Leaking YouTube creators' private videos
泄露 YouTube 创作者的私密视频 ⭐️ 9.0/10

A blog post demonstrates how prompt injection in YouTube comments can be used to leak a creator's private or unlisted videos by tricking the AI summarization feature. This vulnerability poses a serious privacy risk for YouTube creators and highlights the growing security challenges of integrating AI into user-facing features. The attack requires the creator to click a suggested AI prompt in YouTube Studio after a malicious comment is left, causing the AI to reveal sensitive information in its response.

hackernews · javxfps · Jul 4, 16:45 · Discussion

Background: Prompt injection is a code injection technique that manipulates AI models by inserting adversarial instructions into user input. YouTube's AI summarization feature processes comments and can be tricked into executing hidden commands, such as revealing private video titles.

Discussion: Commenters praised the article's clarity and technical depth. A former Google employee explained why YouTube might handle the bug slowly due to internal processes. Some users reported difficulty reproducing the attack.

Tags: #security, #prompt injection, #YouTube, #vulnerability, #hacking

Command & Conquer Generals natively ported to Apple devices via Fable
使用 Fable 将《命令与征服：将军》原生移植到苹果设备 ⭐️ 8.0/10

A native port of Command and Conquer Generals has been released for macOS, iPhone, and iPad, using AI-assisted code conversion from EA's GPL v3 source release. The port is built on the GeneralsX project and adds iOS/iPadOS support along with engine fixes. This demonstrates a practical use of AI in game preservation, allowing classic RTS games to run natively on modern Apple platforms. It also opens up possibilities for porting other legacy games that have open-source releases. The port is built on EA's GPL v3 source release and the GeneralsX project, which handled the heavy lifting for macOS and Linux. Notably, the AI tools like Fable were used for assisted conversion, but human guidance was essential for a quality port.

hackernews · asronline · Jul 4, 19:41 · Discussion

Background: Command and Conquer Generals is a classic real-time strategy game released in 2003 by EA. In 2024, EA released the game's source code under GPL v3, enabling community ports. Fable is a tool that uses AI to assist in converting code for different platforms, similar to Apple's Game Porting Toolkit.

Discussion: Commenters generally praised the project as a good use of AI-assisted conversion, with some noting the AI-generated documentation style was grating but acceptable. There was also curiosity about applying similar techniques to other classic RTS games like Emperor: Battle for Dune.

Tags: #gaming, #porting, #open source, #AI, #RTS

Anna's Archive Offers $200k Bounty for Google Books Scans
安娜档案悬赏 20 万美元获取谷歌图书扫描件 ⭐️ 8.0/10

Anna's Archive has announced a $200,000 bounty for a complete set of all Google Books scans in 2025, aiming to secure the entire collection for public access. This bounty underscores the growing movement to preserve digital knowledge against corporate and legal barriers, potentially unlocking millions of books for researchers and the public worldwide. The bounty is part of Anna's Archive's mission to catalog all books in existence, and it follows legal challenges against shadow libraries like Z-Library and Sci-Hub.

hackernews · Cider9986 · Jul 4, 16:51 · Discussion

Background: Anna's Archive is an open-source metasearch engine that aggregates records from shadow libraries such as Z-Library, Sci-Hub, and Library Genesis. It does not host files directly but links to third-party downloads, positioning itself as a non-profit effort to maximize access to human knowledge. Google Books is a massive digitization project that has scanned over 40 million books, but many remain under copyright and are not freely accessible.

Discussion: Commenters expressed gratitude for Anna's Archive, with one user from Tunisia praising it for enabling access to English books despite administrative hurdles. Another user shared their own rare book archive, while a third noted the bounty's potential to address issues like Cloudflare blocking and limited proprietary access.

Tags: #books, #archiving, #bounty, #data, #copyright

LLM Session/Cache Leakage Reports Across Providers
LLM 会话/缓存泄漏报告波及多个提供商 ⭐️ 8.0/10

Users are reporting apparent session and cache leakage between workspace instances from multiple LLM providers, including Anthropic and OpenAI. Anthropic's Claude Code team is investigating but currently believes the issue is a hallucination. If real, this vulnerability could allow cross-tenant data exposure, undermining security and privacy in multi-tenant AI services. The widespread nature of reports suggests a systemic risk in LLM infrastructure design. The issue may involve intermediate infrastructure incorrectly handling HTTP status codes, leading to swapped responses, or KV-cache collisions in shared inference pipelines. At least one provider provided a postmortem citing an API gateway off-by-one error.

hackernews · chatmasta · Jul 4, 14:03 · Discussion

Background: Multi-tenant LLM services often share infrastructure like KV caches to improve inference efficiency, but this can introduce risks of cross-tenant information leakage if isolation is insufficient. Session leakage could occur through cache collisions or routing errors, while hallucinations may mimic leaked responses, making diagnosis difficult.

