂和混乱的。
所以对你问题的具体回答是——像往常一样,取决于情况。毫无疑问会有公司——已经在发生了——有些公司以为自己在模型之上构建了价值,然后就被下一个模型给干翻了。毫无疑问这在发生。但我认为毫无疑问的是,任何技术要适应现实世界——那个混乱的人类世界——这个过程就是会很混乱和复杂的。不会简单和直接。会很混乱、很复杂,会有大量的公司、大量的产品,甚至是整个行业被建立起来,来帮助所有这些技术真正触达普通人。[哼了一声]
other political issues people have which are many you know this administration has I think a very enlightened view and in particular an enlightened v
iew on AI and in particular on open source AI. Uh and so they're very supportive. Um my read is the chi the Chinese have a very the various Chinese co
mpanies have a very specific reason to do open source which is that they they don't fundamentally they don't think they can sell commercial AI outside
of China right now or at least specifically not not in the US for a combination of reasons and so they they kind of view I think open source AI as a
bit of a loss leader against basically domestic uh you know paid paid services and then kind of you know kind of an ancillary products you know they'r
e they're very excited about it by the way I think it's great I think it's great that they're doing it um you know I think DeepS was like a gift to th
e world um I The great thing about open source open source the the the impact of open source is felt two ways. One is you you get the software for fre
e but the other is you get to learn how it works right and so like
r example I thought this was amazing so open comes out with 01 and it's an amazing technical breakthrough and it's just like absolutely fantastic but
of course they don't explain how it works in detail and then of course they hide the they hide the reasoning traces right [snorts] and and then and th
en and then everybody's like okay this is great but like who's going to be able to replicate this? Are other people going to be able to do this? You k
now, is there secret sauce in there? And then our one comes out and it's just like there's the code and there's the paper and now the whole world know
s how to do it. And then, you know, 3 months later, every other AI model is is adding reasoning. And so, so you get this kind of double like even if t
he Chinese models themselves are not the models that get used, the education that's taken place to the rest of the world, the information diffusion, y
ou know, is incredibly powerful. So, that happens and then I don't know, we'll we'll see, you know, there are a bunch of American, you know, open sour
ce, you know, AI model companies. I mean, look, there's going to be tremendous, you know, there already is. There's, you know, there's going to be tre
, there's tremendous competition, uh, among the primary model companies. You know, there's, depending on how you count, there's like four or five, you
know, big co- model companies now that are, you know, kind of neck andneck in different ways. Um, uh, you know, and and, um, you know, and then obvio
usly both Boax and then Metaware involved, you know, both have huge, you know, huge attempts to, you know, kind to kind of leave frog underway. And th
en you've got, you know, a whole fleet of startups, new companies, including a whole bunch that we're backing that are, you know, trying to come out w
ith different approaches. And then you've got whatever it is. I don't know how how many how many like mainline foundation model companies are there in
China at this point. It's probably six. It's five tigers is what they call it. Uh Quinn is in questionable because there's change in leadership, >
> right? Yeah.
ance and Bite Dance would be like the next tier. They weren't as prominent. They won't have a Yeah. But at least you know C seed dance is very insp
iring and presumably they have more stuff coming and Tencent probably has more stuff coming and so forth and so so like look here would be a thing you
could anticipate which is there are not these markets there are not going to be between the US and China right now there's like a dozen primary finan
cial model companies that are like at scale at some level of like critical mass it's not going to be a dozen in three years right like just because th
ese industries don't bear a dozen it's going to be three you know there's going to be three or four big winners or maybe one or two big winners and so
there's going to be like a whole bunch of those guys that are going to have to figure out alternate strategies um and I think like open source is one
of those strategies. And so I I think you could see like a whole I I think the questions like who's going to do open source I think that could change
really fast. I I think that that that's a very dynamic thing. I think it's very hard to predict what happens and and I think it's very important.
thing in business strategy which is called commoditiz the compliment and it's right and so if you're Jensen it's just kind of obvious of course you wa
nt to commoditize the software and he's and to his enormous credit he's putting enormous resources behind that and so maybe maybe it's literally Nvidi
a and I think that would be great.
onference soon and I got both of them. They got us. They got us.
Yeah, he was in Vienna. Oh, he was in Vienna. And then where is he now?
顺便说一个微妙的点——互联网公司本身从来没有太多债务,因为科技公司通常不靠债务运营。但电信公司靠债务运营——实体基础设施公司靠债务运营。所以像 Global Crossing 这样的公司不仅融了大量股本,还借了大量债务。所以它们是高杠杆的。然后你就看到了——高杠杆、产能过度建设、需求在增长但没有预期那么快——然后轰一声,破产。然后就像酒店行业有句话说的——总是第三任业主才赚钱。得先破产两次,洗掉所有过度乐观的狂热之后,才能到达稳定状态,然后赚钱。
o. And then Yeah. The pie guy, right? The pie guys are European.
警示性的教训是:过度建设是可能发生的。你会陷入这种情况——所有有任何机构资本的人都说"哇,我不知道怎么投资那些疯狂的软件公司,但我肯定可以建数据中心,肯定可以买 GPU,可以部署计算集群和所有这些东西。" 如果你是悲观主义者,你看看这个就会说"哇,这简直就是要复制 2000 年那次崩盘。" 显然那会很糟糕。
我同意的反驳论点是这样的:第一,现在投入资金的公司是最顶级的蓝筹公司。在互联网时代,Global Crossing 就是一个创业者、一家新创公司。但现在大规模部署资金的是 Microsoft、Amazon、Google、Facebook、Nvidia——这些是非常大规模的公司,有大量的现金和从未使用过的债务能力。所以这个过程比当时要机构化得多。
然后第二点是——至少目前为止,每一美元投入到产生运行中 GPU 的东西上,都在立刻转化为收入。你们知道的——每个人都在饥渴地抢算力、抢计算容量,还有所有相关的东西——内存、互连、数据中心空间。所以现在每一美元投到地上都在转化为收入。事实上我还认为有一个有趣的事情正在发生——因为每个人都缺容量,我们今天实际上能用的模型,是劣化版本。
eah, they haven't announced yet any sort of change changed or have they? No, they're they have a company there.
, yeah. Good. Anyways, I think pi and openclaw are very important software things and and I just wanted you to just go off on what you think.
