OrcaSlicer fork restores BambuNetwork support after backlash
OrcaSlicer 分支恢复对 BambuNetwork 的支持 ⭐️ 9.0/10

A community fork of OrcaSlicer (FULU-Foundation/OrcaSlicer-bambulab) has been created to restore full BambuNetwork support for Bambu Lab printers, reversing a firmware update that required cloud-only authentication. This fork empowers users to maintain local control of their printers, resisting vendor lock-in and preserving open-source principles. It reflects growing community backlash against restrictive firmware practices in the 3D printing industry. The fork is based on a prior state of OrcaSlicer that allowed direct LAN printing without cloud authentication. The original firmware update by Bambu Lab required cloud authentication even for local mode, which this fork bypasses.

hackernews · Murfalo · May 12, 21:55 · Discussion

Background: OrcaSlicer is a popular open-source 3D printing slicer. Bambu Lab, a printer manufacturer, released a firmware update that mandated cloud authentication for printing, removing LAN-only functionality. This sparked backlash and led to the creation of this fork to restore previous capabilities.

Discussion: Commenters expressed anger at Bambu Lab for removing functionality, comparing it to theft. Some noted that Bambu Lab initially required cloud auth for LAN mode but backtracked after backlash. Others praised the fork as a necessary response to vendor lock-in.

Tags: #3D printing, #open source, #firmware, #community backlash, #vendor lock-in

CERT Releases Six CVEs for Critical dnsmasq Vulnerabilities
CERT 发布 dnsmasq 六项严重漏洞 CVE ⭐️ 9.0/10

The CERT has released six CVEs for serious security vulnerabilities in dnsmasq, a widely used DNS and DHCP server, affecting millions of devices. The vulnerabilities include remote heap buffer overflows and denial-of-service attacks. Because dnsmasq is embedded in countless routers, IoT devices, and Linux systems, these vulnerabilities pose a widespread risk to network infrastructure and user privacy. The disclosure underscores the urgency of migrating from memory-unsafe languages like C to memory-safe alternatives such as Rust or Go. The vulnerabilities allow remote attackers to execute arbitrary code or cause denial of service through malformed DNS or DHCP packets. One CVE describes an infinite loop causing dnsmasq to stop responding, while another involves a large out-of-bounds write on the heap.

hackernews · chizhik-pyzhik · May 12, 18:12 · Discussion

Background: dnsmasq is a lightweight DNS forwarder and DHCP server commonly used in small networks, embedded devices, and Linux distributions. It provides DNS caching, DHCP, TFTP, and PXE boot services. CVEs (Common Vulnerabilities and Exposures) are identifiers for publicly disclosed security flaws, and CERT coordinates vulnerability disclosures.

Discussion: Community comments reveal frustration over the recurring vulnerabilities, with one user calling this the 'breaking point' for adopting memory-safe languages. Another user sarcastically notes that dnsmasq is used in millions of devices that almost never receive updates, highlighting the update challenge.

Tags: #security, #vulnerabilities, #dnsmasq, #CVE, #memory safety

Google Announces Gemini Intelligence for Android
Google 发布 Android Gemini Intelligence ⭐️ 9.0/10

At the Android Show 2026, Google unveiled Gemini Intelligence, a suite of proactive AI features for Android, including cross-app automation, Rambler speech refinement, generative widgets, and more, set to roll out on Samsung and Pixel devices this summer. This marks a fundamental shift from Android as an app platform to a proactive AI system, meaning the device can anticipate and execute tasks across apps without user initiation. It sets a new standard for mobile AI integration and could reshape how users interact with smartphones. Key features include multi-step cross-app automation (triggered by long press of power button), Chrome-integrated Gemini assistant, enhanced Autofill with contextual understanding, Rambler for converting spoken thoughts into polished text, and Create My Widget for generating custom widgets via natural language. Privacy controls include user confirmation for purchases and opt-in data connections.

rss · meng shao(@shao__meng) · May 13, 00:32

Background: Historically, Android served as an operating system to host apps, with AI features mostly limited to Google Assistant. With Gemini Intelligence, Google is embedding a large language model (Gemini) directly into the OS to enable proactive, screen-aware, and cross-app actions. This builds on the trend of agentic AI, where AI models take actions on behalf of users across multiple applications.

Tags: #AI, #Android, #Gemini, #Automation, #Google

Google Announces Googlebook: Android Laptop with Gemini AI
谷歌发布 Googlebook：搭载 Gemini AI 的 Android 笔记本电脑 ⭐️ 8.0/10

Google has announced Googlebook, a new laptop category that runs on a customized Android desktop mode and integrates deeply with its Gemini AI assistant. The first models are expected to launch later in 2026. This represents Google's bold bet on a unified AI-first computing experience, potentially blurring the lines between mobile and desktop operating systems. If successful, it could challenge traditional laptop paradigms and accelerate the adoption of AI-integrated devices. The Googlebook runs a new operating system codenamed 'Aluminium OS,' replacing ChromeOS, and is designed from the ground up for Gemini Intelligence, offering personal and proactive help. It will be available from various OEM partners for both x86 and ARM architectures.

hackernews · tambourine_man · May 12, 17:37 · Discussion

Background: Android Desktop Mode is a multi-window computing environment built into Android that allows devices to function as desktop workstations when connected to an external display, keyboard, and mouse. Google's Gemini AI is a generative AI platform that can process text, code, images, audio, and video, and is integrated into Android and other Google services. The Googlebook aims to combine these two technologies into a new laptop form factor, succeeding Chromebooks.

