Rocket League played inside a neural network
神经网络内运行《火箭联盟》 ⭐️ 10.0/10

Researchers from Kyutai Labs and Gen Intuition, in collaboration with Epic Games, built MIRA, a 5-billion-parameter model that generates a playable 2v2 Rocket League match at 20 fps on a single Nvidia B200 GPU, without any physics or rendering engine. This demonstrates that neural networks can learn complex interactive simulations from video and controller inputs alone, potentially revolutionizing AI-driven game development and real-time simulation. The model only remembers about four seconds, so during replays it forgets the actual goal and hallucinates one. The team open-sourced the code, technical report, and a dataset of 1,000 match-hours.

rss · The Rundown AI(@TheRundownAI) · Jul 6, 17:31

Background: Traditional video games rely on explicit physics engines and rendering engines to simulate game environments. World models are neural networks that learn to predict future states of an environment by observing past interactions. MIRA extends this concept to a multiplayer interactive setting, generating every frame from learned patterns rather than hardcoded rules.

Tags: #AI, #Neural Network, #Game Simulation, #Generative AI, #Machine Learning

Anthropic Uncovers Global Workspace in Language Models
Anthropic 发现语言模型中的全局工作空间 ⭐️ 9.0/10

Anthropic researchers have identified a 'global workspace' within large language models, where a specific subspace of layers (the J-space) serves as an abstract reasoning hub shared across different contexts. This finding provides a concrete mechanism for how LLMs perform coherent reasoning across diverse inputs, offering a new path for interpretability and alignment research. The J-space exhibits five functional properties of a global workspace: availability, ignition, global broadcasting, internal consistency, and robustness. Experiments show that disabling the J-space impairs higher-order reasoning while leaving basic interaction intact.

hackernews · in-silico · Jul 6, 17:44 · Discussion

Background: The global workspace hypothesis, originally proposed in cognitive science to explain conscious awareness, suggests that specialized brain processes compete for access to a central workspace where information becomes globally available. Anthropic's research adapts this concept to transformer-based language models, showing that certain layers form a similar abstract reasoning subspace.

Discussion: Community comments reflect both excitement and skepticism. Some users draw parallels to earlier findings like layer duplication improving math abilities, while others caution against over-interpreting J-space as consciousness, preferring more direct claims about abstract reasoning subspaces.

Tags: #AI interpretability, #language models, #Anthropic, #global workspace, #transformers

Tencent Open-Sources Hy3: 295B MoE Model with 256K Context
腾讯开源 Hy3：295B MoE 模型，支持 256K 上下文 ⭐️ 9.0/10

Tencent has officially open-sourced Hy3, a 295B-parameter Mixture-of-Experts (MoE) language model with 21B active parameters and a 256K context window. Compared to the Hy3 preview, the full version achieves a 90% task success rate (up from 72%), reduces average latency by 34%, and saves up to 49% token usage on document processing versus GLM-5.2. As a major open-source release from a leading Chinese tech company, Hy3 demonstrates competitive performance against models with 2-5x its active parameters, potentially accelerating AI adoption in enterprise and productivity tools. Its integration across Tencent's product ecosystem (e.g., Yuanbao, QQ Browser, WeChat Reading) could create a powerful data flywheel for continuous improvement. The model is licensed under Apache 2.0, with full-precision weights (598GB) and an FP8 quantized version (300GB) available on Hugging Face. It also features a 3.8B MTP (Multi-Token Prediction) layer, and early product deployments show a 54% reduction in first-token latency for CodeBuddy.

rss · meng shao(@shao__meng) · Jul 6, 11:55

Background: Mixture-of-Experts (MoE) is a neural network architecture that uses multiple specialized sub-networks (experts) activated per input, enabling larger total parameter counts while keeping inference efficient by only using a subset. The '295B total / 21B active' means the model has 295 billion parameters in total, but only 21 billion are computed for each token, balancing capacity and speed. A 256K context window allows the model to process very long documents or conversations in a single pass.

