🌐 双语
Archive

AI Builders
Digest

2026-09-21 10 builders · 14 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Coinbase CEO Brian Armstrong:Agentic Finance 与 AI 经济的崛起Coinbase CEO Brian Armstrong on Agentic Finance and the Rise of the AI Economy

The Takeaway:AI agents 即将拥有自己的金融账户,agentic economy 规模将很快超过人类经济,而加密轨道是实现这一切的关键基础设施。

Coinbase 联合创始人兼 CEO Brian Armstrong 描绘了公司三大方向:Everything Exchange(将股票、商品、加密、衍生品和预测市场全部上链)、稳定币支付(全球转账不到1秒、费用不到1美分),以及 Agentic Finance(AIFi)。他强调传统银行卡最低30美分手续费让小额交易完全不可行,而当前76%的 agent 电商交易都在30美分以下。为解决“AI 被金融排斥”的问题,Coinbase 推出一键 prompt 即可为 agent 创建自托管钱包的工具,并孵化了 X402 协议(已捐赠给 Linux Foundation,Google、Cloudflare、AWS 等参与)。

Armstrong 认为未来会有比人类更多的 agents,专业化 agent 市场将兴起——小型开源模型经过 fine-tuning 后可在特定任务上超越 frontier 模型。他同时分享了 Coinbase 内部的“递归自我改进”实践:每个团队和服务都有“brain”(包含历史事故、金融控制、AB测试、PR记录),agent 修改代码时必须把人类修正反馈回 brain,从而持续提升一次性 PR 通过率。

在预测市场方面,Coinbase 上线数月即达到1亿美元年化收入,季度环比增长超100%。Armstrong 还谈到自己联合创办的抗衰老公司 New Limit,聚焦表观遗传重编程,已在人源化小鼠模型上成功重编程至少一种人类细胞类型,计划明年启动一期临床试验,首批针对肝细胞、血管细胞和免疫细胞。

“我们不想让 AI 成为金融难民。如果要为 AI 提供银行服务,它们也理应拥有金融服务。”
The Takeaway: AI agents will soon have their own financial accounts, the agentic economy will outgrow the human economy, and crypto rails are the essential infrastructure to make it work.

Coinbase co-founder and CEO Brian Armstrong outlines three major bets: the Everything Exchange (stocks, commodities, crypto, derivatives, and prediction markets all on-chain), stablecoin payments (global transfers under one second and under one cent), and Agentic Finance (AIFi). He notes that traditional card networks impose a 30-cent minimum fee, rendering sub-dollar transactions impractical—yet 76% of current agent e-commerce transactions sit under 30 cents. To prevent AIs from being “unbanked,” Coinbase ships a single-prompt tool that gives any agent a self-custodial wallet, and incubated the X402 protocol now under the Linux Foundation with Google, Cloudflare, and AWS contributing.

Armstrong expects more agents than humans in the near future and a marketplace of specialized agents where smaller open-weight models, fine-tuned on proprietary data, outperform frontier models on narrow tasks. Internally, Coinbase is pushing recursive self-improvement: every team and service maintains a “brain” of incidents, financial controls, A/B tests, and PR history. When an agent proposes a change, human corrections are written back into the brain so the system improves over time.

Prediction markets inside Coinbase hit a $100M revenue run-rate within months of launch and grew over 100% quarter-over-quarter. Armstrong also discusses New Limit, the longevity company he co-founded, which uses AI-driven epigenetic reprogramming. It has already reprogrammed at least one human cell type in humanized mouse models and plans a Phase 1 trial next year targeting hepatocytes, vasculature, and immune cells.

“We don’t want the AIs to be unbanked. If we want to bank the AIs, they deserve financial services as well.”
查看原文 →

Peter Yang:个人 Agent 赛道的“扶手椅”分析Peter Yang’s Armchair Take on the Personal Agent Race

产品经理与 AI 教程作者 Peter Yang 给出个人 agent 赛道的清晰排序:Meta 的 Muse 最有可能成为领导者,因为产品直观且全平台强力推广,一旦体验过就觉得 agent 不该活在现有消息应用里;ChatGPT 仍凭10亿用户保持领先,但工作与个人场景难以用同一套 UX 兼顾;Grok Bot 更聚焦工作场景,有潜力成为“多玩家 agentic Slack”;Google 是黑马,因为所有 agent 都依赖 Gmail 和 Calendar 数据,但目前 Spark 仍是 Gemini 的次要入口;真正的下一波解锁是多玩家协作,目前没人真正做好。