Discussion: The community discussion is polarized. Some commenters see the vision of an AI-first, app-less future where the LLM replaces traditional apps, while others criticize the limitations of Android as a desktop OS, questioning the target audience and citing poor UX choices like the top panel copied from macOS. There is also a desire for a 'Google Linux' instead of Android base.

Tags: #Google, #Android, #laptops, #Gemini, #AI

Needle: 26M Parameter Tool-Calling Model Distilled from Gemini
Needle：从 Gemini 蒸馏的 2600 万参数工具调用模型 ⭐️ 8.0/10

Cactus open-sourced Needle, a 26 million parameter function-calling model distilled from Gemini, achieving 6000 tok/s prefill and 1200 tok/s decode on consumer devices. This demonstrates that small, specialized models can outperform larger ones for tool calling, enabling efficient on-device AI for agentic applications on phones, watches, and glasses. Needle uses a Simple Attention Network architecture with no MLP layers, relying solely on attention and gating, and was trained on 200B tokens pretraining plus 2B tokens of synthetic function-calling data.

hackernews · HenryNdubuaku · May 12, 18:03 · Discussion

Background: Traditional transformer models use both attention layers and feed-forward networks (FFNs). Distillation involves training a smaller "student" model to mimic a larger "teacher" model like Gemini. Needle shows that for retrieval-and-assembly tasks like function calling, FFNs can be removed entirely, reducing model size and increasing speed.

Discussion: Commenters expressed interest in practical applications like natural language CLI parsing, while also raising concerns about Google's anti-distillation defenses and the need for a live demo. One researcher confirmed similar findings on removing MLP layers.

Tags: #machine learning, #small models, #tool calling, #distillation, #on-device AI

Scrcpy v4.0 Adds Virtual and Flexible Displays
Scrcpy v4.0 新增虚拟显示和灵活显示 ⭐️ 8.0/10

Scrcpy 4.0 was released, introducing virtual display support via --new-display and flexible display resizing via --flex-display (or -x). This major update enhances scrcpy's screen mirroring capabilities, allowing users to mirror non-primary displays and resize the mirrored window on the fly, which is especially useful for app testing, presentations, and multi-screen workflows. The virtual display feature uses Android's MediaProjection API to create a separate display, while flexible display dynamically updates the device's resolution as the client window is resized.

hackernews · xnx · May 12, 20:50 · Discussion

Background: scrcpy is an open-source tool that mirrors and controls Android devices from a PC via USB or TCP/IP, streaming the screen as H.264 video. It is widely used for its low latency and high performance. Virtual displays are a new capability that allows mirroring a secondary screen or creating a separate display for apps, extending beyond the default device screen.

Discussion: Commenters expressed strong enthusiasm, with one praising the seamless flexible display and another sharing a creative WiFi-sharing setup using a phone in a salad bag. However, a user reported a persistent gesture navigation bug on Samsung phones, which the developer could not replicate, hindering their continued use.

Tags: #scrcpy, #android, #screen mirroring, #open source, #tool

DuckDB Announces Quack Client-Server Protocol
DuckDB 宣布推出 Quack 客户端-服务器协议 ⭐️ 8.0/10

DuckDB has announced the Quack remote protocol, enabling client-server database access with multiple concurrent writers, available in DuckDB v1.5.2 via the core_nightly repository. This protocol addresses DuckDB's previous limitations in remote access and concurrency, allowing it to scale horizontally and be used in multi-user applications, broadening its use cases in data analysis and application backends. Quack is a Remote Procedure Call (RPC) protocol for DuckDB, using the 'quack:' URI scheme, and is designed to be simple to set up, building on proven technologies like Apache Arrow and Flight RPC.

hackernews · aduffy · May 12, 17:54 · Discussion

Background: DuckDB is an embedded, in-process SQL database management system optimized for analytical queries. Previously, it lacked native support for remote connections and concurrent multiple writers, limiting its use in client-server architectures.

Discussion: The community reacted positively, with users expressing relief at solving concurrency and remote access problems. Some noted that 'concurrent writers' likely means serialized writes on the server side, and a few expressed uncertainty about DuckDB's evolving identity.

Tags: #DuckDB, #database, #client-server, #remote protocol

Obsidian Launches New Plugin Review System to Ease Bottleneck
Obsidian 推出新插件审核系统以缓解瓶颈 ⭐️ 8.0/10

Obsidian announced a new community plugin review system that automates checks to replace manual review, addressing the backlog that made plugin submission nearly impossible. This change relieves a critical scaling bottleneck for Obsidian, reducing developer frustration and team burnout while accelerating plugin availability in the ecosystem. The system includes automated checks, but lacks a permissions sandbox, meaning plugins still have full disk and network access, raising security concerns.

hackernews · xz18r · May 12, 15:45 · Discussion

Background: Obsidian is a popular note-taking app with a large plugin ecosystem. Previously, all new plugins required manual review by a small team, leading to long delays and burnout.

Discussion: Community members expressed mixed reactions. Some praised the system for solving the submission bottleneck, while others criticized the lack of a permission sandbox, arguing that plugins still pose security risks like remote code execution.