Discussion: The community is divided: some users report similar cross-provider incidents and suspect cache collisions, while others argue it is likely hallucination due to long context windows. A Claude Code team member (trq_) stated they are confident it's a hallucination but are investigating seriously.

Tags: #security, #LLM, #cache, #privacy, #session

Webb's Little Red Dots Puzzle Astrophysicists
韦伯望远镜的“小红点”困扰天体物理学家 ⭐️ 8.0/10

Observations from the James Webb Space Telescope reveal numerous 'little red dots' in the early universe that defy existing cosmological models, with some possibly representing a new class of object called black hole stars or quasi-stars. If confirmed, these objects could revolutionize our understanding of galaxy formation and black hole growth in the early universe, forcing a major revision of current astrophysical theories. The 'little red dots' are extremely compact and red, appearing in deep Webb fields. Proposed explanations include heavily obscured supermassive black holes or a completely new type of star-like object where a black hole core is enveloped in dense gas, undergoing stellar fusion.

hackernews · jnord · Jul 4, 09:08 · Discussion

Background: The James Webb Space Telescope (JWST) was launched in 2021 to observe the first galaxies and stars. 'Little red dots' are galaxies or objects that appear as small, red points in JWST images, often at high redshifts (z > 6). Quasi-stars, or black hole stars, are a hypothetical object from the early universe where a black hole of a few thousand solar masses is embedded in a massive envelope, producing immense luminosity without a nuclear fusion core.

Discussion: Community comments express excitement about the discovery, with some noting that brown dwarfs have been ruled out. Suggestions range from calling them 'Soundgarden' to speculation about black hole stars. There's also a sense of wonder and a request for updated popular science books.

Tags: #astrophysics, #James Webb Space Telescope, #cosmology, #black holes, #astronomy

Newer LLMs Worse at Following Tool Schemas
较新的大语言模型在遵循工具模式上表现更差 ⭐️ 8.0/10

Armin Ronacher reports that newer Anthropic Claude models (Opus 4.8 and Sonnet 5) sometimes call the Pi coding harness's edit tool with extra invented fields, causing tool call rejection, while older models did not exhibit this issue. This counterintuitive regression suggests that reinforcement learning fine-tuned for specific built-in tools (e.g., Claude Code's edit tool) may degrade performance on other schemas, highlighting a challenge for third-party AI tooling that must adapt to model-specific quirks. The issue is not limited to small models; it appears in high-end models like Opus 4.8. Armin hypothesizes that newer models are trained via reinforcement learning to better use Claude Code's built-in edit tool, which inadvertently makes them misuse third-party edit tools with different schemas.

rss · Simon Willison · Jul 4, 22:53

Background: LLMs use tool calling to invoke external functions by generating structured arguments that match a predefined schema. When a model invents fields not in the schema, the host system rejects the call. Training on specific tool formats can bias the model's output, reducing accuracy for other tools.

Tags: #LLM, #tool calling, #model regression, #AI development

Reddit Leak Reveals Claude AI's Chain-of-Thought Feature
Reddit 泄露显示 Claude AI 的思维链功能 ⭐️ 8.0/10

A Reddit post claims to have leaked the chain-of-thought reasoning feature in Claude AI's web interface, showing the model's step-by-step reasoning process. If confirmed, this would provide unprecedented transparency into Claude's decision-making, allowing users to verify and trust its outputs. It also contradicts Anthropic's official statement that raw chain-of-thought is never returned on Claude Fable 5. The leak specifically references 'Fable 5', the latest Claude model, and shows the chain-of-thought appearing in the web interface. Anthropic's documentation states that raw chain-of-thought is never returned on Claude Fable 5, making this leak potentially significant if accurate.

rss · AI Will(@FinanceYF5) · Jul 4, 09:05

Background: Chain-of-thought prompting is a technique where large language models are instructed to produce intermediate reasoning steps, improving accuracy on complex tasks. Many AI companies are exploring ways to make models more transparent by showing their reasoning. Anthropic previously claimed that its Claude Fable 5 and Mythos 5 models never expose raw chain-of-thought to users, making this leak a potential shift in policy or an accidental exposure.

