与「Microsoft」相关的搜索结果

Cloudflare 要自己做 CA 了～ Cloudflare 宣布将申请成为公共证书颁发机构（CA），已向 Chrome、Apple、Microsoft、Mozilla 四大根证书计划提交申请。

这个开源项目系统整理了 35 家 AI 公司的 AI 工程师面试题，全部来自 “公开报告的面试经历” 来自 @outcome_school 团队 @pallavishekhar_ 开源发布，内容几乎涵盖了 OpenAI、Anthropic、DeepMind、xAI、DeepSeek、Kimi、GLM 等 AI Labs，Cursor、Cognition、ElevenLabs 等 AI Native 团队和 Nvidia、Microsoft、Amazon、Apple 等头部大厂。 覆盖的岗位头衔也非常多：AI Engineer、LLM Engineer、Gen AI Engineer、ML Engineer、Research Engineer、Applied Scientist、FDE、MLOps/LLMOps 工程师等。 开源地址： # 内容架构：一个精心设计的双层结构 第一层：跨公司通用题（Common Questions）。 作者把在多家公司反复出现的题目只列一次，标注"Asked at"哪些公司，按十大主题组织： 1. LLM 内部机制与架构 — attention 缩放因子、KV cache 内存公式推导、MQA/GQA/MLA、FlashAttention、BPE、RoPE/YaRN、Chinchilla scaling laws、MoE、解码采样策略、lost-in-the-middle、RMSNorm、SwiGLU 2. 推理、服务与 GPU 性能 — prefill vs decode、continuous batching、PagedAttention、投机解码、量化（FP16→FP4）、五种并行策略、TTFT/TPOT 指标、H100 上的 roofline 计算、vLLM/SGLang/TensorRT-LLM 选型、“如何把服务成本降 10 倍” 3. RAG 与检索 — 分块策略、BM25 vs 稠密检索、重排序器、HyDE、权限感知检索、ANN 索引、索引新鲜度、答案归因 4. Agent 与工具调用 — ReAct、MCP、工具 schema 设计、多智能体编排、Agent 记忆、循环终止条件、人类审批 5. 微调与对齐 — RLHF/DPO/GRPO/RLVR 全谱系、LoRA/QLoRA 数学、灾难性遗忘、“提示 vs RAG vs 微调”决策框架、蒸馏、reward hacking 6. 评估与可观测性 — LLM-as-judge 及其偏差、幻觉检测、基准污染、Agent 评估、回归门禁 7. 安全与负责任 AI — 提示注入（直接/间接）、OWASP LLM Top 10、护栏、Constitutional AI、红队 8. 多模态与语音 — VLM、语音 Agent 延迟预算、barge-in 打断处理、级联 vs 端到端语音、ASR/TTS 评估 9. AI 系统设计 — 十类高频设计题（千万级文档企业 RAG、代码助手、客服 Agent、Text-to-SQL、LLM 网关、数亿用户聊天服务等） 10. 编码题 — 从零实现 attention、KV cache、BPE、采样；LRU 缓存、令牌桶限流器、异步批处理器、SSE 流解析器、最小 Agent 循环 第二层：35 家公司的专属章节，分为五大梯队： 1. 前沿实验室（12 家）：Anthropic、OpenAI、Google DeepMind、Meta、xAI、Mistral、Cohere、DeepSeek、月之暗面（Kimi）、智谱（GLM）、阿里（Qwen）、Sarvam AI（印度） 2. 大厂 AI 组织：Microsoft、Amazon、Apple、NVIDIA、Tesla，以及一组消费级 ML 公司（Uber/Netflix/LinkedIn/Airbnb/Pinterest/Spotify） 3. AI 基础设施公司：Databricks、Groq、Together AI、Hugging Face、Scale AI、Perplexity 4. AI 原生产品公司：Cursor、Cognition（Devin）、Sierra、Harvey（法律）、Glean、 AI（机器人）、Waymo 5. 前向部署/企业 AI：Palantir # 题目分布透露的行业信号同样值得关注 1. 公司的差异化考察方向，和它的商业模式严丝合缝。 这是最能体现整理功力的地方： · DeepSeek、月之暗面、智谱、Qwen 的题目深度绑定自家论文——MLA、auxiliary-loss-free 负载均衡、Multi-Token Prediction、MuonClip、DualPipe、长上下文扩展、GLM 的 thinking 模式。面试这些公司等于面试它们的论文，还要求 PyTorch 从零实现 MoE 路由。 · Groq 的题全是 SRAM-only 架构下的 roofline 重推演：“没有 HBM，decode 的 roofline 论证哪里变了”、“确定性在 p99 层面到底买到什么”。 · Apple 清一色端侧：3B 模型在手机上跑、PTQ vs QAT、不采集用户内容的前提下用设备信号改进模型、30+ 语言区无法记录用户内容的评估方案。 · CharacterAI 是推理经济学：“我们的服务成本被 KV cache 而非权重主导，降一个数量级，代价是什么”。 · Harvey（法律）和 Abridge（医疗） 考的是领域约束下的工程：200 页信贷协议里第 140 页的条款依赖第 8 页的定义术语怎么检索、生成的病历中出现了患者没提过的药怎么当作安全事故处理、PHI 如何约束整个架构。 · Palantir 的招牌是 "decomposition" 轮：把“一家货运铁路公司每年因机车非计划停机损失数千万”分解成工程计划。 2. 编码轮的形态正在发生实质性变化。 文档里反复出现的一类题，与传统 LeetCode 明显不同： · “实现一个内存 KV 存储：先 SET/GET/DELETE，再加事务 BEGIN/COMMIT/ROLLBACK，包括嵌套事务”（OpenAI、xAI 都问） · “给你一个 LLM 推理引擎的调度器类，其中一个方法是空壳，没有规格没有文档。说说你头三十分钟干什么”（xAI） “重构这 120 行能跑但很乱的代码，不许破坏测试”（OpenAI） · “对 5 万个文档跑 LLM 调用，API 限 100 并发、偶发 429 和超时，把 Python 写出来”（Anthropic） 考察重心从算法记忆转向增量需求下的代码演进能力、并发正确性、真实工程约束下的取舍。 3. “AI 协作轮”作为新题型已经进入正式面试流程。 这是文档里最前沿的信号： · Anthropic 部分机器学习岗有 AI-collaboration 轮：现场给你 Claude，考察的是你如何指挥它和验证它的产出，而不是你自己写。 · Meta 2026 年的流程新增三阶段 AI 辅助编码轮（探索修复 → 实现新功能 → 扩展改进）。 · Cursor 的 onsite 是两天在真实 Cursor 代码库上做一个功能（或 8 小时远程版），明确考核你对 AI 工具的使用效率和自主 scoping 能力。 · Sierra 给你两小时和任意 AI 工具，看你选择做什么。 xAI 有四小时限时产品构建。 4. FDE 成为一级岗位类别。 Anthropic、OpenAI、Databricks、Scale、Together、Sierra、Harvey、ElevenLabs、Palantir 的章节里都有 "Applied and Forward-Deployed Scenarios" 专属题库——典型题目如“企业客户说 Claude 幻觉太多，你是驻场工程师，头 48 小时做什么”。这对应了 AI 公司向企业交付方式的转变：模型能力差距收窄后，落地能力成为差异化。 5. 硬核系统题的普及。 "H100 上 70B 模型 batch size 1 的 roofline 计算"、"估算 70B 模型的 GPU 显存（权重 + KV cache + 激活 + 碎片）"、"p99 延迟在部署后翻倍但模型没变，走一遍诊断”——这类题横跨 NVIDIA、Together、OpenAI、Perplexity 等多家，说明推理性能的量化直觉已成为 AI 工程师的通用素养，而非基础设施工程师的专属。

