🌐 双语
Archive

AI Builders
Digest

2026-11-03 12 builders · 29 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Legora CTO:在看不见数据的情况下构建法律 AgentLegora CTO: Building Legal Agents When You Can't See the Data

The Takeaway:法律工作中最有价值的部分不会消失,真正的护城河是在隐私约束下把 Agent、工作流和行业知识编织成可扩展的产品。

Jacob Lauritsen 是 Legora 的 CTO。这家公司从瑞典地下室起步,已成为增长最快的法律科技之一。他指出,律师真正想要的是“让它能用”,而不是自己去挑 Sonnet 还是 Opus。Legora 把模型选择和任务拆分交给自己,律师只看到结果。

最有意思的是他们的限制:客户数据极度机密,团队几乎看不到任何真实 trace。于是他们用律师+工程师混合的 eval 团队,搭建模拟 due diligence 的 RL 环境,用 heat map 跟踪不同实践领域和任务上的模型表现。模型升级有时会解决 citation 问题(从 70% 到 99%),有时又会在某些任务上退步。

产品上,Tabular Review 成为行业术语——把成千上万份合同扔进去,按列提问(终止期、IP 条款、赔偿金额),百万级单元格也能跑。Intake 和 Lists 则像 Linear,把客户请求变成可编排的工作流。Jacob 认为,未来律师会把更多时间花在战略判断和客户关系上,而不是在文档里翻找红旗。

他留了一句很实在的话:“If you're doing a four hundred million dollar M&A transaction, you're gonna want Legora.”
The Takeaway: The highest-value legal work will not disappear; the real moat is weaving agents, workflows, and domain knowledge into a product that works under extreme privacy constraints.

Jacob Lauritsen is CTO of Legora, the legal-tech company that started in a Stockholm basement and became one of the fastest-growing in the space. Lawyers do not want to choose between Sonnet and Opus; they want the system to just work. Legora therefore owns model routing and task decomposition so users only see results.

The hardest constraint is that client data is so confidential the team almost never sees real traces. They solve this with a mixed eval team of lawyers and engineers, building RL-style environments that simulate full due-diligence projects and tracking model performance across practice areas with heat maps. Model upgrades sometimes magically fix citation accuracy (70 % → 99 %) and sometimes regress on other tasks.

On the product side, Tabular Review has become an industry term: drop thousands of contracts, ask column-level questions (termination periods, IP assignment, severance), and review up to a million cells. Intake and Lists act like Linear for legal work, turning client requests into orchestrated agent workflows. Jacob expects lawyers to spend far more time on strategic judgment and client relationships and far less time hunting for red flags in documents.

His sharpest line: “If you're doing a four hundred million dollar M&A transaction, you're gonna want Legora.”
查看原文 →

最贵的智能时代已经过去The Most Expensive Intelligence Will Ever Be

前 Claude Code 成员 Thariq 用一周时间得出结论:你不需要无限 token 就能做出出色工作,$200 套餐已经非常强大,而且这很可能是智能有史以来最贵的时候。
Former Claude Code member Thariq spent a week and concluded you do not need unlimited tokens to do great work. The $200 plans are already excellent, and this is likely the most expensive intelligence will ever be.
查看原文 →

Agent 应运行在组织可控的隔离环境AI Agents Need Dedicated, Organization-Controlled Environments

Replit CEO Amjad Masad 强调,AI Agent 应该运行在组织完全可控、资源访问受限的专用环境中,只获得完成当前任务所需的最小权限。金融领域已经出现广泛滥用,封禁即将来临。
Replit CEO Amjad Masad argues that AI agents should operate in dedicated, organization-controlled environments with strictly limited access to only the resources each task requires. The same pattern is already widespread in finance and the ban hammer is coming down.
查看原文 →查看原文 →

💰 创业成功案例

WisprFlow 完成 2.8 亿美元 B 轮,估值 20 亿WisprFlow Raises $280M Series B at $2B Valuation

SPC 校友、四次创业者兼 IOI 奖牌得主 Tankots 创办的 WisprFlow 刚刚完成 2.8 亿美元 B 轮融资,估值达到 20 亿美元。他将于 11 月 12 日回到 SPC 分享。
SPC alumni, four-time founder and IOI medalist Tankots, co-founder and CEO of WisprFlow, just closed a $280M Series B at a $2B valuation. He returns to SPC on November 12.
查看原文 →

Legora:从 20 人斯巴达到 100 人仍觉得太小Legora: From 20 Spartans to 100 Engineers Still Feeling Too Small

Jacob 曾放言“我们要一直保持 20 名工程师,打败所有更大的对手”。如今团队已到 100 人,却仍感觉人手不足。激烈竞争和客户需求迫使他们从“说不”转向用核心原语(Lists、Projects、Agent、Tabular Review)快速扩展产品表面。
Jacob once declared they would stay twenty engineers forever and beat far larger competitors. They are now at one hundred and still feel too small. Fierce competition and customer demand forced a shift from pure “say no” to building a few powerful primitives (Lists, Projects, Agent, Tabular Review) that let the product surface expand rapidly.
查看原文 →

