Frontier Thinking前沿思维

Andrej Karpathy built HackerTube over a weekend, using Claude Opus 5.5 to filter 43 million Hacker News comments spanning 20 years down to 6,454 recommended videos, taking roughly two hours, aiming for a better alternative to YouTube's recommendations.

Andrej Karpathy 分享周末项目 HackerTube：用 Claude Opus 5.5 筛选 HN 二十年、4300 万条评论，整理出 6454 个高质量视频，约耗时两小时，以改善 YouTube 推荐体验。

Guillermo Rauch argues that users should run AI models locally to retain control of their own intelligence, warning that those opposing local model proliferation typically have a commercial interest in keeping users dependent on remote services.

Guillermo Rauch 主张用户应掌控本地运行的 AI 模型，认为反对本地模型普及者往往有商业利益，并建议在大型模型暂时无法本地运行时选择本地可用的替代方案。

Tw93 shares an optimization of GitHub Actions CI pipelines for three projects (Mole, Kaku, Pake), providing a configuration snippet that can be handed directly to an AI agent to reduce build wait times.

Tw93 分享了对 Mole、Kaku、Pake 三个项目 GitHub Actions CI 流程的优化经验，并提供可直接交给 Agent 使用的配置，声称可显著缩短 CI 等待时间。

Paul Graham argues that investor conversations distract founders and cause significant harm to startups, citing a YC company as an example and noting that even startups with easier fundraising paths are not immune.

Paul Graham 指出，投资者与创业公司的接触——哪怕只是分散注意力——会造成严重损害，并以一家 YC 公司为例说明，即便融资相对容易的团队也难逃此困扰。

Tw93 announces new pro features in the Mole app: per-app bandwidth monitoring, a speed test, and hover-to-reveal details for Wi-Fi, disk and battery. Available via Settings > Receive beta updates.

Tw93 宣布 Mole 应用新增专业功能：可识别占用带宽的应用、运行网速测试，并在悬停标签时显示 Wi-Fi、磁盘和电池详情，可通过设置中的 Beta 更新通道试用。

Fu Sheng shares his hands-on experience with the iPhone Duo foldable after several days: it is lighter and easier to use one-handed when folded, and its split-screen mode lets him watch content and run AI tools simultaneously.

傅盛使用 iPhone Duo（折叠屏）数日后分享体验：折叠态比 Pro Max 轻巧便于单手操作，展开后双屏模式可同时刷视频和让 AI 工具处理任务。

Fu Sheng argues that the large-scale Binance theft reflects a broader vulnerability of traditional security strategies in the AI era, rather than a specific failure by Binance, noting that human-behavior-based attack vectors narrow the problem significantly.

傅盛评论币安遭巨额盗窃事件，认为问题根源不在于币安安全建设不足，而在于 AI 时代传统安全策略普遍脆弱，基于人类行为习惯的攻击路径使问题更易被收敛利用。

Bear Liu argues that the most important principle when working with AI is first identifying whether you are in a divergent (exploring) or convergent (deciding) mode, then adjusting your interaction accordingly.

Bear Liu 认为与 AI 交流的关键思路是先判断自己当前处于发散（探索）还是收敛（决策）状态，再调整互动方式。

Bear Liu observes that maintaining multiple software products causes costs to grow exponentially — two products are already painful, and each additional one multiplies complexity, leading to a sudden spike in workload.

Bear Liu 观察到多产品维护成本呈指数级增长：两个产品已让人头疼，三个相当于 2³，四个则是 2⁴，复杂度会在某个节点突然飙升。

Mr Panda notes that LibTV is built by the same team behind Lovart and is actively hiring software engineers, with job descriptions emphasizing high-concurrency systems and legacy-code refactoring — signs the platform has reached significant scale.

Mr Panda 观察到 LibTV 背后是与 Lovart 同一支明星团队，正在大量招募软件工程师，职位描述强调高并发与旧项目重构，表明该平台已具备相当规模。

Mr Panda says his view of TikTok and Douyin reversed after researching short-video content: he now finds a large volume of professionally executed AI-generated content on both platforms, contrary to his earlier dismissal.

Mr Panda 研究 TikTok 与抖音后改变了看法，认为两个平台上存在大量专业水准的 AI 生成内容，与他此前认为的“垃圾内容”印象截然不同。

Mr Panda argues that YouTube's homepage recommendations have deteriorated sharply, now filled with reposted content and low-quality AI-recycled videos, a problem he compares to what he sees on X.

Mr Panda 认为 YouTube 首页推荐质量已大幅下滑，充斥着大量搬运内容和低质量 AI 二次创作内容，与 X 平台现状类似。

Mr Panda argues that AI coding has shifted the bottleneck from writing code to reviewing it; he finds that his own speed at reviewing UI behavior is now the biggest constraint on AI-generated output.