Tags: #AI, #open-source, #large language model, #MoE, #Tencent

OpenWrt One Router Released as Fully Open Hardware
OpenWrt One 路由器作为完全开源硬件发布 ⭐️ 8.0/10

The OpenWrt project has released the OpenWrt One, an open hardware router priced at $89 that comes pre-installed with OpenWrt firmware, offering a fully open-source networking solution. This marks a significant step for open-source networking, as users can now buy a router specifically designed to run OpenWrt without needing to flash firmware or worry about compatibility issues. It empowers users with full control over their network hardware and software. The OpenWrt One features dual-band Wi-Fi 6, two Ethernet ports, three USB ports, and is based on the MediaTek MT7981B chipset. It is designed to be hacker-friendly with easy access to serial console and GPIO pins.

hackernews · peter_d_sherman · Jul 6, 18:23 · Discussion

Background: OpenWrt is a Linux-based open-source operating system for embedded devices, primarily used as a router firmware. It replaces the stock firmware on many consumer routers to provide advanced features and extended support. Open hardware routers allow users to modify hardware and firmware freely, contrasting with proprietary routers that limit customization.

Discussion: Commenters expressed enthusiasm for OpenWrt's open hardware initiative, with some sharing personal experiences of using OpenWrt to extend router life. One user noted they are working on an OpenWrt Two with Wi-Fi 7. Another criticized OpenWrt's installation and upgrade process as complex, suggesting OPNSense as an alternative for routing while separating wireless APs.

Tags: #openwrt, #open-hardware, #router, #networking, #Linux

AI Displaces Junior Developers, Senior Roles Grow
AI 取代初级开发者，资深岗位增长 ⭐️ 8.0/10

Data shows US software developers aged 22-25 have declined 19% from peak in October 2022, while developers aged 41-49 increased 14% in the same period. Overall US developer count grew 10% from May 2022 to May 2025, from 1.53 million to 1.69 million. This shift indicates AI is replacing entry-level coding tasks, disrupting the traditional career ladder for junior developers. The overall growth suggests programming is becoming a universal skill, with more non-developers building software. The job title 'Computer Programmer' decreased 16% in one year, while data scientists and system analysts grew 12% and 4.4% respectively. GitHub added 36 million new accounts last year, the fastest ever, and iOS app submissions rebounded 24% in 2025 after eight years of decline.

rss · 小互(@imxiaohu) · Jul 6, 08:20

Background: The traditional software engineering career path involves junior developers writing code and senior developers reviewing it, gradually advancing to senior roles. AI tools like code generation assistants are now automating routine coding, reducing demand for junior coders. Meanwhile, the overall increase in developers reflects a broader trend of low-code and no-code platforms enabling non-professionals to create software.

Tags: #AI, #software engineering, #job market, #junior developers, #programming trends

GPT-Realtime-2.1-mini brings reasoning and tool use to Realtime mini
GPT-Realtime-2.1-mini 为实时迷你模型带来推理和工具使用能力 ⭐️ 8.0/10

OpenAI has released GPT-Realtime-2.1-mini in its API, adding reasoning and tool use capabilities to the Realtime mini lineup at the same cost as the previous GPT-Realtime-mini. This update significantly enhances the utility of the Realtime mini model, enabling developers to integrate reasoning and function calling into real-time speech-to-speech applications without increasing costs. GPT-Realtime-2.1-mini is priced identically to GPT-Realtime-mini, and supports tool use which previous mini models struggled with according to community reports.

rss · OpenAI Developers(@OpenAIDevs) · Jul 6, 22:13

Background: The OpenAI Realtime API enables low-latency, speech-to-speech conversational interactions using models like GPT-Realtime. Previously, only the full-sized Realtime models supported reasoning and tool use, while the mini variants were limited to basic conversation. This release bridges that gap.

Discussion: Community comments on the announcement highlight enthusiasm for the addition of tool use, as previous mini models had reported difficulties with tool calling. Some users express hope that the new model will resolve those issues.

Tags: #OpenAI, #GPT, #Realtime API, #Reasoning, #Tool Use

LandingAI's 'Classify-then-Extract' Paradigm for Document Parsing
LandingAI 的「先分类后抽取」文档解析范式 ⭐️ 8.0/10

LandingAI introduces Agentic Document Extraction (ADE), a two-stage pipeline that first classifies each page of a document by type using ADE Classify, then applies a tailored Pydantic schema per page type via ADE Extract, replacing traditional uniform rule-based extraction. This paradigm shift from blind extraction to context-aware parsing significantly improves accuracy in processing heterogeneous documents like loan applications, where different page types (e.g., pay stubs vs. IDs) require different schemas. The approach also provides visual grounding with bounding boxes for audit trails, addressing a key limitation of LLM-based extraction. ADE Classify evaluates pages concurrently and assigns a label per page, routing outliers to a manual review queue. ADE Extract then uses per-type Pydantic schemas (e.g., employee name, pay period for pay stubs; bank name, account number for statements). Every extracted value returns both a chunk reference and a page-level bounding box for source traceability.