Tags: #obsidian, #plugins, #automation, #community, #security

Amazon, Meta Employees Game AI Usage Metrics
亚马逊和 Meta 员工为应付 AI 指标刷 Token 用量 ⭐️ 8.0/10

Amazon and Meta employees are artificially inflating their AI token consumption, a practice dubbed 'tokenmaxxing,' using internal tools like MeshClaw to perform unnecessary tasks to meet company-wide usage targets. This highlights the perverse incentives created by forcing AI adoption metrics, potentially leading to wasted resources and distorted productivity data, while reflecting broader pressure among tech giants to justify massive AI investments. Amazon's internal tool MeshClaw, inspired by the open-source OpenClaw, creates AI agents that can automate office tasks; employees use it to generate token consumption without real productivity gain. Meta employees reportedly engage in similar token-maxxing behavior.

rss · 宝玉(@dotey) · May 12, 18:18

Background: Tokens are units of data processed by AI models, often used as a proxy for usage. 'Tokenmaxxing' refers to maximizing token consumption to appear more productive. Amazon's MeshClaw is based on OpenClaw, an open-source autonomous AI agent that can browse the web, run commands, and interact with applications. The trend reflects management's focus on quantifiable AI utilization metrics.

Tags: #AI工具, #企业文化, #硅谷动态, #激励扭曲

Death Switch Script Deletes Files If Stolen GitHub Token Revoked
死亡开关脚本：GitHub 令牌被撤销即删除文件 ⭐️ 8.0/10

An attacker has installed a script named 'gh-token-monitor' on a victim's machine that polls the GitHub API every 60 seconds using a stolen token. If it detects the token has been revoked, the script immediately executes 'rm -rf ~/' to delete all files in the user's home directory. This 'death switch' technique turns a victim's defensive action—revoking a compromised token—into a destructive trigger, making incident response extremely dangerous. It highlights a new class of supply chain attacks where malware not only steals credentials but also enforces retaliatory data destruction. The script was discovered during analysis of a GitHub issue related to a supply chain attack on TanStack/router. The attacker had previously released 84 malicious npm versions by poisoning CI caches and abusing OIDC tokens to exfiltrate credentials.

rss · Viking(@vikingmute) · May 13, 01:32

Background: A 'death switch' in malware is a mechanism that triggers a destructive action (e.g., data wipe) when a certain condition is met, often to retaliate against the victim. In this case, the condition is the revocation of a stolen GitHub personal access token. The script continuously monitors the token's validity via the GitHub API; if the token is revoked, it assumes the victim is trying to stop the attack and responds by deleting local files. This technique is particularly insidious because it penalizes the victim for taking a standard security step.

Tags: #security, #malware, #GitHub token, #death switch, #supply chain attack

GB200 GPUs Enable Novel Prefill/Decode Disaggregation for MoE Models
GB200 GPU 为 MoE 模型实现创新的预填充/解码分离 ⭐️ 8.0/10

Perplexity published research on using NVIDIA GB200 NVL72 Blackwell racks to serve the Qwen3 235B MoE model, demonstrating a novel prefill/decode disaggregation method that improves throughput compared to Hopper GPUs. This advancement addresses a key bottleneck in serving large Mixture-of-Experts (MoE) models, enabling higher efficiency and lower latency for inference, which is critical for deploying advanced AI assistants at scale. The method leverages GB200's high-bandwidth memory and interconnect to separate prefill (compute-intensive) and decode (memory-bound) phases across different GPU sets, optimizing resource utilization.

rss · Aravind Srinivas(@AravSrinivas) · May 12, 14:27

Background: Large language models like Qwen3-235B-A22B use a Mixture-of-Experts architecture where only a subset of parameters are activated per token, reducing compute cost. Prefill-decode disaggregation is a serving technique that separates the initial prompt processing (prefill) from the token-by-token generation (decode), allowing each phase to be scheduled on different hardware for better throughput. NVIDIA's GB200 GPU represents a significant upgrade over Hopper in memory bandwidth and scaling capabilities.

Tags: #AI infrastructure, #large language models, #MoE, #GPU serving, #disaggregation

Andreessen: AI Makes Programmers Work Harder, Not Replace Them
安德烈森：AI 让程序员更忙碌而非取代他们 ⭐️ 8.0/10

Marc Andreessen, co-founder of Netscape and venture capitalist, argued in a live discussion that AI is causing programmers to work harder and longer hours, contrary to the Luddite expectation that automation would reduce work. This counters the common fear that AI will displace programmers, suggesting instead that AI boosts productivity and expands the scope of human work, which has significant implications for the future of software engineering and the broader economy. Andreessen observed that programmers are 'working harder than ever,' 'more hours than ever,' and are 'completely exhausted but euphoric,' attributing this to increased marginal productivity. The discussion also touched on topics like AI doomer literature and the future of tech jobs.

rss · a16z(@a16z) · May 12, 19:40

Background: The Luddite fallacy is the belief that automation inevitably leads to mass unemployment, which has been disproven historically as technology creates new jobs. Zero-sum thinking, when applied to technology, assumes that gains from automation come at the expense of workers, but Andreessen argues that AI instead expands work opportunities. This perspective is part of a broader debate on AI's impact on employment and productivity.