ah. So I think in the combination of the two of them I think is one of the 10 most important software. Open claw got all the attention but talk abo
ut pi pi's kind of the end. Yeah pi pi is kind of the architectural breakthrough for those of us who are older. There was this whole thing that was ve
ry important in the world of software basically from like 1970 to I don't know it still is very important but like 19 from 1970 through to like basica
lly the creation of Linux which was basically this this thing used to call like the Unix mindset like so so because there were all these different you
know theories there all these different operating systems and mainframes and and then you know all these windows and Mac and all these things and the
n there was this but kind of behind it all was this idea of kind of the Unix mindset and the Unix mindset was this thing where basically you don't hav
e these like like in the old days like like the operating system that like made the computer industry really work like in the 1960s was this thing cal
led OS 360 which was this big operating system that IBM developed that was supposed to basically run everything and it was this like giant monolithic
architecture in the sky. It was like a you know it was like a giant castle um of software and and by the way it worked really well and they were very
successful with it but like it was this huge castle in the sky but it was this thing it was almost unapproachable which is like you had to be kind of
inside IBM or very close to IBM and you had to really understand every aspect how the system worked and then the Unix guys originally out of AT&T and
then out of out of Berkeley um you know came out and they said no let's have a completely different architecture and the way architecture is going to
work is we're going to have we're going to have a prompt and a and a shell and then and then we're going all the functionality is going to be in the f
orm of these discrete modules and then you're going to be able to chain the modules together and so the it's almost like the operating it operating sy
stem itself is going to be a programming language. Um and then that le led to the the the sort of centrality of the shell. Um and then that led to sor
t of you know basically chaining together Unix tools and then that led to the emergence of these these scripting languages like Pearl where you could
basically kind of very easily do this and then the shells got more sophisticated and then and then and then look like you know that that number one th
at worked and that that was the world I grew up in like I was I was a Unix guy you know sort of from call it 1988 to you know kind of all the way thro
ugh my work and it worked really well. it's in the background. Um, you know, nor normal people don't need to didn't need to necessarily know about it,
but like if you were doing like system architecture, application development, you you knew all about it. Um, and then, you know, it's been in the bac
kground ever since. And, you know, look, your Mac still has a Unix shell, you know, kind of in there and your iPhone still has a Unix shell kind of bu
ried in there somewhere. So, they're kind of in there. And then, you know, the Windows shell is kind of a, you know, sort of a weird derivative of tha
t. But, um, you know, but look, the the internet runs on Unix. Um, and then smartphones. Actually, both iOS and Android are Unix derivatives. And so y
ou know kind of Unix did end up winning but but anyway and then we just started taking that for granted and then and then so so basically the way I th
ink about what happened with pi and then with openclaw is basically what those guys figured out is I always say the great breakthroughs are obvious in
retrospect right which is
there is a like a real conceptual leap but then you look at it sort of the backwards looking and you're just like oh of course
e always the best breakthroughs. Well actually language models themselves are like that. It's just like, oh, next token completion. Oh, of course. Yeah. What other objective mattered?
ly did it, right? And so the conceptual breakthrough is real and deep and powerful and very important. And so the way I think about pi and openclaw is
it's basically marrying the the language model mindset to the to the Unix basically shell prompt mindset. And so it's it's basically this idea that w
hat what so what is an agent, right? And and as you know like many smart people have been trying to figure out what an agent is for for for decades an
d they've had many architectures to build agents in the whole thing. And it turns out what is an agent? So it turns out what we now know is an agent i
s the following. It's so it's a language model. And then above that it's a bash it's a bash shell. So it's it's a Unix shell and then it's and then th
e agent has access has access to to the shell in you know hopeull hope hopefully in a sandbox maybe maybe in a sandbox. So it's it's the model um it's
the shell um and then it's a it's a file system. Um and then the state is stored in files and then you know there's the markdown format for the you k
now for for the files themselves and then and then there's basically what in Unix is called a crown job. There's a loop and then there's a heartbeat f
or there's heartbeat and and the thing basically wakes up wakes up. So it's basically LLM plus shell plus file system plus markdown plus cron and i
t turns out that's an agent and and and every part of that other than the model is something that we already completely know and understand. And in fa
ct it turns out that like the latent power of the Unix shell is like extraordinary because basically like all like there's just like there's just enor
mous latent power in the shell. There's enormous numbers of Unix commands. There's enormous number of command line interfaces into all kinds of things
already in the you know your entire I mean your entire just to start with your computer runs in a shell if you're running a Mac or or a phone your co
mputer your computer's running on a shell uh already and so like the full power of your computer is available at the command line level um and then it
turns out it's really easy to expose other functions as a command line interface and so like this whole idea where we need like MCP and these like pr
o fancy protocols whatever it's like no we don't we just need like a command command line thing so that's the architecture and then it turns out what
is your agent your agent is a bunch of files stored in a file system. And then there's the thing that just like completely blew my mind when I wrote m
y head around it as a result of this, which is like, okay, this means your agent is now actually independent of the model that it's running on because
you can actually swap out a different LLM underneath your agent and your your agent will change personality somewhat because the model is different,
but all of the state stored in the files will be retained,
我知道的是这样的:在未来 3 到 4 年内,基本上所有产能都卖光了。整个供应链已经卖光或正在卖光。所以我们在接下来几年内就是会持续面临慢性供给短缺。市场会有反应,会带来巨额投资涌入新的晶圆厂产能等等——这已经在发生了。在某个时候供应链瓶颈会解锁,至少某种程度上会——那将成为行业增长的另一个加速器,因为产品会更好,所有东西都会变便宜。我知道部署场景、实际用例确实非常引人注目,加上推理和 agent 等等,我知道它们只会从这里变得好得多。我也知道技术进步不会停下来。它在加速——这些突破非常了不起。就算是逐月来看,突破都非常惊人。
所以,我觉得如果你是一个愤世嫉俗的人,你可以回头看 2000 年,能找到相似之处,但我根本无法想象去押注说这会在未来几年以某种方式令人失望。我觉得做那个押注基本上等于自杀。
ht. It's like, right, swapping out a ship and recompiling, but it's it's still it's still your agent with all of its memories um and with all of its c
apabilities. And then by the way you can also swap out the shell. Uh so you can move it to a different execution environment that is also is also a ba
sh shell. By the way you can also switch out the file system right uh and you can and you can and you can swap out the the the heartbeat the crown fra
mework the loop the agent framework itself. And so your agent basically is basically at the end of the day it's just it's just its files. Um and then
and then there's of course yeah it's it's basically it's just the files. [laughter] Um and then by the way as a consequence of that the agent it and t
hen the agent itself it turns out a couple important things. So one is it it's it can migrate itself, right? And so you're you can instruct your agent
migrate yourself to a different uh runtime environment, migrate yourself to a different file system, migrate yourself to a different you swap out the
language model. Your agent will do all that stuff for you. And then there's the final thing which is just amazing which is the agent is the agent act
ually has full introspection. It actually it actually knows about its own files and it can rewrite its own files,
lly no widely deployed software system in history where the the the thing that you're using actually has full introspective knowledge of how it itself
works and is able to modify itself like that that I mean there have been toy systems that have had that but there there's never been a widely deploye
d system that has that capability and then that leads you to the capability that just like completely blew my mind when I wrapped my head around it
which is you can tell the agent to add new functions and features to itself and it can do that
d yourself give yourself a new capability right and so and so literally it's It's like you run into somebody at a party and they're like, "Oh, I have
my open claw do whatever. Connect to my eight sleep bed and it gives me better advice and sleep." And you go home at night and you tell your claw or i
f they're at the party, by the way, you tell your claw, "Oh, add this capability to yourself and your claw will say, "Oh, okay, no problem." And it'll
go out on the internet and it'll figure out whatever it needs and then it'll go out to cloud code or whatever. It'll write whatever it needs and then
the next thing you know, it has this new capability and so you don't even have to like you can have it upgrade itself without even having to without
having to do anything other than tell it that you want it to do that. And so anyway, so the combination of all this is just I mean this is just like a
massive incredible I mean it's just incredible. Like if I if I were if I were 18 like this is 100 this is what I would be spending all of my time on.
