OpenAI Proposes 5% Equity Transfer to US Government for Public Wealth Fund
OpenAI 提议向美国政府转让 5%股份以建立公共财富基金 ⭐️ 9.0/10

OpenAI is exploring a plan to transfer 5% of its equity to the US government to establish a Public Wealth Fund, allowing citizens to directly benefit from AI-driven economic growth. CEO Sam Altman has been in discussions with Trump administration officials since early 2025. This unprecedented proposal could reshape how the benefits of advanced AI are distributed, potentially creating a model for other nations to follow. If implemented, it would give ordinary citizens a direct stake in AI's economic upside, addressing growing concerns about wealth inequality from automation. The proposed Public Wealth Fund would be initially capitalized by contributions from AI companies and the government, investing in diversified assets including AI firms. Dividends would be distributed directly to every citizen, regardless of their initial wealth or access to financial markets.

rss · 宝玉(@dotey) · Jul 2, 05:19

Background: OpenAI, currently valued at $852 billion, is a leading AI research and deployment company. A Public Wealth Fund is a state-owned investment fund that invests for long-term growth, similar to sovereign wealth funds. The proposal aims to ensure that the immense value generated by AI benefits the broader population, not just shareholders.

Tags: #AI Governance, #Public Wealth Fund, #OpenAI, #AI Ethics

One AI agent runs 99.98% of Gumroad
一个 AI 智能体运行 Gumroad 的 99.98%业务 ⭐️ 9.0/10

Sahil Lavingia, founder of Gumroad, stated that a single AI agent named Gumclaw autonomously handles 99.98% of Gumroad's operations, including support, operations, and engineering. This claim suggests a potential paradigm shift in business automation, where a single AI agent could run nearly all aspects of a company, potentially reducing human labor and operational costs if broadly applicable. The agent, Gumclaw, is described as an autonomous AI agent that handles end-to-end operations. The claim is based on actual deployment at Gumroad, a platform for creators to sell digital products.

rss · Sahil Lavingia(@shl) · Jul 2, 15:21

Background: AI agents are autonomous software programs that can perform tasks without human intervention. Gumroad was founded by Sahil Lavingia and has paid out over $500 million to creators. If true, this claim would represent a major milestone in AI-driven automation.

Discussion: Some users enthusiastically endorsed the claim, seeing it as the ideal goal for efficient automation, while others dismissed it as unrealistic or insufficient for complex business operations.

Tags: #AI Agents, #Automation, #Business Operations, #Gumroad, #Sahil Lavingia

Virginia Bans Sale of Geolocation Data Without Consent
弗吉尼亚州禁止未经同意出售地理位置数据 ⭐️ 8.0/10

Virginia has passed a law that bans the sale of geolocation data without the consumer's consent, adding to existing privacy protections in the state. This law is significant as it sets a precedent for other states and strengthens consumer privacy by limiting how tech companies can monetize sensitive location information. The ban applies to the sale of 'geolocation data' without consent, which includes data from GPS, Wi-Fi, and cell towers. Violations could result in enforcement actions by the state.

hackernews · toomuchtodo · Jul 2, 21:03 · Discussion

Background: Geolocation data refers to information that identifies the physical location of a device or entity, often obtained through GPS, Wi-Fi, or cellular triangulation. Such data is collected by various apps and services and can be sold to third parties for advertising or other purposes. Virginia previously passed the Virginia Consumer Data Protection Act (VCDPA), and this new law builds on those protections.

Discussion: Commenters expressed mixed reactions, with some praising the law as a good start but noting it needs real enforcement teeth (jmward01). Others raised concerns about enforcement challenges for out-of-state companies and specific applications like automated license plate readers (smalltorch, dv_dt). One comment linked the law to recent reports of location data being used for anti-abortion ad campaigns (danielrmay).

Tags: #privacy, #geolocation data, #legislation, #data protection, #regulation

American Privacy Emergency Exposed
美国隐私紧急状态被揭露 ⭐️ 8.0/10

Scott Aaronson reveals a secret Trump-era directive that pressured social media companies to provide user data without obtaining warrants. This directive undermines Fourth Amendment protections and sets a dangerous precedent for government surveillance without judicial oversight, threatening civil liberties. The directive was issued by the Trump administration and remained secret; it pressures social media platforms to hand over data, bypassing the traditional warrant process.

hackernews · flowercalled · Jul 3, 00:01 · Discussion

Background: In the United States, the Fourth Amendment generally requires law enforcement to obtain a warrant based on probable cause to search or seize data. This directive allegedly attempted to circumvent that requirement, raising serious privacy concerns.

Discussion: Commenters discussed the implications of differential privacy for census data, questioned the political motivation behind the directive, and noted a missing link to contact legislators. Some comments were off-topic, such as those on parental leave and corporate influence.

