Skip to content

All updates

Today 82 items
Nov 3Tue
  1. Hacker News · AI65

    Survey of 30 Silicon Valley Tech Teams: The Evolution of Technical Hiring and Interviewing in the AI Era

    Two former engineering leaders surveyed recruiting and engineering heads across 30 organizations, including Anthropic, Stripe, and GitHub, releasing a comprehensive report on technical hiring in the AI era. Findings show that generative AI-assisted fraud accounts for up to 80% of remote job applications, rendering traditional LeetCode-style algorithm tests increasingly obsolete. Leading teams are shifting their evaluations toward system design, building projects from scratch live with AI tools allowed, reviewing AI-generated code, and debugging complex sandbox environments to assess candidates' architectural judgment and practical execution skills.

    Why it matters: The report details how Silicon Valley tech teams are redesigning their interview processes after moving away from traditional coding puzzles, offering a practical reference for assessing genuine engineering judgment.

  2. TechCrunch AI59

    建站平台 Squarespace 宣布裁员约 19%Chinese

    网站托管与建站平台 Squarespace 确认将裁减约 19% 的全球员工（涉及约 300 人），以简化运营并推动下一阶段增长。在提示词建站工具冲击传统市场的背景下，该公司正面临来自各类新型 AI 工具的竞争，目前正在筹备推出其 AI 原生平台 Forma。

  4. Hacker News · AI63

    Interview with Hercules Founder Brendan Falk: How AI Is Reshaping Software Development and Custom SaaS Opportunities

    In an interview, Fig co-founder and Hercules founder Brendan Falk discussed the commercial evolution and opportunities in AI-driven software development. He noted that AI is drastically lowering the cost of software development, shifting the market focus from tools for engineers toward empowering non-technical users to build core business systems such as CRMs and ERPs. He added that custom software is designed to drive revenue growth for enterprises rather than simply serving as lower-cost replacements for existing SaaS.

    Why it matters: The interview closely examines the transition from command-line developer tools to a full-stack platform for non-technical users, offering a practical reference for automating the creation of vertical business software.

  5. Ars Technica · AI 筛选29

    在 Google 搜索“Zelda”可体验 1998 年的复古搜索界面Chinese

    Google 推出复古彩蛋，用户搜索“Zelda”可体验 1998 年的极简搜索界面，点击剑形图标可切换返回。该活动疑似为任天堂定于 11 月 5 日登陆 Nintendo Switch 2 的《时之笛》重制版预热，其数字版售价 $60、实体版 $70。复古界面与当前充斥生成式 AI 和赞助内容的现代搜索形成鲜明对比，展现了 Google 核心产品的巨大变化。

  6. Hacker News · AI65

    PaidinToken Launches Local GPU VRAM and Open-Source LLM Compatibility Calculator

    PaidinToken has launched a hardware compatibility lookup tool covering 16 mainstream hardware setups, ranging from the RTX 3060 to RTX 5090, Apple Silicon, and data center GPUs. The tool details the maximum supported model size, quantization recommendations, estimated inference speeds, and one-click Ollama run commands for each configuration, along with calculation methodologies for VRAM and KV cache overhead and an FAQ section.

    Why it matters: By detailing the VRAM capacity limits and recommended quantization schemes for various graphics cards and chips, the tool enables developers to evaluate hardware compatibility before downloading massive model weights.

  7. Hacker News · AI71

    How to Write Evaluations for AI Agents: Core Concepts and a Seven-Step Guide

    EvalShift has released a guide to building AI agent evaluations, systematically outlining the core differences between end-to-end agent evaluations and traditional single-turn LLM evaluations, alongside a seven-step workflow for developing test suites. The guide emphasizes that agent evaluations must rigorously verify external environment states and critical execution rules to prevent fluent text responses from concealing failures to invoke tools. In practice, it recommends prioritizing code assertions for deterministic rules, complemented by validated LLM-as-a-judge for text quality assessment, and using the pass^k metric to gauge production stability.