This is like such an incredible conceptual breakthrough. And again people are going to look at it and they already get this response. People are goin
g to look at it and they're going to say oh where's the breakthrough because these the all of these components were already known before. But but this
is the key the key to the breakthrough was by using all these components that were known before you get all of the underlying capability that's burie
d in there. And so all and so for example, computer use all of a sudden just kind of falls trivial trivial. Of course, it's going to be able to use yo
ur computer. It [snorts] has full access to the shell, right? And then and then you just you give it access to a browser and then you've got the compu
ter and the browser and off and away it goes. And and then you've got all the abilities of the browser also. Um and so and so the capability unlock he
re is profound. My friends who are, you know, deepest into this are having their claw do like like literally like a thousand things in their lives. Th
ey have new ideas every day. They're just like constantly throwing new challenges at the thing. And by the way, it's early and you know, these are, yo
u know, these are prototypes and there's, you know, as you guys know, there's security issues and and so, you know, there's a bunch of stuff to be iro
ned out, but the the unlock of capability is just incredible.
have at least, you know, an agent like this, if not an entire family of agents, and we're going to be living in a world where I think it's almost i
nevitable now that this is the way people are going to use computers.
he next step is your claw talking to my claw. posting on claw Facebook uh posting their jobs on claw LinkedIn and post posting their tweets on claw xi
or whatever you know [laughter] um I do think that that is how uh you know we we get into some danger there in terms of like alignment and whether or
not we want these things to to to run you guys know rent reentum.com
flipped
thoughts on the engineering side. So when you build the browser, the internet, you know, just a bunch of mostly plain text file plus some images and
today the every website and app is like so complex and like somehow, you know, the browser kept evolving to fit that in. Are there any design choices
that were made like early in the browser and kind of like the internet and the protocols that you're seeing agents similar to today's like hey this th
ing is just not going to work for like this type of new compute and we should just rip it out right now.
you a couple. So one is um and we didn't you know to be clear like this this was not you know this was totally different. We didn't have the capabilit
ies we have today but because we have we didn't have the language models underneath this but um we did have this idea that human readability actually
mattered a great deal. Um and and so and specifically in those days it was it was not so much English language but it was there there was a design dec
ision to be made between binary protocols and text protocols. And basically every every every basically old school systems architect that had grown up
between like the 1960s and the 1990s basically said you know the internet is what do you know about the internet? It's star for bandwidth. You you ju
st you have these very narrow straws. You know look people when we did the work on mosaic like people who had the internet at home had a 14 kilobit mo
dem right? So you're you're trying to like hyper optimize every bit of data that that travels over the network. And so obviously if you're going to de
sign a protocol like HTTP, you're going to want it to be binary, you know, highly compressed binary protocol for maximum efficiency. And you're going
to want to have it be like a single connection that persists and you're the last thing you're going to want to do is like bring up and tear down new c
onnections. And you definitely you're not not going to want a text protocol. And so of course we said no, we actually want to go completely the other
direction. It's obviously we only want text protocols. U by the way, same thing in HTML itself. We want HTML to be relatively verbose. you know, we wa
nt the tags to actually be like human readable. Um, we want to use the most inefficient things possible.
nefficient things. You're the original token maxer.
this was actually the the conscious thing which basically says just like assume assume a future of infinite infinite bandwidth, build for that. And t
hen basically what it was is it was a bet that it was a bet that if the system was if the if the latent capabilities of the system were powerful enoug
h and that was obvious enough to people that would create the demand for the bandwidth that would cause the supply of bandwidth to get built that woul
d actually make the whole thing work. And then specifically what we wanted was we wanted everything to be human readable because we at the engineering
level we wanted people to be able to read the protocol coming over the wire and be able to understand it with their with their bare eyes without havi
ng to like disassemble it or whatever, right? Have it converted out of binary, right? And so the the the all the pro, you know, HTTP and everything el
se were were it was always text protocols. And the same thing with HTML and and in many ways some people say that the key breakthrough in the browser
was the view source option um which is every web page you go to you could view source which means you could see how it worked which means you could te
ach yourself how to build right new uh to to build new web pages there was that so human readability um and again human readability in those days stil
l meant technical you know specs you know now it means English language but that there's an incredible latent power in giving everybody who uses the s
ystem the option to be able to drop down and actually understand how it's working and that worked really well for the web and I think it's working rea
lly well for AI That was one. Um what was the other? Um a big part of the idea of web servers was to actually surface the underlying latent capability
of the operating system and to be able to surface the also the underlying latent capability of the database because basically what was a web server?
What what what is a web server fundamentally architecturally? It's it's it's the operating system. So it's it's the operating system's ability to you
know it's running on top of an OS. So it's the OS's ability to manage the file system and do everything else that you want to do and process everythin
g. Um and then of course a lot of early you know lot of websites are front ends to databases. Um and so you wanted to you wanted to unleash the underl
ying latent power of whether it was an Oracle database or some other you know some other Postgress or whatever whatever it was. Um and so a lot of the
function of the web server was to just bridge from that internet connection coming in to be able to unlock the underlying power of the OS and the dat
abase. Uh and again people looked at it at the time and they were like well is this really does this really matter? Like is this important because we'
ve had databases forever and we've always had you know user interfaces for databases and this is just another user interface for a database. It's like
okay yeah fair enough. But on the other side of that, it's just like this is now a much better interface to databases and one that eight billion peop
le are going to use and is going to be like far easier to use and far more flexible and and and and you're not just going to have old databases. Now y
ou have a system where people can actually understand why they want to build, you know, a million times more database apps than they had in the past.
And then the number of databases in the world exploded. And so again, this goes to this thing of like building building in layers. Some of the smartes
t people in the industry look at any new challenge and they're like, "Okay, I'm I need to build a new kind of application. So the first thing I need t
o do is build a new programming language, right? Right? And then the next thing I need to do is build a new operating system, [laughter] right? And th
en the next thing I need to do is I need to build a new chip, right? And they they kind of want to reinvent everything. And I've I've always had maybe
it's just I don't know pragmatic mentality or something or maybe an engineering over science mentality, but it's more like no, you have just like all
of this latent power uh in the existing systems and you don't want to be held back by their constraints, but what you want to do is you want to kind
of liberate that power and open it up. Yeah. And so I I think I think and I think the web did that for those reasons and I think it's the same thin
g now that's happening.
那么三年后一般人每天、每月的 token 成本会是多少?我不知道会怎样。我有朋友——你们大概也有——现在每天花 1000 美元在 Claude token 上跑 OpenClaude——
dcast and we were talking about Rust and you know Rust is memory safe by default. So why are we teaching the model to not write memory unsafe code jus
t use Rust and then you get it for free. How much do you think there's like time to be spent like recreating some of these things instead of taking th
em for granted or be like oh okay Python it's kind of slow on TypeScript
frana.
