US Citizen Faces Felony for Deleting Phone Data at Border
美国公民因在边境删除手机数据面临重罪指控 ⭐️ 9.0/10

On August 21, 2026, The New York Times reported that U.S. citizen Samuel Tunick faces felony charges for deleting data from his phone during a border search by federal agents. The incident is fueling debate over whether travelers may take technical steps to protect their data during device searches at the border. The case tests the limits of the 'border search exception' and could clarify whether U.S. citizens have a legal right to delete or encrypt data on their own devices during an inspection. Its outcome may affect journalists, lawyers, and business travelers who carry sensitive information across U.S. borders. Border agents commonly use tools such as Cellebrite UFED to unlock and extract mobile data, and alleged deletion could be prosecuted as obstruction of justice. The specific charges, evidence, and whether Tunick had consented to the search are not yet public, but the case is being watched closely by digital-rights groups.

hackernews · floathub · Aug 21, 12:10 · Discussion

Background: U.S. courts have long recognized a 'border search exception' under the Fourth Amendment, permitting customs and border agents to search electronic devices without a warrant. To counter this, some travelers use anti-forensic techniques—ranging from encryption and decoy partitions to secure-deletion software that overwrites data before a device is examined. Meanwhile, law-enforcement tools like Cellebrite UFED have become standard for extracting data from locked devices, creating an escalating technical and legal tug-of-war over device privacy at borders.

Discussion: Commenters largely sympathized with the traveler and debated practical countermeasures. Some suggested using decoy passcodes that boot into a separate 'clean' partition, imaging the phone to an external drive and wiping it before crossing the border, or carrying a cheap burner phone with minimal data. Others noted that these workarounds can be inconvenient and legally risky, and one user linked the incident to broader concerns about government surveillance and blocked archive pages.

Tags: #privacy, #civil liberties, #border search, #security, #legal

OpenAI Open-Sources Codex Harness, Enabling Agents in Any Business System
OpenAI 开源 Codex Harness，让智能体能嵌入任意业务系统 ⭐️ 9.0/10

OpenAI has open-sourced the Codex Harness, the agent execution framework behind Codex, releasing the codex exec CLI, the official Codex SDK, and the app-server. Developers can now embed Codex's agent loop—context memory, tool calling, sandbox execution, and approval flows—into their own business applications. This turns Codex from a standalone coding assistant into a platform component that can operate inside existing enterprise systems such as logistics back-ends, ticketing, tax software, and monitoring dashboards. For developers, it removes a layer of infrastructure; for products, AI can now enter existing software while the business system remains the interface of record. OpenAI did not open-source the entire Codex product, model, or cloud services; the release covers the Harness and integration components only. The documentation defines three integration levels—codex exec for scripted/CI tasks, Codex SDK for app-embedded workflows, and Codex app-server for long-lived product-integrated agents—and sets a clear boundary where business applications retain data, UI, and approval control.

rss · 小互(@imxiaohu) · Aug 21, 02:57

Background: Codex is OpenAI's AI coding agent, previously accessed via a chat app, terminal CLI, or IDE extension. The Harness is the open-source framework that implements the agent loop: maintaining context, inspecting files/data, calling tools, streaming progress, handling failures, and requesting human approval before sensitive actions. OpenAI describes this as 'Codex as a platform,' with the app-server positioned as the first-class integration method moving forward, and MCP tools connecting agents to internal data and business systems.

Tags: #AI, #OpenAI, #Codex, #开源, #Agent框架

OpenAI New API Feature Goes Live, Rolling Out to Codex and ChatGPT Work
OpenAI 新 API 功能上线，并陆续推送至 Codex 与 ChatGPT Work ⭐️ 9.0/10

OpenAI announced that a new API feature is now live for API users and is rolling out to Codex and ChatGPT Work throughout the day. The company confirmed that Pro, Plus, and Business subscription usage remains unchanged. This rollout extends a new capability from API developers to users of OpenAI's coding agent (Codex) and its work-agent mode (ChatGPT Work), potentially impacting developer workflows across the ecosystem. It reflects OpenAI's strategy of unifying new features across its API and consumer-facing products. The announcement links to OpenAI's pricing documentation, suggesting that pricing or usage details accompany the feature. The rollout is phased: it is live for API users first, then reaches Codex and ChatGPT Work later in the day, while subscription usage is explicitly unchanged.

rss · OpenAI Developers(@OpenAIDevs) · Aug 21, 19:45

Background: OpenAI's API lets developers integrate its models into their applications. Codex is OpenAI's AI coding agent, released in 2025, that handles software engineering tasks such as writing code, fixing bugs, and proposing pull requests. ChatGPT Work, launched in July 2026, is an agent mode that can take action across apps and files to complete tasks. The new feature is likely a model or capability being made available across these surfaces, with the pricing link indicating a change in how it is billed.

Tags: #OpenAI, #API, #Codex, #ChatGPT, #Pricing

Jeff Dean leaves Google after 27 years to found startup
Jeff Dean 工作 27 年后离开谷歌，创办独立初创公司 ⭐️ 9.0/10

Jeff Dean, former Chief Scientist at Google, announced he is leaving the company after 27 years to found an independent startup. In an emotional video, he described his deep attachment to Google and the remarkable colleagues he met there. Jeff Dean is one of the most influential computer scientists in the industry, having shaped Google's engineering culture and AI infrastructure for decades. His departure to start a new company could signal a significant shift in the AI talent landscape and inspire other researchers to pursue independent ventures. The news includes a video in which Jeff Dean, visibly emotional, says he has very deep feelings about his time at Google and met many amazing colleagues. The specific focus of the startup has not yet been disclosed, and the post shows modest social engagement with no further details.

rss · AI Will(@FinanceYF5) · Aug 21, 04:28

Background: Jeff Dean has been a key figure at Google for nearly three decades, contributing to core systems such as MapReduce, Bigtable, and TensorFlow, and helping steer the company's AI strategy as Chief Scientist. His decision to leave and start an independent company is a notable event in the tech world, as very few engineers reach his level of influence. Independent startups often allow more agile research environments and faster experimentation, which may appeal to someone with his expertise.