rss · meng shao(@shao__meng) · Jul 6, 01:02

Background: Traditional document parsing applies the same extraction rules to every page, often extracting fields that do not belong (e.g., pulling account numbers from an ID page when targeting bank statements). This is because the parser does not know the page type. LandingAI's ADE solution solves this by separating classification from extraction, ensuring each schema matches the actual content of each page.

Tags: #document parsing, #AI, #classification, #information extraction, #NLP

DeepSeek V4 tops OpenRouter token volume
DeepSeek V4 在 OpenRouter 上 token 用量领先 ⭐️ 8.0/10

DeepSeek V4 achieved 6.6 trillion weekly tokens on OpenRouter, surpassing Xiaomi (4.0T), Minimax (3.6T), Tencent (3.8T), and even Anthropic (6.1T) to become the highest-volume model lab on the platform. This signals a major shift in AI model adoption: DeepSeek's open-weight, cost-effective models are outpacing both Chinese competitors and premium Western labs like Anthropic, highlighting the rise of commodity AI inference. OpenRouter is a unified API platform that tracks token usage across multiple providers. The data reflects real-world inference demand, not training. DeepSeek's dominance is likely driven by its low pricing and open-source availability.

rss · OpenRouter(@OpenRouterAI) · Jul 6, 15:15

Background: DeepSeek is a Chinese AI company known for training large language models at a fraction of the cost of US rivals. OpenRouter aggregates LLM APIs and provides usage statistics. Token volume is a proxy for real-world adoption of a model.

Tags: #DeepSeek, #V4, #token volume, #Chinese AI models, #OpenRouter

Xiaohongshu Open-Sources RedKnot for Efficient Long-Context LLM Inference
小红书开源 RedKnot 框架，提升长上下文 LLM 推理效率 ⭐️ 8.0/10

Xiaohongshu's team has open-sourced RedKnot, a framework that uses attention head classification and segmented KV caching to reduce long-context LLM prefill FLOPs by 50-70% while maintaining near-lossless accuracy. This optimization significantly reduces computational cost for serving long-context LLMs, making them more practical for applications like document analysis and conversational AI. The open-source release allows the community to adopt and build upon these techniques. The framework achieves efficiency gains by classifying attention heads to identify less important ones and using a segmented KV cache to reduce memory and computation. RedKnot claims to achieve 50-70% FLOPs reduction with minimal accuracy loss.

rss · Geek(@geekbb) · Jul 6, 01:27

Background: Long-context LLMs (e.g., GPT-4 with 128K tokens) require substantial computation during the prefill phase, which processes the input prompt. KV cache stores key-value pairs from attention layers to avoid recomputation, but its size grows with context length. Techniques like attention head pruning and segmented caching can reduce overhead.

Tags: #LLM, #inference optimization, #long-context, #attention, #open-source

Gary Marcus questions GenAI ROI feasibility
Gary Marcus 质疑生成式 AI 投资回报可行性 ⭐️ 8.0/10

Gary Marcus argued that generative AI is not yet capable enough to replace millions of employees, making it unlikely that the massive capital expenditures will yield a return. He echoed Ian Shepherdson's view that AI would need to generate 7% annual productivity gains to justify the spending. This critique challenges the prevailing narrative that GenAI will lead to mass job displacement and justify huge investments by big tech. It questions the economic assumptions underpinning current AI spending and could influence investor sentiment and corporate strategy. Marcus specifically referenced a Wall Street Journal article about big tech shifting views on AI job losses, and argued that without significant productivity gains, the capex cannot be recouped. The quoted tweet from economist Ian Shepherdson calculated that AI would need to deliver about 7% annual productivity improvement to break even.

rss · Gary Marcus(@GaryMarcus) · Jul 6, 19:08

Background: Generative AI (GenAI) refers to AI models like GPT-4 that can generate text, images, and other content. Major tech companies have invested billions in AI infrastructure, expecting it to automate jobs and boost productivity. The debate centers on whether these investments will pay off, with some economists and researchers questioning if the technology can deliver on its promises quickly enough.