Stunning stat: "Anthropic’s investors are expecting the company to reach a valuation of $2tn when it goes public in the coming weeks. Add in SpaceX, which began trading at $2tn after its IPO in June, and OpenAI, which is considering raising money privately at $1.2tn ahead of a public listing next year, and these companies alone could be worth well north of $5tn. Now compare that with the entire history of IPOs from 1980 to 2025. The 3,365 tech companies that went public in that period were worth a combined $4.1tn when they started trading." While I've written much about AI's transformative potential across sectors, I've always been dubious about how much value hyperscalers can seize enabling that transformation. As I dissect in my recent report on "The AI Trade" ( it's not about user acquisition. OpenAI claims to have over one billion active users across its services and two million businesses using its AI models. Anthropic has claimed to have more than 300,000 business customers. The question is not whether they can bring users to their services, but rather how much average revenue they can generate per customer relative to the price of building and maintaining their models. The cost of compute is inflating at the same time competition is depressing token pricing power. That's a precarious dynamic when so much hinges on the success of two companies. To again quote the report: "It’s hard to overstate how much hinges on these IPOs. As mentioned in the Executive Summary, OpenAI and Anthropic will account for 13% of AWS revenue and 27% of Google Cloud revenue this year. As for Microsoft, OpenAI alone accounts for roughly 70% of its AI-specific revenue ($24.1 billion out of an estimated $34 billion total for the fiscal year ending in June 2026). OpenAI has committed to tens of billions of spending on chips from the likes of Nvidia and Broadcom. Deepening circularity concerns, tech giant earnings growth has been increasingly driven by paper gains in the private-market valuations of Anthropic and OpenAI. In Q2, Amazon, Alphabet, Nvidia, Meta, and Microsoft reported $160 billion in cumulative “other income”, trouncing the $69 billion in “other income” reported in Q1. To quote the FT: “These one-off valuation boosts, derived in large part from enthusiasm around AI, risk distorting the financial picture at a time when investors are closely scrutinizing tech earnings.” If either Anthropic or OpenAI stumble in their IPOs, it’ll hit tech giants on multiple balance-sheet fronts and likely ripple through the US and global economy. A pin-prick popping of the AI bubble may not be our base case, but if a pin is out there, it’s likely the revenue versus spending trajectories of Anthropic and OpenAI." FT link:

