2026-04-12
AI Builders Digest — 2026-04-12
X / TWITTER
Swyx (swyx), AI Builder & Podcaster
Flew straight back from London to do a Broadway cabaret — performed his first solo in ~18 years and only fumbled lyrics once. The thread is mostly personal, but the energy says something about what high-performance builders do to stay human.
从伦敦直飞回来赶 Broadway cabaret 演出，18 年来第一次独唱，只忘了一次词。这个 thread 主要是 personal，但能量说明了一件事：高效 builder 怎么保持人性。
Peter Yang (petergyang), Product at Roblox
Teased an interview with Figma CEO Dylan Field on whether AI can learn design taste — plus tangentially whether design systems constrain creativity and how Figma plans to compete in the AI era. Full episode drops tomorrow.
预告了和 Figma CEO Dylan Field 的访谈，讨论 AI 能否学会 design taste——顺带涉及 design systems 是否束缚创造力，以及 Figma 在 AI 时代怎么打。明天的完整 episode 值得关注。
Thariq (trq212), Claude Code at Anthropic
There's now a TurboTax connector built into Claude Code — Thariq was grateful for the tax-filing procrastination this enabled. The broader signal: agentic tools are increasingly connecting to real-world personal finance infrastructure.
Claude Code 现在内置了 TurboTax connector——Thariq 对此表示非常感谢，因为他正好拖延了报税。更大的 signal：agentic 工具正在越来越多地连接真实世界的个人财务基础设施。
Guillermo Rauch (rauchg), CEO at Vercel
Vercel Sandbox is now the #1 fastest microVM-based sandbox. Beyond lab benchmarks, customers are reporting real-world performance and reliability leadership. Rauch frames it as the foundation for coding agents and general compute parallelism.
Vercel Sandbox 现在是最快的 microVM sandbox。Rauch 的判断：这是 coding agents 和通用计算并行的基础层。这条 thread 的背景是整个沙盒赛道正在加速。
Aaron Levie (levie), CEO at Box
Two sharp threads today:
On datacenter demand: Levie points out that most current AI adoption is on token-efficient chat tools, but coding agents use orders of magnitude more tokens — and they're still only used by a small population. The consumption patterns coming for the rest of knowledge work will be "100s of times greater than we're realizing right now." Expect these capex charts to keep going vertical.
On AI and lawyers: Despite automation fears, the ABA estimates US attorneys grew from ~400K in 1975 to ~1,375K in 2025 — between the PC and internet revolutions. Levie's case: AI will generate so many new legal questions, exotic contract terms, and cross-industry regulatory challenges that demand for lawyers will increase, not decrease.
View tweet — datacenter | View tweet — lawyers
两条有深度的 thread：
关于数据中心需求： Levie 指出，当前大多数 AI 应用还是 token 高效的 chat 工具，但 coding agents 消耗的 token 高出几个数量级——而且现在使用的人还很少。知识工作中即将到来的消费模式会是"比我们意识到的 100 倍还多"。这个判断值得记住。
关于 AI 和律师： 尽管有自动化焦虑，美国律师协会数据显示律师数量从 1975 年的约 40 万增长到 2025 年的约 137.5 万。Levie 的逻辑：AI 会产生大量新的法律问题、复杂的合同条款和跨行业监管挑战，需求反而会增加。
Garry Tan (garrytan), President & CEO at Y Combinator
Garry Tan's counterpoint after three months of open source release used by tens of thousands of agentic engineers daily: thin harness, fat skills. His core argument: if your memory dies when your harness dies, you built the harness too thick. Memory is markdown. Skills are markdown. Brain is a git repo. The harness is a thin conductor.
Y Combinator 总裁 Garry Tan 放出反观点：在开源被数万 agentic engineers 日常使用三个月后，核心原则是 thin harness, fat skills。如果 memory 跟着 harness 一起死，那说明 harness 做太厚了。Memory 是 markdown，Skills 是 markdown，Brain 是 git repo，harness 只是 thin conductor。
Zara Zhang (zarazhangrui), Builder
Two concise reframes on human-AI collaboration:
On what human communication is actually for going forward: Deciding what to work on and what "good" looks like. Keeping each other company and preventing loneliness. Creative brainstorming. Not project syncs, status updates, or handovers.
On collaboration model: The most efficient collaboration is often no collaboration. One person should own something end-to-end and work with agents.
View tweet — communication | View tweet — collaboration
两条关于人机协作的简洁重构：
Human communication 的未来功能： 决定做什么、什么是"好"、互相陪伴对抗孤独、头脑风暴。不是项目同步、状态更新或交接。
协作模型： 最高效的协作往往是不协作。一个人端到端拥有一个东西，然后和 agents 一起工作。
Nikunj Kothari (nikunj), Partner at FPV Ventures
Nikunj called out a pattern he's seeing in VC behavior: the same funds that push founders to take low valuations ("I'm investing for the long term, conviction matters!") will flip six months later and push for highest-possible valuations ("more money, less dilution, helps LP marks"). His advice: backchannel the partner joining your cap table, and try to find missionaries over mercenaries.
VC 圈子里一种 Nikunj 观察到的模式：同一只基金早期压低估值（"我投长期，信念很重要！"），六个月后又推最高估值（"更多钱、更少稀释、有利于 LP 账面"）。建议：找 mission-driven 的投资人，而不是 mercenary。
Peter Steinberger (steipete), OpenClaw Core Team
Three significant OpenClaw updates from Peter:
Strict mode for GPT-5.x: agents.defaults.embeddedPi.executionContract = "strict-agentic" — forces the model to keep working: read code, call tools, make changes, or return a real blocker instead of stopping at "here's the plan." Hit 1009 likes.