    Why it matters: The article breaks down the essential differences between agent evaluations and single-turn text evaluations, delivering a practical seven-step methodology that blends code-based assertions with model-based judges—an ideal reference for engineering teams.

  9. The Verge · AI 筛选42

    BMW 将于 2027 年推出城市驾驶辅助以对标特斯拉 FSDChinese

    BMW 宣布将于 2027 年起将基于导航的部分自动驾驶拓展至城市环境，首发德国市场以对标特斯拉 FSD。该 Level 2 辅助驾驶系统支持在城区变道、转弯与通行环岛，但要求驾驶员在市区道路保持手扶方向盘并随时准备接管。相关传感器已预装于 3 系列、7 系列、X5、iX3 与 iX4 等车型中，后续将通过 OTA 更新激活。

  10. Hacker News · AI13

    我打造了全球最 AI-Pilled 的公司Chinese

    作者发文自述其创立了全球最深度拥抱 AI（AI-Pilled）的公司。该讨论源自 Hacker News 社区，聚焦于深度融合并应用 AI 技术来构建企业。原始发布页面目前未提供更具体的正文与实践细节。

  11. Hacker News · AI62

    TBC Launches First AI Video Model 'mischief 1' Derived from Living Neuron Research

    Biocomputing company TBC has officially launched its AI video model, mischief 1, delivering 5x the generation speed of its base model while maintaining output quality. Built on the open-source Wan 2.2, the model optimizes software by emulating information processing mechanisms observed in laboratory living neurons, increasing parameter count by less than 0.1%. Priced at 0.5 cents per second, it is now available on fal and OpenRouter.

    Why it matters: By translating the computational properties of living neurons into software-level optimizations, the model achieves a 5x speedup and ultra-low pricing without sacrificing visual quality, demonstrating a novel cost-reduction approach distinct from conventional engineering optimizations.

  12. Hacker News · AI74

    A Complete Practical Guide to Building a Domain-Specific BJJ AI Coach

    The author details a complete engineering pipeline for training a specialized Brazilian Jiu-Jitsu model, bjj-coach-7b-v3, and a reranker on a single GPU using video and document libraries. The system uses Whisper for transcription error correction and prepends context headers to enrich data, integrates knowledge graph logic with Qwen2.5-VL-7B to extract silent action frames, and fine-tunes bge-reranker with a 5-way listwise loss before exporting it as an ONNX service.

    Why it matters: The article provides a full vertical deployment guide covering multimodal video parsing, knowledge graph constraints, single-GPU fine-tuning, and zero-VRAM-overhead DPO.

  14. Buzzing HN · 中文精选17

    一位开发主管在 AI Agent 时代的日常与反思：让我来……Chinese

    一位鲜花速递服务的开发主管记录了 AI Agent 接管编程后，人类工程师退化为辅助验证的日常工作。面对持续三周的早晨丢单问题，AI Agent 独立排查出服务器与仓库间约 1 分钟的时钟不同步故障，仅因权限限制交由人工核查。作者借此反思机器接管了有趣的开发工作，将人类推向等待与核对等琐事，引发对角色定位的迷茫。

  15. The Decoder47

    AI 精英或像重商主义者囤积黄金一样囤积智能Chinese

    工程师 Venkatesh Rao 指出，部分 AI 精英正像重商主义者囤积黄金一样将智能视为稀缺资源，试图限制开源模型并利用安全担忧谋求政治影响。他认为 AI 降低了复杂项目的门槛，将促使人们更多思考并激发对创意的需求，甚至自主 AI 智能体也会催生新需求。政策重点应放在保障公平获取 AI，而非保护传统知识岗位。

  16. The Decoder75

    Goldman Sachs Restricts AI Coding Tools for Traders and Bankers, Limiting Access to Engineers

    Goldman Sachs has revoked access to advanced AI coding tools such as Claude Code and GitHub from most non-technical staff, limiting their use strictly to engineers. Traders and bankers had spent months using these tools to monitor trade flows and generate content, with senior executives previously encouraging non-technical employees to experiment with vibe coding. A Goldman spokesperson said the experimental phase has concluded, and the firm is now focusing the tools on those best positioned to drive long-term value, citing considerations around productivity and risk management.