所以需求会非常疯狂。顺便说,agent 这个事情——到目前为止很多瓶颈在 GPU。但我认为 agent 现在也会转化为 CPU 瓶颈——
to be good at programming in every language. And I think they're going to be good at translating from any language to any other language. Like, okay,
so this gets into the coding side of things. I I think we're going through a really fundamental change. And I look, I I grew up, you know, I grew up
handcode, you know, I grew up hand coding. Everything I did was actually everything I did actually was written in C. I was back in the days
sn't even using C++ so or like Java or any of this stuff, right? Uh and so um everything everything I ever did, I was like managing my own memory at t
he level of C. And then I, you know, I'm still from the generation that, you know, I knew assembly language and, you know, I I, you know, um, so I I c
ould drop down and do things uh, right on the ship. And so we we've just we've all all of us we've always lived in a world in which software is like t
his precious thing that like you have to think about very carefully and it's like really hard to generate good software and there's only a small numbe
r of people who can do it and like you have to be very like jealous in terms of thinking about like how do you allocate like what are your engineers w
orking on and how many good engineers do you actually have and how much software can they write and how can how much software can human beings you
know kind of maintain and I think like all those assumptions are being shot right out the window right now like I think they're I I think those days a
re just over and I think the new world is like actually high quality software is just infinitely available
do XYZ like you're just gonna wave your hand and you're gonna get it and then if it's if you don't like the language it's written in you just tell th
e thing all right I want the now I want the rush version um or you know security you know sec we're about to by the way we're about to go through comp
uter security is about to go through the most dramatic change ever which is number one like every single latent security bug is about to be exposed ri
ght so we're going to have like the we're we're set up here for like the computer security apocalypse for a while but but on the other side of it n
ow we have coding agents that can go in and actually fix all the security bugs and so How how are you going to secure software in the future? You're g
oing to tell the tell the bot to secure it and it's going to go through and and fix it all. And so so this thing that was this incredibly scarce resou
rce of high quality software is just going to become a completely fungeible thing that you're just going to have as much as you want, right? And and t
hat has like you know that has like tons and tons of consequences.
t I don't know simple or something or straightforward which is just if you want all your software in rest, you just tell the bot you want all your sof
tware in rest. like things that used to be like hard or even like seem like an insurmountable mountain to get to get through all of a sudden I think b
ecome very easy. I think Brett had a theory that there would be a more optimal language for LMS and so the contention is uh there isn't like just d
on't bother just whatever humans already use LLMs are perfectly capable porting.
然后还有其他动因——比如大型集中式模型提供商到底建立了多少信任?至少在某些情况下,某些人对某些用例会说"我不愿意把一切都交出去"。还有信任问题。还有就是纯粹的价格优化——很多 AI 的使用场景不需要云端的爱因斯坦,只需要一个聪明的本地模型就行了。还有性能问题——你希望你的门把手里有一个 AI 模型来做访问控制。显然,所有带芯片的东西都会有 AI 模型在里面,而且很多都会是本地的。然后还有可穿戴设备——你不想做一个完整的网络往返,你希望你的智能设备是超低延迟的。
k today. I think we're pretty close to being able to ask the AI what would its opt optimal language be and let and let it design it. [laughter] Tru
e. Okay. Here's a question. Are you even going to have programming languages in the future? Or a just going to be emitting binaries? Let's assume for
a moment that humans aren't coding anymore. Let's assume it's all bots. The what levels of intermediate abstraction do the bots even need?
a language model now that actually emits model weights for a new language model, [laughter]
ah. Will the bots literally be emitting not just coding binaries but will they will will they actually be admitting weights for for new models direct
directly and
le very inefficient you're basically inefficient simulation of a simulation in a simulation inside of weights yeah very inefficient but like looks
are already like incredibly inefficient ask I have a favorite thing ask 2 plus 2 equals four right it's just like you know it's like you know it's it'
s like whatever billions and billions of times more inefficient than using your pocket calculator.
开源的影响体现在两方面。一是你免费获得软件。另一个是你能了解它是怎么工作的——论文和代码。比如我觉得这件事太了不起了——OpenAI 发布了 o1,是一个惊人的技术突破,非常了不起,但他们当然不详细解释它怎么工作的,还隐藏了推理链(reasoning traces)[哼了一声]。然后所有人都在说"好的这很好,但谁能复现这个?其他人能做到吗?里面有秘密配方吗?" 然后 R1 出来了——代码和论文都在那里,现在全世界都知道怎么做了。然后 3 个月后,每一个 AI 模型都在加推理功能。所以你得到了这种双重效应——就算中国模型本身不是最终被使用的模型,对世界其他地方的教育、信息扩散,是极其强大的。
然后我不知道——我们会看到的。有一堆美国的开源 AI 模型公司。看,会有巨大的竞争——已经有了。根据你怎么算,现在大概有四五家大的模型公司,在不同方面基本上不相上下。然后显然 Meta 也参与其中,都在做大规模的赶超尝试。然后你有一整支创业公司舰队——新公司,包括我们支持的很多家,在尝试不同的方法。然后中国有多少主流基础模型公司?大概六家。
neral capability. And so anyway, like I I kind of think in 10 years like I'm not sure yeah like I'm not sure there will even be a salient concept of a
programming language um in the way that we understand it today and in fact what we may be doing more and more as a form of interpretability which is
we're trying to understand why the bots have decided to uh structure uh code in the way that they have. I mean if you play it through you don't nee
d browsers then like that's the death of the browser. Well, so I I would take it a step further, which is you may not need user interfaces.
is going to use software in the future? Other bots.
in and out.
er you want. Exactly. [laughter] Isn't that better?
ow, I don't know. Look, like you know, you know all the arguments here.
ow.
ke much better ways for people to spend time than plowing fields.
nds. And look, and I'm not an absolutist and I'm not a utopian and and to be clear, like I I have an 11-year-old and he's learning how to code and lik
e I'm, you know, I think it's still a really good idea to learn how to code and so forth, but I just if you project forward, you just have to think fo
rward to a world in which it's just like, okay, I'm just going to tell the thing what I need and it's going to do it
o do it in whatever way is most optimal for it to do it. Unless I tell it to do it nonoptimally, like if I tell it to do it in Java or in Rust or what
ever, it'll do it, I'm sure. But like, if I'm just going to tell it to do, it's going to do it in whatever way is like the optimal way to do it. And t
hen I and then if I need to understand how it works, I'm going to ask it to explain to me how it works, right? And so it's going to be doing its own i
nterpret. It's going to be the engine of interpretability to explain itself. And I I just am not convinced that that I'm not I'm not convinced that in
that world you have these historical the goals of the abstractions will be whatever the boss need with you, right? Yeah.
curious like if that's true then shouldn't the models providers be building some internal language representation that they can do extreme kind of lik
e RL uh and reward modeling around because it's like today they're kind of like tied to like TypeScript and Python because the users need to write in
that language versus they can have their own thing internally and like they don't need to teach it to anybody they just need to teach their model a
nd I think that's how you get maybe diversion between the models like going back to like the pi open cloud thing it's Oh, I built all the software usi
ng the OpenAI model and now switch to the entropic model, but the entropic model doesn't understand the thing. So, I it feels like there still needs t
o be some obstruction. But maybe not. Maybe that's the lock in that the model providers want to have. I don't
gh cuz why can't the second model just learn what the first model has done? Like Exactly.
ls can now reverse engineer software B, right? Isn't it the whole thing now where people are reverse engineering like Nintendo game binaries? Yeah.
So you you have like I've seen a bunch of reports like this where somebody has like a favorite game from the 1980s
g dead but they have like a binary burned into a chip or something and now they're reverse engineered to get a version that runs on their Mac. Right.
And so if you reverse it this is why I kind of say if you're reversing like x86 binaries then why can't you reverse engineer
And because we're on a Unix based system it has to be reversible because it needs to run on the target. Yeah. Yeah. Yeah. Yeah. Yeah. Basically. An
d so I just I just think it's this thing where it's just like and by the way and everything we're describing is something that human beings in theory
could have done before but just with like but with enormous where but it was just always like cost and labor prohibitive. Reverse engineer like I lear
ned how to reverse engineer
t but now with the model you don't. And so all of a sudden you get you get these things or another way to think about it is so much of human built sys
tems sort of compensate for the human limitations. Yep.
have and it's not that you you won't have abstractions, but you'll have a different kind of abstraction. Yeah.