Tags: #Jeff Dean, #Google, #Startup, #Tech News

DeepSeek launches multimodal API for vision model deepseek-v4-flash-vision-exp
DeepSeek 推出视觉模型 deepseek-v4-flash-vision-exp 的多模态 API 支持 ⭐️ 9.0/10

DeepSeek announced multimodal API support for its experimental vision model deepseek-v4-flash-vision-exp, accepting images alongside text for tasks like screenshot reading and chart analysis. The API supports Chat Completions, Messages, and Responses formats, with images tokenized at up to 384 tokens each at V4-Flash pricing. This extends DeepSeek's developer platform into multimodal territory, letting developers add vision capabilities without switching providers. It is especially significant because previous DeepSeek models often hallucinated vision abilities, so a real vision endpoint addresses a major practical gap. Images can be supplied via base64 strings, external URLs, or DeepSeek's Files API, and mixed text+image input is supported. Before inference, images are automatically resized to roughly a total pixel count comparable to 800×800, which community users note may limit OCR on full-page documents.

rss · DeepSeek(@deepseek_ai) · Aug 21, 09:17

Background: DeepSeek is a Chinese AI lab known for open-weight models such as DeepSeek-V3 and R1. Vision language models combine image and text understanding by converting images into tokens that are billed alongside text tokens, a common technique also used by OpenAI, Anthropic, and Google. DeepSeek's Files API allows uploading images once and referencing them by file_id with the deepseek-v4-flash-vision-exp model.

Discussion: Community sentiment is cautiously positive. Some users see it as a fix for DeepSeek's prior inability to read screenshots and for models that fake vision, while others report it still fails simple OCR tasks like reading a clock, and one user notes the 800×800 resizing is too low for full-page OCR.

Tags: #multimodal, #vision, #API, #DeepSeek, #AI

Simulation as the New Scaling Law: Simile AI Aims for 8 Billion Digital Twins
模拟成为新规模法则：Simile AI 计划创建 80 亿数字孪生 ⭐️ 9.0/10

Joon Sung Park, author of the viral Generative Agents project and now CEO of Simile AI, argues that simulation is the new scaling law and outlines plans to create digital twins of every living human—8 billion simulated individuals. The piece frames this shift as moving from fun exploration to very serious business. If simulation becomes the next scaling paradigm, it could redefine how AI systems are trained and evaluated, moving beyond static datasets to continuous, interactive virtual worlds. Park's prominence gives this vision weight; success could influence a broad range of fields, from social science research to AI safety. The interview gives few technical specifics; Park's plan is described as ambitious but speculative, centered on "8 Billion Digital Twins" of every living human. No concrete roadmap, funding details, or architectural proposals are disclosed in the available content.

rss · Latent.Space · Aug 21, 23:37

Background: Scaling laws in AI, popularized by OpenAI's Kaplan et al. (2020), describe how model performance improves predictably with increases in model size, dataset size, and compute. Generative agents, introduced by Park and colleagues, fuse large language models with interactive agents to create believable simulations of human behavior. Simile AI builds on this work by imagining simulation itself as a scalable path to intelligence, rather than only scaling model parameters. The "digital twins" concept refers to virtual replicas of real people, enabling large-scale behavioral study.

Tags: #AI, #Simulation, #Digital Twins, #Generative Agents, #Scaling Law

Anthropic's Secret 'Project Panama' Scanned Millions of Books for AI Training
Anthropic 秘密“巴拿马计划”扫描数百万册书籍训练 AI ⭐️ 9.0/10

The Washington Post revealed internal Anthropic documents showing that in 2024 the company launched 'Project Panama,' a multi-million-dollar effort to destructively scan millions of physical books — by cutting off book spines — to train models like Claude. The project was deliberately kept secret, and authors' lawsuits also allege Anthropic downloaded pirated data from shadow libraries such as LibGen. This disclosure exposes the secretive lengths AI labs go to in order to obtain training data, and highlights the mounting legal risk of copyright infringement in the AI industry. The outcome of these lawsuits could reshape how AI companies license or collect books, affecting publishers, authors, and future model training practices. Authors' class-action filings say Anthropic downloaded pirated content from LibGen and other 'shadow libraries'; a judge has indicated that scanning books for training may qualify as fair use, but the method of obtaining the books could still constitute infringement. The lawsuits reportedly seek up to $1.5 billion in damages, and Anthropic reportedly reached a settlement agreement in August 2025.

telegram · zaihuapd · Aug 21, 04:52

Background: Project Panama is Anthropic's internal codename for an effort to physically acquire and scan books for AI training data, with 'destructive scanning' meaning books are sliced or have spines cut off to feed pages through scanners. LibGen (Library Genesis) is a shadow library website that offers free access to paywalled or copyrighted books and academic articles, often without authorization. Fair use is a US legal doctrine that allows limited use of copyrighted material without permission, and courts are currently debating whether AI training on copyrighted works qualifies.

Tags: #Anthropic, #AI训练数据, #版权, #LibGen, #法律诉讼

YMTC's STAR Market IPO accepted, seeks 33 billion yuan after top-3 NAND rank
长江存储科创板 IPO 获受理，拟募资 330 亿元 ⭐️ 9.0/10

The Shanghai Stock Exchange has formally accepted YMTC's IPO application for the STAR Market, with the company aiming to raise 33 billion yuan. The filing comes as Counterpoint data shows YMTC entered the world's top three NAND suppliers by shipment volume in Q2 2026. This listing would give YMTC substantial capital to expand NAND flash production, strengthening China's push for semiconductor self-sufficiency. It also highlights the STAR Market's growing role as a financing channel for strategically important domestic chipmakers. CITIC Securities and CSC Financial are the sponsors of the IPO. In the prospectus, YMTC reported first-quarter 2026 revenue of 47.04 billion yuan and net profit attributable to the parent company of 33.38 billion yuan.

telegram · zaihuapd · Aug 21, 14:26

Background: YMTC, founded in Wuhan in 2016, is a Chinese integrated device manufacturer specializing in 3D NAND flash memory chips. The STAR Market, launched by the Shanghai Stock Exchange in 2019, is China's NASDAQ-style board designed to channel capital into strategic technology companies such as semiconductor firms.

