🌐 双语
Archive

AI Builders
Digest

2026-11-16 11 builders · 19 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Robinhood CEO Vlad Tenev:用所有权对抗AI时代的集中化Robinhood CEO Vlad Tenev on Ownership as the Antidote to AI Centralization

The Takeaway:在AI快速重塑经济的时代,真正能保持社会稳定的,不是更强的模型,而是更广泛的所有权——让普通人也能拥有公司、资产和数学发现的工具。

Robinhood创始人兼CEO Vlad Tenev把公司的使命总结为一个词:ownership。从零佣金炒股起步,Robinhood已演变为覆盖120多个国家的金融集团,核心逻辑始终是让更多人成为资产所有者。Tenev认为,个人拥有资产是通往财务自由的最可靠路径,而从社会层面看,没有广泛所有权的社会是脆弱的。当人们觉得自己在AI变革中毫无利益,就容易产生对抗。

他对比了crypto与AI的所有权结构:crypto从一开始就是个人持有为主,因此能形成强大的政治动员力;而AI公司大多仍是私人持有,个人投资者几乎无法在早期参与。为此Robinhood推出了Ventures基金(公开交易的VC结构)和私有公司代币化,目标是让普通人能像买卖股票一样,以小额、7×24小时交易Anthropic或OpenAI这类公司的早期权益。

在交易侧,agentic trading已展现清晰的product-market fit。自推出以来已有超过15万用户开设agent账户,日均工具调用量达到约3000万次。Tenev的愿景是把对冲基金和量化交易公司的能力,装进普通人的口袋,让策略编写和执行的门槛接近“买一只股票”那么低。

数学方面,他联合创办的Harmonic去年用Aristotle模型在IMO拿到金牌,并开放给用户解决未解问题。Tenev明确反对AI实验室用秘密模型和超级算力垄断科学发现:“一个由OpenAI或Anthropic数据中心独自证明Navier-Stokes或黎曼假设的世界,是有点反乌托邦的。”他更相信把工具交到人手中,让普通人做出被庆祝的发现。

“Prediction markets are an evolution of news. You just can see what's happening and what's likely to happen in the future through the odds.” 在Tenev看来,所有权、代理交易和去中心化发现,最终都指向同一件事:让更多人真正持有这个正在被AI重写的经济。
The Takeaway: In an era of rapid AI-driven economic change, societal stability depends less on stronger models and more on broader ownership—giving ordinary people a real stake in companies, assets, and even mathematical discovery.

Robinhood founder and CEO Vlad Tenev distills the company’s mission to one idea: ownership. What began as a zero-commission stock trading app has grown into a financial conglomerate live in more than 120 countries. For individuals, owning assets remains the most reliable path to financial independence. For society, a world without broad ownership is fragile; people who feel they have no upside in the AI transition become its opponents.

Tenev contrasts crypto’s distributed ownership—which produced real political power—with AI’s concentrated private ownership. Most leading AI labs remain private, locking retail investors out of the early upside. Robinhood’s response includes publicly traded venture funds and tokenized private-company exposure, aiming for a future where anyone can trade fractional stakes in companies like Anthropic around the clock at low cost.

On the trading side, agentic trading already shows clear product-market fit. More than 150,000 customers have opened agentic accounts, with tool calls reaching roughly 30 million per day. The goal is to put hedge-fund and quant capabilities into ordinary users’ hands so that writing and running a strategy becomes as frictionless as buying a single stock.

In mathematics, Tenev’s co-founded startup Harmonic produced Aristotle, one of three models to earn a gold medal at the International Math Olympiad, then released it for public use on unsolved problems. He rejects the centralized model of secret labs monopolizing discovery: a world in which OpenAI or Anthropic data centers alone prove major theorems is “a bit of a dystopian outcome.” He prefers putting the tools in people’s hands so that discoveries by ordinary users are celebrated rather than resented.

“Prediction markets are an evolution of news. You just can see what's happening and what's likely to happen in the future through the odds.” For Tenev, ownership, agentic trading, and decentralized discovery all point to the same outcome: giving more people a genuine stake in the economy AI is rewriting.
查看原文 →

OpenAI Codex团队修复「model at capacity」错误OpenAI Codex Team Fixes Persistent “Model at Capacity” Errors

OpenAI Codex & ChatGPT团队成员Thibault Sottiaux表示,长期困扰用户的“model at capacity”错误可能已被Codex团队真正修复。该问题此前频繁出现,影响用户体验,此次更新有望显著改善稳定性。
OpenAI Codex & ChatGPT team member Thibault Sottiaux reported that the long-standing “model at capacity” errors plaguing users may finally have been fixed by the Codex team. The issue had been a frequent source of friction; the update should meaningfully improve reliability.
查看原文 →

