🌐 双语
Archive

AI Builders
Digest

2026-11-06 13 builders · 29 tweets · 1 podcasts · 0 blogs

🔥 热点话题

JEV 创始人谈决策模型、基准疲劳与实验室为何错过JEV Creator on Decision Models, Benchmark Fatigue & Why the Labs Missed It

The Takeaway:真正推动软件进步的是可靠性与实用性,而不是新形状或基准分数;开发者应重新掌握主动权,与模型共建可组合的机器原生智能。

Jogo(TypeSave / JEV 模型创始人兼 CEO)曾在 OpenAI 工作,如今以系统一式决策模型重新定义 AI 与软件的结合方式。他坚信未来绝大多数 LLM 调用将面向机器而非人类,因此需要从认知核心倒推,打造可组合、可靠、可验证的原语,而不是让开发者被聊天机器人或编码代理锁定。

他坦承 JEV 仍未达到理想状态,但坚持透明展示 jaggedness(参差不齐的能力),并呼吁开发者反馈弱点,共同把能力变成生产级工具。基准与 demo 已让人产生疲劳;真正有用的模型应像 SQL 一样可靠、提供商无关。他调侃实验室的最大优势或许是金融工程而非模型本身,并警告对 AGI 的无限估值可能像玛丽·安托瓦内特的“让他们吃蛋糕”。

“人们已经对基准和实际无用性之间的鸿沟产生了极端疲劳,速度比我预想的要快得多。”

他期待更多形状出现,但更在意效用;希望软件与模型共进化,让开发者重新获得 agency,而不是被动接受模型提供商的路线图。
The Takeaway: Reliability and real-world utility matter far more than novel shapes or benchmark scores; developers must reclaim agency and co-build composable, machine-native intelligence with models.

Jogo (founder and CEO of the company behind JEV decision models, formerly at OpenAI) argues that the future of AI is machine-native: the vast majority of LLM calls will be for computers, not humans. He works backward from that cognitive core, building system-one models optimized for discrete, reliable decisions rather than human-pleasing prose. JEV is not yet there, but the north star remains making intelligence trustworthy enough to run forever in production without constant human babysitting.

He rejects the idea that developers should form rigid “theories of mind” about which model is good for which task. Instead he wants models as reliable and provider-agnostic as SQL. Benchmarks and demos have burned goodwill so quickly that people already ignore new releases in this category. He wants developers to complain loudly about jaggedness so the field can fix it collaboratively.

On Anthropic’s valuation and the broader lab landscape he is measured: coding agents are valuable (he estimates a 20-40% average speed-up with large error bars), but much of the expectation baked into trillion-dollar numbers rests on misunderstandings of what current AI can actually do. “The let them cure cancer is going to be like Marie Antoinette’s let them eat cake.”

His final call is to untrain the learned helplessness LLMs have instilled and start building playful, composable software again.
查看原文 →

Claude 推出个人财务功能(美国 beta)Claude launches Personal Finance (US beta)

Anthropic 的 Claude 正式在美国 Pro / Max / Ultra 计划中推出 Personal Finance 功能(beta)。用户可选择连接自己的账户,Claude 用日常语言回答关于余额、交易等问题;访问权限为只读,无法移动、管理或投资资金,且没有任何广告或赞助推荐。

该功能强调安全与透明:它帮助你理解自己的钱,而不是提供投资建议。
Anthropic’s Claude is rolling out Personal Finance in beta for US users on Pro, Max, and Ultra plans. Connect the accounts you choose and Claude answers money questions in everyday language using your own balances and transactions. Access is read-only; Claude cannot move, manage, or invest money, and there are no ads or sponsored recommendations.

The feature is designed purely to help users understand their finances rather than give investment advice.
查看原文 →查看原文 →

OpenAI Codex 工程师:形式化证明与超安全基础设施已进入可行范围OpenAI Codex engineer: formal proofs and hyper-secure infrastructure are now realistic

OpenAI Codex & ChatGPT 团队成员 Thibault Sottiaux 指出,对数千万行代码进行形式化证明已不再是荒谬想法,在设计新系统时应认真考虑。他直言人类在编写安全基础设施方面并不擅长,最终状态将是人人可用的超安全基础设施与软件,但过程中会经历动荡。
Thibault Sottiaux (Codex & ChatGPT at OpenAI) states that formal proofs of tens of millions of lines of code are no longer ridiculous and should be considered when designing new systems. He adds that humans were never very good at writing secure infrastructure; the end state will be hyper-secure infrastructure and software for everyone, though the path there will involve turbulence.
查看原文 →查看原文 →

💰 创业成功案例

Replit:唯一的协作式 AI 编码系统 + 自定义内核挡住 96% 零日漏洞Replit: the only collaborative AI coding system + custom kernel blocks 96% of zero-days