    Why it matters: Goldman's shift from encouraging company-wide vibe coding to restricting access to engineers highlights the practical considerations of major financial institutions tightening tool permissions after evaluating productivity gains against risk management.

  18. MIT News · AI36

    MIT 能源倡议年度研究会议：在关键时刻蓄力能源转型Chinese

    MIT 能源倡议（MITEI）举办年度研究会议，聚焦在人工智能与数据中心快速扩张背景下跨学科推进能源转型与深度脱碳。会议探讨了三里岛 1 号机组计划于 2027 年中重启等核裂变复兴项目、二氧化碳规模化利用、储能混凝土以及电网优化 AI 工具等关键技术路线与商业化落地。

  19. TechCrunch AI61

    Data Center and AI Backlash Becomes US Midterm Issue, Reaching Brooklyn Voters

    Public backlash against AI and data centers has emerged as a key voter issue in the US midterms, with residents across the country voicing strong discontent over backroom tech deals, environmental risks, and job displacement. A report by Data Center Watch indicates that in the first half of 2026, nearly $200B in data center projects were delayed, stalled, or blocked due to political and community opposition. States including Pennsylvania, Georgia, Texas, New York, and Virginia are considering tighter regulatory restrictions and transparency mandates.

    Why it matters: This on-the-ground reporting highlights how public pushback against the data center surge has morphed into a ballot-box issue, shedding light on the societal headwinds facing compute infrastructure deployment.

  20. TechCrunch AI74

    The New York Times and Other Media Outlets Sue Brave Browser Over Alleged Content Scraping and Distribution to AI Companies

    The New York Times and several other media organizations have filed a lawsuit against Brave browser in Manhattan federal court, accusing it of illegally scraping copyrighted content and distributing it to AI clients, including Anthropic, Perplexity, Mistral, xAI, and Cohere. The complaint alleges that Brave used unblockable crawlers to compile an index of over 30 billion web pages, while also bypassing paywalls and replacing ads. The New York Times is seeking an injunction and damages; Brave has not yet responded to requests for comment.

    Why it matters: The lawsuit defines scraping and indexing providers as data distribution intermediaries, underscoring the publishing industry's widening push to hold the AI training and retrieval data supply chain accountable.

  21. Hacker News · AI65

    Open-source tool Claude Tandem launches, enabling multiple Claude accounts on Mac and seamless cross-device project handoffs

    A developer has open-sourced Claude Tandem, a macOS tool that allows users to simultaneously log in to and run multiple independent Claude client and Claude Code accounts on a single Mac. The tool uses profiles to isolate session states and project memory across accounts, featuring badged Dock icons and a menu bar switcher. When an account hits usage limits or when switching Macs, users can migrate project definitions, context summaries, and verified files with a single click via the built-in /handoff skill.

    Why it matters: The tool tackles the challenge of running multiple Claude accounts simultaneously and keeping context in sync across multiple Macs, providing an out-of-the-box solution for developers who need to separate work and personal subscriptions.

  22. Hacker News · AI66

    State of AI breaks down enterprise AI adoption for 2026: Discrepancies across usage, spend, and financial returns

    State of AI analyzed the divergent enterprise AI adoption metrics projected by major research and data organizations for 2026, pointing out that because denominators and definitions vary, there is no single adoption metric. A McKinsey survey showed that nearly 90% of surveyed organizations regularly use AI in at least one business function; Ramp transaction data revealed that 55.9% of its sampled businesses incurred AI-related spending; meanwhile, the U.S. Census Bureau's BTOS sample found that only 19.8% of businesses had used AI in the prior two weeks.

    Why it matters: The piece dissects the methodological differences among McKinsey, Ramp, and the U.S. Census Bureau, establishing a comparative framework to assess real-world enterprise AI deployment and financial returns.