他的个人栈是:Muse 处理私人事务,Grok Bot 处理云端任务,Claude 处理特定场景,ChatGPT 兜底。他建议所有人从现在开始把个人 skills 和 files 设计成可轻松迁移的形式,否则未来一半时间都会花在换工具上。
Product leader and AI educator Peter Yang ranks the personal-agent race: Meta’s Muse is best positioned to win because the app feels native and Meta is pushing it hard across every surface—once you try it, agents living inside existing messaging apps start to feel wrong. ChatGPT still leads on sheer distribution (1B+ users) and model quality, but serving both work and personal use with one UX is hard. Grok Bot is work-focused and could become multiplayer agentic Slack. Google is the dark horse because every agent lives off Gmail and Calendar data, yet Spark remains a secondary tab inside Gemini. The real next unlock is multiplayer: today almost no one can easily loop a spouse or coworker into an agent thread.

Yang’s current stack: Muse for personal, Grok Bot for cloud tasks, Claude for specific jobs, ChatGPT for everything else. His advice: design personal skills and files so they port cleanly between harnesses, or half your time will be spent migrating.
查看原文 →

🛠️ 开发者工具与技巧

Vercel CEO Guillermo Rauch:Agent 的 QA 强度人类永远无法匹敌Vercel CEO Guillermo Rauch: Agents’ QA Thoroughness Humans Can Never Match

Vercel CEO Guillermo Rauch 分享了一次真实体验:他指出一个移动端应用内浏览器渲染问题后,agent 深入到“Mordor 的深处”去复现、模拟、修复、部署并验证,甚至创建了临时 Vercel 部署并扔进 iPhone 模拟器。人类根本无法维持这种强度——我们会耗尽精力、会外推、会寄希望于“应该能行”。未来的软件质量与性能将远超我们现在的认知,而这个转变正在眼前发生。
Vercel CEO Guillermo Rauch describes watching an agent hunt down a mobile in-app browser rendering bug. The agent went to the depths of Mordor—reproducing, simulating, fixing, deploying, and verifying—creating an ephemeral Vercel deployment and throwing it at an iPhone simulator. Humans simply cannot match that intensity: we run out of energy, we extrapolate, we hope it works. The software of the future will reach a quality and performance level we have never experienced, and the shift is already visible.
查看原文 →

Peter Steinberger:Claw 现在可以 FaceTime 你了Peter Steinberger: Your Claw Can Now FaceTime You

OpenClaw 与 OpenAI 相关开发者 Peter Steinberger 宣布两个更新:Claw 现在支持 FaceTime 通话,同时放出了新的 benchmark。这些进展继续把 agent 从文本交互推向更自然的多模态与实时协作形态。
OpenClaw and OpenAI-linked builder Peter Steinberger shipped two updates: Claw can now FaceTime you, and a new benchmark dropped. Both moves push agents further beyond pure text into more natural multimodal and real-time collaboration territory.
查看原文 →查看原文 →

🌍 其他动态

Box CEO Aaron Levie:对 Agent Swarm 几乎无防御Box CEO Aaron Levie: Literally Impenetrable from Agent Swarms

Box CEO Aaron Levie 引用并强调一个观点:当前系统在面对 agent swarm 时几乎无防御。这句话迅速引发大量讨论,反映了业界对 agent 协作攻击面快速扩大的真实担忧。
Box CEO Aaron Levie amplified the observation that current systems are “literally impenetrable from agent swarms.” The quote struck a nerve, highlighting growing industry concern about the expanding attack surface created by coordinated agent activity.
查看原文 →

Thariq:电子游戏给年轻人高风险与责任感Thariq on Video Games Giving Young People High Stakes and Responsibility

Anthropic Claude Code 工程师 Thariq 指出,电子游戏给年轻人提供了现实中难以获得的高风险与责任感,但最终仍需“毕业”回到真实世界。这个观察在 AI 时代尤其有共鸣——模拟环境正在成为训练复杂决策能力的重要场域。
Anthropic Claude Code engineer Thariq notes that video games uniquely give young people high-stakes responsibility they rarely get in real life—yet at some point you still have to graduate. The observation lands especially hard in the AI era, where simulated environments are becoming key training grounds for complex decision-making.
查看原文 →