Tags: #privacy, #surveillance, #civil liberties, #government overreach, #social media

Podman v6.0.0: New Network Stack and Quadlet Enhancements
Podman v6.0.0：新网络栈与 Quadlet 增强 ⭐️ 8.0/10

Podman v6.0.0 introduces a new network stack and Quadlet enhancements, marking a major version release of the open-source container engine. This release further solidifies Podman's position as a leading Docker alternative by improving networking and systemd integration, benefiting users seeking daemonless and rootless container management. The new network stack brings performance and security improvements, while Quadlet enhancements allow for more declarative container management under systemd.

hackernews · soheilpro · Jul 2, 14:23 · Discussion

Background: Podman is a daemonless container engine developed by Red Hat that can run containers with or without root. It uses a fork-exec model instead of a client-server architecture like Docker. Quadlet is a tool that integrates Podman containers with systemd using declarative unit files (.container, .volume, etc.).

Discussion: Commenters generally praised Podman, with many sharing positive migration experiences from Docker. One user highlighted ease of switching from Docker Desktop. However, a user expressed frustration over limited official installation support for Ubuntu, which hinders adoption.

Tags: #containerization, #Podman, #networking, #open source, #Docker alternative

How to Ask Strangers for Help Effectively
如何有效向陌生人求助 ⭐️ 8.0/10

A practical guide published on pradyuprasad.com outlines strategies for asking strangers for help, emphasizing proof of work and respect for their time. This advice is valuable for professionals, especially software engineers, who often need to network and seek mentorship. It provides actionable tips that can improve the success rate of cold outreaches. The guide highlights that showing genuine effort (proof of work) upfront is more important than the length or polish of the request. It also advises brevity and clarity in the ask.

hackernews · FigurativeVoid · Jul 2, 13:19 · Discussion

Background: Asking strangers for help, such as reaching out on LinkedIn or email, can be intimidating and often ignored. The key is to demonstrate that you have done your homework and value the person's expertise, making it easy for them to say yes.

Discussion: Commenters shared personal experiences, noting that proof of work must be deep, not superficial. Some suggested offering to pay for the person's time as a sign of seriousness, which often leads to free help.

Tags: #communication, #career-advice, #professional-development, #networking, #soft-skills

Immich 3.0 Released: Major Update to Self-Hosted Photo Manager
Immich 3.0 发布：自托管照片管理器重大更新 ⭐️ 8.0/10

Immich 3.0, a major version of the open-source self-hosted photo management platform, has been released with significant updates and improvements. This release solidifies Immich as a leading open-source alternative to proprietary photo management services like Google Photos and Apple Photos, emphasizing user privacy and data ownership. Immich 3.0 continues to offer self-hosted photo and video management with features such as album sharing without requiring an account for contributors, and integration with VPNs like Tailscale for secure remote access.

hackernews · hashier · Jul 2, 14:13 · Discussion

Background: Immich is an open-source, self-hosted photo and video management platform that gives users full control over their personal media. It is designed as a privacy-focused alternative to cloud services like Google Photos and Apple Photos, allowing users to store and manage their photos on their own servers.

Discussion: The community is generally positive, with many users calling Immich a 'no-brainer replacement' for Google Photos. However, some express concerns about the lack of end-to-end encryption and iOS sync issues, comparing it to alternatives like Ente.

Tags: #self-hosted, #photo management, #open source, #privacy, #software release

Every team shares five practical Codex workflows
Every 团队分享五条 Codex 实用工作流 ⭐️ 8.0/10

The Every team published an in-depth article detailing five distinct workflows for using Codex, ranging from non-technical builders to long-thread integration, showing how each member leverages Codex for different tasks. This provides real-world, practical insights into how Codex can be used beyond simple coding, demonstrating its versatility for productivity, automation, and knowledge work, which could influence how others adopt AI assistants. The five workflows include Natalia's 'low-friction Claude Code' for non-technical users, Dan's long-running threads with built-in browser and router threads, Katie's local context folder system, Austin's lightweight results-oriented approach, and Kieran's portable cross-agent context system. Common patterns emphasize context over prompts, letting Codex design its own system, delegating repetitive work, and establishing audit feedback loops.

rss · meng shao(@shao__meng) · Jul 3, 00:37

Background: Codex is an AI agent developed by OpenAI that can read and edit files, run tests, and invoke tools in a cloud environment. It functions like a coding agent for software development but is also used for broader productivity tasks. Attio is a modern CRM platform that integrates with Codex for automating customer relationship management.

