Ace System Design at OpenAI, Anthropic, Google & xAI 攻克 OpenAI / Anthropic / Google / xAI 系统设计面试
A curated, deeply-researched, fully bilingual study hub. Browse a verified 真题 arena of real interview questions with sources, then dive into a comprehensive study guide synthesized from eight canonical books on distributed systems, ML infrastructure, and agentic AI. 一个精心策划、深度研究、完全双语的备考平台。浏览经核实的真题竞技场（含出处），然后深入学习由八本核心书籍（分布式系统、ML 基础设施、智能体 AI）综合而成的全面学习手册。
Know your target 精准定位面试风格
Each frontier lab interviews differently. Pick your track — or browse all 100 questions. 每家前沿实验室面试风格各异。选择公司——或浏览全部 100 道真题。
What's inside 内容概览
Four interlocking modules — use them linearly if you're new, or jump around by topic. 四个互补模块——按顺序学习或按需跳转，皆可。
真题 Arena
真题竞技场
100 real interview questions from OpenAI, Anthropic, Google and xAI — filterable by company, category, difficulty, and frequency. Each question links to its source (LeetCode, Blind, PracHub, Glassdoor, GitHub, 小红书, company eng blogs) and opens a deep solution page with architecture diagrams, APIs, data models, trade-offs, and expected follow-ups.
100 道 OpenAI / Anthropic / Google / xAI 真题，可按公司、类别、难度、频率过滤。每题附有出处链接（LeetCode、Blind、PracHub、Glassdoor、GitHub、小红书、各公司工程博客），点开是详细解题页：架构图、API、数据模型、权衡、追问清单。
→Study Guide
学习手册
20 deeply-synthesized topic notes organized into 6 tracks — Foundations, Distributed Systems, Classical Designs, LLM Systems, ML Systems, and Safety. Each note cross-references the 8 canonical books by chapter.
20 篇深度综合的专题笔记，分为 6 个方向：基础、分布式系统、经典题、LLM 系统、ML 系统、安全。每篇都精确引用 8 本权威教材的章节。
→Resources
资源集合
Books ranked by interview leverage. Top blogs (Chip Huyen, Eugene Yan, Anthropic/OpenAI engineering). Recommended courses (ByteByteGo, Educative, Hello Interview, Exponent) and GitHub repos.
按面试权重排序的书单。顶尖博客（Chip Huyen、Eugene Yan、Anthropic/OpenAI 工程）。推荐课程（ByteByteGo、Educative、Hello Interview、Exponent）与 GitHub 仓库。
→About & Roadmap
关于与路线图
Methodology, credibility scoring (S/A/B/C/D), a proven 8-week study plan, and the exact stack used to build this site — so you can fork, extend, and deploy your own.
方法论说明、可信度评级（S/A/B/C/D）、经过验证的 8 周备考计划，以及本站的完整技术栈——你可以直接 fork、扩展、部署自己的版本。
→Most-cited questions 最高频真题
Questions appearing in three or more independent reports — these are the ones you cannot afford to miss. 在三份以上独立面经中出现——这些是你必须拿下的题目。
Streaming tokens, prefill vs decode, KV cache, continuous batching, tail latency control, GPU memory. The canonical Anthropic system-design question. 流式 token、prefill 与 decode 分相、KV cache、连续 batching、尾延迟控制、GPU 显存。Anthropic 最经典的系统设计题。
Billions of requests, 24-hour retries, idempotency, DLQ, per-endpoint ordering, multi-tenant isolation. Interviewers drill into every component's internals. 十亿级请求、24 小时重试、幂等性、死信队列、每 endpoint 顺序、多租户隔离。面试官会对每个组件追问到内部实现。
Balance throughput vs latency SLOs. Flush policies (size, age, length-spread), overload handling, admission control, observability. 在吞吐与延迟 SLO 之间权衡。Flush 策略（大小/时间/长度方差）、过载处理、准入控制、可观测性。
Real-time messaging, channels, presence, delivery reliability, fan-out strategy. 2-week-MVP framing is a trap — scope ruthlessly. 实时消息、频道、在线状态、投递可靠性、fanout 策略。「2 周 MVP」是陷阱——必须果断砍需求。
What they actually evaluate 他们究竟在评估什么
Anthropic engineers publicly list 5 criteria. OpenAI cares about agency and scale. Both want evidence, not buzzwords. Anthropic 工程师公开了 5 条评分维度。OpenAI 看重主动性与规模思维。两家都要证据，不要口号。
1. Abstraction 1. 抽象能力
Can you see through the "AI wrapper" to the core infra problem? Most LLM questions reduce to queues, schedulers, storage.
能否穿透「AI 外衣」看到核心基础设施问题？多数 LLM 题本质是队列、调度、存储。
2. Trade-off articulation 2. 权衡表达
Latency vs throughput, sync vs async — with reasoning that cites SLOs, not vibes.
延迟 vs 吞吐、同步 vs 异步——推理必须引用 SLO，而非感觉。
3. Failure-mode reasoning 3. 故障模式推理
"What if the queue dies? What if load drops to zero? What about partial batch failures?" — proactively proposing these is a senior signal.
「队列挂了怎么办？请求突然归零怎么办？批内部分失败怎么办？」——主动提出这些，是高级工程师信号。
4. Scale reasoning 4. 规模推理
Designs must work under real constraints. "Just add more servers" is a disqualifier.
设计必须在真实约束下成立。「加服务器就行」是一票否决。
5. Driving the conversation 5. 主导对话
You set scope, pick the deep-dive, declare what to skip. Asking "what should I focus on?" = junior.
你来设定范围、选择深入点、声明跳过哪些。问「我该关注什么？」= 级别不够。
6. Safety (Anthropic gate) 6. Safety（Anthropic 门槛）
Every round has a safety lens. Treating it as a checkbox is a reject signal. Prepare 2–3 behavioral stories where you caught a safety issue early.
每一轮都有安全视角。把它当「勾选项」= 拒绝信号。准备 2-3 个「早期发现安全问题」的行为面试故事。