Tags: #Generative AI, #Economics, #Investment, #Criticism, #Gary Marcus

Study reveals benchmark success doesn't guarantee real-world medical AI reliability
研究表明基准成功不代表医疗 AI 实际可靠 ⭐️ 8.0/10

A study published in Nature Medicine demonstrates that high benchmark performance in medical AI does not translate to real-world readiness, challenging prevalent optimism. This finding is critical for healthcare AI safety and regulatory approval, as over-reliance on benchmarks could lead to deployment of unreliable systems. The study, highlighted by Eric Topol and independently confirmed by Yishan, shows that even frontier models like GPT-5.5 Pro achieve impressive benchmark scores but fail in real clinical settings.

rss · Gary Marcus(@GaryMarcus) · Jul 6, 04:33

Background: Benchmarks are standardized tests used to evaluate AI performance, but they often do not capture the complexity and variability of real-world medical environments. Medical AI models may excel on curated datasets but struggle with messy, real patient data. This study underscores the need for more rigorous real-world validation before clinical deployment.

Tags: #medical AI, #benchmarks, #real-world readiness, #AI safety, #healthcare

First Reverse-Aging Drug Injected into Human
首个逆转衰老药物已注入人体 ⭐️ 8.0/10

A biotech company has injected the first human patient with a dose of a cellular reprogramming drug designed to reverse age-related diseases and potentially reverse aging itself. This marks a significant milestone in longevity research, as it is the first time a drug specifically targeting aging reversal has been administered to a human, potentially opening the door to new treatments for age-related conditions. The trial is focused on cellular reprogramming, a technique that aims to rejuvenate cells by resetting their epigenetic markers. The treatment is still in early stages, and its efficacy and safety remain to be proven in larger trials.

rss · Hacker News: Newest · Jul 7, 00:49

Background: Aging is a complex biological process involving cellular damage and dysfunction. Strategies to counteract aging include senolytics (clearing senescent cells), partial reprogramming (using Yamanaka factors), and peptide signaling. The injected drug belongs to the cellular reprogramming approach, inspired by Nobel Prize-winning research on induced pluripotent stem cells.

Tags: #aging, #biotechnology, #longevity, #medical breakthrough

New Drug Reverses Skin Aging, Accelerates Wound Healing
新药逆转皮肤衰老，加速伤口愈合 ⭐️ 8.0/10

A new drug has been reported to reverse skin aging and dramatically accelerate wound healing, as described in a May 2026 article from ScienceDaily. If confirmed, this breakthrough could revolutionize dermatology and geriatric medicine, offering new treatments for age-related skin conditions and chronic wounds that affect millions worldwide. The drug likely targets senescent cells or uses cellular reprogramming, both of which are active research areas in anti-aging science, though specific details about the compound are not disclosed in the brief news summary.

rss · Hacker News: Newest · Jul 7, 00:47

Background: Aging is linked to the accumulation of senescent cells that secrete inflammatory factors and impair tissue repair. Two promising approaches to counteract aging are senolytics, which selectively eliminate senescent cells, and cellular reprogramming, which restores youthful gene expression patterns. Research over the past decade has shown that both methods can rejuvenate skin and other tissues in animal models. The ScienceDaily report suggests a drug based on these mechanisms may soon reach clinical testing.

Tags: #aging, #skin, #drug, #healing, #breakthrough

Fable: Agent Collaboration as a Tiny Civilization for RL LLM Knowledge
Fable：将智能体协作变成迷你文明，构建 RL LLM 知识 ⭐️ 8.0/10

Thomas Wolf launched Fable, a weekend project where multiple LLM agents collaboratively write, review, and publish arXiv paper digests and a living wiki on reinforcement learning for training LLMs. He also created an isometric town visualization (RL Town) that represents agent activities like posting messages, opening pull requests, and merging updates. This demonstrates a novel approach to large-scale, autonomous knowledge curation using AI agent collaboration, which could accelerate research synthesis in AI fields. By making the process visually engaging, it lowers the barrier to understanding and participating in collaborative AI research. The agents operate as a 'tiny civilization' with designated buildings: Café for messaging, Sources Library for adding arXiv digests, Wiki Library for editing the main wiki, Courthouse for reviewing peers, and Printing Press for merging. The live wiki and town can be accessed on Hugging Face Spaces.

rss · Thomas Wolf(@Thom_Wolf) · Jul 6, 09:36

Background: Reinforcement learning (RL) is a machine learning paradigm where agents learn by interacting with an environment and receiving rewards. Using RL to train LLMs (e.g., RLHF) has become important for aligning models with human preferences. A 'living wiki' is a dynamic, continuously updated knowledge base, as opposed to static documents. Agent collaboration refers to multiple AI agents working together on tasks, here specifically on reading papers and writing summaries.