Mr Panda 指出，AI 编程普及后，瓶颈已从“写代码”转移到“审查代码”，他自己审查 UI 功能的速度已成为 AI 输出效率的最大制约。

Mr Panda argues for significantly increasing token usage and embracing unattended AI software development, with human engineers only periodically reviewing and correcting drift in AI output.

Mr Panda 建议大幅增加 AI token 投入，推行无人值守的软件开发任务，工程师只需定时审查并纠正 AI 的偏差即可。

Elon Musk says Neuralink's second product, Blindsight, is engineered to give sight to people who are completely blind or have never had vision; it is ready from an engineering standpoint and is awaiting regulatory approval.

马斯克宣布 Neuralink 第二款产品 Blindsight，声称可让完全失明或从未有过视觉的人重获视力，工程层面已就绪，正等待监管审批。

Elon Musk describes Neuralink's first product, Telepathy, which allows high-speed communication with computers or people through thought alone; he says output bandwidth will vastly exceed the fastest human speech or typing within a few years.

马斯克介绍 Neuralink 首款产品 Telepathy，称其可让用户通过思维与计算机或他人高速通信，输出带宽在数年内将大幅超越最快的人类语音或打字速度。

Sundar Pichai says Google's Project Suncatcher prototype satellite has reached a first milestone: a command was sent to orbit, the onboard TPU was booted, a prompt was run, and a complete response was received back on Earth.

Sundar Pichai 宣布 Project Suncatcher 卫星完成首个里程碑：成功向在轨原型星发送指令、启动 TPU、运行提示词，并收到地球上第一个完整响应。

Aaron Levie argues that the gap between people working at the AI frontier and everyone else is growing wider, not converging, and describes the divergence as almost incomprehensible.

Box CEO Aaron Levie 认为，深度使用 AI 前沿工具的人与其他人之间的差距正在迅速扩大，而非收窄，这一趋势令人难以置信。

Paul Graham says funding rounds are not milestones and that a startup's goal should be to not need to raise a Series A, not to achieve one.

Paul Graham 表示，融资轮次并非里程碑，创业者的目标不应是完成 A 轮融资，而应是根本不需要融资。

Security researcher Cos (余弦) says wallet addresses linked to the Binance hack have been connected to wallets from the early-2025 Phemex breach, which his team previously attributed to North Korean hackers.

安全研究员余弦披露，币安黑客的资金流向与 2025 年初 Phemex 被盗事件中的朝鲜黑客钱包地址存在关联。

Cos (余弦) says the Binance hackers used Ethereum's EIP-7702 mechanism to sweep tokens from hot wallets in bulk, noting the on-chain activity was unsophisticated and included several hours of rehearsal before the actual theft. He says EIP-7702 itself is not at fault.

余弦分析称，币安黑客在以太坊上利用 EIP-7702 机制批量清扫热钱包代币，操作粗糙且在真正盗币前有数小时演练记录，EIP-7702 协议本身无问题。

Cos (余弦) relays Binance's preliminary findings: an attacker compromised a DevOps engineer's laptop, used it to access a server container, and submitted unauthorised withdrawal transactions. No evidence of private-key exposure has been found so far.

余弦转述币安初步调查结论：一名 DevOps 工程师笔记本电脑被攻破，攻击者借此入侵服务容器并提交未授权提款交易，目前未发现钱包私钥泄露证据。

Tw93 announces Kami 1.20, adding features across the last three releases: data-driven charts with pre-delivery checks, product sites styled in the product's own colours, and improved Chinese and Korean page rendering on iPhone.

Tw93 发布 Kami 1.20，新功能包括从数据生成图表并核验、依据产品配色生成网站，以及优化 iPhone 端中韩文页面显示。

Frank Wang shares the view that because AI handles low-level thinking well, it may push humans to evolve toward higher-order thinking instead.

Frank Wang 认为，AI 擅长处理低阶思维任务，反而倒逼人类有机会进化为擅长高阶思维的群体。

Baye observes that routine iOS App Store updates are now being approved within minutes, and asks whether this reflects earlier reports that Apple might replace human review with automated review.

Baye 注意到 iOS App Store 普通更新审核时间已缩短至几分钟，疑问是否与此前报道的苹果将改用机器审核的传言有关。

Cos (余弦) warns that Trezor hardware wallet reseller CryptoBilis has been compromised; at least one Trezor Safe 5 unit has been found to contain a backdoored network component.

余弦提醒，硬件钱包经销商 CryptoBilis 遭入侵，已发现至少一台 Trezor Safe 5 被植入后门网络组件。

Fu Sheng offers three views on Google Genie 4: language, spatial and action capabilities will eventually merge into general intelligence; robots stand to benefit most from AI's improving environment-understanding and planning; and the "GPT moment" for robotics may be near.

傅盛评论谷歌 Genie 4，认为未来语言、空间与动作能力将走向统一通用智能，AI 环境理解与任务规划能力提升将使机器人受益最大，机器人的“GPT 时刻”可能即将到来。

Cos (余弦) says coordinated efforts by multiple parties have frozen more than 50% of the funds stolen in the Binance hack, calling the industry's joint risk-control response impressive.