Tags: #AI, #software engineering, #economics, #productivity, #future of work

Perceptron Mk1 Launches on OpenRouter
Perceptron Mk1 登陆 OpenRouter ⭐️ 8.0/10

Perceptron Mk1, a frontier vision-language model for video analysis and embodied reasoning, is now live on OpenRouter, offering dynamic frame rate up to 2 FPS, 32k multimodal context, and structured spatial outputs like points, boxes, polygons, and clips. This release brings advanced video understanding and spatial reasoning capabilities to a wide audience via a unified API, potentially accelerating applications in robotics, autonomous driving, and video analytics. The model supports dynamic frame rate up to 2 FPS across a 32k multimodal context window and outputs structured spatial primitives as first-class citizens, enabling precise localization and tracking.

rss · OpenRouter(@OpenRouterAI) · May 12, 16:08

Background: OpenRouter is a unified API platform that provides access to hundreds of large language models through a single interface, simplifying integration and cost management. Embodied reasoning refers to AI's ability to understand and reason about physical interactions in real-world environments, combining vision, language, and spatial awareness.

Tags: #vision-language model, #video reasoning, #embodied AI, #OpenRouter, #multimodal AI

FinalDose Launches First Programmable Drug Platform for All Cancers
FinalDose 推出首个可编程药物平台，针对所有癌症 ⭐️ 8.0/10

FinalDose, a Y Combinator-backed startup, launched the first programmable drug platform—a single smart drug molecule that detects diseased cells by their DNA and destroys them, initially targeting all cancers. This approach could revolutionize targeted cancer therapy by offering a universal, DNA-based detection mechanism that reduces off-target effects and adapts to various cancer types, potentially making treatment more effective and personalized. The platform is described as a single programmable molecule that combines DNA sensing and therapeutic action, though specific technical details on the molecular mechanism are not yet publicly disclosed. The startup is starting with all cancers, implying broad applicability.

rss · Y Combinator(@ycombinator) · May 12, 16:00

Background: Programmable drug platforms aim to create intelligent therapeutics that can sense disease biomarkers and respond with targeted action. Traditional cancer treatments like chemotherapy lack specificity, while newer targeted therapies often require custom designs for each cancer type. DNA-based targeting allows the drug to recognize cancer-specific genetic signatures, potentially enabling a one-size-fits-all solution. This field builds on advances in nanomedicine and genetic engineering.

Tags: #Biotech, #Cancer Therapy, #Programmable Drugs, #Y Combinator, #Startup

X's 'For You' Algorithm Shifts Users Politically Right
X 的“为你推荐”算法使用户政治立场右移 ⭐️ 8.0/10

A study published on The Conversation found that using X's 'For You' feed for a few weeks significantly shifts users' political views to the right, including favoring GOP policies, opposing Trump prosecution, and showing pro-Russia sentiment. The effect did not wear off quickly after stopping. This study provides empirical evidence of algorithmic amplification on social media, raising concerns about manipulation of public opinion and democratic processes. It highlights the need for transparency and regulation of recommendation systems. The study explicitly tested the 'For You' feed, which combines in-network and out-of-network content ranked by a Grok-based transformer model. The algorithm's influence was significant and persistent, even after participants stopped using the feed.

rss · kottke.org · May 12, 15:59

Background: X (formerly Twitter) uses a machine learning algorithm to power its 'For You' feed, blending posts from accounts you follow with recommendations from unknown accounts. This algorithm ranks content using a transformer model to predict engagement, and its design can amplify certain viewpoints. The study adds to concerns about algorithmic bias and filter bubbles in social media.

Tags: #algorithmic bias, #social media, #politics, #study, #X algorithm

Vercel's MDXG turns .md files into interactive docs
Vercel 推出的 MDXG 让 .md 文件变为交互式文档 ⭐️ 8.0/10

Vercel has introduced the Markdown Experience Guidelines (MDXG), a specification that enables .md files to function as multi-page documentation with navigation, search, and syntax highlighting, without requiring any changes to the original files. This standard simplifies the creation of rich documentation from plain Markdown files, potentially streamlining developer workflows and making Markdown a more powerful tool for technical writing and knowledge bases. MDXG is hosted on Vercel Labs GitHub and defines how interfaces should present and interact with Markdown documents, leveraging existing Markdown syntax without extensions.

rss · Geek(@geekbb) · May 13, 08:04

Background: Markdown is a lightweight markup language created by John Gruber in 2004, widely used for formatting text in plain-text editors. Traditionally, Markdown files are static and require additional tooling to generate navigable documentation sites. MDXG aims to standardize the interactive presentation of Markdown content across different platforms.

Tags: #Markdown, #Vercel, #Documentation, #MDXG, #Web Development

Unitree Launches GD01 Manned Mecha Robot for $650,000
宇树科技发布 GD01 载人机甲，售价 65 万美元 ⭐️ 8.0/10

Unitree Robotics has unveiled the GD01, a manned mecha robot that can carry a human pilot, with a starting price of $650,000 and a weight of about 500 kg. The company claims it is mass-production ready, moving beyond prototype stage. This announcement marks a significant step from laboratory robotics to real-world civilian applications, potentially revolutionizing personal transportation and heavy-duty operations. It positions Unitree as a leader in commercial mecha technology, competing with other humanoid and quadruped robots. The GD01 features stable bipedal walking and can switch to quadruped mode for rough terrain. It is optionally manned, meaning it can operate with or without a human pilot. The robot is designed as a civilian vehicle, not just a showpiece.

rss · AI Will(@FinanceYF5) · May 13, 01:12

Background: Unitree Robotics, founded in 2016 by Wang Xingxing, is known for its high-performance quadruped robots like the Go1 and B2. The GD01 is their first human-carrying mecha, building on their expertise in legged locomotion. Mecha robots are a staple of science fiction, but practical implementations have been limited to experimental machines. Unitree's claim of production readiness distinguishes the GD01 from earlier prototypes.