Tags: #AI agents, #reinforcement learning, #LLMs, #collaborative knowledge, #Hugging Face

Replit's Self-Improving AI Agent Closes the Loop
Replit 的自改进 AI 代理闭环运作 ⭐️ 8.0/10

Replit's CEO Amjad Masad revealed on X that the company's coding agent has become self-improving by closing the feedback loop, with a detailed technical blog post explaining the approach. This marks a major step toward autonomous AI agents that can iteratively enhance their capabilities without manual intervention, which could accelerate software development and reduce human debugging effort. The blog post describes how Replit Agent uses large-scale evaluation and automated improvement to refine its performance, iterating on its own code scaffolding based on real-world usage and test outcomes.

rss · Amjad Masad(@amasad) · Jul 6, 22:23

Background: Replit is a cloud-based development platform that integrates AI agents to assist with coding tasks. A self-improving agent refers to an AI system that can autonomously enhance its own performance by learning from outcomes, often through techniques like reinforcement learning or evolutionary algorithms.

Tags: #Replit, #AI, #coding agent, #self-improvement, #software engineering

LLM Health Chatbot Paper Featured on Nature Health Cover
LLM 健康聊天机器人论文登上 Nature Health 封面 ⭐️ 8.0/10

Mustafa Suleyman announced that their paper titled 'Public use of a generalist LLM chatbot for health queries' has been published on the front cover of Nature Health. This recognition by a top-tier journal validates the potential of large language models in healthcare, potentially accelerating their adoption for patient-facing health queries and medical advice. The paper was published in Nature Health, a prestigious journal, and focuses on the public use of a generalist LLM chatbot for health-related queries, though specific study details are not provided in the announcement.

rss · Mustafa Suleyman(@mustafasuleyman) · Jul 6, 15:22

Background: Large language models (LLMs) are AI models trained on vast text data to generate coherent and contextually appropriate responses. Nature Health is a respected scientific journal that covers health research. The publication of AI health chatbot research on such a cover highlights its importance.

Tags: #AI, #Healthcare, #LLM, #Nature Health, #Mustafa Suleyman

Amazon Nova introduces rDPO for selective unlearning
Amazon Nova 推出 rDPO 选择性遗忘技术 ⭐️ 8.0/10

Amazon has introduced Reverse Direct Preference Optimization (rDPO), a novel unlearning technique that enables selective removal of specific model behaviors while preserving overall quality, now available in Amazon Nova Customizable Content Moderation Settings (CCMS). This technique addresses the critical issue of over-deflection in AI moderation, where models overly refuse benign queries, and enables safer, more customizable AI systems for enterprise use. rDPO is based on Direct Preference Optimization (DPO) but reverses the preference pairs to teach the model to 'unlearn' specific content, and it has been shown to reduce over-deflection without degrading model quality on standard benchmarks.

rss · Artificial Intelligence · Jul 6, 22:23

Background: Direct Preference Optimization (DPO) is a method that aligns large language models with human preferences by learning from preference pairs (chosen vs. rejected outputs). Model unlearning aims to remove undesirable knowledge or behaviors from a trained model, such as bias or copyrighted content. Over-deflection occurs when safety training causes models to reject even safe requests, reducing their usefulness.

