🌐 双语
Archive

AI Builders
Digest

2026-11-12 16 builders · 32 tweets · 1 podcasts · 0 blogs

🔥 热点话题

自动化后的伟大人类工作:亲密视角优于聚合趋势Great Human Work After Automation: The Intimate View Beats Aggregate Trends

The Takeaway: 即使 AI 能力飙升并抬高所有人的基线,人类独特的视角与责任仍是价值的真正来源,模型只会成为可分配的资源。

Every CEO Dan Shipper 在 Thesis 大会闭幕演讲中指出,OpenAI 的 GDPval 基准显示 AI 在专家任务上击败人类的比例从 2024 年的 12.4% 升至 2026 年约 83%。从聚合视角看,人类工作似乎走向零。但亲密视角揭示另一面:他用 Claude 在一周内完成数据科学家、游戏开发者、营销、工程与销售的工作量,却仍依赖 30 人团队。同一份客户调查数据,他发现团队销售机会并自建 CRM;增长工程师 Yash 看到首月流失因邮件过多;营销负责人 Douglas 发现读者因周围人对 AI 的不理解而孤独;编辑 Jack 做成黏土动画游戏。生成可能性很便宜,但排名与判断很难。实验显示 Claude 在 400 次尝试中能覆盖他的 Slack 回复 93%,却只能在 60% 的情况下将其排进前 10%。团队中每个人的责任缩小了搜索空间,个人视角完成了筛选。模型会进步,人也不会冻结:2027 年的 Yash 会带着新经验使用 2027 年的模型。默认答案成为新基线,真正的 alpha 是用你的视角把模型推得更远。这就是分配经济:价值不再是你知道多少,而是你如何分配与管理智能资源。

“I pick Yash every day, every time.”
The Takeaway: Even as AI capabilities soar and raise the floor for everyone, human perspective and responsibility remain the real source of value; models become allocatable resources.

Every CEO Dan Shipper, closing the Thesis conference, notes that OpenAI’s GDPval benchmark shows AI beating human experts on economically valuable tasks rising from 12.4% in 2024 to roughly 83% in 2026. The aggregate view suggests human work heads toward zero. The intimate view tells a different story. In one week he used Claude to analyze a customer survey, build a live team game, spot a teams sales opportunity, vibe-code a CRM and start outreach—work that once required five specialists and $200k. Yet Every still runs with a 30-person team. Given the same survey and models, different people extracted different high-value insights: Yash spotted first-month churn from excess email and shipped a fix while building the live-stream player; Douglas saw reader loneliness around AI enthusiasm; Jack turned results into a Claymation game show. Generation is cheap; ranking is hard. An experiment showed Claude could surface answers matching Shipper’s Slack style 93% of the time across 400 samples, but ranked his actual response in its top 10% only about 60% of the time. Each teammate’s responsibility narrows the search space; their perspective ranks what matters. Models will improve at ranking, but people are not frozen: 2027 Yash will use 2027 Claude with another year of experience. Defaults become the new baseline; alpha lives in how you push past them with your own lens. This is the allocation economy: you will be judged on how well you allocate and manage intelligence, not how much you know.

“I pick Yash every day, every time.”
查看原文 →查看原文 →

Jensen Huang 相关传闻引发 AI 圈激烈讨论Jensen Huang Rumors Spark Intense Debate Across AI Circles

Peter Yang 表达了对 Jensen Huang 的深深失望:从洗碗工到全球最有价值公司 CEO 的励志故事、41 年婚姻,如今却传出与间谍及 NVIDIA 高管的绯闻,让许多人难以接受。他同时强调,出轨令人反感,但解雇需经董事会正式调查并让股东参与,Jensen 确实建立了划时代的公司。

Replit CEO Amjad Masad 质疑时机:为何多年前就未彻底掩盖的旧闻突然成为国际地缘新闻?谁受益、为何是现在?如果开放模型最强拥护者被拿下,那些希望 AI 封闭、中心化、受控的人可能受益。中期选举与即将到来的 IPO 窗口都值得关注。

Y Combinator 总裁 Garry Tan 认为董事会罢免 Jensen 极不明智,他的日常领导对这家未来丰裕引擎至关重要,任何过渡都可能导致市值大幅崩盘。

Matt Turck 也转发了相关法方高管的第二则故事。
Peter Yang voiced deep disappointment in Jensen Huang. The arc from dishwasher and waiter to CEO of the world’s most valuable company, plus a 41-year marriage to his college sweetheart, made the reported affairs with a spy and an NVIDIA executive especially hard to process. He still draws a line: the cheating is appalling, yet any board action should follow proper investigation and shareholder involvement, because there is no denying Jensen built a generational company.