然后 Unix 的人——最早来自 AT&T,后来来自 Berkeley——出来说"不,我们要一个完全不同的架构"。这个架构的工作方式是:我们会有一个提示符(prompt)和一个 shell,然后所有功能都以离散模块的形式存在,然后你可以把模块串联起来。所以操作系统本身几乎就是一种编程语言。然后这就带来了 shell 的核心地位,然后是 Unix 工具的串联,然后出现了像 Perl 这样的脚本语言,shell 变得更加强大——这套东西是管用的。那就是我成长的世界——从大约 1988 年起我就是一个 Unix 人。它工作得非常好。普通人不需要知道它,但如果你在做系统架构、应用开发,你就完全了解它。然后它一直在后台运行。你的 Mac 里还有一个 Unix shell。你的 iPhone 里还埋着一个 Unix shell。Windows 的 shell 也是某种奇怪的衍生物。互联网跑在 Unix 上。智能手机——iOS 和 Android 都是 Unix 的衍生物。所以 Unix 最终赢了。但无论如何,我们后来就把这些当成了理所当然。
然后——基本上,我对 Pi 和 OpenClaude 发生的事情的理解是:那些人想明白的事情——我总说,最伟大的突破在事后看来都是显而易见的——
a close and you can pick whichever one. So, just talking about protocols, was it you or someone else? I forget my internet history who said that like
the biggest mistake that we didn't figure out in the early days was payments. Yes.
chance now. I don't think we're going to figure it out. I don't know. Like what's your take?
ng to happen for sure. Yeah.
我对 Pi 和 OpenClaude 的理解是:它基本上是把语言模型思维和 Unix shell/提示符思维结合到了一起。所以它的核心思想是:什么是 agent?很多聪明人几十年来一直在试图搞清楚什么是 agent,他们有各种架构来构建 agent。结果是什么?我们现在知道 agent 是这样的:它是一个语言模型。在它上面是一个 bash shell——一个 Unix shell。然后 agent 可以访问 shell,希望是在沙箱里。所以就是:模型,加上 shell,加上文件系统。状态存储在文件中。然后有 markdown 格式用于文件本身。然后基本上就是 Unix 中所谓的 cron job——有一个循环,有一个心跳(heartbeat),它会定期醒来。
所以基本上就是 LLM + shell + 文件系统 + markdown + cron。结果这就是一个 agent。而除了模型之外的每一个部分,都是我们已经完全了解和理解的东西。事实上,Unix shell 的潜在力量是非凡的——因为有大量的 Unix 命令,大量的命令行接口连接到各种东西。你的整个计算机——如果你在跑 Mac 或手机——你的计算机就运行在一个 shell 上,你计算机的全部能力在命令行层面就可以使用。然后把其他功能暴露为命令行接口其实非常容易——所以那些什么 MCP 啊、什么花哨的协议之类的——不,我们不需要那些,我们只需要一个命令行工具。
这就是架构。然后你的 agent 就是一堆存储在文件系统里的文件。然后有一件事彻底让我头脑爆炸——当我想明白之后——这意味着你的 agent 现在实际上独立于它运行的模型,因为你可以在 agent 下面换一个不同的 LLM。你的 agent 会因为模型不同而稍微改变"性格",但存储在文件中的所有状态都会被保留——
he form of crypto stable coins and crypto and this is I I think this is the grand unification basically of AI and crypto is what's about to happen now
. Um I think AI is the crypto killer app I think is where where this is really going to come out. Um and then the other is just it I mean it's just I
think it's now obvious it's like obviously AI agents are going to need money and it's already happening right if you've got a if you got a claw and yo
u want it to buy things for you you have to give it money in some form. I would say the adoption is probably like 0.1% if if that. But yeah. Oh, to
day. Yeah. Yeah. Yeah. But think think forward like where is it going
t I think I we we do is it's the William Gibson quote which is the future is already here. It just isn't distributed isn't isn't distributed yet. M
y friends who are the most aggressive use users of of of OpenClaw just like have given their claws bank accounts
然后作为这个的结果,agent 本身有几个重要特性。第一,它可以自我迁移。你可以指示你的 agent 迁移到另一个运行环境,迁移到另一个文件系统,换掉语言模型。你的 agent 会替你做所有这些。然后最后还有一件太了不起的事——agent 实际上拥有完全的自省能力(full introspection)。它真的知道自己的文件在哪里,而且能重写自己的文件——
nd and not only have they done it, it's obvious that they needed to do it because it's obvious that they needed to be able to spend money on their beh
alf. Yeah. Yeah. It's just completely obvious and so and again like so the number of people who have done that today to your point is like I don't
know probably 5,000 or something but
然后这就引出了一个能力,当我想明白的时候彻底炸了我的脑子——你可以告诉 agent 给自己添加新功能,它能做到。
open cloud by the way if you don't give it a bank account it's just going to break into your [laughter] it's break it's going to break into your bank
account anyway and take your money so you might as you might as well do it you might as well do it
lly love the phenomenon I love the yolo um I'm not doing it myself to be clear but I love the people that are just like what is it dangerously dangero
us which by the way is a Facebook thing.
总之,所有这些的结合就是——这简直太不可思议了。如果我 18 岁的话,这百分之百是我会把所有时间花在上面的东西。这是一个难以置信的概念突破。人们会看着它说"突破在哪里?所有这些组件以前都已经有了。" 但关键就在于——通过使用所有这些已知的组件,你释放了所有埋藏在其中的底层能力。所以比如说,计算机使用(computer use)突然变得微不足道。它当然能用你的电脑——它有完全的 shell 访问权限 [哼了一声]。然后你给它一个浏览器——然后你就有了电脑加浏览器,然后它就可以大展身手了。然后你还有浏览器的所有能力。
所以这里释放的能力是深远的。我那些最深度使用这些东西的朋友们正在让他们的 Claude 在生活中做一千件事。他们每天都有新点子,不断地给它扔新的挑战。当然了,这还很早期,这些是原型,你们知道还有安全问题等各种东西需要解决,但能力的释放是不可思议的。
我毫不怀疑,世界上每个人都会有至少一个这样的 agent,如果不是一整个 agent 家族的话。我们将生活在一个——我认为几乎不可避免的——这就是人们使用电脑的方式的世界里。
erous so that you are aware when you enable the flag that you are opting into a dangerous thing.
f course that makes it enticing. Sam
y see the future is to find the people who are doing that. [laughter] There's a mand, you know, log everything, you know, just watch it. Watch the
logs.
然而我们当然说"不,我们要完全走另一个方向。我们只要文本协议。" 顺便说,HTML 也是一样——我们希望 HTML 相对冗长,我们希望标签是人类可读的。我们希望用——
Let it try everything. Let it unlock everything. By the way, that's how you're going to find all the good stuff it can do. By the way, that's also how
you're going to find all the flaws. I think the people who turn that on for bots are like they're like martyrs to the progress of human civilizati
on. Like I feel very bad for their descendants that their bank accounts are going to get looted by their bots in the first like 20 minutes. But I thin
k the contribution that they're making to the future of our species is amazing. He's like gentleman science, you know.
klin out with a trying to trying trying to get lightning to strike [laughter] his his balloon and see seeing if he gets electrocuted.
nas Sulk with the polio vaccine, [laughter] right? Injecting. Yes. So, yes, I I I I think we should have like a we should have like flags and like we
should have like monuments to [laughter] the people that just let OpenCloud run their lives. More anecdotes or like what what are the craziest or i
nteresting things that people listening to this should go go home and do? I mean, this is this is the this is the the extreme thing is just like the s
traight yolo like just yeah turn your light.
group chat just lit up.