Tags: #robotics, #humanoid, #mecha, #unitree, #commercial

Linux Hit by Second Severe Vulnerability in Two Weeks
Linux 两周内遭遇第二个严重漏洞 ⭐️ 8.0/10

A new severe security vulnerability has been reported in the Linux kernel, following a similar critical flaw disclosed just two weeks earlier. This rapid succession of high-severity vulnerabilities underscores ongoing challenges in Linux security and may erode trust among enterprise users relying on the OS for critical infrastructure. The specific CVE identifier and technical details of the vulnerability have not yet been disclosed, but it is described as 'severe' and likely to require urgent patching.

rss · Hacker News: Newest · May 13, 13:20

Background: Linux is the world's most widely used open-source operating system kernel, powering everything from servers and smartphones to embedded devices. Severe vulnerabilities, especially those allowing privilege escalation or remote code execution, require immediate attention and can impact millions of systems.

Tags: #Linux, #security, #vulnerability, #CVE

Nous Research Introduces Lighthouse Attention for Efficient Long-Context Pretraining
Nous Research 推出 Lighthouse Attention 加速长上下文预训练 ⭐️ 8.0/10

Nous Research has introduced Lighthouse Attention, a subquadratic wrapper for standard scaled dot-product attention (SDPA) that speeds up long-context pretraining. The wrapper uses a hierarchical, gradient-free selection layer to compress and decompress queries, keys, and values symmetrically, and can be removed before deployment so the model retains vanilla attention inference. Most efficient-attention methods either change the deployment architecture or sacrifice quality. Lighthouse Attention avoids both by being a training-only wrapper that can be cleanly removed, potentially making long-context pretraining significantly faster without any inference-time cost. The method wraps SDPA with a hierarchical, gradient-free selection layer that preserves left-to-right causality. Preliminary LLM experiments show faster total training time and lower final loss compared to full-attention baselines, and the wrapper is removed during a short recovery phase near the end of training.

rss · elvis(@omarsar0) · May 12, 15:36

Background: Scaled dot-product attention (SDPA) is the core mechanism in Transformer models, but its quadratic complexity with sequence length makes long-context pretraining expensive. Subquadratic attention variants aim to reduce this complexity, but often require architectural changes at inference time. Lighthouse Attention proposes a novel approach: a subquadratic wrapper that is used only during training and removed before deployment, preserving vanilla SDPA at inference.

Tags: #attention, #pretraining, #efficiency, #long-context, #machine learning

Microsoft unveils multi-model agentic security system MDASH
微软发布多模型智能体安全系统 MDASH ⭐️ 8.0/10

Microsoft announced a new multi-model agentic security system codenamed MDASH, which uses over 100 specialized AI agents across frontier and custom models to discover and validate vulnerabilities. The system achieved top performance on the CyberGym benchmark and has already helped find and fix 16 vulnerabilities, including four critical remote code execution flaws, ahead of Patch Tuesday; customers can now sign up for a private preview. This represents a significant advancement in AI-powered cybersecurity, demonstrating how multi-agent systems can outperform single models and traditional tools in automating vulnerability discovery. If successful, it could dramatically reduce the window between vulnerability discovery and patching, enhancing security for millions of Windows users and setting a new standard for AI-driven defense. The system, built by Microsoft's Autonomous Code Security team, uses an ensemble of frontier models (like GPT-4, Claude) and distilled custom models. Agents collaborate in a debate-style process to confirm exploitability, reducing false positives. MDASH has shown superior performance on CyberGym, a benchmark from UC Berkeley with 1,507 real-world vulnerabilities.

rss · Satya Nadella(@satyanadella) · May 13, 00:01

Background: Agentic AI systems use multiple specialized AI agents that work together to accomplish complex tasks, such as code analysis and vulnerability exploitation. The CyberGym benchmark, developed by UC Berkeley, evaluates AI agents on their ability to discover and exploit real-world software vulnerabilities across 188 large projects. Traditional vulnerability detection often relies on static analysis or single AI models, which can miss complex flaws or generate high false-positive rates.

Tags: #AI, #security, #multi-agent, #Microsoft, #vulnerability detection

Andrew Ng: No AI Jobpocalypse, Fears Overblown
吴恩达：AI 不会导致大规模失业，恐慌被夸大 ⭐️ 8.0/10

Andrew Ng published an article arguing that fears of AI-induced mass unemployment are overblown, citing strong hiring in software engineering and a healthy 4.3% U.S. unemployment rate. This counter-narrative is significant because it challenges the dominant AI jobpocalypse story, potentially influencing public policy and investment decisions, and encouraging more people to pursue AI skills. Ng notes that frontier AI labs and SaaS companies have incentives to exaggerate AI's impact, and businesses may blame AI for layoffs that are actually due to pandemic-era overhiring.

rss · Andrew Ng(@AndrewYNg) · May 12, 16:25

Background: The AI jobpocalypse narrative suggests that AI will automate many jobs, leading to widespread unemployment. Andrew Ng, a prominent AI leader, argues that historical technology waves have created more jobs than they destroyed, and AI will follow the same pattern.