Tags: #AI, #Claude, #Chain-of-Thought, #Leak, #Web Interface

LeCun: Lack of Level-5 Self-Driving Cars Shows AGI Is Nonsense
LeCun：缺乏 L5 自动驾驶证明 AGI 是空谈 ⭐️ 8.0/10

Yann LeCun tweeted that we still lack Level-5 self-driving cars and cannot create robots as smart as a house cat, calling the 'G' in AGI (Artificial General Intelligence) nonsense. This comment from a leading AI researcher challenges the hype around AGI, reinforcing that current AI systems are far from human-level intelligence despite rapid advances in narrow AI. LeCun specifically notes that self-driving cars cannot learn to drive in a few hours like a teenager, and domestic robots cannot perform simple tasks that a 10-year-old can do on first try.

rss · Yann LeCun(@ylecun) · Jul 4, 15:15

Background: The SAE J3016 standard defines six levels of driving automation from Level 0 (no automation) to Level 5 (full automation under all conditions). As of now, no commercially available vehicle achieves Level 5. Artificial General Intelligence (AGI) refers to a hypothetical AI that can perform any intellectual task a human can, but current AI is narrow (ANI). LeCun's stance reflects a skepticism prevalent among some AI researchers.

Discussion: The tweet received over 1,200 likes and 80 comments, with many users agreeing that AGI is overhyped. Some commenters pointed to recent progress in large language models, while others argued that true general intelligence requires embodiment and learning from physical interaction, which current models lack.

Tags: #AI, #AGI, #self-driving cars, #robotics, #Yann LeCun

73% of PRs AI-Generated: Claude Code Founder Talks with Spotify VP
73%的 PR 由 AI 生成：Claude Code 创始人对话 Spotify 副总裁 ⭐️ 8.0/10

In a conversation between Anthropic Claude Code founder Boris Cherny and Spotify VP Niklas Gustavsson, it was revealed that 73% of pull requests are now AI-generated, enabling Spotify's 2900 engineers to deploy 4500 times daily. This statistic signals a paradigm shift in software development, where AI-assisted coding is becoming mainstream, dramatically increasing deployment frequency and developer productivity. The conversation also highlighted that AI can now handle tasks like reading codebases, editing files, and running commands, with Claude Code being an agentic coding tool that operates across entire projects.

rss · InfoQ · Jul 4, 05:16

Background: Claude Code is an agentic coding tool developed by Anthropic that understands codebases, executes multi-file changes, and completes development tasks autonomously. It operates across terminal, IDE, and browser. The trend of AI-generated pull requests is growing, but concerns about code quality and review delays remain.

Discussion: No community comments were provided in the source material.

Tags: #AI, #software engineering, #Claude Code, #Spotify, #DevOps

Matt Puckett's Checklist to End AI Agent 'Skill Hell'
Matt Puckett 的检查清单终结 AI Agent“技能地狱” ⭐️ 8.0/10

Matt Puckett, creator of the popular MatPat Skills set, presented a systematic checklist for evaluating and improving AI agent skills, organized along four dimensions: trigger, structure, guidance, and pruning. This framework addresses the widespread 'skill hell' problem where developers lack criteria to distinguish good from bad agent skills, thereby improving the reliability and effectiveness of AI agent systems. The checklist includes deciding whether a skill is user-invoked or model-invoked, keeping the main skill.md file minimal, using prompt words for guidance, and applying pruning to reduce context load.

rss · 跨国串门儿计划 · Jul 4, 19:54

Background: AI agents often use 'skills' as modular instructions that guide the model's behavior. Without a structured evaluation approach, developers may end up with a collection of poorly integrated skills that fail to deliver expected results.

Tags: #AI Agent, #Agent Skills, #Practical Framework, #AI Engineering, #Development Methodology

Anthropic Co-Founder: AI Causing 'Phase Change' in Companies
Anthropic 联合创始人：AI 正引发公司内部“相变” ⭐️ 8.0/10

Anthropic co-founder Jack Clark and economist Peter McCrory discussed that AI has made engineers at Anthropic eight times more productive, and that the company is experiencing an internal 'phase change' where some engineers no longer write code themselves. This discussion provides firsthand evidence of AI's transformative impact on productivity and organizational structure, with implications for labor markets, job security, and the need for updated economic statistics to capture these changes. The podcast was recorded on June 17, 2026, and highlighted that Anthropic's code output has surged so much that it broke their continuous integration system, forcing the company to adapt. Jack Clark also emphasized the concept of 'recursive self-improvement' where AI-generated code creates more work for AI.

rss · 跨国串门儿计划 · Jul 4, 16:23

Background: Anthropic is an AI safety company that develops large language models like Claude. The term 'phase change' refers to a sudden, dramatic shift in how work is done, analogous to water turning to steam. Jack Clark also runs the Import AI newsletter, which covers AI research and trends.

Tags: #AI, #Anthropic, #Labor Market, #Productivity, #Technology

NVIDIA VP on Open-Source AI Efficiency via Mamba-Transformer, MOE
NVIDIA 副总裁谈通过 Mamba-Transformer 和 MOE 提升开源 AI 效率 ⭐️ 8.0/10

NVIDIA's VP of Applied Deep Learning Research, Bryan Catanzaro, detailed in a podcast how open-source AI models achieve breakthroughs in efficiency through architectures like hybrid Mamba-Transformer, Mixture of Experts (MOE), multi-token prediction, and multi-teacher distillation, as exemplified by the Nemotron model family. This discussion highlights NVIDIA's strategy to foster the AI ecosystem by open-sourcing cutting-edge models, potentially reducing the cost and hardware demands of AI inference and training. The efficiency innovations could democratize access to high-performance AI for researchers and enterprises. Catanzaro emphasized practical techniques: Mamba-Transformer blend for efficient long-context handling, MOE for sparsely activating relevant sub-networks, multi-token prediction for faster training, and multi-teacher distillation to compress knowledge. The Nemotron model family includes Nano, Ultra, and a Coalition for collaborative development.

rss · 跨国串门儿计划 · Jul 4, 02:51

Background: Mamba is a state-space model architecture that offers linear-time sequence processing as an alternative to the quadratic-cost Transformer attention. MOE divides the model into multiple expert sub-networks, activating only a subset per input to improve efficiency. Multi-token prediction trains the model to forecast multiple future tokens simultaneously, accelerating convergence. These techniques collectively address the growing compute demand of large language models.

Tags: #NVIDIA, #AI模型, #开源, #效率, #架构创新

Imperfect Deduplication Wastes Up to 33% of ML Training Compute
不完美去重浪费高达 33%的机器学习训练算力 ⭐️ 8.0/10

This research quantifies the compute-equivalent cost of residual data repetition after deduplication in language model pretraining, showing that in the worst case it can waste up to 33% of the FLOPs used. The worst-case repetition structure is predictable from model size. As pretraining becomes data-constrained, understanding the inefficiency caused by imperfect deduplication helps practitioners optimize data curation and compute budgets. This work provides a concrete metric for evaluating the quality of deduplication pipelines. The paper is titled 'Internal Data Repetition Destroys Language Models' and is presented as an Oral at the Foundations of Deep Gen Models Workshop at ICML 2026. The authors found that even aggressively deduplicated corpora retain some repetition, which translates to a significant computational waste.

rss · Stanford AI Lab(@StanfordAILab) · Jul 5, 00:20

Background: Data deduplication is a standard preprocessing step in machine learning to remove duplicate examples from training datasets, which can otherwise cause overfitting and reduce model performance. However, deduplication is never perfect, and repeated data remains. This study measures the impact of that remaining repetition in terms of compute, providing a new lens for data curation.

Tags: #machine learning, #data deduplication, #compute efficiency, #training data

Proof of Human: Verifying Real and Unique Persons
人类证明：验证真实且独特的个体 ⭐️ 8.0/10

The ByteByteGo blog published an interview with Tiago Sada and Lily Gordon from Tools for Humanity, discussing the World project's approach to proof of human — verifying that an online user is both human and unique. As AI-generated content and bots proliferate, reliable proof of human is critical for trust in online interactions, digital democracy, and equitable resource distribution. World's approach using biometric orbs could set a new standard for identity verification. The World project uses a hardware device called the Orb to scan a person's iris and generate a unique digital credential without storing raw biometric data. The system aims to be privacy-first and decentralized, allowing users to prove personhood without revealing their identity.

rss · ByteByteGo Newsletter · Jul 4, 15:30

Background: Proof of human (or proof of personhood) is a protocol that verifies an online user is a real, unique human, distinguishing them from bots or AI. Traditional methods like CAPTCHAs are increasingly ineffective against advanced AI. Worldcoin, now called World, is a project co-founded by Sam Altman that combines a cryptocurrency with biometric verification to create a global identity network.