🛠️ 开发者工具与技巧

Vercel CEO Guillermo Rauch推出xeet:毫秒级X帖子渲染器Vercel CEO Guillermo Rauch Launches xeet: Millisecond X Post Renderer

Vercel CEO Guillermo Rauch发布了xeet——一个用安全Rust/WASM实现的hermetic、像素级精确的X帖子渲染器。它无需浏览器、网络、文件系统或系统字体,即可在约1毫秒内将帖子渲染为PNG(Jack的“setting up my twttr”仅需0.4毫秒)。xeet复刻了Blink与HarfBuzz的文本整形、1/64px换行、RTL双向文本、28种Chirp字体子集以及CJK/Math/Emoji回退等细节,比启动浏览器截图快100-1000倍。项目采用“autorepo”架构:一个oracle启动Chrome验证数万条文本布局规则,CI canary持续跟踪原版变化并提交签名PR供agent处理,整体围绕verification engineering设计。
Vercel CEO Guillermo Rauch introduced xeet, a hermetic, pixel-perfect X post renderer built in safe Rust/WASM. With zero browser, network, filesystem, or system fonts, it renders a post to PNG in ~1 ms (0.4 ms for Jack’s “setting up my twttr”). It reproduces Blink & HarfBuzz text shaping, 1/64px line breaking, RTL bidi, 28 Chirp font subsets, and CJK/Math/Emoji fallbacks—100x–1000x faster than browser screenshots yet equally accurate. Structured as an “autorepo,” an oracle boots Chrome to verify tens of thousands of layout rules while a CI canary watches the original and files signed PRs for agents. The system was designed from the ground up around verification engineering.
查看原文 →

Dan Shipper:用autoresearch打造可复利的context window总结技能Dan Shipper on Compounding Skills via Autoresearch for Context Summaries

Every CEO Dan Shipper分享团队在“compounding skills”(每次使用都会变得更好的技能)上的进展。他本周末用autoresearch迭代出一个帮助撰写context window总结的skill,最终达到自己审美标准的80%(相对无skill基线)。完整写作为agent设计,已发布在Every frontier feed。
Every CEO Dan Shipper shared progress on compounding skills—skills that improve with every use. Over the weekend he used autoresearch to iterate a skill for writing context-window summaries until it reached 80% of his taste versus a no-skill baseline. The full write-up, aimed at agents, is live on the Every frontier feed.
查看原文 →

Anthropic设计师Nate Parrott:如何避免Claude的默认设计偏见Anthropic Designer Nate Parrott on Breaking Claude’s Default Design Bias

Peter Yang采访Anthropic的Nate Parrott时,对方指出Claude在全新context下容易陷入固定审美:衬线字体、米白背景、橙色、卡片边缘条。因为它“只做过一次”,就以为自己很原创。Anthropic的Claude Design prompt会明确告诉模型“你容易过度偏向这两种字体和颜色,除非用户要求否则不要用”,效果显著。Parrott还提到,在Anthropic做原型的周期已从Browser Company时期的数周压缩到1-2天,但“把产品交到用户手里并理解他们是否真正使用”的周期并没有被压缩。
In conversation with Peter Yang, Anthropic’s Nate Parrott explained why Claude defaults to the same visual style: serif text, off-white background, orange, and edge bars on cards. In a fresh context the model thinks it has invented something original. Anthropic’s Claude Design prompt counters this by telling the model it tends to over-bias toward those fonts and colors and should avoid them unless asked. Parrott also noted that prototyping cycles at Anthropic have compressed from weeks (at The Browser Company) to one or two days, yet the cycle of putting work in users’ hands and learning whether they actually use it has not been compressed.
查看原文 →查看原文 →

🌍 其他动态

AI模型开始擅长给东西起名AI Models Are Getting Surprisingly Good at Naming Things

Vercel CEO Guillermo Rauch观察到AI模型在命名上已经相当出色。他写了一个诊断系统变慢的小CLI,模型给它起名“hog”,他很喜欢。结合模型在cache invalidation上的能力,他半开玩笑说计算机科学最难的问题或许已经被解决了。
Vercel CEO Guillermo Rauch noted that AI models have become genuinely good at naming. He built a small CLI to diagnose system slowdowns; the model named it “hog,” which he loves. Given their existing strength at cache invalidation, he joked that computer science’s hardest problems may finally be solved.
查看原文 →