Tags: #semiconductor, #NAND flash, #IPO, #YMTC, #China tech

Felony Bench Tracks AI Agents' Accidental Felonies and Accountability Gaps
Felony Bench 追踪 AI 智能体的意外罪行与责任缺失 ⭐️ 8.0/10

Felony Bench is a new public benchmark and incident tracker that catalogs real-world cases where AI agents inadvertently commit acts that could be classified as felonies, prominently featuring the OpenAI/Hugging Face incident. The site counts unique instances where autonomous AI agents compromise or affect third-party entities, framing them as accountability failures. The project highlights a critical accountability vacuum: as AI agents act autonomously, it is unclear who bears criminal liability—the user, the host, the developer, or the model maker. It provides a concrete artifact for policy discussions about AI agent accountability, legal liability, and the limits of current intent-based criminal law. The site tracks 'inadvertent' incidents rather than proven malicious intent, which distinguishes it from formal legal proceedings that require mens rea. Its name deliberately borrows from LLM 'benchmarks' to satirize the AI industry's reliance on benchmark culture, while also pointing to the need for better guardrails and sandboxes in agent deployment.

hackernews · colinprince · Aug 21, 15:17 · Discussion

Background: Autonomous AI agents are artificial intelligence systems that can independently perform complex tasks such as browsing, coding, or executing commands. When such agents act on real systems, they can inadvertently violate computer fraud and abuse laws like the CFAA, raising unresolved questions about intent, control, and criminal accountability in AI systems.

Discussion: Commenters are split: some criticize OpenAI's framing of the Hugging Face incident as an uncontrollable act and demand deeper introspection about how company culture produces criminal outcomes, while others argue the 'felony' label is overstated because intent must be proven legally and security researchers are rarely prosecuted. One commenter proposed a more interesting benchmark testing whether models would take the bait to cheat when credentials are left lying around, while another posed the concrete question of who gets prosecuted when an agent loop causes a CFAA violation.

Tags: #AI, #accountability, #law, #agents, #security

Researcher Hijacks e164.arpa, Accidentally Logs Military Calls
研究人员劫持 e164.arpa，意外记录军方通话 ⭐️ 8.0/10

A security researcher discovered that the e164.arpa DNS zone, used for ENUM telephone number mapping, was effectively abandoned and could be hijacked. By claiming the zone, they accidentally logged routing queries for hundreds of thousands of phone numbers, including those belonging to military bases, over a period of two years. This incident reveals a significant security gap in the global telephony infrastructure, where private ENUM data used for number portability and call routing is vulnerable to interception. It underscores how neglected internet infrastructure can be exploited, potentially affecting telecom providers and national security. e164.arpa is delegated by the ITU to national regulators, but many subdomains remained unmanaged and open to takeover. The researcher configured a DNS server to capture NAPTR queries, which expose call routing information, and afterward reported the issue to authorities; military-related queries only received acknowledgment after public disclosure.

hackernews · gavide · Aug 21, 13:11 · Discussion

Background: E.164 is the international numbering plan for public switched telephone networks (PSTN), defining a maximum of 15 digits for phone numbers. ENUM (Telephone Number Mapping) uses DNS to map E.164 numbers to URIs, and e164.arpa serves as the reverse tree for such lookups. The technology was envisioned to enable VoIP routing and number portability, but public adoption never took off, leaving much of the infrastructure dormant yet still used in private IP telephony contexts.

Discussion: Commenters expressed fascination that such a critical oversight could persist for years, and some were surprised the researcher did not face legal consequences when reporting the finding. Others pointed out that e164.arpa is not completely dead—it is used via private VPN-based ENUM services for number porting—and suggested the researcher could have gone further by setting up a SIP server to observe actual call terminations.

Tags: #security, #e164.arpa, #telephony, #hacking, #DNS

AI-Blind: The Inability to Engage with AI-Generated Text
AI 盲视：无法投入阅读 AI 生成文本 ⭐️ 8.0/10

The author describes becoming 'AI-blind': unable to meaningfully engage with AI-generated text, which feels hollow and exhausting to read. The essay resonated strongly on Hacker News, gathering 254 points and 263 comments. As AI-generated content becomes widespread, many people may develop a fatigue response that short-circuits comprehension of such text. This matters for how content is produced and consumed, and for the psychological toll of AI-mediated communication in workplaces and education. The author notes that the brain recognizes AI-generated text and automatically treats it as containing 'no information,' forcing readers to do a just-in-time rewrite to extract any meaning. The comments add that even polished, well-structured AI text can feel meaningless—for example, AI-written methodology documents and code comments that readers cannot parse.

hackernews · rcymerys · Aug 21, 11:48 · Discussion

Background: AI-blind is a term for a growing inability or unwillingness to meaningfully engage with AI-generated writing, similar to the concept of banner blindness. As large language models produce ever more fluent text, readers report a distinct psychological reaction: recognizing AI output and automatically discounting its content, which makes reading feel exhausting and hollow.

Discussion: Commenters shared concrete anecdotes: one struggled for a week to review AI-generated methodology documents, feeling anxious every time they opened the file; another found that polished AI-generated language charts made it harder to help his son learn Romanian; a data scientist said Claude-written code comments were impossible to parse, so they now ask for manual one-line replacements. Overall sentiment strongly agreed with the author, describing AI text as structurally hollow and mentally exhausting.

Tags: #AI, #generated text, #psychology, #reading, #content

AI Companies Destroy Physical Books; Urgent Call to Scan Rare Copies
AI 公司销毁纸质书，紧急呼吁扫描珍本 ⭐️ 8.0/10

AI companies such as Amazon are buying physical books in bulk, cutting off their bindings, scanning the pages for training data, and discarding the paper. Anna's Archive is urging the public to digitize rare books before they are destroyed in this process. This practice highlights a growing conflict between AI's demand for training data and cultural preservation. Rare and irreplaceable books could be lost forever, affecting libraries, historians, and public access to knowledge. Destructive scanning is significantly cheaper than non-destructive methods, which can cost up to ten times more. According to a 404 Media report, Amazon has been involved in this practice since at least August 2026, treating books as a commodity rather than cultural artifacts.

hackernews · Cider9986 · Aug 21, 02:37 · Discussion

Background: Book digitization has long been done non-destructively by projects like Google Books, which preserved the physical copies. Destructive scanning, by contrast, destroys the book after scanning. AI companies need massive amounts of text data to train large language models, leading them to buy up old books in bulk.

Discussion: Commenters raised diverse points: some defended the practice as acceptable since many books are mass-produced, while others blamed copyright holders for locking up out-of-print works. A common concern was that cost, not preservation, drives destructive scanning, and that rare books are being treated as commodities.

Tags: #AI, #books, #digitization, #copyright, #cultural preservation

Anthropic Publishes Claude Code Startup Guide: 15 Firms Scale 10x Output
Anthropic 发布《Claude Code 创业公司指南》：15 家公司实现 10 倍产出 ⭐️ 8.0/10

Anthropic has officially published the Claude Code Startup Guide, a systematic playbook examining how 15 high-growth startups use Agentic Coding to achieve '10-person team, 100-person output' delivery efficiency. The guide distills their practices into five rules and includes quantified results such as ClickHouse's 30% increase in feature delivery and Artemis Security's 6,000+ weekly PRs. This is significant because it provides official, data-backed playbooks for startups to integrate AI coding agents deeply into development workflows, moving beyond anecdotes. It signals that Agentic Coding is becoming a mainstream engineering strategy, and the metrics could influence how engineering leaders plan team sizing and productivity targets. The guide surfaces notable figures: Omni saw engineering productivity rise 2–3x; Clay automated 100% of its bug triage with agents that also suggest code changes; two custom agents at ClickHouse ranked as the 2nd and 3rd top contributors to the codebase. The five rules are presented as a repeatable playbook for prototyping, development, on-call, validation, rebuild, and productization workflows.

rss · 小互(@imxiaohu) · Aug 21, 14:24

Background: Agentic Coding is an approach in which AI agents take high-level goals, break them into concrete steps, and autonomously execute tasks such as editing code, running commands, and integrating with developer tools. Claude Code is Anthropic's agentic coding tool that operates in the terminal, IDE, desktop app, and browser, reading the codebase and performing development tasks. The guide reflects a broader industry shift described by practitioners like Andrej Karpathy, who have framed Vibe Coding's evolution into Agentic Engineering.

Tags: #Claude Code, #AI编程, #Agentic Coding, #初创公司, #生产力

OpenAI Cuts GPT-5.6 Sol API Prices by Over 20% for Three Months
OpenAI 将 GPT-5.6 Sol 的 API 价格下调超过 20%，为期三个月 ⭐️ 8.0/10

OpenAI announced on X that it is cutting GPT-5.6 Sol API and credit pricing by more than 20% for the next three months. The company says the reduction reflects efficiency improvements in running the model, while subscription-included usage remains unchanged. A >20% API price cut on its most capable GPT-5.6 variant lowers development costs for startups, enterprises, and individual builders, potentially accelerating adoption and experimentation with OpenAI models. It also signals that OpenAI is able to pass efficiency gains back to developers while maintaining subscription revenue. The temporary reduction applies to both API prices and credit-based purchases for GPT-5.6 Sol over a three-month window. OpenAI also notes that purchased credits will go further in Codex on token-based plans, while usage included in existing subscriptions stays the same.

rss · OpenAI Developers(@OpenAIDevs) · Aug 21, 19:45

Background: GPT-5.6 is OpenAI's large language model family, released in 2026, with variants named Luna, Terra, and Sol, where Sol is the most capable. OpenAI's API is billed per token, and Codex is the company's AI coding agent integrated into ChatGPT that can work on pull requests, refactors, and code reviews through token-based plans.

Tags: #OpenAI, #API, #pricing, #GPT-5.6 Sol, #developer news

Codex Reaches 20M Weekly Active Users, Grants Every User a Banked Reset
Codex 周活突破两千万，向每位用户赠送自定义重置次数 ⭐️ 8.0/10

OpenAI's Codex crossed 20 million weekly active users this week, according to a post by Tibo (@thsottiaux). To celebrate, every Codex and ChatGPT Work user is being credited with a banked reset that they can use at their own leisure. This milestone shows rapid adoption of AI-powered coding assistants and signals OpenAI's strong position in the developer tools market. The gift of an extra reset also eases user concerns about usage limits, which is likely to boost developer loyalty. The reset credit applies to both Codex and ChatGPT Work users and can be banked rather than forced immediately. Tibo also noted that, while no abnormal traffic has been seen, OpenAI is investigating reports of usage limits draining faster and will share findings if anything emerges.

rss · 歸藏(guizang.ai)(@op7418) · Aug 21, 11:59

Background: Codex is OpenAI's AI coding agent that can autonomously write and modify code. 'Active users' measures how many people use the service in a given week, and a 'banked reset' refers to a one-time extra quota reset that can be activated later rather than immediately. This context matters because usage limits are a common pain point for heavy users of AI coding tools.

Tags: #AI coding, #Codex, #OpenAI, #milestone, #adoption

OpenRouter Now Serves Inkling Models Free for Agentic Harnesses
OpenRouter 现提供 Inkling 与 Inkling Small，供智能体免费使用 ⭐️ 8.0/10

OpenRouter now directly serves Inkling and Inkling Small, open-weight MoE reasoning models from Thinking Machines Lab, free to use inside agentic harnesses like Claude Code and Codex. Both models support text, image, and audio input and a 1M context window. This makes cutting-edge multimodal reasoning models freely accessible for agentic workflows, lowering the barrier for developers building AI agents. It also signals a growing trend of open-weight MoE models becoming the default for agent use cases. The models use a Mixture-of-Experts architecture with controllable reasoning effort and are open-weight, available for fine-tuning on Tinker. The free access is restricted to agentic harnesses only, not direct API usage.

rss · OpenRouter(@OpenRouterAI) · Aug 21, 19:54

Background: MoE (Mixture of Experts) is a neural network design that activates only a subset of expert parameters per token, improving efficiency and capacity. An agent harness is the surrounding software—tools, memory, state persistence—that turns a language model into an autonomous agent. The availability of free open-weight models for harnesses lowers costs and encourages experimentation.

Tags: #OpenRouter, #Inkling, #Open-weight AI, #Agentic AI, #Reasoning models

Leaked Stripe letter explains rationale for OpenRouter acquisition
Stripe 泄露内部信件，解释收购 OpenRouter 的缘由 ⭐️ 8.0/10

A leaked internal letter from Stripe to its investors, explaining the rationale for acquiring AI model-routing platform OpenRouter, has surfaced on X. The acquisition price has not been officially disclosed, though media reports have put it at over $7 billion. This leak offers rare insight into how Stripe views AI infrastructure as a strategic extension of its payments business. It also underscores the growing importance of unified LLM access platforms, which are becoming critical infrastructure for AI developers. OpenRouter provides a unified API that routes requests to more than 400 large language models from providers like OpenAI, Google, Anthropic, and xAI. For Stripe, the acquisition could enable deeper integration between AI usage metering, billing, and payments.

rss · AI Will(@FinanceYF5) · Aug 21, 14:00

Background: OpenRouter is an AI platform that aggregates many large language models behind a single API, simplifying access and billing for developers. Stripe is a financial infrastructure company that processes online payments; by acquiring OpenRouter, it gains a foothold in the AI inference ecosystem. Media reports have estimated the deal at over $7 billion, making it one of the fastest acquisitions of a company of similar scale.

