HackerNews中文版

### 停止构建 GPT 包装器。相反，构建一个世界模型。 --- *Foundry 存在的原因：* 每个没有 API 的 SaaS 应用、企业工具和内部工作流程都将很快被浏览器代理自动化——而不是传统的 API 集成。数十万亿美元的全球劳动依赖于目前没有任何 AI 系统能够可靠自动化的基于浏览器的工作流程。但现在的浏览器代理就像 GPT-1：令人印象深刻的演示，但在实际应用中脆弱且不可靠。要将浏览器代理从 GPT-1 的演示扩展到 GPT-4 的可靠性——并解锁这一巨大的经济飞跃——它们需要专门构建的基础设施： - *现实的网页模拟*：能够模拟复杂任务、模糊的用户界面和非确定性的交互。 - *强大的强化学习训练循环*：能够有效地从稀疏奖励和真实用户行为中学习。 - *严格的评估基准*：与现实世界的结果和真实的商业影响相关联。还没有人构建过这样的系统。我们将率先行动。我们正在构建的正是这一点——从基本原则出发专为浏览器代理设计的核心基础设施。没有增量工具，没有围绕 GPT 的包装——只有每个人将来都会依赖的基础 AI 堆栈。我们的创始团队来自 Scale AI 和 Meta。获得种子资金，总部位于旧金山（全职在场）。 --- *我们正在寻找：* 一位在软件工程方面表现出色的创始工程师——但现在希望从零开始构建核心的机器学习系统和强化学习基础设施。这不是另一个提示工程的工作。 *你将会：* - 构建一个浏览器代理训练环境——稀疏奖励、非确定性、混乱的 DOM 树。 - 每天解决生产边缘的困难机器学习问题。 --- *你可能：* - 是一位优秀的工程师（Python, TypeScript），职业生涯早期（0-2 年工作经验）。 - 交付速度极快——实习、侧项目、开源软件、真实研究——你自己构建的东西。 - 具备真实的机器学习经验：训练模型、构建基础设施、进行实验。 - 有强大的开源贡献、我们会认可的出版物，或两者兼具。 - 渴望挑战，想要提升自己，想要构建一些真正新颖的东西。 --- *加分项：* - 你曾构建过强化学习代理、浏览器自动化（Puppeteer, Playwright）或评估工具。 - 你厌倦了包装 API 和微调 HuggingFace 模型——准备做一些更深入的事情。 --- *加入的理由：* - 真实的机器学习基础设施，真实的用户，真实的技术复杂性。 - 早期股权、具有竞争力的薪酬和巨大的成长潜力。 - 总部位于旧金山，每周 5 天，全职在场。我们正在构建核心基础设施。其他人将来会使用的机器学习堆栈——这就是我们今天要交付的内容。 --- 如果你读到这里，你可能明白了。直接联系——让我们一起构建。 [点击这里查看职位](https://www.ycombinator.com/companies/foundry/jobs/uwi8b6I-founding-engineer-ml-x-swe) 邮箱：manil@foundryrl.com

查看原文

### Stop building GPT wrappers. Build a world model instead.---*Why Foundry exists:*Every SaaS app, enterprise tool, and internal workflow without an API will soon be automated by browser agents—not traditional API integrations. Tens of trillions of dollars in global labor depend on browser-based workflows that no AI system can reliably automate today.But browser agents right now are like GPT-1: impressive demos, but brittle and broken in practice. To scale browser agents from GPT-1 demos to GPT-4 reliability—and unlock this enormous economic leap—they need purpose-built infrastructure:- *Realistic web simulation*: capable of modeling complex tasks, ambiguous UIs, and non-deterministic interactions. - *Robust RL training loops*: that learn effectively from sparse rewards and authentic user behavior. - *Rigorous evaluation benchmarks*: tied to real-world outcomes and genuine business impact.No one's built it yet. We're going first.We’re building exactly this—core infrastructure designed from first principles, specifically for browser agents. No incremental tools, no wrappers around GPT—just the foundational AI stack everyone else will depend on later.Our founding team is ex-Scale AI & MetaSeed-funded, SF-based (fully in-person)---*What we’re looking for:*A founding engineer who’s cracked at software engineering—but who now wants to ship core ML systems and RL infrastructure from scratch.This is not another prompt-engineering job.*You’ll:*- Build a browser agent gym—sparse rewards, non-determinism, messy DOM trees. - Solve hard ML problems at the edge of production—every single day.---*You probably:*- Are a killer engineer (Python, TypeScript) early in your career (0–2 YOE). - Ship insanely fast—internships, side projects, OSS, real research—things you’ve built yourself. - Have real ML exposure: trained models, built infra, ran experiments. - Have strong OSS contributions, publications we’d recognize, or both. - Want a grind. Want to get better. Want to build something genuinely novel.---*Bonus points:*- You’ve built RL agents, browser automation (Puppeteer, Playwright), or eval tooling. - You’re tired of wrapping APIs and fine-tuning HuggingFace models—ready to do something deeper.---*Why join:*- Real-world ML infrastructure, real users, real technical complexity. - Early equity, competitive pay, and massive growth trajectory. - SF-based, 5 days/week, fully in-person.We're building core infra. The ML stack everyone else will use later—that’s what we’re shipping today.---If you’ve read this far, you probably get it.Reach out directly—let’s build.<a href="https://www.ycombinator.com/companies/foundry/jobs/uwi8b6I-founding-engineer-ml-x-swe">https://www.ycombinator.com/companies/foundry/jobs/uwi8b6I-f...</a>Email at manil @ foundryrl dot com

Foundry（YC F24）正在招聘 – 来为网络构建一个世界模型吧