Native Codex as harness: OpenClaw harness can now be swapped via plugins — Codex as harness owns threads, resume, compaction, and app-server execution. Enabled with plugins.entries.codex.enabled and agents.defaults.embeddedHarness = { runtime: "codex", fallback: "none" }. Using Codex should increase agentic mode for longer-horizon tasks, though may weaken personality.
Harness-as-plugins architecture: Makes it trivial to replace the default Pi harness with Anthropic's SDK or other custom implementations.
View tweet — strict mode | View tweet — codex harness | View tweet — plugins
OpenClaw 团队 Peter Steinberger 放出的三条重要更新：
Strict mode： 配置 executionContract = "strict-agentic" 后，GPT-5.x 必须持续工作——读代码、调工具、做修改——不能停在"这是计划"就结束。1009 个赞说明这个需求很普遍。
Codex 作为 harness： OpenClaw 的 harness 现在可以插件化替换，Codex 接管 thread、resume、compaction 和 app-server 执行。预计会增强 agentic mode 能力，但可能削弱 personality。
Harness-as-plugins 架构： 可以轻松换成 Anthropic SDK 或其他实现。
Aditya Agarwal (adityaag), General Partner at SouthParkCommons
Aditya cut through the agent-vs-LLM-loop debate with a clean frame: the real difference between a true agent and simply running an LLM in a loop is intelligent long-horizon memory management. He credits Claude Code's 3-tier memory architecture as the most interesting implementation he's seen.
Aditya 对 agent vs LLM loop 之争给出了一个清晰的判断：真正的 agent 和简单跑 LLM loop 的核心区别是 intelligent long-horizon memory management。他点名了 Claude Code 的 3-tier memory architecture 是目前见过最有趣的实现。
PODCASTS
No Priors Podcast — "The Agentic Economy: How AI Agents Will Transform the Financial System with Circle Co-Founder and CEO Jeremy Allaire"
The Takeaway: Stablecoins like USDC aren't just a crypto phenomenon — they're becoming the payment layer for an AI agentic economy, and the infrastructure being built today (especially Circle's new ARC blockchain) is designed specifically for machine-to-machine commerce at micro-scale.
Jerermy Allaire co-founded Circle in 2013 with a thesis that seems prescient now more than ever: the dollar needed a protocol layer on the internet — one that was programmable, globally accessible, and as frictionless as email. USDC is backed by short-duration US Treasuries (~13-day average maturity) and cash held at institutions like Bank of New York Mellon. Under the Genius Act, this full-reserve model is now codified into federal law.
The interesting pivot: Allaire argues that the "agentic economy" — where AI agents consume services from each other, pay in micro-transactions, and coordinate economic activity autonomously — is not science fiction. USDC is already settling transactions for fractions of a penny. The key blocker until recently was cost; with ARC, Circle claims transactions can happen for a millionth of a penny. That's the number that changes everything.
Allaire is particularly bullish on what he calls "ARC as an economic operating system" — blockchains as the coordination layer for machine-driven economic activity, with programmable corporate forms and contracts that didn't exist before. He's also watching the RWA (real-world assets) tokenization wave: Circle's own tokenized stock is apparently the most active tokenized stock on the market right now, ahead of Tesla or S&P index products.
On crypto vs. the agentic future: "We're moving from the early adopter era, which was mostly speculation... to now very squarely because of stablecoins into the real economic activity side of this."
On what AI agents actually need from financial infrastructure: "We don't have an infrastructure that can support billions or trillions of transactions. We don't have an infrastructure that can work globally, interoperably, instantly, that can be programmed through software layers by arbitrary pieces of software. That doesn't exist."
On proof-of-work and inference: Allaire is intrigued by the idea of tying GPU inference compute as a form of productive proof-of-work — the work itself is the inference, rather than burning energy for its own sake.
核心结论： USDC 这类稳定币不只是 crypto 现象——它们正在成为 AI agentic economy 的支付层，而 Circle 新建的 ARC 区块链从一开始就是为机器对机器的微交易经济设计的。
Circle 联合创始人 Jeremy Allaire 2013 年的核心 thesis 现在看来异常有前瞻性：美元需要一个 internet 协议层——可编程、全球可访问、和发邮件一样无摩擦。USDC 由短期美国国债（约 13 天平均期限）和纽约梅隆等机构的现金托管支持。Genius Act 已将这种 full-reserve 模型写入联邦法律。
有趣的转折：Allaire 认为"agentic economy"——AI agents 互相消费服务、以微交易支付、协调经济活动——不是科幻。USDC 已经在以不到一分钱的价格结算交易。之前的核心障碍是成本；ARC 上，Circle 声称交易成本可以低至百万分之一美分。这个数字改变了一切。
Allaire 特别看好他所说的"ARC 作为经济操作系统"——区块链作为机器驱动经济活动的协调层，带有以前不存在的可编程企业形式和合约。他也在关注 RWA（现实资产代币化）浪潮：Circle 自己的代币化股票目前是市场上最活跃的代币化股票，超过了特斯拉或 S&P 指数产品。
关于 crypto 与 agentic 未来： "我们正在从早期采用者时代——主要是投机——非常明确地因为稳定币转向真实经济活动层面。"
关于 AI agents 实际需要什么样的金融基础设施： "我们没有能够支持数十亿或数万亿交易的基础设施。我们没有一种基础设施能够全球运作、互操作、即时工作、由任意软件层编程。这不存在。"
关于工作量证明和推理： Allaire 对将 GPU 推理计算作为生产性工作量证明形式的可能性很感兴趣——工作本身就是推理，而不是为了耗能而耗能。
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders