First macOS Kernel Exploit Bypasses MIE on M5
首个绕过 MIE 的 M5 macOS 内核漏洞 ⭐️ 9.0/10

Researchers presented the first public macOS kernel memory corruption exploit on Apple M5 silicon that successfully bypasses Memory Integrity Enforcement (MIE) during a meeting at Apple Park. This exploit demonstrates that even Apple's latest hardware-backed memory safety defense can be circumvented, highlighting critical security gaps in modern macOS versions and raising the bar for future security research. The exploit is a data-only kernel exploit that corrupts kernel data structures without modifying code, making it harder to detect. It was shared with Apple in a laser-printed report during a meeting on May 18, 2026.

rss · Michael Tsai · May 18, 17:50

Background: Apple silicon is Apple's custom ARM-based processor line used in Macs and other devices. Memory Integrity Enforcement (MIE) is a hardware-backed security feature introduced by Apple to prevent memory corruption attacks. A data-only exploit manipulates kernel data rather than injecting malicious code, bypassing typical code integrity checks.

Tags: #security, #macOS, #exploit, #kernel, #Apple

Meta's AIRA Agents Autonomously Outperform Llama 3.2
Meta 的 AIRA 代理自主发现超越 Llama 3.2 的架构 ⭐️ 9.0/10

Meta released a paper introducing AIRA, a dual-agent system that autonomously discovers neural architectures outperforming Llama 3.2 at 350M, 1B, and 3B parameter scales within a 24-hour compute budget. This achievement demonstrates that AI-driven architecture search can surpass hand-designed models like Llama 3.2, potentially revolutionizing how neural networks are designed and reducing reliance on human expertise. The system splits the search into two agents: AIRA-Compose explores macro-architecture using an ensemble of 11 agents, while AIRA-Design implements low-level mechanisms. This separation of strategy and implementation outperforms single end-to-end agents on real search problems.

rss · elvis(@omarsar0) · May 18, 18:00

Background: Neural architecture search (NAS) is a process of automating the design of neural network architectures, traditionally computationally expensive. Llama 3.2 is Meta's state-of-the-art open-source large language model. AIRA leverages LLM-based agents to reason and search over architectural components like Attention, MLP, and Mamba, enabling efficient discovery.

Tags: #Meta, #neural architecture search, #AI agents, #Llama, #deep learning

Cloudflare and Stripe Enable AI Agents to Create Accounts and Deploy
Cloudflare 与 Stripe 让 AI 代理自主创建账户并部署 ⭐️ 9.0/10

Cloudflare and Stripe have launched a protocol that allows AI agents to autonomously create cloud accounts, register domains, start subscriptions, and deploy to production, with Stripe handling identity verification and payment with a $100/month default cap. This marks a significant paradigm shift in AI-cloud interaction, as no other major cloud provider offers comparable agent-driven account provisioning, potentially enabling fully automated workflows for startups and developers. The protocol includes a $100/month default spending cap via Stripe, and Cloudflare is offering $100,000 in credits to new startups incorporating via Stripe Atlas. The integration allows any platform with signed-in users to work with Cloudflare similarly.

rss · InfoQ · May 18, 09:41

Background: AI agents are software programs that can autonomously perform tasks such as account creation and deployment. Cloudflare is a cloud services provider offering CDN, DNS, and compute, while Stripe is a payment processing platform. Previously, creating cloud accounts and deploying required manual human steps; this protocol automates the entire process for AI agents.

Tags: #AI Agents, #Cloudflare, #Stripe, #Cloud Computing, #Automation

Anthropic acquires SDK generator startup Stainless in acqui-hire
Anthropic 收购 SDK 生成器初创公司 Stainless 作为人才收购 ⭐️ 8.0/10

Anthropic has acquired Stainless, a startup that provides an SDK generator for APIs, as an acqui-hire and will shut down all hosted Stainless products, including the SDK generator, effective immediately. This acquisition highlights Anthropic's need for top engineering talent to advance its Claude platform and agent-to-API connectivity, while also demonstrating the challenges of the SDK generator market amid rising AI-powered code generation tools. Stainless's SDK generator allowed developers to auto-generate SDKs in multiple languages from OpenAPI specs. Existing users and SDKs will not be supported going forward, and new signups are halted immediately.

hackernews · tomeraberbach · May 18, 17:01 · Discussion

Background: An SDK generator automates the creation of software development kits for APIs, simplifying integration for developers. Acqui-hire refers to acquiring a company primarily for its talent rather than its products, which are often discontinued after acquisition.

Discussion: Community sentiment is mixed: some congratulate Stainless's team but regret the product shutdown, noting the rise of AI-powered code generation weakening the SDK generator market. Others criticize the abrupt discontinuation as unfair to existing users, and raise concerns about Anthropic building walled gardens through acquisitions.

Tags: #acquisition, #AI, #SDK, #Anthropic, #startup

Files.md: Open-source Alternative to Obsidian Note-Taking App
Files.md：Obsidian 的开源替代品 ⭐️ 8.0/10

Files.md is an open-source note-taking application that serves as an alternative to Obsidian, supporting Markdown files with a distinct workflow. It was shared on Hacker News and received high engagement. This project highlights the growing demand for open-source alternatives to proprietary tools like Obsidian, especially among users who value data ownership and customizable workflows. It reflects the community's desire for transparent and flexible note-taking solutions. Files.md is not a feature-parity clone of Obsidian; it offers its own approach to managing notes and knowledge, which may appeal to users seeking a different paradigm. The project is hosted on GitHub under an open-source license.

hackernews · zakirullin · May 18, 13:33 · Discussion

Background: Obsidian is a popular proprietary note-taking app that uses local Markdown files and offers a plugin ecosystem. While free for personal use, it is not open-source. Other open-source alternatives like Joplin exist, but Files.md introduces a novel workflow that differentiates it from existing tools.

Discussion: Commenters noted that Obsidian feels like it should be open-source, highlighting the community's trust in open models. Some users are building native alternatives, while others mentioned Joplin as a simpler open-source option. There was also discussion about Files.md's unique workflow not being a direct Obsidian replacement, which some found more interesting.

Tags: #open-source, #note-taking, #markdown, #obsidian-alternative, #productivity

Elon Musk loses lawsuit against OpenAI
埃隆·马斯克起诉 OpenAI 败诉 ⭐️ 8.0/10

A jury ruled that Elon Musk waited too long to bring his claims against Sam Altman and OpenAI, dismissing the lawsuit due to the statute of limitations. The ruling sets a precedent that challenges to OpenAI's for-profit conversion may be time-barred, potentially protecting the company from similar lawsuits and bolstering its path to IPO. The jury found that Musk could have brought the same lawsuit as early as 2019 or 2021 over similar Microsoft deals, making his 2023 complaint untimely under the three-year statute of limitations.

hackernews · nycdatasci · May 18, 17:38 · Discussion

Background: Elon Musk co-founded OpenAI in 2015 as a non-profit AI research organization but left in 2018. He later sued OpenAI and its CEO Sam Altman, arguing that the company had abandoned its original non-profit mission by pursuing profit and partnering with Microsoft. The lawsuit centered on OpenAI's 2023 deal with Microsoft, but the jury determined that earlier comparable deals should have triggered the claim.

Discussion: Commenters widely noted the legal significance of the statute of limitations ruling, with some speculating that Musk's true goal was to damage OpenAI's reputation and IPO prospects rather than win outright. Others raised concerns about the precedent for non-profits transferring IP to for-profit entities.

Tags: #OpenAI, #Lawsuit, #Elon Musk, #Sam Altman, #AI Industry

FBI Seeks Nationwide Access to License Plate Reader Data
FBI 寻求全国范围的车牌读取器数据访问权限 ⭐️ 8.0/10

The FBI has announced a plan to purchase nationwide access to automatic license plate recognition (ALPR) data, raising significant privacy concerns. This move could enable mass surveillance of all vehicles across the United States without judicial oversight, impacting every driver's privacy and potentially setting a precedent for law enforcement data acquisition. The plan involves purchasing data from private companies that operate ALPR systems, which can track vehicle locations and movements over time; the FBI aims to use this data for investigations without obtaining warrants.

hackernews · cdrnsf · May 18, 19:28 · Discussion

Background: Automatic License Plate Recognition (ALPR) technology uses cameras and software to automatically capture and store license plate information, often with time, date, and GPS location. These systems are widely deployed by police and private companies, creating vast databases of vehicle movements. Privacy advocates argue that warrantless access to such data violates Fourth Amendment protections against unreasonable searches.

Discussion: Commenters express deep concern, with some suggesting technical solutions like digital license plates that change daily to thwart mass tracking. Others criticize the lack of legal liability for personal data, noting that personal data is currently an asset rather than a liability, which incentivizes collection. There is also skepticism that political will exists to protect privacy rights.

Tags: #privacy, #surveillance, #law enforcement, #data security, #ALPR

Iran launches Bitcoin-backed insurance for Strait of Hormuz
伊朗启动比特币航运保险 ⭐️ 8.0/10

On May 18, 2026, Iran launched a Bitcoin-backed shipping insurance service named 'Hormuz Safe' for vessels transiting the Strait of Hormuz. This novel use of Bitcoin as collateral for geopolitical shipping insurance could challenge the dominance of the US dollar in international trade and provide Iran with a tool to circumvent financial sanctions. The service, called 'Hormuz Safe', charges premiums in Bitcoin and uses Bitcoin as a reserve asset; it was unveiled by Iran's Ministry of Roads and Urban Development on May 18, 2026.

hackernews · srameshc · May 18, 17:25 · Discussion

Background: The Strait of Hormuz is a narrow waterway connecting the Persian Gulf to the Gulf of Oman, through which about 20% of global oil passes. Iran has threatened to close the strait in response to sanctions, making insurance for ships traversing it extremely expensive or unavailable. Bitcoin-backed insurance offers an alternative to traditional marine insurance markets that may be unwilling to cover risks under US pressure.

Discussion: Comments expressed skepticism about the insurance's effectiveness against US naval power, but also noted the strategic and financial implications of using Bitcoin to bypass dollar hegemony. One commenter saw it as a potential face-saving exit for the US, while another highlighted the significance of Bitcoin becoming a global reserve-like commodity.

Tags: #Bitcoin, #Iran, #Shipping, #Geopolitics, #Insurance

Cursor releases Composer 2.5, matching Opus at 30x lower cost
Cursor 发布 Composer 2.5，性能媲美 Opus 价格低 30 倍 ⭐️ 8.0/10

Cursor has released Composer 2.5, its self-developed coding model that achieves performance comparable to Claude Opus 4.7 in coding benchmarks, with input cost 10x lower and output cost 30x lower than Opus. The model improves its ability to handle long-context tasks, follow complex instructions, and collaborate smoothly. This release significantly reduces the cost of frontier-level coding AI, potentially democratizing access for individual developers and small teams. Cursor's move challenges dominant closed-source models and signals a shift toward specialized, cost-efficient coding models. Composer 2.5 is built on Kimi K2.5 and trained with three key innovations: targeted text-feedback reinforcement learning to address credit assignment in long rollouts, synthetic data generation (25x more than Composer 2), and infrastructure optimizations like Muon optimizer with distributed orthogonalization. The model scores within the same range as Opus 4.7 on coding evaluations, with a maximum difference of less than 1 point.

rss · 小互(@imxiaohu) · May 19, 02:02

Background: Cursor is an AI-native code editor that integrates coding assistants. Its Composer model generates and edits code autonomously. Opus (Claude Opus 4.7) by Anthropic is among the most capable models for complex coding tasks but remains expensive at roughly $15/M input tokens and $75/M output tokens. Composer 2.5 aims to deliver similar capability at a fraction of the cost, making advanced AI coding assistance more accessible.

Tags: #AI, #coding model, #Cursor, #cost efficiency, #Composer 2.5

Alibaba's Qwen3.7 Preview Lands on Arena Benchmarks
阿里 Qwen3.7 预览版登陆 Arena 基准测试 ⭐️ 8.0/10

Alibaba's Qwen3.7 Preview (Max and Plus variants) has been added to the Arena platform for text and vision benchmarks, achieving #13 overall in Text Arena and #16 in Vision Arena. This release marks a strong showing for Alibaba in the competitive AI model landscape, placing the company as the #6 lab in text and #5 in vision, signaling its growing capabilities in both language and multimodal tasks. The Qwen3.7 Max Preview ranks #7 in Math, #9 in Expert, #9 in Software & IT, and #10 in Coding within the Text Arena, while the Plus Preview handles vision tasks. These models are preview versions, with a full release expected soon.

rss · Arena.ai(@lmarena_ai) · May 18, 15:42

Background: The Arena is a crowdsourced benchmark platform where users vote on model outputs to compute Elo ratings and leaderboards. It evaluates models across text and vision modalities, providing a community-driven measure of model quality. Qwen is Alibaba's series of large language models, with Qwen3.7 being the latest iteration.

Tags: #AI, #Alibaba, #Qwen3.7, #Arena, #benchmark

HTML is the new Markdown, says Claude Code developer
Claude Code 开发者：HTML 是新的 Markdown ⭐️ 8.0/10

Claude Code core developer Thariq (@trq212) argues that HTML is replacing Markdown as the preferred format for human-LLM interaction due to its visualization and interactivity advantages. This shift could significantly improve human oversight and collaboration in AI workflows, enabling richer interfaces for planning, editing, and design system management. Thariq suggests three practical workflows: interactive HTML for brainstorming, disposable micro-apps for editing specific content, and living design systems that are human-and-machine readable.

rss · meng shao(@shao__meng) · May 19, 01:08

Background: Markdown is a lightweight markup language commonly used for formatting plain text. It is favored for LLM interactions due to its simplicity and readability by both humans and AI. However, as tasks grow larger, static Markdown documents become cumbersome to review, lacking interactive elements. HTML, in contrast, supports rich visual layouts and interactive components, making it more suitable for complex collaborative workflows.

Tags: #LLM, #Markdown, #HTML, #Claude Code, #AI interaction

Claude Code Dev Log Prompt Externalizes AI Decisions
Claude Code 开发日志提示词外化 AI 决策 ⭐️ 8.0/10

Thariq, a core developer of Claude Code, shared a prompt he uses frequently that instructs the AI to maintain an implementation-notes.html file documenting design decisions, deviations, tradeoffs, and open questions during coding. This prompt addresses a fundamental problem in AI-assisted coding: the dilemma between over-specification and hidden decisions. It provides a lightweight way to make AI's implicit choices auditable, reducing rework and improving code review quality. The prompt asks the AI to record four categories: design decisions (where the spec was ambiguous), deviations (intentional departures from the spec), tradeoffs (alternatives considered), and open questions. It uses HTML or Markdown as a lightweight, PR-compatible format.

rss · meng shao(@shao__meng) · May 19, 00:41

Background: Claude Code is an agentic coding tool by Anthropic that reads codebases, edits files, runs commands, and integrates with development tools. In AI-assisted development, a common challenge is that AI agents make many decisions not fully captured in specifications. Traditional approaches either over-specify (unrealistic) or let the AI operate freely (hidden decisions). This prompt offers a middle path by externalizing those decisions.

Tags: #Claude Code, #AI协作编码, #提示词工程, #开发日志, #软件工程

Former OpenAI Hardware Lead on AI's Shift to Physical World
前 OpenAI 硬件负责人谈 AI 转向物理世界 ⭐️ 8.0/10

Caitlin Kalinowski, who built hardware teams at OpenAI and Meta, shared in an interview why AI companies are increasingly investing in hardware and robotics, citing that AI capabilities behind a keyboard will soon saturate. This insight signals a strategic shift in the AI industry from digital-only advancements to physical world applications, which could accelerate developments in robotics, manufacturing, and space exploration. Kalinowski has a background in Apple's MacBook engineering and led Meta's AR/VR hardware program before building OpenAI's robotics and hardware team. She left OpenAI after the company's deal with the Department of Defense.

rss · Lenny Rachitsky(@lennysan) · May 18, 14:31

Background: The rise of large language models and AI chatbots has led to rapid progress in digital AI capabilities. However, Kalinowski argues that this acceleration is vertical and will hit a ceiling, pushing AI labs to explore the physical world through hardware, robotics, and real-world sensing. This shift mirrors the historical pattern where technological breakthroughs eventually extend beyond the digital realm to impact the physical environment.

Tags: #AI hardware, #robotics, #OpenAI, #Meta, #strategy

What's the SOTA for File Search? Jerry Liu Polls Community
文件搜索的 SOTA 是什么？Jerry Liu 向社区提问 ⭐️ 8.0/10

Jerry Liu, a prominent figure in AI, asked on Twitter what the current state-of-the-art is for file search and retrieval, listing options including grep, BM25, vector search, hybrid search, and SQL. This question sparks debate on the most effective retrieval method for RAG systems and enterprise search, influencing tooling choices in the AI ecosystem. The tweet presents a multiple-choice-like query, reflecting that no single approach is universally accepted; the high engagement (20 replies, 52 likes) indicates community interest.

rss · Jerry Liu(@jerryjliu0) · May 18, 23:37

Background: BM25 (Best Matching 25) is a ranking function that estimates document relevance based on term frequency and document length. Hybrid search combines multiple techniques (e.g., keyword and vector) to improve retrieval accuracy. These methods are critical for modern AI applications like Retrieval-Augmented Generation (RAG).

Tags: #file search, #retrieval, #SOTA, #BM25, #vector search

xAI releases Grok Imagine Video with three generation modes
xAI 发布 Grok Imagine Video，支持三种生成模式 ⭐️ 8.0/10

xAI has launched Grok Imagine Video, a new AI model that generates 1-15 second video clips at 480p or 720p resolution, 24 fps, across seven aspect ratios, with three modes: text-to-video, image-to-video, and reference-to-video using up to seven reference images for character, style, and setting consistency. This release marks xAI's entry into the competitive AI video generation space, offering a unique multi-reference consistency feature that could enable more controlled and personalized video creation, potentially impacting content creators and businesses seeking consistent visual narratives. The model is accessible via OpenRouter under the name 'x-ai/grok-imagine-video', and supports resolutions up to 720p with 24 fps. The reference-to-video mode can ground on up to seven images for maintaining character, style, and setting consistency across the generated clip.

rss · OpenRouter(@OpenRouterAI) · May 18, 18:56

Background: xAI, founded by Elon Musk, develops the Grok chatbot and AI models. Video generation models like this typically use diffusion or transformer architectures trained on large datasets to produce realistic motion and visuals. The reference-to-video approach is a relatively new capability that aims to preserve identity and style across frames, which is challenging for many current models.

Tags: #AI, #video generation, #xAI, #Grok, #OpenRouter

Firefox Uses AI Opus 4.6 to Fix 22 Security Bugs
火狐使用 AI Opus 4.6 修复 22 个安全漏洞 ⭐️ 8.0/10

The Firefox team has used Anthropic's Claude Opus 4.6 AI model to identify and fix 22 latent security vulnerabilities in the browser, with fixes included in Firefox 148. This demonstrates a practical and effective application of frontier AI for browser security, potentially reducing the window of vulnerability exploitation and showcasing collaboration between a major browser vendor and an AI company. The collaboration with Anthropic involved scanning Firefox's codebase with Opus 4.6, which led to fixes for 22 security-sensitive bugs in Firefox 148. The effort has been ongoing since February 2026, as part of a broader initiative to harden the browser.

rss · Michael Tsai · May 18, 17:50

Background: Opus 4.6 is Claude's most capable model released in February 2026, excelling at code review, debugging, and agentic tasks across large codebases. Traditional vulnerability discovery relies on manual auditing or fuzzing, but AI models can analyze code contextually to find subtle bugs. This approach represents a growing trend of using AI for proactive security hardening.

Tags: #Firefox, #Security, #AI, #Vulnerability, #Browser

LangChain Unveils SmithDB for Agent Observability
LangChain 推出用于智能体可观测性的 SmithDB ⭐️ 8.0/10

LangChain announced SmithDB, a new purpose-built distributed database for agent observability that powers core LangSmith workloads, promising up to 12x faster performance and full portability backed by object storage. SmithDB addresses a critical bottleneck in agent development by making observability tools much faster and portable, enabling developers to debug and iterate on AI agents more efficiently across self-hosted and multi-cloud environments. SmithDB consists of three components: object storage for durable trace data, a small Postgres metastore for segment metadata, and stateless services for ingestion, query, and compaction.

rss · LangChain(@LangChainAI) · May 18, 16:38

Background: Observability for AI agents involves tracing the steps and decisions of LLM-based agents to debug and monitor their behavior. SmithDB is a distributed database designed specifically to handle high-volume trace data efficiently, unlike general-purpose databases that become bottlenecks.

Tags: #LangChain, #observability, #AI agents, #tracing, #performance

Gary Marcus: Pure LLMs Are Just Autocomplete
Gary Marcus：纯大语言模型仅是自动补全 ⭐️ 8.0/10

Gary Marcus argued on X that pure large language models (LLMs) are essentially autocomplete, and that recent AI progress, such as Claude Code, actually stems from integrating classical symbolic techniques and tools to compensate for LLM weaknesses. This critique challenges the dominant narrative that scaling LLMs alone drives AI progress, emphasizing the need for neurosymbolic approaches. It sparks debate on the true source of recent breakthroughs and the future direction of AI research. Marcus referenced Claude Code as an example where symbolic methods are used, and responded to a tweet by Bindu Reddy who claimed critics who called AI 'just autocomplete' have been proven wrong. Marcus insists understanding progress requires recognizing the shift away from pure LLMs.

rss · Gary Marcus(@GaryMarcus) · May 18, 07:44

Background: Pure LLMs (large language models) are neural networks trained on vast text data to predict next words, often described as 'autocomplete' by critics. Symbolic AI, also known as Good Old-Fashioned AI, uses logic, rules, and explicit knowledge representation, contrasting with the statistical pattern matching of neural networks. Gary Marcus is a prominent AI researcher and frequent critic of purely statistical AI, advocating for hybrid neurosymbolic systems.

Tags: #GaryMarcus, #LLM critique, #AI progress, #symbolic AI, #autocomplete

Vercel makes all firewall mitigations free for everyone
Vercel 对所有用户免费提供全部防火墙缓解措施 ⭐️ 8.0/10

Vercel announced that all firewall mitigations, including custom rules, are now free. This extends beyond DDoS and system-level mitigations to any rule users configure. This move significantly reduces costs for developers and organizations using Vercel, especially those facing large-scale attacks. It enhances security accessibility and may set a new industry standard for cloud platform pricing. Starting today, users are not charged for requests that are denied, challenged, or rate-limited by Vercel Firewall. Vercel absorbs the computational and network costs for any size of attack or traffic mitigation.

rss · Guillermo Rauch(@rauchg) · May 19, 01:37

Background: Vercel is a cloud platform for frontend developers, offering hosting and serverless functions. Firewall mitigations protect applications from malicious traffic, and previously, custom rules could incur additional costs. This announcement removes those charges.

Tags: #Vercel, #Firewall, #DDoS, #Cloud Computing, #Security

Cursor and SpaceXAI Train Massive Model with 10x Compute
Cursor 与 SpaceXAI 以十倍算力训练巨型模型 ⭐️ 8.0/10

Cursor and SpaceXAI announced a collaboration to train a significantly larger AI model from scratch, using 10x more total compute than previous efforts, leveraging the Colossus 2 supercomputer with a million H100-equivalent GPUs. This partnership marks a major leap in AI capability, potentially leading to more powerful models that could benefit the broader AI ecosystem, including Cursor's coding assistant and SpaceXAI's applications. The model is being trained from scratch using 10x more compute, which is made possible by Colossus 2, the first gigawatt-scale AI training cluster, featuring a million H100-equivalent GPUs. Cursor and SpaceXAI are combining their data and training techniques.

rss · Cursor(@cursor_ai) · May 18, 16:43

Background: SpaceXAI is a division of SpaceX focused on AI, formed after xAI was folded into SpaceX in 2026. It operates the Grok chatbot and the social network X. Colossus 2 is a massive supercomputer built by xAI/SpaceXAI, now the world's first gigawatt-scale AI training cluster.

Tags: #AI, #Large Language Models, #Training, #SpaceXAI, #Cursor

TurboQuant Compression Technique Arrives in Qdrant
TurboQuant 压缩技术现已集成到 Qdrant ⭐️ 8.0/10

Qdrant 1.18 introduces TurboQuant, a new rotation-based vector quantization method from Google Research that achieves recall similar to scalar quantization (SQ) at roughly twice the compression ratio, and outperforms binary quantization (BQ) under the same storage budget. This advancement allows vector search systems to significantly reduce memory and storage costs without sacrificing retrieval quality, making high-dimensional vector search more practical for production deployments at scale. TurboQuant ships in Qdrant 1.18 with extensions to work on real embeddings; for existing collections, enabling it requires a PATCH request specifying quantization_config with turbo settings.

rss · Qdrant(@qdrant_engine) · May 18, 10:30

Background: Vector databases use quantization techniques like scalar quantization (SQ) and binary quantization (BQ) to compress high-dimensional vectors, reducing memory usage at the cost of some retrieval accuracy. TurboQuant, a rotation-based method, aims to improve the accuracy-compression trade-off by applying a learned rotation before quantization.

Tags: #vector search, #quantization, #Qdrant, #compression, #machine learning

AI Will Eat Hardware, Says Top Expert Caitlin Kalinowski
顶级硬件专家谈 AI 将吞噬硬件 ⭐️ 8.0/10

A podcast episode featuring Caitlin Kalinowski (ex-OpenAI, Meta, Apple) discusses how AI is poised to revolutionize hardware design, supply chains, and robotics, while sharing insights on VR/AR failures and future trends. Kalinowski's unique experience across top tech companies provides rare insight into the convergence of AI and hardware, which will impact everything from consumer electronics to robotics and national security. Kalinowski highlights that hardware "compiles" only five times before mass production, making iteration far harder than software. She warns of supply chain vulnerabilities like a single magnet bottleneck and predicts that AI-driven robotics will move beyond digital tasks into the physical world.

rss · 跨国串门儿计划 · May 18, 06:37

Background: The podcast is a clone of Lenny's Podcast, featuring Caitlin Kalinowski who led hardware teams at Apple (MacBook Air, Mac Pro), Meta (VR/AR), and OpenAI (robotics). She discusses how VR/AR technologies, despite not becoming mainstream, are laying groundwork for robotics. AI is beginning to transform hardware engineering tools like CAD and PCB layout, but still lacks a true "world model" for physical understanding.

Tags: #AI, #hardware, #VR/AR, #robotics, #podcast

Blackstone and Google Joint Venture to Build TPU Cloud
黑石与谷歌合资构建 TPU 云 ⭐️ 8.0/10

Blackstone and Google announced a joint venture to create a new cloud infrastructure based on Google's Tensor Processing Units (TPUs), aiming to provide AI and machine learning cloud services. This partnership combines Blackstone's capital with Google's TPU technology, potentially increasing competition in the AI cloud market and offering more specialized hardware for machine learning workloads. TPUs are Google's custom-designed ASICs for accelerating machine learning tasks, and the joint venture will likely leverage Google's latest TPU generations. No specific financial terms or timeline were disclosed.

rss · The Keyword · May 19, 01:00

Background: Tensor Processing Units (TPUs) are custom chips developed by Google for neural network processing. They were first used internally in 2015 and became available via Google Cloud in 2018. TPUs accelerate training and inference of large AI models, making them a key component in AI infrastructure.

Tags: #TPU, #Cloud Computing, #AI Infrastructure, #Google, #Joint Venture

Anthropic Announces Managed Agents and Proactive Workflows at Code with Claude 2026
Anthropic 在 Code with Claude 2026 上宣布托管代理和主动工作流 ⭐️ 8.0/10

At the Code with Claude 2026 event in San Francisco, Anthropic announced managed agents, proactive workflows, and the concept of a capability curve for AI coding tools. These announcements signal a major step toward autonomous AI coding assistants that can handle long-horizon tasks and anticipate developer needs, potentially accelerating software development and transforming team workflows. Managed agents are hosted services that maintain stable interfaces for extended autonomous operations, proactive workflows allow AI to initiate actions without explicit user prompts, and the capability curve describes predictable improvement trajectories for AI models.

rss · InfoQ · May 18, 13:14

Background: AI-assisted coding tools like Claude Code help developers write, review, and debug code. Managed agents extend this by running complex tasks over long durations. Proactive workflows enable AI to suggest or execute actions based on context. The capability curve provides a framework for understanding how AI performance scales with investment and time, aiding strategic planning.