Replit CEO Amjad Masad questioned the timing. Why do old stories that were never fully covered up suddenly become international and geopolitical news? Who benefits and why now? If the premier champion of open models is taken out, anyone who wants AI closed, centralized and controlled might gain. Midterms and upcoming IPOs both look relevant.

Y Combinator President Garry Tan called any board move to oust Jensen ill-advised. His day-to-day leadership is central to an engine of future abundance; a transition without massive market-cap damage is hard to imagine.

Matt Turck also highlighted a second story involving French executives.
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →

💰 创业成功案例

当下该建的两类 AI 公司:扩散 vs 前沿Two Kinds of AI Companies Worth Building Right Now: Diffusion vs Frontier

General Partner at SPC Aditya Agarwal 指出,现在基本有两类值得建的公司。

第一类是 AI 扩散:今天的“垂直 SaaS”。无论是 FDE 公司、rollup 还是水平产品,核心都是把 AI 推到工作真正发生的地方。Profound、Cognition、Harvey 等都在此列,会诞生大量优秀公司。

第二类是 AI 前沿:用 AI 加速科学进步,再从进步中构建产品。这类更难被看见,但 SPC 看到越来越多案例:Externa 为蛋白质增加数百种新化学构件,让药物设计匹配疾病而非迁就自然;Epana Bio 用基因组学与 AI 设计针对自身免疫病细胞起源的疗法;VenomicsAI 从毒液中做 AI 加速药物发现;Equivariant 为半导体做多物理场建模。

第二类外部讨论很少,却可能贡献模型驱动的大部分回报。
SPC General Partner Aditya Agarwal lays out two kinds of companies worth building right now.

1. Diffusion of AI — today’s vertical SaaS. Whether FDE model, rollup or horizontal product, the job is the same: get AI diffused right next to where the work happens. Profound, Cognition, Harvey and others sit here; many strong companies will be built.

2. Frontier of AI — use AI to accelerate scientific progress, then build products from that progress. Harder to see from the outside, yet SPC is seeing more of it: Externa adds hundreds of new chemical building blocks to proteins so drugs can be engineered to the disease instead of settling for nature; Epana Bio uses genomics and AI to design autoimmune therapies aimed at the cellular origin; VenomicsAI runs AI-accelerated drug discovery from venoms; Equivariant does multi-physics modeling for semiconductors.

Category 2 is little discussed externally. It should be, because that is where the vast majority of model-driven returns will come from.
查看原文 →

🛠️ 开发者工具与技巧

Google Labs 推出 Tabletop:从“应该做”到“正在做”的共享空间Google Labs Launches Tabletop: Shared Spaces That Move Crews from ‘We Should’ to ‘We’re Doing’

Google Labs 发布新实验 Tabletop。把链接和文件丢进一个共享空间,拉上伙伴,然后向 Tabletop 要任何需要的东西,即可即时生成定制空间:互动公路旅行地图与晚餐决定轮盘、实时 pickleball 对阵表与周边投票板、拖拽式座位表,避免坐到老板旁边。

创始工程师 Roman Nurik 展示了自家使用场景:互动家务表与存钱罐、雨天寻宝游戏实时计分、随性 LEGO 地图、奇幻书主题纽约行程。美区 18 岁以上可加入 waitlist。
Google Labs introduced Tabletop, a new experiment built to move a crew from “we should totally do that” to “we’re actually doing this.” Drop links and files into one shared space, add your people, and ask Tabletop for whatever you need to spin up custom spaces on the fly: interactive road-trip maps and dinner-decider wheels, live pickleball brackets and merch voting boards, drag-and-drop seating charts so you don’t end up next to the boss.