Tags: #Codex, #AI-assisted development, #workflow, #productivity, #real-world practices

Three LLM Interaction Paradigms: From Chatbots to Embedded AI
LLM 交互三范式：从聊天机器人到嵌入式 AI ⭐️ 8.0/10

A new analysis delineates three LLM interaction paradigms: web chatbots, standalone AI apps, and organizationally embedded AI, exemplified by Claude Tag and Glean Agents. Claude Tag shifts from per-user to per-channel AI with persistent context, while Glean Agents introduce identity, memory, proactivity, and accountability for enterprise agents. This progression matters because organizational knowledge is scattered across tools, and channel-level context alone is insufficient for truly useful AI agents. The shift to embedded, cross-system agents with governed context layers could unlock significant productivity gains for enterprise teams. Claude Tag operates in Slack channels with admin-governed access, remembering channel history and planning future tasks. Glean Agents connect to enterprise systems like Jira, Confluence, and GitHub, enforcing runtime permissions and supporting proactive workflows such as automated incident response.

rss · meng shao(@shao__meng) · Jul 2, 14:08

Background: The first two paradigms—web chatbots and standalone AI apps—limit AI to individual interactions or isolated use cases. The emerging third paradigm embeds AI as a persistent, team-level participant that spans organizational tools. A critical challenge is building a unified, governed context layer that integrates data from multiple systems with proper permissions.

Tags: #LLM, #Enterprise AI, #Claude, #AI Agents, #Collaboration

Why AI Struggles with Design: Codex Lead Explains
AI 为何不擅长设计：Codex 负责人解读 ⭐️ 8.0/10

OpenAI Codex lead Andrew Ambrosino explained that AI struggles with design because labs prioritize coding (which accelerates AI research and is easier to grade) and because great design requires novelty and cultural understanding, which current models lack. This analysis highlights fundamental barriers to AI creativity, affecting industries that rely on design. It underscores the need for new evaluation metrics and training approaches beyond coding-focused benchmarks. Ambrosino identified two core issues: practical prioritization (coding is easier to grade) and structural challenges (novelty and culture). He believes practical issues will be solved, but novelty and cultural adaptation will take longer.

rss · Lenny Rachitsky(@lennysan) · Jul 2, 18:10

Background: OpenAI Codex is an AI system that translates natural language into code, used in a desktop app for agentic coding. Unlike coding, design involves subjective judgment, cultural context, and originality, making it harder to automate. Current AI models tend to replicate existing patterns rather than generate truly novel outputs.

Tags: #AI, #design, #creativity, #OpenAI, #machine learning

Apple Hide My Email Vulnerability Exposes Real Addresses
Apple 隐藏邮件漏洞暴露真实地址 ⭐️ 8.0/10

A vulnerability in Apple's Hide My Email tool allows anyone to uncover a user's real email address, and Apple has not patched it for over a year, as reported by security researchers and 404 Media. This vulnerability undermines a key privacy feature of iCloud+, affecting millions of users who rely on it to protect their personal email. The lack of a timely fix raises concerns about Apple's commitment to user privacy and security. The vulnerability is reportedly trivial to exploit and does not require special access, yet Apple has not issued a patch despite being notified over a year ago. The exact mechanism has not been publicly detailed to avoid further exploitation.

rss · Michael Tsai · Jul 2, 19:44

Background: Hide My Email is a feature available to iCloud+ subscribers that generates unique, random email addresses to forward to a user's real inbox, allowing them to keep their actual email private. This vulnerability could leak the real address through the email forwarding process, potentially enabling spam, tracking, or phishing attacks.

Discussion: On Reddit and other forums, users express frustration with Apple's slow response, with some questioning the reliability of Apple's privacy features. Many recommend using alternative services like SimpleLogin for more robust protection.

Tags: #security, #privacy, #Apple, #vulnerability, #email

Apple Rebuilds Spotlight Search with LLM and RAG
苹果用 LLM 和 RAG 重建 Spotlight 搜索 ⭐️ 8.0/10

Apple announced that it has rebuilt the search infrastructure for Spotlight, Photos, and Mail using large language models (LLMs) and retrieval-augmented generation (RAG), implementing a new SpotlightSearchTool and LanguageModelSession. This integration marks a major shift in Apple's software capabilities, bringing advanced LLM-powered search to core apps across all platforms, significantly improving search relevance and user experience. The rebuilt search leverages Core Spotlight with delegate-based hydration patterns, using SpotlightSearchTool and LanguageModelSession to enable retrieval-augmented generation directly on device.

rss · Michael Tsai · Jul 2, 19:39

Background: Retrieval-augmented generation (RAG) is a technique that allows LLMs to retrieve information from external sources before generating responses, reducing hallucinations and enabling access to updated information. Core Spotlight is Apple's framework for indexing app content for system-wide search. By combining these, Apple's new search infrastructure can index and retrieve data from apps and generate intelligent results.

Tags: #Apple, #Spotlight, #LLM, #Search, #AI

Fur and Neural Sound Added to RL Creatures Using Fable 5
用 Fable 5 给 RL 训练生物添加毛发和神经声音 ⭐️ 8.0/10

A project uses Fable 5 to add fur and sounds generated directly from neural network activations to creatures trained with reinforcement learning via JAX and MuJoCo, rendered with three.js and WebGPU. This integration demonstrates a novel combination of reinforcement learning, neural audio synthesis, and real-time web-based rendering, pushing the boundaries of interactive AI-generated creatures. The sounds are not sampled but synthesized from neural activations, and the rendering uses three.js with WebGPU for high-performance graphics in the browser.

rss · AI Will(@FinanceYF5) · Jul 2, 06:01

Background: Reinforcement learning (RL) trains agents through trial and error. JAX is a numerical computing library, and MuJoCo is a physics simulator for robotics and biomechanics. three.js is a 3D graphics library for the web, and WebGPU is a modern GPU API for browsers. Fable 5 is a state-of-the-art AI model developed by Anthropic.