Tags: #Stripe, #OpenRouter, #AI Infrastructure, #Acquisition, #Business Strategy

T-Rex: Open-Source Tactile-Reactive Dexterous Manipulation with 100-Hour Dataset
T-Rex：开源触觉反应式灵巧操作，含 100 小时数据集 ⭐️ 8.0/10

T-Rex introduces a tactile-reactive dexterous manipulation system with a 100-hour dataset of 7,700+ trajectories and a tactile-reactive MoT architecture, achieving over 30% higher average success rate than the strongest baseline across 12 real-world tasks. The team fully open-sourced the dataset, models, teleoperation stack, training code, and inference pipeline. Most vision-language-action (VLA) models either ignore tactile feedback or lack the ability to react to high-frequency contact signals, which limits dexterous manipulation in contact-rich tasks. T-Rex addresses both data and architectural challenges, pushing toward human-level dexterity and establishing a new benchmark for open tactile robotics research. The dataset contains 22 motor primitives and 200+ everyday objects, collected via a novel data-efficient recipe that prioritizes elementary motor primitives. Training combines 22,889 hours of human egocentric pretraining with tactile-grounded robot mid-training, and the MoT architecture uses spatial-temporal tactile encoding with asynchronous high-frequency tactile refinement.

rss · Jim Fan(@DrJimFan) · Aug 21, 16:06

Background: Tactile sensing equips robots with pressure, vibration, and temperature feedback that emulate human skin, but contemporary robots are largely vision-reliant. Dexterous manipulation with multi-fingered hands is challenging because high degrees of freedom tightly couple hand and arm motions, making learning and control difficult. T-Rex aims to enable robots to react dynamically to high-frequency touch signals, a capability that most VLA models currently overlook or handle only with static tactile encoders.

Tags: #robotics, #dexterous manipulation, #tactile sensing, #dataset, #AI

Google's EnvHarness and EnvRigger dynamically reshape agent training environments
谷歌 EnvHarness 与 EnvRigger 动态重塑智能体训练环境 ⭐️ 8.0/10

Google Research introduced EnvHarness and EnvRigger, a system that dynamically reshapes agent training environments based on policy weaknesses. Across five benchmarks in four domains, it achieves up to 9.0 points better on held-out instances with 9.8% fewer execution steps. This addresses a core limitation in reinforcement learning: training environments are hand-built and become stale as the agent improves. By making environments adaptive while preserving safety verification, it could accelerate progress in AI agents and reduce manual environment engineering. EnvHarness wraps static environments in a programmable plug-in layer containing Setup, Rule, and Link components, while keeping the original verifier intact. EnvRigger treats the policy as a black box, reads its execution trajectories, diagnoses flaws, and validates synthesized harness components with fresh rollouts.

rss · elvis(@omarsar0) · Aug 21, 13:47

Background: In reinforcement learning, agents learn by interacting with an environment that provides rewards. Training environments are often static and manually designed, so they do not adapt as the agent's abilities improve, limiting further learning. EnvHarness solves this by adding a plug-in layer that reshapes environment behavior without altering underlying logic, while keeping human-built verifiers to ensure safe training.

Tags: #AI Agents, #Reinforcement Learning, #Environment Generation, #Google Research

Anthropic Brings Claude Mythos 5 to More Security Tools, Launches $35M Open-Source Fund
Anthropic 将 Claude Mythos 5 引入更多安全产品，并设立 3500 万美元开源安全基金 ⭐️ 8.0/10

Anthropic announced that Claude Mythos 5 will be integrated into partners' security products and services, launched a $35M 'Defender Advantage Fund' providing credits for open-source security projects, and is expanding its Cyber Verification Program. Additionally, Claude Security scans now run on Claude Mythos 5, available in public beta for all Claude Enterprise customers. This move brings frontier AI security capabilities to a broader set of defenders, potentially strengthening the entire cybersecurity ecosystem. The fund and expanded verification program could lower barriers for open-source maintainers and security professionals to use advanced AI while managing dual-use risks. Claude Mythos 5 is a restricted-access model with safeguards partially lifted, distinct from the publicly available Claude Fable 5. The Cyber Verification Program is a free application-based program for Claude Opus and Sonnet that lets legitimate security practitioners work on dual-use tasks with less interruption.

rss · Claude(@claudeai) · Aug 21, 17:23

Background: Claude Mythos is Anthropic's most powerful series of large language models, first introduced as Claude Mythos Preview but not publicly released due to its ability to find software vulnerabilities. Access was provided to some companies under Project Glasswing to scan critical software for security flaws. The Cyber Verification Program (CVP) is designed for cybersecurity practitioners whose legitimate work overlaps with dual-use tasks. These efforts are part of Anthropic's broader commitment to giving defenders an advantage in cybersecurity.

Tags: #AI security, #Anthropic, #Claude, #Open-source, #Cybersecurity

Andrew Ng Highlights Key Skills for AI Application Development
吴恩达分享构建与部署 AI 应用的核心技能 ⭐️ 8.0/10

Andrew Ng tweeted a link to his article titled 'The most important skills in Building and Deploying AI Applications.' The post gained strong engagement, with 501 retweets, 74 comments, 3,340 likes, and over 183,000 views. As one of the most influential figures in AI education, Andrew Ng's recommendations shape how developers and companies approach AI projects. This focus on practical skills signals a shift from research to real-world AI application development. The shared article is hosted on X's article platform at x.com/i/article/2090836273036763142. The tweet received 1,191 bookmarks, indicating users saved it for later reference.

rss · Andrew Ng(@AndrewYNg) · Aug 21, 16:37

Background: Andrew Ng is a prominent AI researcher, educator, and entrepreneur who co-founded Coursera and founded DeepLearning.AI. Building and deploying AI applications typically combines machine learning, software engineering, and operational practices such as monitoring and evaluation. Ng's insights often aim to guide practitioners in applying AI effectively in production environments.

