AI Builders Digest — March 23, 2026
X / TWITTER
Andrej Karpathy (Former Director of AI at Tesla, OpenAI founding team)
One common issue with personalization in all LLMs is how distracting memory features can be. A single question from months ago about some topic keeps resurfacing as if it's a deep interest, mentioned perpetually with undue emphasis. Karpathy observes this across all LLM implementations, suggesting it might stem from training bias where context window information is often relevant, causing models to overfit to anything retrieved via memory features.
https://x.com/karpathy/status/2036836816654147718
所有 LLM 的个性化功能中，一个常见问题是记忆功能的干扰性。几个月前关于某个话题的一个问题会不断被当作深层兴趣提起，被过度强调。Karpathy 观察到这个问题出现在所有 LLM 实现中，认为可能源于训练过程中的偏差——上下文窗口中的信息通常与任务相关，导致模型过度依赖任何通过记忆功能检索到的内容。
https://x.com/karpathy/status/2036836816654147718
Box CEO Aaron Levie
Jevons paradox is happening in real time. Companies, especially outside tech, are realizing they can now afford software projects previously out of reach because AI makes them feasible. Marketing teams at large companies will have engineers automating workflows. Engineers in life sciences and healthcare will automate research. Small businesses will hire engineers for the first time to build better digital experiences.
As long as AI agents require humans who understand how to prompt, review when agents go off track, guide them back, maintain systems, and fix bugs, we'll still need humans managing these agents. This is why advice against going into engineering is wrong—the world is increasingly made of software, and those who understand it best will be in a strong economic position.
https://x.com/levie/status/2036832183131033977
杰文斯悖论正在实时发生。公司，尤其是科技公司之外的，正在意识到他们现在可以承担以前无法企及的软件项目，因为 AI 让这些项目变得可行。大公司的营销团队会有工程师帮助自动化工作流。生命科学和医疗健康领域的工程师会自动化研究。小企业会第一次雇佣工程师来构建更好的数字体验。
只要 AI agent 还需要人类来理解如何提示、在 agent 偏离时审查、引导回来、维护系统、修复 bug，我们仍然需要人类来管理这些 agent。这就是为什么建议不要从事工程工作是错误的——世界正越来越多地由软件构成，那些最理解软件的人将处于有利的经济地位。
https://x.com/levie/status/2036832183131033977
Anthropic's Cat Wu (Claude Code team)
Auto mode is a step change improvement in the Claude Code UX, balancing autonomy and safety. Almost everyone on the team uses this as a daily driver. Now available to Claude for Team users with claude --enable-auto-mode, then Shift + Tab to enter the mode.
https://x.com/_catwu/status/2036852880624541938
自动模式是 Claude Code 用户体验的一个跨越式改进，平衡了自主性和安全性。团队几乎所有人都在日常使用这个功能。现在 Claude for Team 用户可以通过 claude --enable-auto-mode 启用，然后 Shift + Tab 进入该模式。
https://x.com/_catwu/status/2036852880624541938
Anthropic's Thariq (Claude Code team)
iMessage is now available as a channel for Claude Code, expanding integration options for users.
https://x.com/trq212/status/2036959638646866021
iMessage 现在可以作为 Claude Code 的渠道使用，为用户扩展了集成选项。
https://x.com/trq212/status/2036959638646866021
Vercel CEO Guillermo Rauch
Every company will become an AI factory where the unit of production is the token. However, tokens create usage tracking and billing headaches very unlike SaaS. Vercel's AI Gateway now solves the metering problem across models and providers with one /v1/report API call.
https://x.com/rauchg/status/2036963706576527623
每家公司都会成为 AI 工厂，其生产单位是 token。然而，token 会带来与 SaaS 截然不同的使用追踪和计费难题。Vercel 的 AI Gateway 现在通过一个 /v1/report API 调用解决了跨模型和提供商的计量问题。
https://x.com/rauchg/status/2036963706576527623
Replit CEO Amjad Masad
Apple seems happy with apps made with Replit as evident by the high rate of acceptance.
https://x.com/amasad/status/2037004600893472936
苹果似乎对用 Replit 制作的应用很满意，这一点从高接受率可以看出。
https://x.com/amasad/status/2037004600893472936
Cursor Design Lead Ryo Lu
As agents make it easy to add features, design matters more, not less. The role is no longer just pushing pixels—it's deciding what should exist, how it fits together, how humans stay in control, and how intelligence feels clear, trustworthy, and useful. Taste, craft, and judgment have always been the bottleneck. The game is not who ships fastest, but who makes the right thing for humans.
