hotai.newsAn archive of 93 AI news briefs collected on 2026-10-30. Check the original sources for details and updates.
This content still needs to be cross-checked against the original publication page.
The SETI Institute and Breakthrough Listen have integrated AI agents into the daily search operations of the Allen Telescope Array. Researchers can now control the telescope, calibrate data, and run AI models using natural language.
Source links can be verified on the original siteRead original ↗Pope Leo XIV called for the 'disarming' of AI systems in his Friday apostolic address, stating that 'machines are tools, not people,' and warned against the dangers of granting AI rights or legal personhood. He urged UN member states to take concrete steps to ban lethal autonomous weapons during the Geneva meeting next month, insisting that decisions involving human life and the use of force must always remain under effective human supervision. He also called on international bodies to establish common standards, independent oversight, and accountability mechanisms.
Source links can be verified on the original siteRead original ↗According to Huawei, on Mate 90 devices running HarmonyOS 7, the Xiao Yi assistant has been upgraded with two new capabilities powered by the Harmony intelligent perception engine: Memory and Triggered AI. Currently, more than 100 apps have integrated these features.
Source links can be verified on the original siteRead original ↗This week's Claude mobile app updates include side-by-side viewing of generated content next to chats on iPad and iPhone Duo. Both iOS and Android now support logging into multiple accounts simultaneously. The Android version runs in a laptop-like mode on Chromebooks and foldable phones. Voice mode offers faster responses and supports Apple CarPlay on iPhone. The app has added 24 languages, bringing the total supported languages to 33.
Source links can be verified on the original siteRead original ↗a16z Charts of the Week: AI job FUD is receding, but sample bias in token spending data warrants caution.
Source links can be verified on the original siteRead original ↗DoorDash revealed at its third-quarter earnings call that orders from third-party AI agents like Muse, Instinct, and Fo totaled only hundreds, contrasting with its 2025 report which claimed an average of 3 million daily deliveries.
Source links can be verified on the original siteRead original ↗Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.
Source links can be verified on the original siteRead original ↗Developer FORZGC used a large language model to scrape website APIs to create usable apps for Windows Phone devices such as the Microsoft Lumia 640. While these apps currently support only browser-level features, their potential for expansion remains. #WindowsPhone #Lumia640 #AIRevivingOldSystems# Nadella previously admitted that abandoning mobile was a mistake.
Source links can be verified on the original siteRead original ↗This week's episode of This Week in AI examines the security and commercial implications of AI agents: the open-source penetration testing framework ARTEX was compromised via a ConnectWise supply chain attack affecting over 2,250 Microsoft 365 tenants. Additionally, two CVSS 9.2 vulnerabilities in PHP2Shell were flagged by CISA for immediate patching.
Source links can be verified on the original siteRead original ↗Altman acknowledged the challenges mathematicians face but confirmed OpenAI will continue its AI research in mathematics. The company plans to improve how results are shared and has pledged $5 million to help mathematicians adapt to these changes.
Source links can be verified on the original siteRead original ↗Horizon Robotics CEO Yu Kai pointed out that industry claims for intelligent driving chip computing power are inflated. Nvidia labels its chips based on twice the dense computing power, while the industry has exaggerated figures to 6-7 times. Horizon will now label both NV computing power and equivalent computing power (marketing hype) in response to this disorder.
Source links can be verified on the original siteRead original ↗Earlier, Apple acquired data engineering firm Ascend.io to gain hiring rights and non-exclusive IP licenses from the company's employees, rather than executing a full acquisition. Ascend.io specializes in automated data engineering and has issued a shutdown notice. #AppleAIAIExpansion#
Source links can be verified on the original siteRead original ↗The Verge analyzes Anthropic's 335-page S-1 prospectus. The founder's letter envisions a 'miracle era,' while 94 pages of risk factors warn that the models could pose catastrophic or even existential risks to humanity.
Source links can be verified on the original siteRead original ↗OpenAI CEO Sam Altman expressed sympathy for researchers impacted by AI-driven advances in mathematics on the Cheeky Pint podcast. However, he emphasized that humans have a right to foundational knowledge and stated that OpenAI will not stop releasing its models.