然后具体来说,我们希望一切都是人类可读的,因为在工程层面我们希望人们能直接阅读传输过来的协议内容,用肉眼理解它,而不需要反汇编或者从二进制转换出来。所以 HTTP 和其他一切都是文本协议。HTML 也是一样——有些人说浏览器的关键突破是"查看源代码"(View Source)选项——每个网页你都可以查看源代码,这意味着你能看到它是怎么工作的,这意味着你能自学如何构建新网页。
所以人类可读性——当时的人类可读性仍然意味着技术规格,现在意味着英语(自然语言)——但给予每一个系统用户能够深入了解系统如何运作的选择,这里面蕴含着难以置信的潜在力量。这对 web 非常有效,我认为对 AI 也一样有效。那是第一个。
另一个是什么来着……Web 服务器的一个重要理念是释放操作系统的底层潜在能力,以及释放数据库的底层潜在能力。因为 web 服务器从根本架构上来说是什么?它是操作系统——它运行在操作系统之上,所以是操作系统管理文件系统和做所有其他事情的能力。然后很多早期网站是数据库的前端——你想释放底层的力量,不管是 Oracle 数据库还是 Postgres 还是别的什么。所以 web 服务器的很大一个功能就是从互联网连接桥接到底层操作系统和数据库的力量。
人们当时看着它会说"这真的重要吗?因为我们一直有数据库,一直有数据库的用户界面,这不就是数据库的另一个用户界面吗?" 嗯好吧是的。但从另一面看——这是一个好得多的数据库接口,80 亿人会使用的,远更容易用、远更灵活的。而且你不只是有旧数据库了——现在你有了一个系统,人们能理解为什么他们想构建比过去多一百万倍的数据库应用。然后世界上数据库的数量就爆发了。
这又回到了在层次上构建的理念。行业里一些最聪明的人面对任何新挑战都会说"好的,我要构建一种新应用。所以第一件事是构建一种新的编程语言,然后下一件事是构建一个新的操作系统 [笑],然后下一件事是构建一块新芯片。" 他们想从头发明所有东西。我一直有一种——也许是实用主义的心态,或者是工程重于科学的心态——更像是"不,你已经有了所有这些潜在力量在现有系统中。你不想被它们的限制束缚住,但你想要做的是解放那些力量,把它们开放出来。"
is just absolutely amazing. The number of stories on I'm just don't want to violate people's you know obviously personalize but um
of the things OpenCloud is really good at is hacking into all the stuff in your LAN.
et of Like,
我觉得这个会发生很大变化,因为我不觉得模型在乎用什么语言编程。我觉得它们会擅长用每一种语言编程。我觉得它们会擅长从任何语言翻译到任何其他语言。好,这就涉及到编程方面的事情了。我觉得我们正在经历一个非常根本的变化。看,我成长的时代——所有东西都是我亲手写的 C 代码。我那时候连 C++ 都没用,Java 什么的更不用说了。所有我做过的东西,我都在 C 的层面管理自己的内存。然后我还是会汇编语言(Assembly Language)那一代人,可以直接在芯片上做事情。
我们一直生活在一个软件是这种珍贵东西的世界里——你得非常仔细地思考,生成好的软件非常难,只有少数人能做到,你必须非常精打细算地考虑工程师们在做什么,你有多少好工程师,他们能写多少软件,人类能维护多少软件。我觉得所有这些假设正在被统统打碎。那些日子已经结束了。新世界是:高质量软件实际上是无限供应的。[哼了一声] 如果你需要新软件做 XYZ,你只要挥挥手就能得到。如果你不喜欢它是用什么语言写的,你就告诉它"好,现在我要 Rust 版本"。
顺便说,计算机安全即将经历有史以来最剧烈的变化。第一,每一个潜在的安全漏洞都即将被暴露出来——我们接下来要面对一段计算机安全的末日期。但在另一面,现在我们有了编码 agent,它们可以进去修复所有安全漏洞。未来你要怎么保障软件安全?告诉机器人去保障安全,它会进去把一切都修好。所以这个曾经极其稀缺的高质量软件资源,将变成一种完全可替代的东西,你想要多少就有多少。
这有大量的连锁后果。
he things. And then my my my friends who are most aggressive at this are having OpenClaw take over everything in their house.
heir security cameras. It takes over their their, you know, their whatever their their access control systems. It takes over their webcams. I have a f
riend whose claw watches him sleep. Put a webcam in your bedroom. Put the put the claw put the claw on a loop. Have it wake up frequently and have it
watch just tell it watch me sleep. And and I' I've seen the transcripts and it's literally like Joe's asleep. This is good. This is good that Joe's as
leep because, you know, I have I have his health data and I know that he hasn't been getting enough sleep and so it's really good that he's getting as
leep. I really hope he gets his full whatever, you know, 5 hours of REM sleep. D uh Joe's moving. Joe's moving. Um uh Joe might be waking up. This is
a real If Joe wakes up now, he's going to ruin his sleep cycle. Oh, okay. It's okay. Joe just rolled over. Okay, he's gone back to bed. Okay, good. Al
l right. Okay, I can relax. This is fine.
now is just like very [laughter] focused right it's just like this is like his reason for existence is to watch Joe sleep and then and I was talking t
o my friend who did this like you know on the one hand it's like all right this is weird and creepy um and I need to I need maybe this is taking over
my life and then the other thing is like you know what if I had a heart attack in the middle of the night this thing literally would like freak out an
d call 911 like there's no question this thing would figure out how to like alert medical authorities and like pro probably summon SWAT teams and like
do whatever would be required [laughter] to save my life right And so it's like you know like yeah like that's happening. What else? Um give um th
ere's a company unit that makes the robot dogs. Um then and I actually have one at home which is it's actually really fun with the Chinese companies t
he Chinese companies are so aggressive at adopting uh new technology but they don't always like let's said take the time to really
ge it and maybe think it all the way through and so so the at least the industry dog I have so it it has a old nonLM just control system which by the
way is not very good it markets well but in practice it's not that good it has trouble with stairs and so forth and so it's not quite what it should b
e but then the language model thing comes out the voice so they they add so they add LLM capability and then they add a voice mode to it. Um, but but
that LM capability is not at all connected to the control system. [laughter] So, so you've got this schizophrenic dog that like is a complete idiot wh
en it comes to climbing the stairs, but it will happily teach you quantum mechanics, [laughter] right, in like a plumbing English accent, right? Like
it's just like absolutely amazing intelligence. Yeah. Yeah. Talk about now obviously what's going to happen in the future is is they're going to conne
ct together, but but right now it's it's and so right now it's not that useful. And so I I have a friend who has one of these who had his claw basical
ly hack in and rewrite the code. Write new firmware, write new firmware for the for the unit robot.
kids. You could do that before after like the motion.
whenever there's an issue in the thing now the claw just like rewrites the code, you know, you goes in does does the code and is kind of goes to your
thing here. So so like all of a sudden this is why we want to think about AI code. AI coding is not just like writing new apps. is also going in and
rewriting all the old stuff that should have worked that never worked. And so like I I think I think basically I think the internet the internet of is
basically over like I I think everything there's the potential here where like all these devices in your house that have been like basically marginal
or you know basically dumb you know like all of a sudden they might all get really smart. Now you have smart home
are horror movies in which this is of which this is the premise and so you have to decide if you want this but but but this is the first time I can sa
y with confidence I now know how you could actually have a smart home with 30 different kinds of things with chips and internet access where it actual
ly all makes sense and all works together and it's all coherent and the whole thing and to have that unlock without a human being having to go do any
of that work like yeah I'm waiting for a sorry Mark I can't let you open that fridge door you know Like, [laughter]
e doing, you know, d I think you can do this, but you know, this is a real Are you [laughter] really, you know, are you really sure?