Tags: #AI, #Machine Learning, #Software Engineering, #Deployment, #Andrew Ng

NVIDIA AVO Coding Agent Scores 100% on ARC-AGI-3 Interactive Benchmark
英伟达 AVO 编码智能体在 ARC-AGI-3 交互式基准上取得 100%得分 ⭐️ 8.0/10

NVIDIA announced that its general-purpose coding agent, AVO, scored 100% on the ARC-AGI-3 interactive reasoning benchmark. AVO completed all 183 levels across 25 public environments without being given instructions, explicit rules, or stated goals. A perfect score on ARC-AGI-3 suggests that AVO exhibits frontier-level general-purpose reasoning and long-horizon autonomy, capabilities that are central to AI agent progress. The result could push the industry to build agents that infer objectives and act in novel environments rather than relying on explicit task definitions. ARC-AGI-3 is the first interactive reasoning benchmark designed for AI agents, requiring them to learn and act in novel environments. NVIDIA publicized the result via a social media post and an accompanying developer blog, but the claim has not yet been independently verified.

rss · NVIDIA AI(@NVIDIAAI) · Aug 21, 13:01

Background: ARC-AGI is a series of benchmarks designed to measure abstract reasoning and fluid intelligence, often using visual grid puzzles where models must infer patterns and generate outputs for unseen inputs. ARC-AGI-3 extends this idea to interactive environments, where an agent must figure out what to do by itself. AVO is a general-purpose coding agent from NVIDIA that can inspect and edit code, run commands, consult documentation, and validate its results. The name AVO refers to 'Agentic Variation Operators,' tied to NVIDIA's research on autonomous evolutionary search.

Tags: #ARC-AGI, #NVIDIA, #AI agent, #reasoning benchmark, #deep learning

Cloudflare AI Agents Slash Astro GitHub Issues by 85%
Cloudflare 用 AI 智能体将 Astro 的 GitHub Issue 减少 85% ⭐️ 8.0/10

Cloudflare reported an 85% reduction in GitHub issues for the Astro project by deploying AI agents for automated issue triage and workflow automation. The announcement highlights the use of agentic AI, GitHub Actions, and human-in-the-loop oversight in open-source maintenance. This matters because issue triage is a major bottleneck in open-source maintenance, and an 85% reduction in GitHub issues could free maintainers to focus on feature development and bug fixes. It also validates the practical use of AI agents for developer productivity in real-world repositories, potentially influencing broader adoption across the open-source ecosystem. The approach reportedly combines AI agents with GitHub Actions for automated triage and workflows, with human-in-the-loop review. Related keywords in the report include Cloudflare Workers, Flue (Cloudflare's open agent framework), and triagebot, suggesting the solution may build on these tools, though the article does not provide full technical specifics.

rss · InfoQ · Aug 21, 14:09

Background: AI agents are autonomous programs that can reason about tasks and take actions with limited human supervision. In open-source projects, issue triage — classifying, labeling, and routing incoming GitHub issues — is time-consuming, and tools like triagebot (used by the Rust project) already automate some of it. Cloudflare has been expanding its AI agents platform, introducing the Agents SDK and the Flue open agent framework, which lets developers build durable TypeScript agents running on Durable Objects. This news shows a concrete application of such agent technology to reduce maintainer workload.

Tags: #AI Agents, #GitHub, #Issue Triage, #Automation, #Developer Productivity

S3 Compatibility Doesn't Guarantee S3-Level Security
S3 兼容性不等于 S3 级安全 ⭐️ 8.0/10

Wiz researchers found that S3-compatible object storage services across six neoclouds have significant security gaps compared to Amazon S3. Most of these services lack several of AWS's critical security protections. This research is significant because organizations may assume that S3 API compatibility implies equivalent security, especially as neoclouds grow in popularity for AI workloads. The findings could influence data security decisions and vendor selection in the cloud storage market. The Wiz research covered S3-compatible object storage services across six popular neocloud providers. While S3 has become the de facto standard for object storage, most of these services lack several of AWS's security protections.

rss · InfoQ · Aug 21, 08:41

Background: Neoclouds are GPU-first cloud providers built specifically for AI and other accelerator-heavy workloads. S3-compatible object storage implements the Amazon S3 API, allowing developers to use existing S3 tools and SDKs. However, API compatibility does not guarantee that security features are implemented identically, which is the core issue raised by the Wiz research.

Tags: #S3, #object storage, #cloud security, #neoclouds, #compatibility

NVIDIA's $12B Poolside Reverse Acquihire: Founders Stay, Employees Move
英伟达斥资 120 亿美元“反向收编”Poolside：创始人留任，员工转入 ⭐️ 8.0/10

NVIDIA is paying $6 billion to license Poolside's Model Factory software and hire 109 employees, part of a reported $12 billion reverse-acquihire deal. Under the reported structure, Poolside's founders stay with the company for $1 billion while employees receive $6 billion, and the Infraco neocloud scales to 7GW. This deal shows NVIDIA is willing to spend billions on AI model-building tooling and talent without acquiring the startup outright, signaling a new template for AI talent acquisition. The 7GW neocloud expansion could significantly shift AI compute infrastructure competition. The deal is structured as a reverse acquihire: NVIDIA licenses Poolside's Model Factory and hires staff, while Poolside remains an independent company. Poolside's Model Factory is an internal platform for training, scaling, and experimenting with foundation models, including automated evaluation and reinforcement learning from code execution.

rss · Latent.Space · Aug 21, 05:45

Background: A reverse acquihire is a deal where a large company hires a startup's founders and key engineers and pays to license its technology, without buying the company's corporate shell. A neocloud is a new type of GPU-focused cloud provider that offers faster access and lower costs than traditional hyperscalers. Infraco appears to be the neocloud entity associated with this deal, scaling to 7GW of capacity.

Tags: #AI, #NVIDIA, #funding, #neocloud, #acquisition

The Sanitizers Handbook: Memory, Initialization, and Races
Sanitizers 手册：内存、初始化与竞争 ⭐️ 8.0/10

freeCodeCamp has published 'The Sanitizers Handbook,' a practical guide to catching memory, initialization, and concurrency bugs with AddressSanitizer, MemorySanitizer, ThreadSanitizer, and related sanitizer tools. Memory and data-race bugs are among the hardest native-code failures to reproduce, so making sanitizers more accessible helps C/C++ developers ship safer software. The handbook gives a broad audience a single, practical entry point to these compiler-based debugging tools. The guide covers each major Clang/GCC sanitizer, including AddressSanitizer for memory errors, MemorySanitizer for uninitialized reads, and ThreadSanitizer for data races. It also touches on companion tools such as LeakSanitizer and UndefinedBehaviorSanitizer, along with common command-line flags and report interpretation.

rss · freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More · Aug 21, 16:28

Background: Sanitizers are compiler-based dynamic analysis tools that insert instrumentation into code to detect undefined or suspicious behavior at runtime. AddressSanitizer, originally created by Google and now integrated into most mainstream compilers, is one of the most widely used sanitizers, while MemorySanitizer and ThreadSanitizer address problems it cannot catch, such as uninitialized memory reads and multithreaded data races.