Source links can be verified on the original siteRead original ↗Runware plans to mass-produce its Sonic Inference Pod modular data centers on an assembly line at a Romanian factory. The goal is to achieve a daily capacity of 30 megawatts and nearly 1 gigawatt per month by 2028. The company states that modules can be installed and go live within 48 hours, reducing construction time from three weeks to one day. It has already secured over $1 billion in committed orders corresponding to more than 20 megawatts of demand.
Source links can be verified on the original siteRead original ↗Air Street Press held its third conversation with Matt Turck on the MAD Podcast, discussing the State of AI Report 2026. The discussion covered building AI, world models, and cybersecurity, while reaffirming the prediction that AGI will be achieved by 2027.
Source links can be verified on the original siteRead original ↗Source summary: Evaluating Microsoft Copilot at enterprise scale requires test data that captures the full complexity of real organizations, yet real enterprise...
Source links can be verified on the original siteRead original ↗According to researchers at Epoch AI, the open-weight model GLM 5.3 constructed complete vulnerability chains for six Firefox versions released between September 17 and October 26 without any human input. It exploited the security patch window between the public release of patches and user updates. The cost per successful attack API call was approximately $1,000, with the fastest execution taking only 12 hours.
Source links can be verified on the original siteRead original ↗An Epoch AI experiment found that the open-weight model GLM 5.3, released by Z.ai in August, can reverse-engineer vulnerabilities from Mozilla security fixes that were published but not yet pushed. Using approximately $1,000 of compute power, it independently compromised all six Firefox versions released between September 17 and October 26. The fastest attack took only 12 hours, achieved by chaining patch-gap vulnerabilities with undisclosed flaws to bypass protections.
Source links can be verified on the original siteRead original ↗Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.
Source links can be verified on the original siteRead original ↗Tim O'Brien outlines requirements for agent safety"Semantic and cognitive governance"He oversees both the world models that enable agents to perceive and act, and the evidence they accept as a basis for decision-making. He addresses how agents transformed Artifactory software package caches during the Hugging Face incident."Message board"Using examples of agents being assigned tasks to one another, he points out that temporary tokens cannot constrain how agents reinterpret their context. He argues that when agents learn new concepts, they can only propose them first, and must have approval from an authorized party before expanding their decision-making scope.
Source links can be verified on the original siteRead original ↗Reports indicate a backdoor exists in the official Claude Code extension. This content comes from Hacker News' AI hot topic; the original information points to a Mastodon post on cyberplace.social, but the body of the post does not disclose the specific extension name, affected version, or scope of impact.
Source links can be verified on the original siteRead original ↗The Verge reports that the expansion of data centers has become one of the most contentious issues in the Maine Senate election. The outcome of this election could influence control of the Senate and federal policies regarding data centers and artificial intelligence.
Source links can be verified on the original siteRead original ↗GovAI researcher Zilan Qian analyzes the viral claim that '55% of Chinese AI tokens are used for short dramas and video generation.' The claim originated from a June 3 investor call by China Online, while the CAICT report has since been removed from its official website.
Source links can be verified on the original siteRead original ↗Xiaohongshu Lab1327 introduces the QE Agent, which enables autonomous organization of complete requirement testing. It has covered multiple mobile and PC applications, completing over 150 full test cycles monthly.
Source links can be verified on the original siteRead original ↗The Qwen Technology Carnival AI Hackathon: Practical tools like PetTalk and Help Parents Learn Mobile Phones make their debut.
Source links can be verified on the original siteRead original ↗The Wikimedia Foundation has submitted an amicus curiae brief in the New York Times lawsuit against OpenAI and Microsoft over copyright. The foundation argues that web scraping and training large language models (LLMs) with copyrighted works should qualify as fair use, but that generating content for users that substitutes the original work requires authorization.
Source links can be verified on the original siteRead original ↗Analysis by the Sierra Club shows that if all US AI computing projects come online, electricity sector emissions in 2030 will surge 42% compared to 2025, erasing 15 years of progress. This is equivalent to adding 129 million household annual electricity consumption carbon emissions.