told, you know, you told me last night you really don't want me to let you do this. So, you know, I'm sorry, but the fridge door is locked. [laughter]
can pass the test, I will open the fridge door for you.
t. That's the last piece that we got to figure out.
world right now where we've known these asymmetries exist and we we society have been unwilling to grapple with them and I think they're both tipping
right now and and they're they're they're the same thing. It's the virtual world version is the physical world version. So, the virtual world version
is is the bot problem. We're just like, you know, the internet internet is just like a wash in bots. Internet's a wash in fake people. It has been for
ever. Um, by the way, a lot of that has to do with lack of money, you know, and so this, you know, this is this is this
are the same thing and corporations are people too, you know. So, [laughter] interesting. Yeah. Yeah. Yeah. Okay.
f human.
, look, the bot I mean, every social media user knows this. The bot the bot problem is a big problem. You know, the bot the bot problem has been a big
problem forever. It's it's a huge problem. And it's never really been confronted directly like at any point. By the way, the physical world version o
f this is the drone the drone problem. Um, right. And so we we've known for, you know, we've known for 20 years now that the asymmetric threat both in
mil military in actual military conflict, but also in just like security like like you know, security on the home front, the big threat is is the che
ap attack drone, right? The the cheap the cheap suicide, you know, drone with a bomb. And we've known that forever. And by the way, like, you know, it
's very disconcerting how like every, you know, every office complex in in the c, you know, in the world is like unprotected from drone attacks. um ev
ery every stadium, every school, every prison like it like okay, we've known that we've never done anything about it. Yeah. One possibility is just
leave leave them unprotected forever and live in a world of like asymmetric terrorism forever. Or the other is take the problem seriously and figure
out the set of techniques and technologies required to to be able to deal with that whether those are lasers or jammers or early warning systems or yo
u know
nomic asymmetries. These are economic asymmetries, right? Because it's really cheap to field a bot, but it's very hard to tell something a bot. It's v
ery cheap to field a drone. It's very hard. It's very expensive to defend against a drone. But you see what I'm saying is it's it's the it's the virtu
al version of the problem and it's the physical version of the problem. Uh the virtual version of the problem, what what we need quite literally is pr
oof of human. The reason is because you're you're not going to have proof of bot. The the especially now the bots are too good. The the bots can pass
the touring test. And if the bots can pass the touring test, then you can't you can't screen for bot. You can't have proof of not a bot. But what you
can have is you can have proof of human. You can have, you know, cryptographically validated this is definitely a person and this is and then you can
have cryptographically validated this is definitely like something that a person said. This video is real, right?
do you think Alex Blania with world do you think he's got it or is there an alternative?
think many people will try. We're one of the key you know participants in in the world in the world project and yeah so we're partisans but yeah I I t
hink so we think world is exactly correct and and the reason is it it has it has to be it it has to be proof of human. It it has because you can't do
proof of not bot. You have to do proof of human to do proof of human. You you need you need biological validation. You needed to start with this was a
ctually a person, right? Because otherwise you have bots signing up as fake people, right? So you you have to have like something you have to have a b
io biometric and then you have to have cryptographic validation and then the ability to do to do to do the lookup. And then by the way, the other thin
g you need which they you also need selective disclosure. Um so you need to be able to do proof of human without revealing all the underlying informat
ion. By the way, another thing you need you're going to need proof of age, right? because there's all these laws in all these different countries now
around you need to be 13 or 16 or 18 or whatever to do different things and so you're you're going to need you know sort of validated a proof of age u
m you know to be able to legally operate right and so that that's coming and then you're going to want like proof of credit score and you know proof o
f like you know hundred other that's a tricky one. It it is a tricky one, but you're you're going to there there's no reason like if somebody's che
cking on your credit, somebody shouldn't give you an example. Somebody shouldn't need to know your name in order to be able to find out whether you're
credit worthy, right?
e privacy problem at large, which is I I only need to prove what I need to prove at that moment. So like you're going to need that and I I think their
their architecture makes sense. So that needs to get solved. I think language models have tipped the bots are now too good. uh and and so they're und
etectable. And so as a consequence, we now need to go confront that problem directly. And then like I said, and then the other problem is we we need t
o go actually confront the drone problem. The Ukraine conflict has really unlocked a lot of thinking on that. And now the um and now the the the th
e Iran situation is also unlocking that. And so I think there's going to be just like this incredible explosion of of both drone counter drones.
r drones are better than their drones as long [laughter] as as long as you keep it that way.
不久以前 99% 的人类都在犁后面推犁。
neak in one more question. Um I'm trying to tie together a lot of things that you said over the year. So at the Milkin Institute debate with Teal whic
h is amazing. Um you talked about the lag between a new technology and kind of like the GDP um impact of it.
rgeoa capitalism and how you know this kind of managerial class was needed because of this complexity and I think if you bring AI into the fold you ha
ve like much higher leverage of people. So, like if you have, you know, the Musk industries um and you give Elon AGI, you can run a lot more things uh
at once. That's right.
thing and you're like absolutely not.
e now people are taking that seriously. So I'm just curious like how you're seeing the structure of organization changing especially when you invest i
n early stage companies and um yeah just like how the impact of work structure and uh all of that is playing out.
看,我不是绝对主义者,也不是乌托邦主义者——说清楚一下,我有一个 11 岁的孩子,他在学编程,我觉得学编程仍然是个好主意。但如果你向前展望,你就必须思考一个世界——在这个世界里——"好,我只要告诉它我需要什么,它就去做"——然后它用对它来说最优的方式去做。除非我告诉它用非最优方式——如果我告诉它用 Java 或 Rust,它当然会用。但如果我只是告诉它去做,它会用最优方式。然后如果我需要理解它怎么工作的,我会让它解释给我听。它会成为自我解释的引擎——自己做可解释性。
我就是不确定在那个世界里,你还会有这些历史性的抽象层——那些抽象的目标会变成机器人需要的东西。
there's a whole bunch of top yeah [laughter] we could by the way we would be happy to spend more time but we could we could spend more time on all th
at. So just for people who haven't followed this, so the this this this term managerial comes from this thinker in the 20th century, James Burnham who
um is one of the great kind of 20th century political thinkers um societal thinkers and he sort of said as and he was writing in like the 1940s 1950s
um and he said kind of that the whole history of capitalism up until that point had been in two phases. Number one had been what he called bgeoa capi
talism which was think of it as like name on the door like Ford Motor Company because Henry Ford runs the company. Um, and Henry, it's like a dict dic
tatorial model and Henry Ford just like tells everybody what to do. And he said the problem with boogeoa capitalism is it doesn't scale because Henry
Ford can only tell so many people to do so many things and then he runs out of time in the day. And so um he said the second phase of capitalism was w
hat he called managerial capitalism which was the creation of a professional class of managers
be whatever experts in any particular field but are trained to be experts in management. And then that led to you know the importance of like Harvard
business you know business schools and management consulting firms and all these things. And then you look at every big company today and like most o
f the executives at most of the Fortune 500 companies are not domain experts in whatever the company does and they're certainly not the founders of th
ose companies but they're professional managers. And in fact in the course of their careers they'll probably manage many different kinds of businesses
. They'll rotate around and they might work in healthcare for a while and then work in financial services and then go work in something else you know
come work in tech. And what Burnham said is he said that transition is absolutely required because the the the problem with boogea capitalism is is it
doesn't scale. Henry Ford doesn't scale. And so if you're going to run capitalist enterprises that are going to have millions to billions of customer
s, um you're going to need to they're going to be operating a level of scale and complexity that's going to require this professional management class
. And he said, look, the professional management class has its downsides. Like they're not necessarily experts at doing the thing. They're not as inve
ntive. You know, they're not going to create the next breakthrough thing. But he's like, whether you think that's good or bad or whatever is what's
going to be required. And basically that's what happened right and so he wrote that book originally in like 1940 you know over the course of the next
50 years basically managerialism well I mean today up till today manager managerialism basically took over everything and you know what I'm describin
g is basically how all big companies run and how all governments run and how large scale nonprofits run and kind of everything you know everything run
s
hich is to say Elon Musk or or the next or the next Elon Musk or the next Steve Jobs. the next Bill Gates, the next Mark Zuckerberg. And so we we we w
e start these companies in the old model, right? We we we start them out as as as as in the Henry Ford model.