Tags: #AI, #job market, #Andrew Ng, #automation, #software engineering

Demis Hassabis Announces $2.1B Funding for Isomorphic Labs
Demis Hassabis 宣布为 Isomorphic Labs 融资 21 亿美元 ⭐️ 8.0/10

Demis Hassabis, CEO of Isomorphic Labs and DeepMind, announced $2.1 billion in new funding to accelerate AI-powered drug discovery, building on the AlphaFold breakthrough. This substantial investment signals strong industry confidence in AI-driven drug discovery, potentially revolutionizing how new medicines are developed and reducing the time and cost of bringing treatments to patients. Isomorphic Labs, founded in 2021 as a spin-off from DeepMind under Alphabet, uses AlphaFold technology to predict protein structures and discover new drug targets. The $2.1B funding will turbocharge their mission to 'reimagine drug discovery and one day solve all disease'.

rss · Demis Hassabis(@demishassabis) · May 12, 13:50

Background: AlphaFold is an AI system developed by DeepMind that predicts 3D protein structures from amino acid sequences with high accuracy, achieving a breakthrough in biology. This work earned Demis Hassabis a share of the 2024 Nobel Prize in Chemistry. Isomorphic Labs was founded specifically to apply AlphaFold's capabilities to drug discovery, aiming to dramatically accelerate the process of finding new medicines.

Tags: #AI, #Healthcare, #Drug Discovery, #Funding, #AlphaFold

Jina AI Releases v5-omni Multimodal Embedding Model
Jina AI 发布 v5-omni 多模态嵌入模型 ⭐️ 8.0/10

Jina AI has released v5-omni, a multimodal embedding model that supports text, image, audio, and video. The model is now available on Elastic Inference Service, HuggingFace, and Jina API. This release enables unified vector representations across four modalities, simplifying search and retrieval tasks in multimodal AI applications. It is available on major platforms, lowering the barrier for developers to integrate multimodal embeddings. v5-omni is a single model covering text, image, audio, and video, with an accompanying arXiv paper and blog post providing technical details. It is accessible via HuggingFace, Jina API, and Elastic's managed inference service.

rss · Jina AI(@JinaAI_) · May 12, 15:44

Background: Multimodal embeddings represent different data types (e.g., text, images) in a shared vector space, enabling similarity search across modalities. Jina AI specializes in neural search and embeddings, and Elastic Inference Service provides a managed platform for deploying such models.

Tags: #multimodal embeddings, #Jina AI, #v5-omni, #AI model release, #HuggingFace

Perplexity Optimizes Qwen on Blackwell for Faster Inference
Perplexity 在 Blackwell 上优化 Qwen 实现更快推理 ⭐️ 8.0/10

Perplexity published a research paper detailing how they host the post-trained Qwen3 235B model on NVIDIA GB200 NVL72 Blackwell racks using disaggregated prefill/decode, Blackwell-native quantization, custom kernels, and rack-scale NVLink to achieve faster answers and lower serving costs. This demonstration shows that NVIDIA's Blackwell platform significantly outperforms Hopper for high-throughput inference on large Mixture-of-Experts models, positioning it as a leading platform for production-scale LLM serving. The optimizations by Perplexity provide a blueprint for other organizations to achieve cost-effective, low-latency inference. The solution leverages prefill/decode disaggregation where separate GPU resources handle prefill and decode phases, and uses Blackwell's native support for sub-8-bit data types (like MXFP6 and MXFP4) for efficient quantization. Custom CUDA kernels and rack-scale NVLink interconnects further minimize latency and maximize throughput.

rss · Perplexity(@perplexity_ai) · May 12, 14:17

Background: Large language model inference involves two phases: prefill (processing the input prompt) and decode (generating tokens). Disaggregated serving separates these phases across different GPUs to improve utilization and reduce latency. NVIDIA's Blackwell architecture introduces native support for low-precision formats like FP4, which can double throughput compared to Hopper. Rack-scale NVLink enables high-bandwidth communication across multiple GPUs in a chassis.

Tags: #NVIDIA, #Blackwell, #large model inference, #Perplexity, #Qwen

AI-Accelerated Teams Face Production Reliability Risks
AI 加速团队面临生产可靠性风险 ⭐️ 8.0/10

The Stack Overflow Blog article discusses how AI-accelerated development teams are increasingly causing production outages and highlights best practices from teams that maintain high reliability. As AI tools speed up code generation, the risk of introducing subtle bugs and breaking production grows, making reliability a critical competitive differentiator for software teams. The article contrasts common failure patterns (e.g., over-reliance on AI without review) with countermeasures such as systematic testing, observability, and human-in-the-loop validation specific to AI-generated code.

rss · Stack Overflow Blog · May 12, 16:41

Background: AI-accelerated development refers to using large language models and code generation tools to produce software faster. While this boosts velocity, it also introduces challenges because AI may generate incorrect or insecure code that escapes traditional quality checks, leading to production issues.

Tags: #AI, #software engineering, #DevOps, #best practices, #production reliability

Mathematics of Backlogs for Queue Recovery in Distributed Systems
分布式系统中队列恢复的积压数学 ⭐️ 8.0/10

Rajesh Kumar Pandey's article on InfoQ provides practical mathematical formulas for calculating backlog drain time, consumer headroom, and auto-scaling triggers in distributed systems, covering failure modes such as retry amplification and metastable states. These formulas give engineers a systematic way to plan capacity and prevent common failures like retry storms and metastable collapses, which can lead to costly outages in large-scale systems. The article includes equations for drain time based on queue length and processing rate, defines consumer headroom as spare capacity for absorbing spikes, and proposes auto-scaling triggers based on queue growth rate rather than absolute length to avoid oscillation.

rss · InfoQ · May 13, 09:00

Background: In distributed systems, backlogs occur when producers outpace consumers, leading to queued requests. Retry amplification happens when failed requests trigger cascading retries, overwhelming the system. Metastable states are failure patterns where a system remains in a degraded state even after the initial trigger is removed. These concepts are critical for understanding the failure modes addressed in the article.