🛠️ 开发者工具与技巧

Vercel Sandbox:有状态才是真正有用的 Agent 环境Vercel Sandbox: Stateful Workspaces Beat Stateless Benchmarks

Vercel CEO Guillermo Rauch 指出,无状态沙箱是错误的基准。Agent 只有带着身份、技能、仓库和记忆一起工作才真正有用。Vercel Sandbox 团队在 19 个全球区域保持延迟与吞吐 SLO 的同时,每周发布新功能,并承受两位数百分比的周增长需求。
Vercel CEO Guillermo Rauch argues that stateless sandboxes are the wrong benchmark. Agents are only useful when they carry their full workspace—identity, skills, repos, memories. The Vercel Sandbox team is delivering latency and throughput SLOs across 19 global regions, shipping weekly, under double-digit weekly growth in customer demand.
查看原文 →查看原文 →

写作编辑的最佳模型被 harness 拖累The Best Writing Model Is Held Back by Its Harness

Peter Yang 对比 ChatGPT 与 Gemini 的文案修改建议:ChatGPT 几乎全是差建议,Gemini 明显更好。但 Antigravity 的浏览器操作仍不可用,只能把 Gemini 的修改手动贴回 ChatGPT,希望 Antigravity 尽快补齐。
Peter Yang compared ChatGPT and Gemini copy-edit suggestions. ChatGPT’s were almost entirely poor; Gemini’s were actually strong. Unfortunately Antigravity still lacks working browser use, forcing him to paste Gemini’s edits back into ChatGPT. He hopes the harness improves fast.
查看原文 →

Capy Captains 让问题修复吞吐量一夜提升 5 倍Capy Captains Increased Issue-Fixing Throughput 5× Overnight

YC 总裁 Garry Tan 表示,Capy Captains 让他协调跨 captain 和子线程的能力一夜之间提升约 5 倍,是真正的解锁。同时他注意到小型模型的新用法:直接操纵权重。
YC President Garry Tan reports that Capy Captains raised his issue-fixing throughput roughly 5× overnight by enabling coordination across captains and subthreads. He also flags an interesting new use of smaller models: direct manipulation of weights.
查看原文 →查看原文 →

群聊是糟糕的多人 AI 体验Group Chat Is a Terrible Multiplayer AI Experience

OpenAI Codex 产品负责人 Nan Yu 直言:除了极少数情况,群聊是糟糕的多人 AI 体验,几乎保证产出 slop。
OpenAI Codex product lead Nan Yu states that, except in rare circumstances, group chat is a terrible multiplayer AI experience and is literally guaranteed to produce slop.
查看原文 →

Weight-space programming 已进入《Exhalation》第二段Weight-Space Programming Has Reached Paragraph 2 of Exhalation

Swyx 观察到,随着 weight-space programming 的出现,我们现在已经走到 Ted Chiang《Exhalation》的第二段。
Swyx notes that with weight-space programming we have now reached paragraph 2 of Ted Chiang’s “Exhalation.”
查看原文 →

🌍 其他动态

OpenAI 的 Git-fu 与模型自动选择OpenAI’s Git-fu and Automatic Model Selection

OpenAI Codex & ChatGPT 的 Thibault Sottiaux 为团队终于能公开讨论他们的 Git 能力感到兴奋,并透露 “Your dot will choose the best model for the task”,根据工作不同部分自动调用不同模型。
OpenAI Codex & ChatGPT’s Thibault Sottiaux is excited the team can finally talk about their Git capabilities, and notes that “Your dot will choose the best model for the task,” drawing on different models for different parts of the work.
查看原文 →查看原文 →

创业公司模糊发推营收到正式融资的时间大约 11 天Vague-Posting Revenue to Fundraise Kickoff Takes ~11 Days

FPV Ventures 合伙人 Nikunj 统计(n=17):创业公司从模糊发推营收数字,到启动原本说“明年早些时候才考虑”的融资,平均只需要大约 11 天。
FPV Ventures partner Nikunj observed (n=17) that the pipeline from a startup vaguely posting revenue numbers to kicking off a fundraise they “were totally not going to do until early next year” is roughly 11 days.
查看原文 →

Argon 表现亮眼,Antigravity 仍停留在 2025Argon Looks Strong While Antigravity Feels Stuck in 2025

Every CEO Dan Shipper 认为 Argon 之所以好用,部分原因是它本身出色,部分原因是 Antigravity 仍停留在 2025 年,缺少那些疯狂的新功能和复杂度。
Every CEO Dan Shipper attributes part of Argon’s strength to the product itself and part to the fact that Antigravity is still stuck in 2025 and lacks the newer crazy features and complexity.
查看原文 →查看原文 →