Tags: #reinforcement learning, #neural sound, #simulation, #rendering, #MuJoCo

Google Research Launches TabFM: Zero-Shot Foundation Model for Tabular Data
谷歌研究院发布 TabFM：表格数据零样本基础模型 ⭐️ 8.0/10

Google Research has introduced TabFM, a foundation model specifically designed for tabular data classification and regression. It enables zero-shot predictions on unseen tables in a single forward pass without any training or fine-tuning. TabFM represents a significant step toward simplifying machine learning on tabular data, which is ubiquitous in industry. It eliminates the need for time-consuming hyperparameter tuning and feature engineering, potentially accelerating many data science workflows. TabFM is scikit-learn compatible and supports both classification and regression tasks. The model is pretrained on a large corpus of tabular data and can generate predictions in a forward pass without task-specific training.

rss · AI Will(@FinanceYF5) · Jul 2, 05:13

Background: Tabular data, organized in rows and columns, is the most common format in business and science. Traditional machine learning approaches like XGBoost require careful hyperparameter tuning and feature engineering. Foundation models, which are large-scale pretrained models, have shown remarkable zero-shot capabilities in other domains like NLP and computer vision, and TabFM extends this paradigm to tabular data.

Tags: #tabular data, #foundation model, #Google Research, #machine learning, #zero-shot learning

Browser Use CLI 3.0 Turns Any Model into SOTA Browser Agent
Browser Use CLI 3.0 将任何模型转变为最先进的浏览器代理 ⭐️ 8.0/10

Browser Use CLI 3.0 introduces direct Chrome DevTools Protocol (CDP) control via a new browser-harness, enabling any AI model to function as a state-of-the-art browser agent while being 6 times smaller and using fewer tokens. This release significantly lowers the barrier for creating powerful browser automation agents, as developers can now leverage any model without specialized training or large token budgets, potentially accelerating applications in web testing, data extraction, and AI-driven workflows. The CLI uses a self-healing harness that allows agents to edit their own helpers mid-task, and it supports running on cloud browsers or a real Chrome instance. Direct CDP control replaces higher-level abstractions, resulting in a 6x smaller footprint and reduced token usage.

rss · Browser Use(@browser_use) · Jul 2, 15:10

Background: The Chrome DevTools Protocol (CDP) is a remote debugging protocol that allows tools to instrument, inspect, and debug Chromium-based browsers. Browser automation frameworks typically use higher-level APIs like Playwright or Puppeteer, but Browser Use CLI 3.0 bypasses these with direct CDP calls for greater efficiency and flexibility. The browser-harness component is a thin, self-healing harness built directly on CDP, enabling LLMs to complete any browser task by editing their own helpers as needed.

Tags: #browser automation, #AI agents, #CLI, #CDP, #open-source

First precision editing study in human embryos identifies master gene
人类胚胎首次精准编辑研究识别出主控基因 ⭐️ 8.0/10

Researchers at the University of Cambridge used base editing, a precise genome editing technique, for the first time in viable human embryos to study early development and identified a master gene that controls the first cell fate decisions. This breakthrough offers unprecedented insights into human embryogenesis, potentially improving understanding of developmental disorders, infertility, and early pregnancy loss. The study used base editing to introduce specific point mutations without causing double-strand breaks, focusing on a gene that regulates the differentiation of inner cell mass and trophectoderm.

rss · Hacker News: Newest · Jul 3, 02:03

Background: Base editing is a precise genome editing technique that allows direct conversion of one DNA base to another without requiring double-strand breaks, unlike traditional CRISPR-Cas9. This reduces unintended edits and improves safety for applications in development and disease modeling. Previous human embryo genome editing studies used CRISPR-Cas9, which can introduce off-target effects.

Tags: #genetics, #embryo development, #precision editing, #CRISPR, #research

LangSmith Engine: An AI agent that fixes other agents
LangSmith Engine：一个能修复其他 AI Agent 的 AI Agent ⭐️ 8.0/10

LangChain launched LangSmith Engine, a self-improving agent that analyzes production traces, clusters failures, and autonomously opens pull requests to fix them. It uses a team of specialized sub-agents—screener, verifier, etc.—delegated by a main model, akin to an org chart. This shift from monolithic agents to team-based architectures could dramatically improve reliability and scalability of AI agents in production. LangChain, a key player in the agent ecosystem, is pushing the frontier of autonomous self-healing systems. Engine generates an 'agent overview document' for context and uses a sandbox to test fixes before shipping. It clusters millions of traces into condensed views and can even run on itself for meta self-improvement.

rss · LangChain · Jul 2, 14:00

Background: LangSmith Engine is an always-on agent that monitors production traces of other AI agents, identifies failure patterns, and proposes code fixes via pull requests. Its architecture uses a main orchestrator agent that delegates to four sub-agents (screener, verifier, etc.), inspired by organizational hierarchies. This approach contrasts with single-agent systems by enabling specialized, cost-effective processing of large-scale trace data.