  23. Hacker News · AI72

    Tests Show AI Assistants Scrape Standalone Sites 10x More Often Than Driving User Traffic

    Canonry's 30-day log monitoring of a headless Shopify store revealed that ChatGPT and Claude scraped pages 1,264 times to answer user queries, yet generated only 122 referral visits—meaning AI read volume was roughly 10 times actual user click-throughs.

    Why it matters: By using real server-side logs to quantify the disparity between AI retrieval and actual referral traffic, the source article provides transferable methods for standalone sites to optimize routing and defend against rogue crawlers.

  25. Hacker News · AI43

    cpuTopBars 与 devTopBars：macOS 菜单栏系统与 AI 监控工具Chinese

    macOS 菜单栏工具 cpuTopBars 与 devTopBars 开启公测，支持实时监控硬件状态及 Claude Code 编程会话。devTopBars 配合开源插件可展示 Claude Pro 或 Max 计划的 5 小时与每周用量上限，并追踪各运行会话的状态。软件还包含网络地址显示与阻止系统休眠功能，目前通过 TestFlight 提供测试，后续将上架 Mac App Store。

  26. Hacker News · AI60

    Terminal Tool Arvre Launched to Visualize and Manage Claude Code Branched Sessions as Trees

    A developer has released Arvre, an open-source terminal tool that visually displays Claude Code sessions and forks as tree diagrams or flowcharts. Users can inspect parent-child session hierarchies, quickly resume specific sessions, or copy resume commands directly within the terminal. Because Claude Code does not log parent links, Arvre identifies branching relationships by comparing shared message IDs, reading data directly from local directories.

    Why it matters: By analyzing message history, the tool solves the problem of untracked session branches, making it well-suited for developers who frequently fork sessions to test coding approaches and manage workflows.

  28. The Verge · AI 筛选16

    Apple Watch Ultra 4 降价 100 美元，追平历史最低价Chinese

    Apple Watch Ultra 4 在 Amazon 迎来 100 美元优惠，多款配色降至 699.99 美元，追平历史最低价。该智能手表搭载苹果全新 S11 芯片，支持包含 Live Rewind 在内的多项 Apple Intelligence 功能，单次充电续航最长可达 50 小时。设备还提供全新 Readiness 评分，并具备 100 米防水与 IP6X 级防尘能力。

  29. Hacker News · AI17

    Ask HN：有人打算仅用 AI 报税吗？Chinese

    Hacker News 用户发帖询问是否有人打算完全仅靠 AI 来完成报税。作者表示去年尝试时模型能力就已接近无需人工检查并节省了数小时，鉴于今年模型能力的大幅飞跃，完全交由 AI 报税已毫无疑问。

  30. Hacker News · AI80

    Oracle Plans $20B Stock and Bond Offering to Fund AI Infrastructure

    Oracle filed regulatory documents outlining plans to raise a total of $20 billion through $10 billion in common stock and $10 billion in senior notes to finance its AI infrastructure buildout. Although the funding plan was pre-announced during its Investor Day, shares fell over 6% on Monday amid market concerns regarding its debt load. As of August 31, Oracle's total debt reached $125.34 billion; the company plans to bridge the gap using vendor financing and customer prepayments, expecting free cash flow to turn positive in the second half of 2028.

    Why it matters: The source article provides an in-depth breakdown of Oracle's financing structure and current debt position, offering insights into the financial leverage and cash flow pressures hyperscalers face when expanding AI infrastructure.

  31. Buzzing HN · 中文精选72

    Weight-Space Programming: LLMs Can Directly Read and Write Neural Network Weights to Optimize Qwen3.5-0.8B

    A research team has introduced Weight-Space Programming, discovering that frontier LLMs can natively read and directly modify neural network weights without gradients or environmental feedback.

    Why it matters: The study demonstrates the feasibility of frontier models reading and writing network weights without gradients, opening up new avenues for interpretability research and retraining-free fine-tuning.