Tags: #distributed systems, #capacity planning, #queue management, #backpressure, #auto-scaling

AWS WorkSpaces Lets AI Agents Control Legacy Apps Without APIs
AWS WorkSpaces 让 AI 代理无需 API 操作遗留应用 ⭐️ 8.0/10

AWS announced that Amazon WorkSpaces can now serve as managed virtual desktops for AI agents in public preview. Agents authenticate via IAM and operate legacy applications using computer vision and input simulation without requiring APIs. This capability solves the challenge of integrating legacy desktop applications that lack APIs, enabling enterprises to automate workflows with AI agents. It could significantly reduce manual effort in environments where modernizing legacy systems is costly or infeasible. The service is in public preview and uses IAM for authentication. Reflex benchmarks show that vision-based agents consume 45x more tokens than API-based agents, highlighting potential cost considerations for enterprises.

rss · InfoQ · May 13, 07:31

Background: AI agents typically require APIs to interact with software, but many legacy desktop applications lack modern APIs, making automation difficult. AWS WorkSpaces provides virtual desktop infrastructure; this new feature allows agents to 'see' the screen and simulate mouse and keyboard input using computer vision and input simulation, effectively bridging the gap. The high token consumption of vision-based agents compared to API-based agents is a key trade-off.

Tags: #AWS, #AI Agents, #Legacy Applications, #Computer Vision, #Virtual Desktops

AI coding agents expand attack surface beyond source code
AI 编码代理将攻击面扩展到源代码之外 ⭐️ 8.0/10

Google Cloud published a blog post detailing how AI coding agents introduce new attack vectors through files like agent instructions, runtime settings, and extension packages, and announced the use of semantic analysis via VirusTotal Code Insight to detect malicious configurations. This expands the software supply chain attack surface to non-code files that influence AI agent behavior, requiring security teams to rethink defense strategies as attackers exploit trusted files in development environments. The attack surface is categorized into four areas: what executes, what instructs, what connects, and what extends. Google Threat Intelligence integrates agentic capabilities to link these artifacts to broader threat campaigns at scale.

rss · Cloud Blog · May 12, 16:00

Background: AI coding agents are autonomous tools that operate in IDEs, editors, and terminals, with access to local files and command execution. They rely on configuration files (e.g., AGENTS.md, .env) that influence behavior but are often trusted without scrutiny, creating new attack surfaces that traditional source code scanning fails to address.

Tags: #AI security, #software supply chain, #developer tools, #AI agents

How Figma Upgraded Data Pipeline from Multi-Day Latency to Real-Time
Figma 如何将数据管道从多日延迟升级为实时 ⭐️ 8.0/10

Figma's engineering team revamped its data pipeline to achieve real-time latency, moving from multi-day batch processing to streaming architecture. The upgrade addressed scaling challenges as the platform's user base and data volume grew exponentially. This shift enables Figma to deliver real-time analytics and insights, improving user experience and operational decision-making. It also provides a valuable reference for other companies grappling with similar scalability and latency issues in data pipelines. The pipeline likely leveraged PostgreSQL read replicas and PgBouncer connection pooling to near-real-time replication, combined with streaming technologies. The migration tackled bottlenecks that caused data freshness delays of multiple days.

rss · ByteByteGo Newsletter · May 12, 15:31

Background: Data pipelines often process information in batches (e.g., overnight), causing significant latency for analytics and reporting. Real-time pipelines process data continuously, enabling immediate insights. Figma, a collaborative design platform, generates massive amounts of user interaction data that requires fast processing for features like team analytics and usage monitoring.

Tags: #data engineering, #data pipeline, #real-time, #system design, #scalability

Databricks Catalog Commits GA Unifies Open Table Formats and Catalogs
Databricks Catalog Commits 正式发布，统一开放表格式与目录 ⭐️ 8.0/10

Databricks announced the general availability of Catalog Commits, which unify open table formats like Delta and Iceberg with open catalogs such as Unity Catalog in the lakehouse architecture. This advancement simplifies lakehouse management by enabling atomic, multi-table transactions across different table formats and catalogs, reducing operational complexity and improving data consistency for enterprises. Catalog Commits require tables to be Unity Catalog managed (Delta or Iceberg) and Databricks Runtime 16.4 or above is needed to read, write, or create tables with catalog commits enabled.

rss · Databricks · May 12, 15:00

Background: Open table formats like Delta Lake and Apache Iceberg bring ACID transactions and schema evolution to data lakes, while open catalogs like Unity Catalog provide centralized metadata management. Catalog Commits bridge these two layers, allowing atomic commits that span both the table format and the catalog, ensuring consistency across the lakehouse ecosystem.