Tags: #AI, #software development, #Anthropic, #Claude, #AI-assisted coding

Navigation API Reaches Baseline, Replacing History API
Navigation API 达到基线，取代 History API ⭐️ 8.0/10

The Navigation API has become baseline and is now available in all major browsers as of January 2026, providing a modern replacement for the History API with unified events and better error handling. This simplifies client-side navigation in single-page applications, reducing bugs and improving developer experience. It marks a significant evolution in web platform capabilities for routing. Key features include a unified navigate event, automatic URL updates, and integrated error handling, addressing long-standing issues with the History API.

rss · InfoQ · May 18, 06:33

Background: The History API allowed manipulation of browser session history but lacked a unified event model and clear error handling. Client-side routing in SPAs often relied on workarounds. The Navigation API provides a more robust and standardized approach.

Tags: #Navigation API, #History API, #Client-side Routing, #Web Development, #Browser APIs

Google A/B Tests Infrastructure Fleet-Wide Safely
谷歌在大规模基础设施上安全进行 A/B 测试 ⭐️ 8.0/10

Google published a detailed methodology for safely conducting A/B experiments on low-level infrastructure components like memory allocators and kernel schedulers across its entire fleet. This enables Google to validate performance and efficiency improvements at massive scale without risking fleet-wide outages, providing a model for other large-scale systems engineering teams. The methodology emphasizes four pillars: application-level vs machine-level experimentation, maintaining a balanced setup, ensuring binary hermeticity, and selecting the right performance metrics.

rss · Cloud Blog · May 18, 16:00

Background: A/B testing is commonly used for user-facing changes like UI tweaks, but applying it to low-level infrastructure is risky because a bug could bring down many machines. Memory allocators manage dynamic memory allocation, while kernel schedulers allocate CPU time to processes, and both are critical for system performance.

Tags: #A/B testing, #infrastructure, #Google, #experimentation, #systems engineering

GPT-3 Paper Review: Few-Shot Learning Breakthrough
GPT-3 论文解读：少样本学习突破 ⭐️ 8.0/10

This article reviews the landmark GPT-3 paper, which demonstrates that large language models can perform a wide range of tasks with only a few examples, without fine-tuning. GPT-3's few-shot learning capability significantly reduces the need for task-specific training data, potentially lowering the barrier to applying AI in many domains. GPT-3 has 175 billion parameters and was trained on a diverse corpus of text. The paper shows that simply scaling up model size and data leads to emergent few-shot abilities.

rss · freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More · May 18, 20:29

Background: Few-shot learning is a machine learning approach where models learn to make predictions from very few labeled examples, mimicking human ability to learn from limited data. GPT-3 is a transformer-based language model that, without task-specific fine-tuning, can perform tasks like translation, question answering, and text generation by being given a few examples in the input prompt.

Tags: #GPT-3, #few-shot learning, #language models, #AI paper review

SpaceX Starship V3 First Flight Key to Planned IPO
SpaceX 押注 Starship V3 首飞为 IPO 铺路 ⭐️ 8.0/10

SpaceX has secretly filed for an IPO in April with a target listing in June, and the first flight of the Starship V3 is seen as a critical validation step to support the company's public offering narrative. The success of Starship V3 directly influences how capital markets perceive SpaceX's growth story, as Starship is central to the company's future launch capacity, satellite deployment, and higher launch frequency plans. Starship V3 is about 5 feet (1.5 meters) taller than previous versions and features the new Raptor 3 engine, which is sleeker, more powerful, and more reliable; but the program has faced setbacks, including the failure of the first four Block 2 upper stages in 2025.

telegram · zaihuapd · May 18, 13:45

Background: Starship is a two-stage, fully reusable super heavy-lift launch vehicle under development by SpaceX, intended as the successor to the Falcon 9 and Falcon Heavy rockets. It consists of the Super Heavy booster and Starship spacecraft, both powered by Raptor engines burning liquid methane and liquid oxygen. The rocket has undergone multiple test flights since April 2023, with several failures and partial successes, and is designed to reduce launch costs through reusability and mass production.

Tags: #space, #spacex, #starship, #ipo