We have a limited window to strengthen cyber defenses, and together with organizations including @AnthropicAI, @awscloud, @Google, @Microsoft, and @Oracle, we're calling for a global effort to give defenders the tools, resources, and support to protect the infrastructure we all depend on. If we act decisively, we can turn today's AI advances into lasting improvements in security and make our digital world safer for everyone.

BREAKING: Grok 4.6 from SpaceXAI is now available in Microsoft Foundry Models. 🤖⚡ Organizations can now build with Grok 4.6 on Foundry, compare frontier AI models, run workload-specific tests, deploy managed endpoints, and use enterprise security and governance controls. Grok 4.6 is expanding further into the enterprise AI ecosystem.

Nvidia acquires Hugging Face for ~$12.9bn. But who gets the 💰? My usual breakdown below 👇 *Investors are sharing ~$7.3bn of profits on <$400m invested. A cool ~20x blended.* *1) The single biggest winner here? Lux Capital.* 👑 Lux is taking the crown for the most $ returned in this transaction. The two rounds they led (Series A & Series C) generated ~$2.6bn combined. They also participated in the Series B. Congrats to @wolfejosh and team. The $15.0m Series A alone will return ~132.5x the capital invested, or ~$2.0bn. The most lucrative round in absolute $ returned. *2) Backing an exceptional team paid off with a >1,000x return for Betaworks and the angels* 📈 The story few people know: Hugging Face started life as a chatbot for lonely teenagers... You sent it a selfie and a sad emoji, it told you it understood. Quite far from where it landed! That is why backing an exceptional team in @ClementDelangue, @Thom_Wolf and @julien_c is always the right thing to do. It paid over 1,000x for Betaworks, Kevin Durant and Thibaud Elziere. *3) Founders and team are crushing it: not taking too much dilution along the way paid off* 👏 I estimate they will share ~$5bn in proceeds, assuming no secondary was taken along the way. A life changing outcome for them and many employees. Taking a 15% attributed option pool at exit, this would mean ~$2bn for the team and >$3bn to share for the three co-founders. Legendary! *4) Salesforce Ventures is cementing its reputation as one of the very best corporate VCs.* 🥇 Salesforce is adding Hugging Face to its long list of successful exits: Snowflake, Zoom, DocuSign, nCino, Auth0... Very few investors can claim as many successes, let alone corporates. The $235m round they led returned $675m in just three short years. Meanwhile, Salesforce's own stock has been flat for the period. Capital well invested! *5) This is a little bit of a heartbreak: but I wanted to congratulate @joinstationf here:* Hugging Face walked into StationF as a 4-person team on day one, June 2017. Back when it was a chatbot for teenagers. They went through multiple corporate programs (Microsoft, Ubisoft, then Naver). Unfortunately they were not investing or taking equity at the time - but they contributed to making it possible. And so for that: hats off. And of course, congratulations to all others involved: Addition, Sequoia, Coatue and many others. No one will be left out!

BREAKING: Grok 4.6 from SpaceXAI is now available in Microsoft Foundry Models. Build with Grok 4.6 on Foundry, where organizations can compare frontier AI models, run workload-specific tests, deploy managed endpoints, and operate with enterprise security and governance controls.