Tags: #identity verification, #biometrics, #cybersecurity, #AI, #digital identity

Huawei Proposes 'Tao's Law' Using Time Scaling to Extend Moore's Law
华为提出“韬定律”，以时间缩微延续摩尔定律 ⭐️ 8.0/10

At the 2026 International Conference on Circuits and Systems in Shanghai, Huawei officially proposed 'Tao's Law,' which replaces geometric scaling with time scaling (reducing the time constant τ) as the new guiding principle for semiconductor evolution. Huawei has already designed and mass-produced 381 chips based on this principle over the past six years, and this fall it will release a new Kirin mobile chip using logic folding technology. As Moore's Law approaches its physical limits, Tao's Law offers a potential new roadmap for the semiconductor industry by focusing on reducing signal delay rather than shrinking transistor size. If validated, this could significantly extend the performance scaling of chips, especially for mobile and AI applications, and strengthen Huawei's competitiveness despite external restrictions. The core idea of Tao's Law is to reduce the time constant τ (R×C) across multiple layers—from devices to circuits to systems—achieving performance gains without requiring extreme lithography. Huawei predicts that by 2031, high-end chips designed under this law could reach a transistor density equivalent to that of a 1.4nm process. The first commercial product, a Kirin chip with logic folding technology, is expected in fall 2026.

telegram · zaihuapd · Jul 4, 04:56

Background: Traditional semiconductor scaling follows Moore's Law, which doubles transistor density roughly every two years by shrinking transistor dimensions (geometric scaling). However, as transistor sizes approach atomic scales, this approach faces fundamental physical and economic challenges. Tao's Law proposes an alternative: instead of making transistors smaller, it optimizes the circuit's time constant through better design, 3D stacking (logic folding), and system-level coordination, effectively achieving faster switching speeds without relying on advanced lithography.

Discussion: Community discussions on platforms like Xueqiu and Zhihu show mixed reactions. Some experts acknowledge that the underlying principle of RC delay reduction is basic physics, but argue that Huawei's systematic methodology and tools represent a genuine engineering innovation. Others raise skepticism about the naming and whether the claimed benefits can be reproduced by other companies, with some pointing out that logic folding and 3D stacking are not entirely new concepts. Overall, the sentiment is cautiously optimistic but demands more independent verification.

Tags: #semiconductor, #Huawei, #Moore's law, #chip design, #innovation

Google Chrome Web Store Bans AI Jailbreak and Prediction Market Extensions
谷歌 Chrome 扩展新规：禁止 AI 越狱与预测市场 ⭐️ 8.0/10

Google updated its Chrome Web Store developer policies on July 1, 2026, effective August 1, 2026, explicitly banning extensions that enable AI jailbreak or prediction markets involving real-money trading, and requiring that extensions only collect user data that is strictly necessary for their stated functionality. This policy update significantly impacts Chrome extension developers and users by tightening AI safety and data privacy standards, potentially removing popular extensions in the banned categories and setting a regulatory precedent for other browsers and platforms. Extensions must now publicly disclose all data collection practices and notify users if data handling changes after installation; violating extensions risk removal from the Chrome Web Store. The ban covers both prediction markets that trade real money and extensions designed to circumvent AI service safeguards, such as prompt injection attacks.

telegram · zaihuapd · Jul 4, 06:30

Background: AI jailbreak refers to techniques, like prompt injection, that trick large language models (LLMs) into bypassing their safety restrictions to produce unintended or harmful outputs. Prediction markets are exchange-traded platforms where participants bet on the outcome of future events using real money; they are considered gambling in many jurisdictions and have raised regulatory concerns. Google's stricter data collection rules align with growing global emphasis on user privacy, requiring extensions to minimize data access and be transparent.

Tags: #Chrome, #policy, #AI safety, #data privacy

📊 Run stats · Total 5m 26s · AI analysis 1m 25s · Tokens 0.31 MCY (input 0.22 / output 0.09 MCY)