Replit CEO Amjad Masad 强调 Replit 是目前唯一真正协作的 AI 编码系统。同时他公布新沙箱设计:通过自定义 Kernel 构建,Replit 能够避免 96% 的 Linux 零日漏洞,大幅提升生产环境安全性。
Replit CEO Amjad Masad highlights that Replit is the only collaborative AI coding system available. He also details a new sandbox design: a custom Kernel build lets Replit avoid 96% of Linux zero-day vulnerabilities, significantly hardening the production environment.
查看原文 →查看原文 →

Box 与 NVIDIA 合作:为 AI Agent 提供企业级上下文检索Box partners with NVIDIA for advanced context retrieval for AI agents

Box CEO Aaron Levie 预测企业将拥有比人类多 100 倍的 AI Agent。关键在于让这些 Agent 安全地获取非结构化数据(合同、研究材料、设计文件等)中的上下文。Box 与 NVIDIA 合作,将词法与语义搜索结合,在严格遵守企业安全、访问控制与数据治理的前提下,为 Agent 提供最相关的信息。
Box CEO Aaron Levie forecasts 100× more AI agents than humans inside the enterprise. Success hinges on giving those agents secure access to the critical context locked in unstructured data—documents, contracts, research, designs. Box is partnering with NVIDIA to deliver advanced context retrieval that combines lexical and semantic search while fully respecting enterprise security, access controls, and data governance.
查看原文 →

Vercel:从 X 订阅到 Bot/Cursor 构建、部署、变现的完整闭环Vercel: the complete loop from X subscription to Bot/Cursor build, deploy, and monetize

Vercel CEO Guillermo Rauch 描绘了一个几乎完整的创业循环:获取 X 订阅 → 用 Bot/Cursor 构建 → 部署到 Vercel(从 0 扩展到月球)→ 在 X 上发布与分发 → 用 X Money / Creators 变现。他还展示了 SpaceXSI 与 Vercel 的合作,以及 Vercel Schedules(运行时可定义、无限扩展、沙箱外的增强版 Cron)如何让 Agent 在夜间巩固记忆、早晨提醒、按天触发任务。
Vercel CEO Guillermo Rauch outlines a nearly complete founder loop: get an X subscription → build with Bot/Cursor → deploy to Vercel (scales from 0 to the moon) → launch and distribute on X → monetize with X Money/Creators. He also highlights the SpaceXSI partnership and Vercel Schedules—runtime-defined, infinitely scalable, sandbox-external cron on steroids that let agents consolidate memories at night, remind users in the morning, and message after N days of inactivity.
查看原文 →查看原文 →查看原文 →

🛠️ 开发者工具与技巧

Agent 实用提示:基于位置的洗衣提醒Practical agent prompt: location-based laundry reminders

前 Claude Code 工程师 Thariq 分享了他最喜欢的 Agent 提示之一:基于位置的提醒。例如“跟踪我上次洗衣的时间。如果超过 5 天,在工作日我到家 30 分钟后提醒我,并持续催促直到我确认或推迟。”这类长期、有状态的记忆与触发是当前 Agent 最有价值的杀手级应用。
Former Claude Code engineer Thariq shares one of his favorite agent prompts: location-based reminders. Example: “Keep track of when I’ve done my laundry. If it’s been more than 5 days, remind me 30 minutes after I get home on a weekday and bug me until I confirm or delay.” Long-horizon, stateful memory and triggers remain among the highest-value uses of agents today.
查看原文 →

YC 总裁:高金融长周期基准是通往 AGI 的路径之一YC President: long-horizon PE/IB benchmark is on the path to AGI

Y Combinator 总裁兼 CEO Garry Tan 指出,首个针对私募股权/投行工作的长周期基准已经出现。他认为解决高金融问题是通往 AGI 的路径之一。同时他呼吁两党合作,针对大科技滥用市场力量采取行动,为下一代创始人打开市场,让 AI 红利广泛扩散而非进一步集中。
Y Combinator President & CEO Garry Tan notes the arrival of the first long-horizon benchmark for private-equity and investment-banking work, arguing that solving high finance sits on the path to AGI. He also calls for bipartisan action against Big Tech’s abuse of market power so that thousands of startups can compete, lower costs, and spread AI gains widely rather than concentrating more power in the companies that already control distribution.
查看原文 →查看原文 →

🌍 其他动态

其他值得关注的短暂动态Other brief but notable updates

Aditya Agarwal(SPC 普通合伙人)提醒创始人:有人愿意以某个估值投资,并不意味着那就是公司的正确估值。Peter Steinberger 公开质疑 Apple 关于 Safari 性能与兼容性的宣传,称实际体验仍远落后于 Chrome。Swyx 澄清近期与 Sequoia 的沟通误会,强调与更广泛的重组无关。
Aditya Agarwal (General Partner at SPC) reminds founders that just because someone is willing to invest at a certain valuation does not mean it is the right valuation for the company. Peter Steinberger questions Apple’s claims about Safari performance and compatibility, arguing the real-world experience still lags Chrome significantly. Swyx clarifies a recent communication gap with Sequoia was not related to the broader restructuring announced this week.
查看原文 →查看原文 →查看原文 →