  32. Hacker News · AI31

    AI 时代的无我数学Chinese

    针对 AI 带来的替代焦虑，学者 Karthik C. S. 等人撰文探讨了 AI 时代的“无我数学”理念。文章指出，若仅将数学工作视为量化产出与个人能力的证明，AI 必然引发价值危机；但数学真理独立于研究者存在，形式化推理也无法直接替代对深刻规律的理解。面对冲击，研究者不应执着于在能力上胜过机器，而应回归对客观秩序的探索本身。

  33. Hacker News · AI57

    从云优先到 AI 优先：是生产力飞跃还是单纯消耗 Token？Chinese

    文章对比了 2016 年工程师说服管理层上云与 2026 年管理层催促工程师引入 AI 编程智能体的角色反转，剖析了企业落地 AI 工具的阻力与隐患。作者指出管理层与工程师在追求交付速度与承担代码维护后果上存在认知错位，过度依赖自动化生成容易削弱团队对系统的理解并增加审查负担。企业应当挑选积压的遗留迁移等受控任务试点，关注审查、修复、Token 账单等全流程真实成本，而非仅看生成速度或席位开通量。

  34. The Verge · AI 筛选58

    Anthropic 更新 Claude Code 提示词以限制克隆既有电子游戏Chinese

    Anthropic 更新了 Claude Code 在 Opus 5.5 上的系统提示词，限制该工具被用于直接复刻既有电子游戏。Anthropic 团队成员 Addy Osmani 确认此调整旨在防范 IP 侵权，但强调用户仍可开发原创游戏、特定类型游戏或基于个人资产的项目。团队表示将持续微调以平衡创作与合规，并鼓励用户反馈被误拦截的自主设计内容。

  35. Ars Technica · AI 筛选61

    Anthropic Introduces Anti-Game-Cloning Rules for Claude, Yet May Struggle to Curb AI Copycat Wave

    Anthropic recently added restrictions to its system prompts prohibiting Claude models from cloning specific existing video games, aiming to mitigate potential legal liabilities. While the rules permit the creation of original mechanics and genre-alike games, they explicitly forbid recreating specific pre-existing titles. These guardrails have been confirmed in the system prompt bundles for Claude Code 2.1.314 and Claude Opus 5.5 released on October 30.

    Why it matters: Anthropic's attempt to mitigate copyright infringement risks via system prompts highlights the compliance challenges and preventative limitations code-generation tools face when replicating commercial assets.

  36. Ars Technica · AI 筛选43

    Apple 计划 2027 年推出 $1,000 家用机器人：支持 FaceTime 通话与自动开关灯Chinese

    Apple 正在推进代号为 J595 的 AI 家用机器人研发，该设备可能命名为 HomeView Ultra，计划于 2027 年推出且定价或在 $1,000 以上。该机器人能够自主移动，目前支持 FaceTime 通话以及在人员进房时自动开关灯，研发若遇延期可能推迟至 2028 年上市。工程师正尝试整合更多 AI 技术以拓展其功能，从而支撑较高的产品定价。

  37. Claude 官方博客（网页）75

    Claude Platform Launches Public Beta for Message Threads in Messages API

    Claude Platform has launched a public beta for message threads in the Messages API. Long-horizon agents no longer need to resend the entire conversation history in each turn; transmitting only incremental messages significantly reduces time-to-first-token (TTFT).

    Why it matters: By resolving the latency overhead caused by repeatedly transmitting massive historical context in extended conversations, this feature offers an immediate optimization for coding agents that rely heavily on multi-turn calls.

  38. The Decoder70

    SpaceX Cuts Grok Bot Usage Quotas to Manage Soaring Operating Costs

    SpaceX recently tightened Grok Bot usage quotas, cutting allowances in half for some users. Elon Musk responded that operating costs relative to revenue had become unsustainable. Grok Bot supports tasks such as web browsing and drafting email replies, sharing quota pools with Grok and Cursor; in early October, its weekly paying users reached 800,000.

    Why it matters: The report highlights the compute cost pressures that agent products face under heavy usage, offering readers insight into the broader industry trend of model providers tightening flat-rate quotas.