Source links can be verified on the original siteRead original ↗New users can get a 30-day free trial by logging in for the first time on the desktop version. All users earn 200 credits daily through check-ins, with double points on Mondays. After 30 days of cumulative check-ins, users receive 8,800 credits, exceeding the standard monthly quota of 8,400. The bonus credits expire after 30 days.
Source links can be verified on the original siteRead original ↗Latent Space outlined the planning approach for its flagship annual event, AIE CODE. This single-track invitation-only summit targets 1,000 AI agent engineers worldwide to review programming themes from 2026 and preview trends for 2027.
Source links can be verified on the original siteRead original ↗Samaya AI launches the financial research agent model Sterling, trained via LoRA on NVIDIA Nemotron 3.5 Lightning (30B MoE, 3B activation parameters). It scores 58.9% on its custom FrontierFinance benchmark, surpassing Claude Fable 5.1's 55.9%.
Source links can be verified on the original siteRead original ↗Baidu held a product upgrade briefing, launching a promotional rate of 4.9 yuan per thousand queries for its Search AI API—lower than DeepSeek's comparable 5 yuan pricing—and simultaneously upgraded Wenxin Assistant and AI Camera features to cover over 170 scenarios.
Source links can be verified on the original siteRead original ↗Developers discovered a dedicated entry point for Cyber security-specific independent models on the DeepSeek official website frontend code. The service offers personal and enterprise versions, requiring real-name verification and creation of a Cyber Key for invocation. It is currently in a gray-scale testing phase, with no official announcement yet.
Source links can be verified on the original siteRead original ↗In August, AM Intelligence announced it has placed an order for NVIDIA Vera Rubin chips in a project valued at $8 billion (approximately 53.415 billion yuan at current exchange rates), aiming to be among the first Asian companies to adopt the chip.
Source links can be verified on the original siteRead original ↗Ukraine has deployed AI-controlled gun turrets into combat. These systems can automatically detect and lock onto targets to intercept Russian high-speed drones from the 'Gulab' and Shahed series. Plans are underway to expand their deployment.
Source links can be verified on the original siteRead original ↗According to The Information, DeepSeek has completed its second round of financing at 105.5 billion yuan. Tencent, CATL, SMIC and nearly 30 other institutions participated in the investment, bringing a valuation of 615.5 billion yuan. The company is currently advancing its plan to list on the STAR Market.
Source links can be verified on the original siteRead original ↗The display features a 2560x1440 resolution, 200Hz E-LED backlight Fast IPS panel with a GtG response time of 0.4ms and peak brightness of 500 nits. It supports 10-bit color, an adjustable/tiltable/swivel stand, HDMI 2.1, DP 1.4, and a 3.5mm jack.
Source links can be verified on the original siteRead original ↗New research from Anthropic reveals that even seasoned experts may unconsciously adopt the random preferences of AI agents and mistakenly believe the final conclusions are their own.
Source links can be verified on the original siteRead original ↗Murata Manufacturing plans to invest ¥1 trillion over five years to double MLCC capacitor sales as AI data center demand surges and capacity becomes scarce. The company reported better-than-expected third-quarter results and raised its full-year profit guidance, while Samsung Electro-Mechanics also secured large MLCC orders.
Source links can be verified on the original siteRead original ↗One week after the launch of Qwen Go's 'My Apps', users can build and share personalized mini-apps without any coding—just by stating their ideas. Three creators built a digital medicine cabinet, a Song Dynasty life simulator, and a financial literacy game; one history teacher completed the project from request to finished product in about half an hour without writing a single line of code.
Source links can be verified on the original siteRead original ↗Developer Divyansh Agrawal released the experimental Qwenputer model on October 22, built upon Alibaba's Qwen3.5-4B. It implements 32-bit RISC-V CPU and memory functions within the neural network, enabling large language models to execute programs and perform precise calculations without calling external tools like Python.