Founding engineer Roman Nurik walked through his family’s actual uses: interactive chore charts and piggy banks, rainy-day scavenger hunts with live scoring, on-the-whim LEGO maps, fantasy-book-themed NYC trips. Waitlist is open for US users 18+.
查看原文 →查看原文 →

AI 原生公司基础设施实况:803 家子处理商数据洞察What AI-Native Companies Actually Use: Insights from 803 Sub-Processor Lists

FPV Ventures 合伙人 Nikunj Kothari 抓取了 803 家公司的子处理商列表,观察 AI 原生公司的基础设施选择。与 3 月数据对比:44% 的公司列出两家以上模型提供商(此前 33%),Anthropic 增速约为 OpenAI 的两倍;86 对互相把对方列为子处理商(如 Anthropic + Cloudflare、Vercel + SpaceX AI);33% 列出推理提供商(此前 22%),Modal 持续领先;中国模型主要通过视频路径出现(如字节)而非文本;TypeSafe AI 从 0 快速成为多公司选择。子处理商列表是滞后指标,却是确认真实使用的可靠方式。完整数据与洞察见 deploygraph.com。
FPV Ventures partner Nikunj Kothari scraped sub-processor lists from 803 companies to see what AI-native teams actually run. Compared with March: 44% now list two or more model providers (up from 33%), with Anthropic growing roughly twice as fast as OpenAI; 86 pairs name each other as subprocessors (Anthropic + Cloudflare, Vercel + SpaceX AI); 33% list an inference provider (up from 22%), Modal still crushing; Chinese models appear mainly via video (ByteDance etc.) rather than text; TypeSafe AI went from zero to many companies in short order. Sub-processor lists lag, yet remain a sure way to confirm real usage. Full data and insights at deploygraph.com.
查看原文 →

Codex 氛围与开放权重模型动态Codex Vibes and Open-Weights Momentum

OpenAI Codex 与 ChatGPT 团队的 Thibault Sottiaux 发帖称“Thursday is feeling codexy”,并暗示“也许也是一个不错的重置时机”,此前已很久没有重置,引发大量讨论。

Vercel CEO Guillermo Rauch 提醒:创造出色产品仍需要大量人类专注、努力、动力与爱,以及注意力跨度。同时分享 Reflection 的 Beam 模型在 vercel.com 上免费至 11 月 12 日:美国开放权重、零数据保留。另有 fx 启动时间基准与高度优化的纯汇编版本竞争的有趣结果。
OpenAI Codex & ChatGPT’s Thibault Sottiaux posted “Thursday is feeling codexy to me. Perhaps also a good time for a banked reset,” after a long stretch without resets, drawing heavy engagement.

Vercel CEO Guillermo Rauch issued a PSA: creating wonderful products still requires extraordinary amounts of human focus, effort, motivation and love, plus an attention span. He also noted Beam by Reflection is free on vercel.com until Nov 12—American open weights, zero data retention. Separately he shared a benchmark where fx boot competes with an ultra-optimized pure-assembly version.
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →

🌍 其他动态

编码智能体的未来体验与苦涩教训的甜面The Coming Experience of Coding Agents and the Sweet Side of the Bitter Lesson

前 Anthropic Claude Code 的 Thariq 将在 AIE Code 做“闭幕”演讲(大家已经累了或赶飞机),主题是“the sweet side of the bitter lesson”,探讨明年使用编码智能体的体验以及我们如何塑造它。欢迎尚未离场的人明天下午 5 点到场。
Former Anthropic Claude Code builder Thariq is giving the “closing keynote” at AIE Code (when everyone is tired or flying out). The talk is titled “the sweet side of the bitter lesson” and covers how the experience of using coding agents could feel next year and how we get to shape it. Anyone still around is invited at 5 pm.
查看原文 →查看原文 →

艺术家用 Claude 生成动画装置Artist Mariko Mori Uses Claude for Computer-Generated Animation

Claude 官方账号分享:Mariko Mori 三十年来不断突破艺术边界。在回顾展“All That Shines”中,她使用 Claude 编写代码生成计算机动画,并作为新的视频装置展出。
Claude’s official account highlighted Mariko Mori, who has pushed the boundaries of art for three decades. For her retrospective “All That Shines,” she used Claude to write code for a computer-generated animation presented as a new video installation.
查看原文 →

开源已够难,别再掺政治Open Source Is Hard Enough Without the Politics

Peter Steinberger(OpenClaw 与 OpenAI 相关)看到一些评论后表示:开源已经够难了,不需要再掺政治。
Peter Steinberger (OpenClaw / OpenAI) reacted to certain comments: “reading some of these comments makes my brain hurt. oss is already hard enough without politics.”
查看原文 →

技术完成最重要工作的标志When Technology Finishes Its Most Important Work

Zara Zhang 转发并强调一句:“Technology only finishes its most important work once the audience forgets the technology.”
Zara Zhang highlighted the line: “Technology only finishes its most important work once the audience forgets the technology.”
查看原文 →