Tags: #data engineering, #lakehouse, #open table formats, #catalog, #Databricks

Daily tech roundup: supply chain attack, AI trends, EU regulation
每日科技综述：供应链攻击、AI 趋势、欧盟监管 ⭐️ 8.0/10

A curated roundup of ten notable tech stories from Hacker News on May 13, 2026, covering a major supply chain attack on TanStack, platform trust issues with Bambu Lab, AI shifting development to systems languages, and new EU regulations for minors. This roundup highlights critical cybersecurity vulnerabilities in open-source CI/CD pipelines, growing tensions between hardware companies and the open-source community, and regulatory shifts that could reshape online platforms. It also signals a potential transition in software engineering roles as AI models favor type-safe languages. The TanStack attack exploited GitHub Actions pull_request_target to inject malicious versions into 42 npm packages. Bambu Lab is criticized for enforcing cloud dependency and legal pressure on community firmware. Other stories include GitLab's restructuring, a GUI screenshot retrospective, and a UCLA stroke recovery drug candidate (DDL-920).

rss · HackerNews每日摘要 on SuperTechFans · May 12, 23:07

Background: Hacker News is a social news website focusing on computer science and entrepreneurship, where users submit and vote on stories. This particular daily roundup condenses the day's most upvoted submissions, allowing readers to quickly grasp key developments across cybersecurity, open-source, AI, regulation, and hardware.

Tags: #HackerNews, #tech news, #cybersecurity, #AI, #regulation

US Commerce Dept Removes AI Safety Testing Agreement Details
美国商务部删除 AI 安全测试协议细节 ⭐️ 8.0/10

The US Commerce Department's Center for AI Standards and Innovation (CAISI) removed web pages detailing agreements with Google, xAI, and Microsoft for pre-release security testing of AI models. The original links returned 404 errors and redirected to the CAISI homepage, with no explanation provided. This removal raises concerns about government transparency in AI safety regulation, especially as such testing is critical for responsible AI development. It could undermine public trust and accountability in the oversight of frontier AI models. The removed pages were part of CAISI, the US counterpart to the UK's AI Security Institute, and involved pre-release vulnerability testing of frontier AI models. The Commerce Department and the White House did not respond to requests for comment.

telegram · zaihuapd · May 12, 13:38

Background: AI safety institutes were established internationally following the 2023 UK AI Safety Summit, with the US creating CAISI to evaluate and ensure safety of advanced AI models. The deletion of testing protocol details may signal a policy shift or lack of coordination, amid ongoing debates about AI regulation and industry self-governance.

Tags: #AI Safety, #US Government, #Transparency, #Regulation, #Big Tech

SpaceX and Google Discuss Orbital Data Centers
SpaceX 与谷歌磋商轨道数据中心合作 ⭐️ 8.0/10

Google is in talks with SpaceX to launch orbital data centers under its Project Suncatcher initiative, aiming to prototype solar-powered satellites with custom AI chips by 2027. This could shift cloud computing to space, addressing terrestrial energy constraints while enabling low-latency processing for Earth observation and global AI workloads. SpaceX plans to use upgraded Starlink satellites with large solar arrays for built-in computing power, and Google has partnered with Planet Labs for satellite development. The project is still in early-stage talks.

telegram · zaihuapd · May 12, 16:28

Background: Orbital data centers are proposed space-based facilities that use space solar power to overcome Earth's energy and cooling limitations for AI computing. Historical concepts include the Strategic Defense Initiative's Brilliant Pebbles and recent military architectures like the Proliferated Warfighter Space Architecture. Google's Project Suncatcher is a research moonshot to network solar-powered satellites with Tensor Processing Units into an orbital AI cloud.

Tags: #space, #cloud computing, #data centers, #Google, #SpaceX

Xiaomi Releases OneVL: One-Step Latent Reasoning for Autonomous Driving
小米发布 OneVL：一步式潜空间推理框架 ⭐️ 8.0/10

Xiaomi has released OneVL, a one-step latent space reasoning framework that unifies vision-language-action (VLA) models and world models for autonomous driving, and has fully open-sourced it. This framework achieves state-of-the-art results on multiple benchmarks while reducing latency to as low as 0.24 seconds, which is only 5.4% of autoregressive VLA models. This could accelerate the deployment of more efficient and safer autonomous driving systems. OneVL uses latent space chain-of-thought (CoT), encoding physical causality with visual latent tokens and driving intent with language latent tokens. During inference, it removes auxiliary decoders used in training for predicting future frames and readable chain-of-thought, enabling one-step parallel generation.

telegram · zaihuapd · May 13, 10:33

Background: Vision-Language-Action (VLA) models combine visual perception, language understanding, and action prediction for robotics and autonomous driving. World models allow autonomous systems to synthesize sensor data and predict future scenarios. Latent space reasoning moves thinking into an abstract, non-language space for efficiency. OneVL unifies these concepts.

Tags: #小米, #自动驾驶, #VLA, #世界模型, #潜空间推理

Ma Huateng Addresses Tencent's AI Progress at Shareholder Meeting
马化腾在股东会回应腾讯 AI 进展 ⭐️ 8.0/10

At Tencent's shareholder meeting, CEO Ma Huateng admitted that the company initially lagged in AI but has now 'boarded the ship,' though it needs to accelerate to keep pace with competitors. This candid acknowledgment from a top tech leader signals a strategic shift and increased urgency in AI investment, which could reshape China's AI competitive landscape and affect Tencent's product ecosystem. Ma described the journey as initially finding a leaking ship, then successfully boarding but unable to sit down, urging faster progress. He emphasized that Tencent would not blindly chase trends but leverage its own strengths.

telegram · zaihuapd · May 13, 13:09

Background: Tencent, a Chinese internet giant, has been perceived as lagging in the AI race compared to rivals like Baidu and ByteDance. The company is now ramping up talent acquisition and internal training to build AI capabilities, focusing on integrating AI into its existing products like WeChat and gaming.

Tags: #腾讯, #AI, #马化腾, #战略, #行业动态