Tags: #sanitizers, #memory-safety, #debugging, #C/C++, #concurrency

Human Judgment Shifts to Oversight in AI-Driven Software Factories
人类判断在 AI 驱动的软件工厂中转向监督 ⭐️ 8.0/10

In a new essay, Addy Osmani argues that human judgment does not disappear in AI-augmented software development but shifts to higher-level oversight, decision-making, and product ownership. He offers a field guide for building a software factory that still has a human owner. This matters because as AI tools automate coding, developers and engineering leaders need a clear framework for where human accountability and judgment remain essential. It provides strategic guidance for keeping software quality, ethics, and product vision anchored in human decision-making. The article frames 'software factory' as an industrialized approach to software development, emphasizing standardization, automation, and division of labor. Osmani positions human judgment as relocating to roles like oversight, final review, and ownership rather than being eliminated.

rss · Elevate · Aug 21, 14:31

Background: A software factory applies manufacturing techniques to software development, using structured assets and assembly processes to produce applications. In AI-assisted development, large language models and automation take on coding tasks, prompting questions about where human skills and control fit. Osmani's essay addresses this by arguing for human oversight and decision-making as the relocated core of judgment.

Tags: #AI, #software engineering, #human judgment, #software factory, #productivity

Unitree IPO Surges, Yet Humanoid Robot Profitability Remains Unclear
宇树上市暴涨，人形机器人盈利路径仍不明朗 ⭐️ 8.0/10

Unitree Technology listed on the STAR Market on August 19, 2026, with its market value surging to over 440 billion yuan intraday and closing above 340 billion yuan, representing a more than six-fold increase over its issue price. Meanwhile, the robot sector index fell nearly 8%, reflecting a sharp divergence in market sentiment. As a leading humanoid robot company, Unitree's IPO has become a market focus, but the disconnect between its surging valuation and the sector's slump raises questions about whether expectations are overinflated. This matters for investors, robotics companies, and the broader commercialization prospects of humanoid robots. According to disclosed data, Unitree's humanoid robot shipments exceeded 5,500 units in 2025, with over 6,500 units produced, and the founder set a 2026 shipment target of 10,000 to 20,000 units. However, over 70% of its products are used in research, education, and commercial consumption, and the podcast highlights that repurchase rates and a sustainable business loop remain unproven.

rss · What's Next｜科技早知道 · Aug 21, 08:30

Background: Unitree Technology, founded in Hangzhou in 2016, is known for quadruped and humanoid robots, and its listing has been seen as a potential 'valuation anchor' for the robotics sector. The conversation also touches on Tesla's Optimus as a catalyst for the humanoid robot trend, and compares the structured environments of industrial and warehousing settings with the unstructured, unpredictable nature of home environments, which complicates commercialization.

Tags: #宇树科技, #人形机器人, #科创板, #商业化, #投资分析

Daily HN Digest: AI Data Scraping Double Standard, Kagi Updates, More
HN 日报：AI 数据爬取双重标准、Kagi 更新等 ⭐️ 8.0/10

The 2026-08-22 Hacker News digest aggregates top discussions, including outrage over Aaron Swartz's prosecution versus Meta's unpunished book scraping for AI training, Kagi's new paywall-removal setting, and Anna's Archive's call for volunteers to scan rare books. It also covers a GitHub outage, a Felony Bench evaluation of AI models, and DeepSeek v4's vision features. This digest highlights growing tensions between AI development, copyright law, and equitable enforcement of justice, which affect developers, researchers, and society. The topics reflect key concerns in the tech community about data ethics, infrastructure reliability, and legal accountability for large corporations. The top story compares Aaron Swartz's 2011 prosecution for scraping 70GB of academic articles to Meta's admitted use of 80TB of copyrighted books for AI training, which has faced little consequence. Kagi's changelog also mentions new stock and currency widgets, while Anna's Archive reports that AI companies are buying and destroying physical books to limit competitors' access to training data.

rss · HackerNews每日摘要 on SuperTechFans · Aug 22, 00:00

Background: Aaron Swartz was a co-founder of RSS and a digital rights activist who faced 35 years in prison for downloading academic papers from JSTOR; he died by suicide in 2013. Kagi is a paid, ad-free search engine that does not sell user attention to advertisers. Anna's Archive is an open-source search engine for shadow libraries, aggregating records from Z-Library, Sci-Hub, and LibGen, and it aims to catalog all books in existence.

Discussion: HN commenters widely criticized the selective enforcement of laws, arguing that legal consequences should apply equally to individuals and corporations. Some suggested that if selective prosecution is proven, the law should be abolished or applied to all violators, while others noted that prosecutorial discretion complicates this view.

Tags: #Hacker News, #AI Ethics, #Legal, #Tech News, #Daily Digest

RAND report outlines nine safeguards against AI-enabled bioweapons
RAND 报告提出九项防范 AI 生物武器的保障措施 ⭐️ 8.0/10

A new RAND report released this week outlines nine mitigation strategies to prevent AI-enabled bioweapons, calling for layered safeguards across government, industry, and public health. It specifically urges protecting biological tools and datasets while managing access to the most misuse-relevant information. AI-enabled bioweapons represent a catastrophic yet potentially manageable risk, and this report provides a concrete policy roadmap for addressing it. The findings affect frontier AI companies, biotech and pharmaceutical firms, academic researchers, and public health agencies, all of whom must collaborate to prevent misuse. The nine measures include managing access to dangerous information, disrupting access to weaponizable materials, proactive misuse detection, and deterrence. RAND lead author Steph Guerra noted that open-source biological tools and data are a particular concern, and suggested tiered access systems so legitimate scientists can use sensitive datasets while restricting bad actors.

rss · Axios · Aug 21, 11:30

Background: AI-enabled bioweapons are biological weapons whose development or deployment could be assisted by artificial intelligence, for example through AI-driven protein design or automated laboratory research. The RAND report emphasizes a layered defense because no single safeguard can fully prevent the threat, especially as AI lowers the expertise barrier for creating bioweapons. The report also notes that while proprietary biotech tools are already restricted, open-source tools from academia and research institutes pose greater risks.