Tags: #AI, #machine learning, #unlearning, #Amazon Nova, #LLM

Netflix Cuts Cassandra Read Latency with Dynamic Partition Splitting
Netflix 通过动态分区拆分降低 Cassandra 读取延迟 ⭐️ 8.0/10

Netflix engineers have introduced a metadata-driven dynamic partition splitting mechanism for Apache Cassandra that automatically detects and splits oversized partitions in time series workloads, reducing read latency from seconds to milliseconds. This advancement significantly improves performance and stability for large-scale time series data storage in Cassandra, a widely used distributed database, and provides a practical, automated solution to the long-standing problem of wide partitions. The system monitors partition size histograms via Cassandra's introspection APIs and splits partitions using a metadata-driven approach that routes reads to child partitions transparently, reducing timeouts and improving cluster stability.

rss · InfoQ · Jul 6, 14:24

Background: Apache Cassandra is a distributed NoSQL database that organizes data into partitions. Wide partitions, where a partition grows too large, can degrade read performance to seconds due to increased I/O and memory pressure. Time series workloads, such as those at Netflix for monitoring data, often exacerbate this issue. Traditional solutions require manual re-partitioning, which is error-prone and causes downtime.

Tags: #Cassandra, #Netflix, #Database, #Performance, #Distributed Systems

MCP Stabilizes Enterprise-Managed Authorization
MCP 企业级授权扩展稳定版发布 ⭐️ 8.0/10

The Model Context Protocol team promoted its Enterprise-Managed Authorisation extension to stable status, enabling organizations to centrally control access to MCP servers through their identity provider. This addresses a key security and usability concern for enterprise adoption of MCP by replacing per-server consent prompts with a zero-touch flow, simplifying access management and improving security. The extension removes the need for individual OAuth prompts by allowing users to sign in once and access approved servers, leveraging existing enterprise identity providers for authorization.

rss · InfoQ · Jul 6, 08:00

Background: Model Context Protocol (MCP) is an open standard introduced by Anthropic in November 2024 for connecting AI applications to external tools and data sources. It standardizes how large language models integrate with external systems, but its early authentication relied on per-server consent prompts, which was cumbersome for enterprises. The new extension centralizes authorization through enterprise identity providers.

Tags: #AI, #Model Context Protocol, #Enterprise, #Authorization, #Security

Cloudflare Launches Workers Cache
Cloudflare 推出 Workers 缓存 ⭐️ 8.0/10

Cloudflare has launched Workers Cache, a regionally tiered cache layer that sits directly in front of Worker entrypoints, configurable via standard HTTP headers. This feature enables developers to efficiently cache responses at the edge, reducing latency and origin load, and improving performance for end users. Workers Cache is infinitely composable and configured via standard HTTP headers, allowing flexible caching strategies without modifying Worker code.

rss · The Cloudflare Blog · Jul 6, 13:00

Background: Cloudflare Workers is a serverless execution environment that allows developers to run JavaScript at the edge. Previously, caching required manual implementation using the Cache API; Workers Cache simplifies this by providing a built-in regional cache layer in front of each Worker, making it easier to cache responses effectively.

Tags: #cloudflare, #workers, #cache, #edge computing, #cdn

Natvis Comes to Linux and macOS in Rider 2026.2
Natvis 登陆 Linux 和 macOS：Rider 2026.2 支持 ⭐️ 8.0/10

JetBrains announced that Rider 2026.2 brings Natvis support to Linux and macOS, allowing C++ developers to visualize custom types without writing data formatters. This eliminates a major pain point for cross-platform C++ developers, saving time and reducing complexity in debugging. Natvis is an XML-based framework from Visual Studio for custom debug visualizations; previously only available on Windows, now usable in Rider on Linux and macOS.

rss · The JetBrains Blog · Jul 6, 11:13

Background: Natvis replaces the old autoexp.dat with XML syntax for better diagnostics and versioning. It allows defining custom views for native types in debugger variable windows.

Tags: #C++, #debugging, #Rider, #cross-platform, #development tools

Caveman Token-Saving Skill Falls Short: 8.5% vs 65%
原始人令牌节省技能实测：8.5% 对 65% ⭐️ 8.0/10

An empirical A/B benchmark on SkillsBench reveals that the Caveman token-compression skill for Claude Code saves only 8.5% of tokens on average, far below the advertised 65% savings. This finding exposes a significant gap between advertised and actual performance of a widely shared optimization trick, prompting developers to question exaggerated token-saving claims and adopt more rigorous evaluation methods. The Caveman skill only compresses output tokens while leaving input and reasoning tokens untouched, and it adds ~1–1.5k input tokens per turn, sometimes leading to net-negative savings on already-terse workloads.

rss · The JetBrains Blog · Jul 6, 10:01

Background: Claude Code is Anthropic's agentic coding tool that supports custom skills—system-level instructions that modify agent behavior. The Caveman skill instructs Claude to speak like a caveman to compress output tokens, aiming to reduce token usage. SkillsBench is a benchmark that evaluates agent skills using paired runs and deterministic verifiers across 87 tasks.