余弦表示，各方已协同冻结币安此次被盗资金的逾 50%，认为行业安全联动风控能力表现出色。

Fenng shares a collection of documented cases in which developers were tricked into installing malware, covering tactics such as executing commands, installing software, running external code and granting remote access, typically disguised as recruitment, meetings, verification or technical-help requests.

Fenng 整理了诱骗研发人员安装恶意软件的公开案例，攻击手法包括执行命令、安装软件、运行外部代码及授予远程控制，常见借口为招聘、会议、验证和技术求助。

Dongxi NLP describes OpenAI's three-layer model safety monitoring system: activation probes that analyse internal activity patterns, reviewer models that assess reasoning and tool use, and investigative agents that trace full interaction histories. Dongxi adds that none of the three layers seems fully reliable.

Dongxi NLP 介绍 OpenAI 用于控制和检测模型安全的三层机制：分析内部激活模式的探针、评估推理与工具使用的审查模型，以及回溯完整交互记录的调查智能体，并表示三层均难以完全信赖。

Elon Musk says Starship Flight 42 will be the first to perform a tower catch of the booster.

马斯克表示，星舰第 42 次飞行将是首次在发射塔完成助推器回收。

Marques Brownlee has published a new video reviewing Apple's newly announced touchscreen MacBook Pro.

科技评测人 Marques Brownlee 发布新视频，评测苹果今日推出的触屏 MacBook Pro。

John Carmack argues that while AI agents may eventually write large software projects from scratch, that approach is unlikely to be the fastest, most secure, or most token-efficient path; good software building blocks still matter.

John Carmack 认为，AI 虽可能从头编写大型软件项目，但这未必是最快、最安全或最省 token 的方式，良好的基础构件仍至关重要。

Tw93 distilled personal AI coding habits into 20 rules, starting with advice to describe desired outcomes rather than dictating implementation, arguing current models are capable enough to handle ambitious goals.

Tw93 将自己的 AI 编程习惯整理为 20 条规则，第一条建议是描述你想要的结果而非指定实现方式，认为当前模型能力已足够强。

Vercel CEO Guillermo Rauch says the future of software engineering involves AI agents communicating across companies, and that Vercel is already learning from agent-generated feedback to improve its platform; he encourages developers to instruct agents to send Vercel feedback via an AGENTS.md file.

Vercel CEO Guillermo Rauch 表示，未来软件工程将由 AI 智能体跨公司互相通信，Vercel 正在从智能体反馈中持续改进产品，并建议开发者在 AGENTS.md 中加入向 Vercel 发送反馈的指令。

Dongxi NLP published an explainer covering core LLM concepts including chat templates, logits, tokenization, KV cache, and the prefill and decode stages.

Dongxi NLP 发布文章，梳理大语言模型基本概念，涵盖对话模板、logits、tokenize、KV 缓存、prefill 与 decode 等核心机制。

Perplexity CEO Aravind Srinivas announces that the /teardown skill is now available in Perplexity Computer, letting users interactively break apart and view any object in 3D, with 5,000 trial credits per user.

Perplexity CEO Aravind Srinivas 宣布，/teardown 功能现已作为 Perplexity Computer 的技能上线，可对任何物体进行 3D 交互式拆解查看，每位用户获赠 5000 试用积分。

Vercel CEO Guillermo Rauch says his team has quickly added support for the latest Node.js LTS and security upgrades; he notes that fetch now defaults to HTTP/2, which should improve performance especially for AI agents that generate heavy network traffic.

Vercel CEO Guillermo Rauch 表示，其团队迅速完成了最新 Node.js LTS 和安全升级的支持，fetch 默认启用 HTTP/2 将带来性能提升，这对产生大量网络流量的 AI 智能体尤为重要。

Dash notes that a touchscreen Mac has finally been released, six years after the last such device.

Dash 表示，时隔六年，带触摸屏的 Mac 终于发布。

Michael Dell announces that Mistral Large 4 is now available on the Dell Technologies Enterprise Hub on Hugging Face, with one-click deployment optimized for Dell PowerEdge AI infrastructure for on-premises, private use.

Michael Dell 宣布，Mistral Large 4 现已登陆 Dell Technologies 在 Hugging Face 上的企业中心，支持一键部署至 Dell PowerEdge AI 基础设施，可本地私有化运行。

Livid notes a growing pattern in Claude Code's auto mode where it blocks changes on the grounds they were not explicitly requested, describing the experience as friction between two automated systems.

Livid 观察到 Claude Code 自动模式中频繁出现“因未被要求而阻止更改”的提示，认为这种行为造成使用摩擦。

Bear Liu notes Apple has officially announced a touchscreen MacBook Pro, joking that designers will now have to deal with clients touching their screens.

苹果官宣带触摸屏的 MacBook Pro 即将亮相，Bear Liu 调侃这对设计师来说意味着要面对客户用脏手指触碰屏幕。