https://x.com/ryolu_/status/2036886854805709097
随着 agent 让添加功能变得容易，设计变得更加重要，而不是相反。这个角色不再只是推像素——而是决定应该存在什么、如何组合在一起、人类如何保持控制、以及智能如何感觉清晰、可信和有用。品味、工艺和判断一直是瓶颈。游戏不是谁发布得最快，而是谁为人类做对了东西。
https://x.com/ryolu_/status/2036886854805709097
Y Combinator CEO Garry Tan
One of the most important things about this new age is you have to use tokens aggressively to create something remarkable. You have to let it rip. If you do, and you have agency and taste, the result will be remarkable. Token credits for AI is a big part of making startups accessible regardless of where you grew up or whether your family has money.
https://x.com/garrytan/status/2037055498974093629
这个新时代最重要的事情之一是你必须大胆地使用 token 来创造卓越的东西。你必须放手去做。如果你这样做，而且你有能动性和品味，结果会是卓越的。AI 的 token 额度是让初创企业变得可及的重要组成部分，无论你在哪里长大或者你的家庭是否有钱。
https://x.com/garrytan/status/2037055498974093629
FirstMark VC Matt Turck
What VCs do on company boards, an evolution:
- 2016: governance, guidance, support
- 2021: cheerleaders
- 2026: Anthropic and OpenAI sales representatives
https://x.com/mattturck/status/2036747586468499889
VC 在公司董事会上的角色演变：
- 2016：治理、指导、支持
- 2021：啦啦队
- 2026：Anthropic 和 OpenAI 的销售代表
https://x.com/mattturck/status/2036747586468499889
OpenClaw's Peter Steinberger
New OpenClaw beta is out with better MS Teams integration, OpenWebUI and more!
https://x.com/steipete/status/2036824286988816737
新的 OpenClaw beta 版本发布，改进了 MS Teams 集成、OpenWebUI 等功能！
https://x.com/steipete/status/2036824286988816737
Every CEO Dan Shipper
The rules of professional product development are being rewritten in real time. PMs and designers can ship software as easily as engineers. Software is no longer just built for humans—it's also built for agents as first-class citizens.
Dan interviewed Mike Krieger (Instagram cofounder, now at Anthropic Labs) on the AI & I podcast about building agent-native products, the trap of building too much too fast with agents, how Anthropic Labs structures product teams, and why you need to throw out your product every 3-6 months.
https://x.com/danshipper/status/2036827118915485942
专业产品开发的规则正在实时重写。PM 和设计师可以像工程师一样轻松地发布软件。软件不再只是为人类构建——它也为作为一等公民的 agent 而构建。
Dan 在 AI & I 播客上采访了 Mike Krieger（Instagram 联合创始人，现在在 Anthropic Labs），讨论了构建 agent-native 产品、使用 agent 过快构建过多的陷阱、Anthropic Labs 如何组建产品团队，以及为什么你需要每 3-6 个月抛弃你的产品。
https://x.com/danshipper/status/2036827118915485942
Roblox Product Lead Peter Yang
The biggest time-waster at most tech companies isn't meetings or Slack. It's trying to predict what your leader wants. Peter breaks down how a VP at Meta solved this by building an AI skill that reviews docs in his voice, including the full skill files so you can set it up today.
https://x.com/petergyang/status/2036812870365917284
大多数科技公司最大的时间浪费不是会议或 Slack。而是试图预测你的领导想要什么。Peter 详细介绍了一位 Meta 的 VP 如何通过构建一个以他的声音审查文档的 AI skill 来解决这个问题，包括完整的 skill 文件，让你今天就可以设置。
https://x.com/petergyang/status/2036812870365917284
Linear Head of Product Nan Yu
Nan wishes we could reclaim design to mean "a plan for arranging elements to accomplish a particular purpose." We got overly obsessed with arranging elements and forgot about the purpose.