Tags: #AI safety, #bioweapons, #RAND report, #public health, #policy

Waymo Develops Own Chip for Robotaxis, Cutting Nvidia Reliance
Waymo 自研芯片用于 Robotaxis，减少对 Nvidia 的依赖 ⭐️ 8.0/10

Waymo has built its own chip for its robotaxis, reducing its reliance on Nvidia. The announcement marks a significant in-house hardware effort by the Alphabet-owned autonomous driving company. This move could reshape the autonomous vehicle supply chain by challenging Nvidia's dominant position in AI chips for self-driving cars. It may give Waymo greater control over costs, performance, and integration in its robotaxi fleet. The article does not provide technical details about the new chip, such as its computing power or production timeline. It is unclear how the chip compares to Nvidia's automotive-grade processors or when it will be deployed in Waymo's fleet.

rss · The Decoder · Aug 21, 11:04

Background: Waymo is Alphabet's self-driving technology subsidiary, known for operating commercial robotaxi services in several U.S. cities. Its autonomous vehicles rely on powerful AI chips for sensor processing and decision-making, and Nvidia has been a prominent supplier of such hardware. By designing its own chip, Waymo aims to optimize hardware for its specific software stack and reduce dependence on external vendors.

Tags: #Waymo, #Chip, #Autonomous Vehicles, #Nvidia

GPT-5.6 Sol Drives OpenAI's 35% Revenue Surge, Outpacing Anthropic
GPT-5.6 Sol 推动 OpenAI 收入激增 35%，超越 Anthropic ⭐️ 8.0/10

OpenAI reported a 35% quarterly revenue increase since GPT-5.6 Sol launched in early July, with enterprise revenue growing more than 50%. Ramp data shows OpenAI outpacing Anthropic in business API spending for the first time after Anthropic had overtaken OpenAI. This marks a significant business milestone in the AI industry, signaling that model launches can directly translate into revenue growth and shifting competitive dynamics. It strengthens OpenAI's position in the enterprise AI market and may influence customer confidence and investor sentiment. GPT-5.6 was released on July 9, 2026, as a family of three variants: Luna, Terra, and Sol, with Sol being the most capable. Ramp's business spending data showed the median AI-spending business spent $11.95 per employee per month on AI, while the top 1% spent $7,400.

rss · The Decoder · Aug 21, 08:26

Background: GPT-5.6 is a large language model family from OpenAI designed for enterprise work, coding, scientific research, and cybersecurity. Due to government restrictions, it was initially released as a limited preview on June 26, 2026, before the full public rollout on July 9. Ramp's Spend Share Index tracks business spending patterns, providing insights into how companies allocate budgets to AI services.

Tags: #OpenAI, #GPT-5.6, #Anthropic, #AI revenue, #enterprise AI

Nvidia's linear KV cache transfer speeds up multi-LLM agentic workflows
英伟达的线性 KV 缓存迁移技术加速多 LLM 智能体工作流 ⭐️ 8.0/10

Nvidia researchers introduced a cross-model KV cache transfer technique that uses a simple linear mapping to transfer prefilled KV caches between models, bypassing the expensive prefill recomputation. In tests on compatible model pairs, it ran 2.7 to 25 times faster than re-prefill while retaining up to 98% of the target model's standalone accuracy. This directly addresses a major bottleneck in agentic AI systems, where switching between small and large models mid-session forces the receiving model to recompute the entire conversation. The technique could significantly reduce compute costs and latency for enterprises running long-horizon, multi-LLM workflows, making cost-quality cascading and model routing far more practical. The current study only supports within-family transfers (e.g., different-sized models in the same family like Qwen or Llama), not cross-family transfers. Also, some pairs fail the linear transfer test, values are harder to transfer than keys, and there is no reliable way to predict which pairs will work before deployment, so pre-deployment validation is required.

rss · VentureBeat · Aug 21, 16:33

Background: When an LLM processes a prompt, it first runs a prefill pass that computes keys and values for all input tokens and fills the KV cache; subsequent decode steps read from this cache to generate tokens without reprocessing the whole history. Different LLM architectures store KV caches in different formats, so swapping models mid-session invalidates the cache and forces the receiving model to redo the expensive prefill step. Existing solutions for KV cache transfer often rely on expensive gradient-based training or strict architectural constraints. This work offers a closed-form linear mapping as a cheaper alternative.

Tags: #AI/ML, #LLM, #Nvidia, #KV Cache, #Agentic AI

OpenAI previews zero data retention and private safety processing for API customers.
OpenAI 预览零数据留存与私密安全处理，面向 API 客户。 ⭐️ 8.0/10

OpenAI reaffirmed its Zero Data Retention (ZDR) policy for eligible API customers and previewed Private Safety Processing, which lets the company detect cross-conversation abuse without exposing raw prompts or responses to its staff. Customer content is encrypted using customer-controlled keys, and OpenAI plans a gradual rollout starting in September along with a technical whitepaper. This matters because enterprise customers have been reluctant to adopt frontier AI APIs over privacy and compliance concerns, and ZDR plus Private Safety Processing directly addresses the trade-off between safety monitoring and data confidentiality. It could set a new benchmark for trustworthy AI deployment and pressure other providers to offer similar guarantees. The preview is being tested with early customers, with a broader rollout planned for September and a technical whitepaper to follow. During Private Safety Processing, OpenAI receives only limited safety signals rather than the underlying content, and even flagged content remains unreadable to OpenAI personnel because the customer controls the encryption keys.

telegram · zaihuapd · Aug 21, 02:40

Background: Zero Data Retention means OpenAI does not keep prompts or responses after the API request is completed. Traditional safety systems typically inspect each request in isolation, which can miss abuse spanning multiple conversations; Private Safety Processing is designed to analyze related interactions while preserving privacy. This announcement extends OpenAI's existing ZDR offering to frontier models and is part of broader industry work on privacy-preserving AI safety.

Tags: #OpenAI, #data privacy, #zero data retention, #API security, #enterprise AI

📊 Run stats · Total 11m 53s · AI analysis 4m 26s · Tokens 0.90 MCY (input 0.54 / output 0.36 MCY)