Tags: #AI agents, #LangChain, #agent architecture, #self-improving, #AI engineering

Deep Agents Adds Programmatic Subagents with Six Patterns
Deep Agents 新增编程性子代理，支持六种模式 ⭐️ 8.0/10

LangChain announced that Deep Agents, an open-source and model-agnostic agent harness, now supports programmatic subagent orchestration. The update includes six common usage patterns such as fan-out, adversarial verification, and recursive loops. This feature allows developers to orchestrate subagents using code (e.g., JavaScript loops and branches) instead of relying on the LLM to choose one subagent at a time, leading to more reliable and scalable multi-agent workflows. It addresses context window limitations and enables complex, multi-step pipelines. The six patterns include fan-out, adversarial verification, recursive loops, and more, detailed in a blog post and a video by Sydney Runkle. A separate post by Hunt Lovell explains how the code interpreter was made secure without requiring a sandbox.

rss · Harrison Chase(@hwchase17) · Jul 2, 14:52

Background: Deep Agents is an open-source harness for building AI agents that can use various models. Subagents are specialized agents that handle specific subtasks, and programmatic orchestration lets the main agent dispatch them via code, ensuring deterministic control. This approach contrasts with tool-based subagent calls where the model decides when to invoke each subagent.

Tags: #LangChain, #AI Agents, #Open Source, #Subagents, #Deep Agents

Open-source realtime voice demo from Hugging Face and Cerebras
Hugging Face 与 Cerebras 联合推出开源实时语音演示 ⭐️ 8.0/10

Thomas Wolf, co-founder of Hugging Face, announced a fully open-source realtime speech-to-speech demo built in collaboration with Cerebras, including both models and code. The demo is hosted on Hugging Face Spaces and accompanied by a technical blog post. This release significantly lowers the barrier for developers to experiment with and deploy realtime speech-to-speech AI, as it provides fully open-source components that can be forked and customized. It demonstrates the growing capability of open-source AI to achieve competitive performance with proprietary systems. The demo is described as 'raw, no cut, no speed-up' to emphasize its realtime performance. It leverages Cerebras' wafer-scale processors for inference acceleration, enabling low-latency voice interaction. The code and models are available for anyone to replicate or improve.

rss · Thomas Wolf(@Thom_Wolf) · Jul 2, 23:31

Background: Speech-to-speech AI converts spoken input directly into spoken output, enabling applications like realtime translation or voice cloning. Prior open-source efforts often suffered from high latency or limited quality, making them impractical for realtime use. Cerebras provides specialized hardware that can reduce inference latency significantly compared to traditional GPUs.

Tags: #speech-to-speech, #open-source, #realtime, #AI, #Hugging Face

CubeSandbox: Tencent's Open-Source RustVMM+KVM Sandbox
腾讯开源 CubeSandbox：基于 RustVMM+KVM 的轻量级硬件虚拟化沙箱 ⭐️ 8.0/10

Tencent has open-sourced CubeSandbox, a production-grade hardware virtualization sandbox built on RustVMM and KVM, achieving 60ms cold start and under 5MB memory overhead, supporting thousands of concurrent instances per node. This release significantly lowers the resource barrier for secure code execution environments, making hardware-level isolation practical for high-density scenarios like AI agent sandboxing and multi-tenant serverless computing. It demonstrates the viability of Rust-based virtualization for performance-critical cloud infrastructure. CubeSandbox is licensed under Apache 2.0 and provides a complete sandbox-as-a-service stack, including pre-built templates for running untrusted code. Its ultra-low overhead allows packing thousands of sandboxes on a single machine, which is critical for scenarios requiring rapid scaling and strict isolation.

rss · 机器之心SOTA模型 · Jul 2, 05:30

Background: Hardware virtualization sandboxes use hypervisor-level isolation (e.g., KVM) to separate guest OS from host, providing stronger security than process-level sandboxes. RustVMM is a set of Rust crates for building custom virtual machine monitors, offering memory safety and performance. CubeSandbox combines these technologies to create lightweight, secure sandbox environments.

Tags: #virtualization, #sandbox, #open source, #Tencent, #RustVMM

Kent Beck: Human Aspects Harder Than Code in AI Era
Kent Beck：AI 时代人的部分比代码更难 ⭐️ 8.0/10

In a podcast interview, Kent Beck argues that the human aspects of software engineering—trust, communication, and collaboration—are more challenging than code itself in the AI era, sharing reflections on his career and the evolution of practices like TDD and XP. As AI coding tools accelerate code generation, Beck's insights from 50 years of experience remind the industry that technical solutions alone are insufficient; focusing on human factors is critical to maintaining quality and trust in software development. Beck introduced his '3X' model (explore, expand, extract) for AI-assisted development and noted that TDD has been misused as a 'moral club' rather than a tool for safety, while also reflecting on the decline of Smalltalk and the origins of Agile.

rss · 跨国串门儿计划 · Jul 2, 22:09

Background: Smalltalk is a pure object-oriented programming language created at Xerox PARC in the 1970s. Test-driven development (TDD) and Extreme Programming (XP) are software methodologies pioneered by Beck to improve code quality through iterative testing and human-centric practices. In the AI era, the rapid generation of code by tools like large language models makes trust and human oversight even more essential.

Tags: #software engineering, #AI, #Kent Beck, #agile, #TDD

SwiftUI at WWDC 2026: New Document Protocol and Performance Boost
WWDC 2026：SwiftUI 新增文档协议并提升性能 ⭐️ 8.0/10

At WWDC 2026, Apple introduced a new SwiftUI Document protocol for efficient disk access and snapshot-based updates, along with performance improvements including AsyncImage caching, lazy state initialization for Observable types, and enhanced APIs for reordering items and swipe actions. These updates significantly streamline document-based app development in SwiftUI and address long-standing performance pain points, such as the absence of built-in AsyncImage caching and premature state initialization. Developers can now build more responsive apps with less boilerplate and fewer third-party dependencies. The new Document protocol replaces the older FileDocument protocol with snapshot-based updates for better efficiency. AsyncImage now caches images automatically in iOS 27 and later, and a new @LazyState attribute or similar mechanism enables lazy initialization of Observable types, reducing unnecessary view updates.

rss · InfoQ · Jul 2, 16:00

Background: SwiftUI is Apple's declarative UI framework for building apps across all Apple platforms. Prior to this update, SwiftUI's DocumentGroup used the FileDocument protocol which required manual file handling. AsyncImage lacked built-in caching, forcing developers to rely on third-party libraries like Kingfisher. Similarly, @State would initialize Observable instances immediately, causing performance issues.

Tags: #SwiftUI, #WWDC, #Apple, #iOS Development, #Performance

Apple Extends Private Cloud Compute to Google Cloud with Enhanced Security
苹果首次将 Private Cloud Compute 扩展至 Google Cloud ⭐️ 8.0/10

Apple has announced it will extend its Private Cloud Compute (PCC) to Google Cloud for the first time, utilizing NVIDIA Blackwell GPUs, Intel Trust Domain Extensions (TDX), and Google's Titan security chip to protect user data. This marks Apple's first use of an external cloud provider for its private AI compute, setting a new precedent for cross-cloud confidential computing and potentially influencing industry security standards. Apple maintains an independent append-only hardware ledger and dual-vendor attestation roots, and notably AWS and Azure are not part of this collaboration.

rss · InfoQ · Jul 2, 10:04

Background: Private Cloud Compute is Apple's infrastructure for processing AI requests from users while preserving privacy. Intel TDX provides hardware-level isolation for virtual machines, and Google's Titan chip is a custom security chip for verifying platform integrity. The dual-vendor attestation roots ensure verification through multiple independent hardware vendors.

Tags: #Apple, #Google Cloud, #Private Cloud Compute, #NVIDIA Blackwell, #security

Netflix's Prioritized Load Shedding in Envoy Proxies
Netflix 在 Envoy 代理中实现优先级负载削减 ⭐️ 8.0/10

Netflix engineers presented a method using Envoy sidecar proxy to prioritize user-initiated requests over non-critical traffic during extreme traffic spikes, incorporating automated chaos testing and retry storm mitigation. This approach enhances reliability for large-scale microservices by ensuring critical traffic survives extreme loads, preventing cascading failures and service degradation. The prioritized load shedding mechanism allows user-initiated requests to draw capacity from non-critical traffic; the presentation also covered automated chaos load testing, configuration generation, and retry storm mitigation strategies.

rss · InfoQ · Jul 2, 09:20

Background: Load shedding is a traffic management technique that drops less important requests to protect critical services during overload. Envoy is a popular proxy that handles network traffic in microservices environments. Retry storms occur when repeated retries amplify load during failures, worsening service recovery.

Tags: #load shedding, #reliability, #Netflix, #Envoy, #traffic management

Google Disrupts NetNut Residential Proxy Network
谷歌捣毁 NetNut 住宅代理网络 ⭐️ 8.0/10

Google, in coordination with the FBI and Lumen, disrupted the NetNut residential proxy network (also known as Popa) by disabling accounts, sharing intelligence, and enabling Google Play Protect to block apps using NetNut SDKs, reducing the available pool of devices by millions. NetNut is one of the largest residential proxy networks, with at least 2 million devices; its disruption significantly degrades the infrastructure used by cybercriminals for credential stuffing, ad fraud, and other malicious activities, demonstrating ongoing public-private collaboration in cybersecurity. Google disabled Google accounts and services used by NetNut for malware command and control (C2), shared technical intelligence on NetNut SDKs and C2 infrastructure with law enforcement and platforms, and ensured Google Play Protect automatically warned users and disabled apps containing NetNut SDKs.

rss · Cloud Blog · Jul 2, 14:00

Background: Residential proxy networks like NetNut hijack consumer devices (PCs, smartphones) by installing software that routes traffic through those devices, providing anonymity to buyers. These networks are used for web scraping, geo-unblocking, and cybercrime. Google previously disrupted the IPIDEA proxy network in January 2026, a similar operation targeting a large botnet-for-proxy service.

Tags: #cybersecurity, #proxy network, #threat intelligence, #Google, #law enforcement

QuasiMoTTo: Correlated Sampling Reduces Inference Compute Cost
QuasiMoTTo：关联采样降低推理计算成本 ⭐️ 8.0/10

Researchers introduced QuasiMoTTo, a method that uses correlated samples instead of independent parallel attempts to reduce the cost of scaling inference compute, achieving 25-47% fewer samples in test-time scaling and 50% fewer training steps in reinforcement learning. This addresses a key inefficiency in scaling inference compute for large language models, potentially saving significant computational resources and accelerating research in test-time scaling and reinforcement learning. QuasiMoTTo generates correlated samples that have higher coverage, are marginally exact draws from the LLM, and can be generated in parallel. The method explores the design space of correlated samplers.

rss · Stanford AI Lab(@StanfordAILab) · Jul 2, 17:44

Background: Scaling inference compute involves generating multiple outputs for a single input to improve performance, e.g., in test-time scaling or reinforcement learning. Independent sampling often wastes compute by rediscovering the same solutions. Correlated sampling aims to increase diversity and coverage with fewer total samples.

Tags: #inference compute, #scaling, #correlated sampling, #efficiency, #AI

GitHub's secret scanning team reaches inbox zero in 9 months
GitHub 秘密扫描团队九个月达到收件箱清零 ⭐️ 8.0/10

GitHub's security team successfully triaged and resolved over 20,000 secret scanning alerts across 15,000 repositories, achieving inbox zero in nine months. This case study provides a practical, battle-tested methodology for managing large volumes of security alerts, demonstrating how organizations can systematically reduce noise and prioritize remediation, which is critical for DevSecOps practices. The team separated signal from noise by building custom remediation workflows, and they focused on high-priority alerts first, automating response for known patterns.

rss · The GitHub Blog · Jul 2, 16:00

Background: Secret scanning is a GitHub feature that scans repositories for hardcoded credentials like API keys and tokens. It alerts repository owners when secrets are detected, helping prevent security breaches. However, large organizations often face alert fatigue due to the volume of notifications.

Tags: #security, #secret scanning, #GitHub, #DevSecOps, #alert management

Cloudflare to Block Mixed-Use AI Crawlers Default Starting Sept
Cloudflare 9 月起默认拦截混合用途 AI 爬虫 ⭐️ 8.0/10

Cloudflare announced that starting September 15, 2026, it will default-block 'mixed-use' crawlers used for search, AI chat, and training from accessing pages with ads, and specifically criticized Google for exploiting a loophole. This policy shift gives website owners more control over AI data scraping and could force AI companies to negotiate fair compensation for using publisher content, potentially reshaping the economics of web content consumption. The blocking applies to 'mixed-use' crawlers that serve multiple purposes like indexing for search and training AI models, but websites can still allow Googlebot for search while blocking other uses via granular policies.

telegram · zaihuapd · Jul 2, 05:37

Background: Web crawlers are programs that automatically browse and index web pages for search engines or AI training. AI companies have been scraping content without permission, leading many websites to block AI crawlers. However, some crawlers like Googlebot are used for both search indexing and AI training, making it hard for sites to allow search while blocking AI use. Cloudflare's new AI Crawl Control provides tools for website owners to monitor and manage crawler access granularly.

Discussion: The provided comment highlights that many sites block AI crawlers but not Google search, allowing Google to exploit this loophole for AI training, which aligns with the news's criticism of Google.

Tags: #Cloudflare, #AI crawlers, #Google, #web scraping, #content policy

Meituan to Cover Occupational Injury Insurance for All Riders from July 2026
美团 7 月起全国骑手职业伤害险全覆盖 ⭐️ 8.0/10

Starting July 1, 2026, Meituan will fully cover occupational injury insurance for all its delivery riders nationwide, ensuring every order and every rider is insured. The company has already paid over 30 billion RMB in cumulative premiums for more than 10 million riders by June 2026. This policy significantly expands labor protections for millions of gig workers in China, setting a precedent for platform companies to provide comprehensive insurance coverage. It addresses long-standing concerns about the vulnerability of delivery riders and could influence industry-wide standards. The pilot program, initiated in July 2022, now extends to all 31 provinces and Xinjiang Production and Construction Corps. Riders do not need to pay; the platform pays per order monthly, covering scenarios such as falls and vehicle collisions during deliveries.

telegram · zaihuapd · Jul 2, 10:44

Background: China's gig economy has grown rapidly, with millions of riders working for platforms like Meituan, often lacking traditional labor protections. In 2021, regulators pushed platforms to improve worker welfare, leading to the pilot occupational injury insurance for new employment forms. The pilot initially covered a limited number of platforms and regions, and its nationwide expansion marks a major step toward universal coverage.

Tags: #gig economy, #labor insurance, #Meituan, #policy change, #platform workers

Citibank bans GPT-5.5; companies curb AI use due to soaring costs
花旗禁用 GPT-5.5；多家公司因 AI 成本飙升限制员工使用 ⭐️ 8.0/10

Citibank has completely disabled access to advanced AI models like GPT-5.5 and Claude Opus 4.6/4.7 as of June 24, citing excessive consumption of AI credits. Atlassian's monthly AI spending surged from $5 million to over $15 million between August 2025 and May 2026, prompting the company to end unlimited usage and introduce cost-tracking dashboards. This trend reveals that the pay-per-use pricing model for AI tools can lead to unpredictable and rapidly escalating costs for enterprises, forcing even large companies to restrict employee access. It signals a critical challenge for enterprise AI adoption: balancing productivity gains with budget control. Adobe has decided not to renew its unlimited Claude usage contract, which expired on June 30. Amazon previously shut down an internal leaderboard that encouraged AI use, and employees later discovered previously unknown token usage caps. Consulting giant Accenture is packaging AI cost management as a new business opportunity for its clients.

telegram · zaihuapd · Jul 2, 13:59

Background: AI services like ChatGPT and Claude are typically charged based on token consumption—the amount of text processed or generated. Companies purchase AI credits or tokens, and advanced models consume far more tokens per request, leading to higher costs. Without careful monitoring, enterprise AI spending can spiral out of control, as seen in the Atlassian example where monthly costs tripled in under a year.

Tags: #AI, #Cost Management, #Enterprise, #Technology News, #Business

PS3 & PS Vita Stores Closing in 2027; Archivists Race to Save Games
PS3 与 PS Vita 商店 2027 年关闭，档案员紧急抢救游戏数据 ⭐️ 8.0/10

Sony announced it will permanently close the PlayStation Store for PS3 and PS Vita in July 2027, prompting digital archivists and the RPCS3 emulator team to urgently back up digital-only games before they become inaccessible. This closure threatens permanent loss of many digital-only titles that lack physical releases, raising critical concerns about digital preservation and the future of gaming history in an increasingly all-digital ecosystem. The RPCS3 team recommends using the no-intro.org database to coordinate preservation efforts, which catalogs cryptographic signatures, file sizes, and serial numbers to track which games are backed up. Once the store shuts down, there will be no legal way to obtain these games.

telegram · zaihuapd · Jul 2, 15:04

Background: RPCS3 is a free, open-source PlayStation 3 emulator that allows PC users to play PS3 games. No-Intro is a community-driven database that catalogs ROMs and digital games, providing metadata (DAT files) for verification and organization. Digital preservation efforts often rely on such databases to ensure completeness and authenticity of archived software.

Tags: #digital preservation, #PS3, #PS Vita, #gaming history, #emulation

Anthropic in Talks with Samsung to Manufacture Custom AI Chip
Anthropic 与三星洽谈制造自研 AI 芯片 ⭐️ 8.0/10

Anthropic has begun developing its own AI chip and is in early-stage discussions with Samsung Electronics about potential manufacturing partnerships. This move aims to gain more control over the computing infrastructure for its Claude model. This development signals increased vertical integration in the AI industry, as major players like Anthropic seek to reduce reliance on external chip suppliers such as NVIDIA. If successful, it could lead to more optimized hardware-software co-design for AI workloads. The project is still in its early stages, and Anthropic is entering the custom chip race later than some competitors. The discussions with Samsung involve using Samsung's foundry services for chip production, similar to how other companies manufacture their designs.

telegram · zaihuapd · Jul 2, 15:57

Background: AI companies traditionally rely on third-party hardware like NVIDIA GPUs for training and inference. To improve efficiency and reduce costs, some are developing custom chips tailored to their specific algorithms. OpenAI and Google have similar initiatives, and Samsung is a major semiconductor foundry that manufactures chips for various clients.

Tags: #AI, #hardware, #chips, #Anthropic, #Samsung

📊 Run stats · Total 11m 18s · AI analysis 3m 55s · Tokens 0.65 MCY (input 0.45 / output 0.20 MCY)