er or a or a or a founder with with colleagues, but you know, there's a founder CEO. Um and then we basically bet that we basically bet that the start
up is going to be able to do things, specifically innovate in ways that the big incumbents in that industry are not going to be able to do. And so it'
s a bet that by basically by relighting this sort of name on the door, you know, kind of thing, this new innovative thing with like a king monarchical
uh uh political structure u that they're going to be able to innovate in a way that the incumbent is not going to be able to because the incumbent is
is being run by managers, right? And and and and by the way, and of course, venture being what it is, sometimes that works, sometimes it doesn't. But
but we're constantly doing that. But I've always viewed it my entire life as like we're like raging against the dying of the light. Like we're we're
[clears throat] we're we're sort of constantly trying to fight off managerialism just basically swamping everything and everything getting basically b
oring and gray and dumb and old, right? And we're trying to keep some level of energy vitality in the system. AI is the thing that would lead you to t
hink, wow, maybe there's a third model, right? And and maybe may and way to think about it would be maybe it's a combination of the two. Maybe the
new Henry Ford or the new Elon or the new Steve Jobs plus AI is the best of both, right? Because it's it's it's sort of the spark of genius of the nam
e on the door model, the Henry Ford model, but then it's give that person AI superpowers to do all the managerial stuff and let the boss do the manage
rial stuff. That may be the actual secret formula. And we've never even known that we wanted this because we never even thought it was a possibility.
But I mean, you know this that what is the thing that these bots are really good at? They're really good at doing paperwork. Like they're really good
at filling out forms. Like they're really good at writing reports. They're really good at reading. They're really good at doing all the managerial wor
k. Like they're amazing at it. And so yeah, so I I think I think the I 100% I think the answer the answer very well might be to get the best best of b
oth worlds by doing this. And then the challenge is going to be twofold. challenge is going to be for the innovators to really figure out how to lever
age AI to actually do this,
gure out like, okay, what does that mean? Because now they're going to they're going to be facing a different kind of insurgent competitor that has a
different set of capabilities than they're used to. And so this really I think is going to force a lot of big companies to kind of figure out innovati
on, either say figure out innovation or die trying.
t SpaceX is like the growth is like so fast and like instead of having these companies kind of like peter out in growth and impact, they can kind of l
ike keep going if not accelerating. That's for sure the hope. Um the the the challenge and and you know and look the AI utopian view is of course a
nd and and that's going to be the future of the economy and it's going to grow 10x and 100x and thousandx and we're training this regime of like much
higher economic growth forever and consumer cornucopia of everything and it's going to be great and I and I hope that's true. I hope that's that's lik
e the you know that's the current kind of utopian vision. I hope that's true. The problem is goes back again. The real world is really messy. Um, and
I'll give you an example how the real world is really messy. It requires 900 hours of professional certification training to become a hairdresser in t
he state of California. Um, so it's like 35% of the economy, something like that. You have to get some sort of professional certification to do the jo
b, [snorts] which is to say that the professions are all cartels, right? And so you have to get licensed as a doctor, you have to get licensed as a la
wyer, you have to get licensed as a you have to get into a union. Um, by the way, to to work for the government, you need to be you you have both civi
l service protections and you have public sector unions. You have two layers of insulation
hing ever [laughter] changing. I'll give you another example. The the doc work the doc workers went on strike a couple years ago because there, you kn
ow, robotics, you know, if if you go look at a modern doc like in Asia, it's all robots. If you go to American doc, it's like all still guys dragging
strike dragging stuff by by hand. The doc workers went on strike. It turns out there are 25,000 doc workers working on on docs in America. It turns ou
t they have incredible political power because it's a it's it's one of these unified blocks of things.
ments from the doc owners to not implement more automation. We learned a couple things in that. So number one, we learned that even a union as small a
s 25,000 people still has like tremendous political stroke. We also learned that they it actually turns out the doc workers union has 50,000 people in
it because there's 20 they have 25,000 people working in the docks, they have 25,000 people during full paychecks sitting at home from prior union ag
reements. From prior union agreements. I'll give you another great example. There are government agencies. is there are federal government agencies
where the employees right of have civil service protections and they're in public sector unions. There are entire federal government agencies that st
ruck new collective bargaining agreements during COVID where not only are they have their jobs guaranteed in perpetuity, but they only have to report
to work in an office one day per month. And so there are entire office buildings in Washington DC that are empty 29 out of 30 days of the year that ar
e still operating and are still we're all still paying for it. And so and then what they do, it turns out what the employees do is they're very they'r
e very smart in in in this way. And so they figure out they come in on the last day of a month and the first day of the next month.
're so they're in their they're in the office 2 days per 60 days,
u see you see where I'm heading with this. Like this is like locked in, right? This is like locked in in a way that has nothing to do with like people
say capitalist. It's like anti- capitalistic. It's like it's it's basically it's restrictions on trade. It's restrictions on the ability to like chan
ge the workforce. And so so much of our economy is is is you know the I'm I'm describing the entire healthare system. I'm describing the entire legal
profession. I'm describing the entire housing industry. I'm describing the entire education system. Right? K through 12 schools in the United States,
they're a literal government monopoly. How are we going to apply AI in education? The answer is we're not because it's a literal government monopoly.
It is never going to change the end. And there is nothing to do. By the way, you can create an entirely new school system. Like that's the one thing y
ou can do is you can do what Alpha School is doing. you can create an entirely new school system. Other than that, you're not going to go in and chang
e what's happening in the American classroom like K through 12. There's no chance. The teachers are 100% opposed to it. It's 100% not going to happen.
So, so you see what I'm saying is like there's this like massive slippage that's going to take place.
far too optimistic,
billion people all of a sudden are going to change how they behave. And it's just like no. So much of how the existing economy works is just is just l
ike wired in. And so we're going to be lucky as a society. We're going to be lucky if AI adoption happens quickly, right? Because if it doesn't, what
we're just going to have is stagnation.
we're truly living in an age of science fiction coming to real life. Yes. Yes. Could not be more exciting. Really with you guys. Awesome. Good. Th
ank you. Hey, [music]