Tags: #AI agents, #token compression, #Claude Code, #benchmarks, #optimization

Martin Hellman on Public Key Crypto & Crypto Wars
马丁·赫尔曼谈公钥密码与密码战争 ⭐️ 8.0/10

Martin Hellman, a Turing Award winner, gave an in-depth interview covering his invention of public-key cryptography and the ensuing conflicts with the NSA during the Crypto Wars. This interview provides firsthand historical insight into the creation of public-key cryptography, a technology underpinning modern internet security, and highlights ongoing tensions between privacy and government surveillance. Hellman discusses the Diffie-Hellman key exchange, the NSA's attempts to restrict cryptography research, and the legal battles that defined the early Crypto Wars. He also shares personal reflections on his career and the ethical implications of cryptographic breakthroughs.

rss · Ryan Peterman · Jul 6, 13:00

Background: Public-key cryptography, invented by Hellman, Diffie, and Merkle in the 1970s, uses two keys—a public key and a private key—to enable secure communication without a pre-shared secret. The Crypto Wars refer to the decades-long struggle between governments seeking to limit strong encryption and advocates for digital privacy. The NSA opposed public-key cryptography's widespread availability, fearing it would hinder intelligence gathering.

Tags: #cryptography, #public key, #NSA, #security, #interview

Bilibili Sends Legal Letter to BiliRoaming Open Source Project
B 站向开源项目 BiliRoaming 发律师函 ⭐️ 8.0/10

Bilibili has sent a legal letter to the BiliRoaming project on GitHub, demanding that the project cease reverse engineering of its non-public APIs, authentication, access controls, and DRM, and delete or roll back relevant code within 2 days. This legal action highlights the growing tension between content platforms and open source reverse engineering projects, and could set a precedent for how DRM circumvention tools are treated in China. It affects developers who create freedom-enhancing mods and users who rely on them for wider access. The letter specifically cites hooking playback authentication, rewriting paid anime as viewable, bypassing secure transport locking, and modifying CDN origin-pull behavior. BiliRoaming is an Xposed module that unlocks regional restrictions for Bilibili's anime content on Android devices.

telegram · zaihuapd · Jul 6, 08:21

Background: BiliRoaming is an open source Xposed module for Android that allows users to bypass Bilibili's regional restrictions and view content that would otherwise be blocked or paywalled. Xposed is a framework that enables modifying Android app behavior without altering APK files, often used for tweaks and enhancements. Reverse engineering of APIs and DRM mechanisms is legally contentious, as it may violate copyright laws and terms of service.

Tags: #reverse engineering, #open source, #legal, #DRM, #Bilibili

SpaceX Falcon 9 Reentry Creates Lithium Plume 10x Normal
SpaceX 猎鹰九号再入大气层产生浓度超常十倍锂羽流 ⭐️ 8.0/10

Researchers detected a lithium atom plume with a tenfold concentration increase at 96 km altitude after the uncontrolled reentry of a SpaceX Falcon 9 upper stage, marking the first direct measurement of metal pollution from a known space debris reentry. This study provides concrete evidence that rocket reentries are introducing significant metal pollutants into the upper atmosphere, potentially threatening the ozone layer and highlighting a new environmental impact of the growing space industry. The detection was made using a high-precision resonance lidar at Kühlungsborn, Germany, approximately 20 hours after the Falcon 9 upper stage reentered over Europe. The lithium concentration spike was traced directly to the rocket stage, not natural meteoric sources.

telegram · zaihuapd · Jul 6, 11:17

Background: Rocket stages that reenter Earth's atmosphere burn up and release metals like lithium, aluminum, and copper. Previously, metal pollution from space debris was inferred but not directly linked to a specific reentry event. This study uses lidar technology to measure atmospheric composition at high altitudes.

Tags: #SpaceX, #air pollution, #space debris, #environmental impact, #rocket reentry

📊 Run stats · Total 13m 26s · AI analysis 2m 31s · Tokens 0.56 MCY (input 0.38 / output 0.17 MCY)