https://x.com/thenanyu/status/2037042617213481410
Nan 希望我们能重新定义设计，让它意味着"为达成特定目的而安排元素的计划"。我们过度痴迷于安排元素，却忘记了目的。
https://x.com/thenanyu/status/2037042617213481410
South Park Commons GP Aditya Agarwal
Observations from Hill Valley Forum:
- Hardtech/deeptech requires three to tango: builders + investors + policymakers
- Insane energy around drones, energy, robotics, atoms+bits
- This energy is oddly not very Silicon Valley pilled—it's in disparate places around the US
- Support from Washington is very real—every congressman and senator wanted to be part of building something amazing
https://x.com/adityaag/status/2036824253970923925
从 Hill Valley Forum 的观察：
- 硬科技/深科技需要三方共舞：建设者 + 投资者 + 政策制定者
- 无人机、能源、机器人、原子+比特领域有疯狂的能量
- 这种能量奇怪地不太硅谷化——它分布在美国各地的不同地方
- 来自华盛顿的支持非常真实——每位国会议员和参议员都想参与构建令人惊叹的东西
https://x.com/adityaag/status/2036824253970923925
Claude (Anthropic)
Your work tools in Claude are now available on mobile. Explore Figma designs, create Canva slides, check Amplitude dashboards, all from your phone.
https://x.com/claudeai/status/2036850783526719610
你在 Claude 中的工作工具现在可以在移动端使用。从手机上探索 Figma 设计、创建 Canva 幻灯片、查看 Amplitude 仪表板。
https://x.com/claudeai/status/2036850783526719610
PODCASTS
Training Data: Biology's Waymo Moment — Ginkgo Bioworks CEO Jason Kelly
The Takeaway: AI models combined with autonomous robotic labs can dramatically accelerate scientific discovery, potentially 10-100x the speed of traditional research.
Jason Kelly, CEO of Ginkgo Bioworks, shares a groundbreaking vision for how AI will transform experimental science. Unlike previous tech revolutions (internet, social media) that had minimal impact on biotech, AI is fundamentally changing how we do science.
The key insight: Most science spending (95%+) goes to overhead—people, lab space, equipment—while less than 5% goes to actual reagents. This is backwards. In an AI-driven autonomous lab, 90% of costs would go to reagents (the actual usage-based pricing of science), representing a 10x increase in data per dollar.
Ginkgo's recent OpenAI collaboration demonstrated this potential: an AI model running experiments on an autonomous lab beat the state-of-the-art in cell-free protein synthesis by 40% after just 6 rounds of experimentation. The model didn't need to be super intelligent—it just needed to design experiments like a scientist and run them 24/7.
The vision: Instead of isolated scientists working in separate labs, sharing results every 1-2 years through papers, imagine 100 AI scientists running on one autonomous lab, sharing raw experimental data daily, learning from each other's failures and successes in real-time. Combined with 10x better resource utilization, this could revolutionize how we do science—from drug discovery to materials science.
Kelly sees this as biology's "Waymo moment"—moving from manual lab work (like driving cars) to autonomous labs (like self-driving cars). The technology is tractable: integrate lab equipment, solve liquid handling, and let AI models handle the experimental design and interpretation.
https://youtube.com/watch?v=g45Alfg7diw
核心观点： AI 模型结合自主机器人实验室可以大幅加速科学发现，可能达到传统研究速度的 10-100 倍。
Ginkgo Bioworks CEO Jason Kelly 分享了一个关于 AI 如何改变实验科学的突破性愿景。与之前对生物技术影响微乎其微的技术革命（互联网、社交媒体）不同，AI 正在从根本上改变我们做科学的方式。
关键洞察：大多数科学支出（95% 以上）用于运营成本——人员、实验室空间、设备——而实际试剂支出不到 5%。这是本末倒置的。在 AI 驱动的自主实验室中，90% 的成本将用于试剂（科学的实际基于使用的定价），代表着每美元数据量的 10 倍增长。
Ginkgo 最近与 OpenAI 的合作展示了这种潜力：一个在自主实验室中运行实验的 AI 模型，仅经过 6 轮实验，就在无细胞蛋白质合成方面比最先进水平提高了 40%。该模型不需要超级智能——它只需要像科学家一样设计实验并 24/7 运行它们。
愿景：不是孤立的科学家在分开的实验室工作，每 1-2 年通过论文分享结果，而是想象 100 个 AI 科学家在一个自主实验室上运行，每天分享原始实验数据，实时从彼此的失败和成功中学习。结合 10 倍更好的资源利用率，这可能会彻底改变我们做科学的方式——从药物发现到材料科学。
Kelly 将这视为生物学的"Waymo 时刻"——从手动实验室工作（如开车）转向自主实验室（如自动驾驶汽车）。这项技术是可行的：集成实验室设备、解决液体处理问题，让 AI 模型处理实验设计和解释。
https://youtube.com/watch?v=g45Alfg7diw
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders