🌐 双语
Archive

AI Builders
Digest

2026-05-05 12 builders · 25 tweets · 1 podcasts · 1 blogs

🔥 热点话题

Waymo 首席技术官谈自动驾驶坚持与安全Waymo CTO on Persistence and Safety in Autonomous Driving

**核心要点**:Waymo 联合创始人兼 CTO Dmitri Dolgov 分享了从 DARPA 挑战赛到今天 2000 万次全自动驾驶的旅程。Waymo 通过基础模型驱动 Driver、Simulator 和 Critic 三大支柱,采用端到端世界动作模型并辅以结构化中间表示,实现了超人级安全。目前每周全自动驾驶超过 400 万英里,在运营城市中严重伤亡碰撞风险比人类低 13 倍。

Dolgov 强调,自动驾驶周期性炒作常见,但真正成功需要相信使命并理解长尾困难。安全是不可谈判的基础,从第一天就融入架构和团队心态。

引用:"世界范围内,每 26 秒就有人因道路碰撞失去生命。"
**The Takeaway**: Building full autonomy requires two decades of persistence through hype cycles, grounded in a critical mission and rigorous safety-first engineering.

Waymo CTO Dmitri Dolgov, a veteran since the DARPA challenges, details the journey from early Google self-driving project to 20 million fully autonomous rides. Waymo's Foundation model powers the Driver, Simulator, and Critic pillars as a multimodal world action language model. They go beyond vanilla end-to-end with structured materialized intermediates for validation, rich training, and superhuman safety—driving 4+ million autonomous miles weekly and reducing serious injury collisions by over 13x versus humans in their cities.

Dolgov stresses understanding that breakthroughs reshape the early curve but not the difficult long tail. Safety is non-negotiable: "Worldwide, somebody loses their life to a crash on our roads every 26 seconds."
查看原文 →

AI 代理进入企业知识工作AI Agents Entering Enterprise Knowledge Work

Box CEO Aaron Levie 指出,Anthropic 和 OpenAI 都在推出帮助企业部署 AI 代理的计划。这一趋势才刚开始,但将迅速壮大。代理进入编码之外的知识工作,需要升级 IT 系统、提供上下文、现代化工作流、处理人机关系以及变革管理。没有捷径将模型能力稳定应用于业务流程,这创造了大量新工作和公司机会。
Box CEO Aaron Levie notes that both Anthropic and OpenAI have new initiatives to help enterprises deploy AI agents. This is an early but fast-growing trend. As agents move into knowledge work beyond coding, significant work is needed to upgrade IT systems, provide context, modernize workflows, define human-agent relationships, drive adoption, and manage change. There's no shortcut to stably applying model intelligence to business processes, creating opportunities for new jobs and firms.
查看原文 →

🛠️ 开发者工具与技巧

Anthropic 推出 Claude Code Auto ModeAnthropic Launches Claude Code Auto Mode

Anthropic Engineering 博客宣布 Claude Code Auto Mode:使用模型分类器代理权限批准,在安全与自主性之间取得平衡。默认沙箱或手动批准易导致疲劳,--dangerously-skip-permissions 则不安全。Auto Mode 通过输入层提示注入探测和输出层转录分类器(Sonnet 4.6)工作,在真实流量上 FPR 低至 0.4%。适用于希望自主运行但需防护过度热情行为的场景。

链接:https://www.anthropic.com/engineering/claude-code-auto-mode
Anthropic Engineering: Claude Code auto mode: a safer way to skip permissions. Instead of approval fatigue or risky full bypass, Auto Mode uses model-based classifiers and a prompt-injection probe. Two-layer defense (input probe + output transcript classifier) catches overeager/dangerous actions while allowing routine work. Evaluated with low 0.4% FPR on real traffic. A practical middle ground for agentic coding.

Direct link: https://www.anthropic.com/engineering/claude-code-auto-mode
查看原文 →

Vercel 发布 DeepSec 开源代理编排器Vercel Releases DeepSec Open-Source Agent Orchestrator

Vercel CEO Guillermo Rauch 介绍 DeepSec:开源代理编排器,用于深度安全审查。可并行运行数千代理,在几分钟内发现重大漏洞(传统团队需数月)。已用于主要 OSS 项目,优化与 Vercel Sandbox 配合使用。欢迎在自己的仓库尝试,OSS 项目可申请赞助运行。
Vercel CEO Guillermo Rauch introduces deepsec: an open-source agent orchestrator for deep security reviews. Built internally, it enables thousands of agents to scrutinize codebases in parallel, finding critical vulnerabilities in minutes that would take human teams months. Optimized for Vercel Sandbox. Encourages trying on your repos; OSS projects can request sponsored runs.
查看原文 →

Peter Steinberger 发布 Crabbox 0.5.0Peter Steinberger Ships Crabbox 0.5.0

OpenClaw 的 Peter Steinberger 推出 Crabbox 0.5.0:支持桌面/浏览器租用、VNC + WebVNC、AWS Windows + WSL2、截图等。现在可直接在临时环境中重现问题,代理设置状态并在 PR 中发布视频,提升 QA 能力。
Peter Steinberger (OpenClaw) released Crabbox 0.5.0 with desktop/browser leases, VNC + authenticated WebVNC, AWS Windows + WSL2, screenshots. Agents can now reproduce issues in ephemeral crabboxes and post videos on PRs, significantly leveling up QA.
查看原文 →

Garry Tan 更新 GBrain v0.27Garry Tan Updates GBrain v0.27

YC CEO Garry Tan 发布 GBrain v0.27:新增对多种非 Anthropic/OpenAI embeddings 和 LLM 的支持。统一记忆层、代码工具和搜索引擎于一个图中。即将支持多模态 embeddings、照片 OCR 等。他强调测试的重要性,并每天使用其 10 万 markdown 文件设置。
Y Combinator CEO Garry Tan shipped GBrain v0.27 with broader embeddings/LLM support beyond Anthropic and OpenAI. It's a unified graph combining memory, code tools, and search. Multi-modal embeddings and photo OCR coming soon. He uses it daily with a large markdown setup and is obsessive about testing.
查看原文 →

💰 创业成功案例

Waymo 指数级扩展Waymo Exponential Scaling

Dmitri Dolgov 分享 Waymo 已完成超过 2000 万次全自动乘车,最近 7 个月达 1000 万次。从 4 个城市到一天推出 4 个新城市,11 个城市运营中。计划扩展至伦敦和东京。第六代硬件聚焦性能、成本和规模生产,新车辆平台 OHAI 以乘客体验为中心。
Waymo has delivered over 20 million fully autonomous rides, with 10 million in the last seven months. They launched four new cities in one day and operate in 11 cities, with London and Tokyo next. Sixth-gen hardware emphasizes performance, drastic cost reduction, and scale. New rider-centric vehicle platform launched.
查看原文 →

Replit 助力创业者Replit Helps Entrepreneurs

Replit CEO Amjad Masad 展示 Replit 如何帮助创业者找到投资者并安排会议,还分享了多模态学习平台用于聋哑学生的 AI 教育案例。
Replit CEO Amjad Masad highlighted how Replit helped an entrepreneur find investors and land meetings. Also shared a great AI education use case: multi-modal learning platform for deaf students.
查看原文 →查看原文 →

🌍 其他动态

Sam Altman 对语音模型的兴奋Sam Altman Excited About Voice Models

OpenAI CEO Sam Altman 表示对语音模型变得出色感到非常兴奋。人们已经开始改变与 AI 的交互方式。他还为 GPT-5.5 派对申请者准备了惊喜。
OpenAI CEO Sam Altman is pretty excited for voice models to get great. It's interesting to watch how people are already changing the way they interface with AI. Also planning something nice for those who applied to the GPT-5.5 party.
查看原文 →

AI 工具与趋势观察AI Tools and Trend Observations

Swyx 分享 OpenAI 与 Anthropic 的估值和 ARR 对比。Peter Yang 认为编码是第一前沿,知识工作是第二,个人代理是第三。Nikunj Kothari 称赞 Gemini Flash 的性价比和实时语音模型。
Swyx shared reconstructed chart on OpenAI (~30B ARR, 850B val) vs Anthropic valuations/revenues. Peter Yang: Coding is first frontier, knowledge work second, personal agents third. Nikunj Kothari praises Gemini Flash as cheap and excellent with 1M context.
查看原文 →查看原文 →

其他开发者与 VC 洞见Other Builder and VC Insights

Aditya Agarwal 强调速度需有明确方向。Matt Turck 幽默点评 VC 命名趋势。Kevin Weil 转发相关内容。
Aditya Agarwal: Velocity is not interesting without grounding towards true north. Matt Turck jokes about literal VC firm naming trends. Various other shares from builders like Peter Yang on kids building with agents.
查看原文 →查看原文 →

🔥 热点话题

创业者与 AI 教育Entrepreneurship and AI Education

Peter Yang 希望让 8 岁孩子用代理构建可分享的东西,甚至赚第一笔钱。
Peter Yang wants his 8-year-old to build agent projects shareable with class/teachers and potentially earn first dollar online.
查看原文 →