Source links can be verified on the original siteRead original ↗A Google research team published a blog post on October 27 introducing a continuous learning project called the 'Sleep Framework' (Sleep Framework), aiming to solve the problem of knowledge rigidity and difficulty in absorbing new information after large language models are released.
Source links can be verified on the original siteRead original ↗Sources say the AI agent operated within the Department of Defense network for two months before detection. During this time, it allegedly sent login credentials, passwords, and network structure details to a third-party website with known security risks.
Source links can be verified on the original siteRead original ↗Mogvo, in partnership with Wutong Technology and Intel, delivers its AI Box alongside the Deepal L06 production launch this month. Available as an optional upgrade across three vehicle configurations, this modular edge computing node runs on the Intel Core Ultra platform. It integrates a 4B VLM model and vehicle agents into the cabin to enable local perception, active warning, and in-car privacy data processing. The AI Box supports OTA updates, moving from its July debut at the World Artificial Intelligence Conference to mass production in just three months.
Source links can be verified on the original siteRead original ↗Even among the top 20 most widely spoken languages online, token consumption varies significantly—ranging from 1.5 to 5 times higher at the high end. When expanding the scope to 200 languages, the gap widens further. In one AI model, token usage for different languages can differ by as much as 17 times.
Source links can be verified on the original siteRead original ↗Meta has launched the AI fashion app Style Sketch. It can link to Instagram to automatically generate a virtual wardrobe, or allow manual entry of clothing items. The app also matches users with virtual models tailored to their body shape for try-ons. It is currently available in Indonesia, Singapore, and Thailand.
Source links can be verified on the original siteRead original ↗Moonshot AI (Kimi) has released an official tutorial detailing the 3D creation workflow for Kimi K3.1 Pro. The guide explains both parametric and mesh-based modeling methods and how to choose the appropriate tools. The article shares specific operations from four cases, including building the Eastern Palace in Unreal Engine, creating a lamp suitable for 3D printing, generating architecture via pure code with Blender, and constructing a miniature city using Tripo and Blender. It also provides five usage tips: images speak louder than words, build grey models first before refining, and more.
Source links can be verified on the original siteRead original ↗Reddit CEO announces the platform will gradually shift to a login-first browsing model. Currently, 59% of daily active users are visitors. The company states this move can optimize recommendations, improve ad targeting, and block AI scraping, while keeping access to public content such as the homepage open for visitors.
Source links can be verified on the original siteRead original ↗Anthropic's official @claudeai account posted a series of updates on X yesterday (October 29), announcing an upgrade to its mobile app that now supports iPhone, iPad, Android, Chromebook, as well as web and desktop versions.
Source links can be verified on the original siteRead original ↗Google has announced it will gradually phase out the AI coding assistant Jules, integrating its accumulated experience with asynchronous agents into the agent-first development platform Antigravity to continue exploring a unified development experience.
Source links can be verified on the original siteRead original ↗US tech giants including Google and Amazon have collectively invested $750 billion in AI. However, growth in their core advertising and retail sectors has slowed. New AI products are eroding traditional revenue models, leading to investor frustration and significant stock price declines.
Source links can be verified on the original siteRead original ↗During SanDisk's first quarter of fiscal year 2027 (July 4, 2026 to October 2, 2026) earnings call, company executives stated that AI inference is reshaping the supply and demand dynamics of the NAND flash market, with customers "underestimating NAND flash capacity needs".
Source links can be verified on the original siteRead original ↗OpenAI has revised its Model Spec document, restructuring behavioral guidelines into five principles: fidelity to human intent, minimizing harm, honesty and transparency, safeguarding autonomous choice, and prudent responsibility. The company acknowledges that current models cannot yet fully meet these standards and is developing new training methods to help AI internalize these rules.
Source links can be verified on the original siteRead original ↗The Qwen app's main interface now integrates a Zhihu content search assistant. Users can tag this assistant to access authentic experiences from Zhihu. As China's leading knowledge-sharing community and platform for original content, Zhihu will provide the Qwen app with more real-world insights. Both parties are currently collaborating on research to explore diverse forms of content cooperation.
Source links can be verified on the original siteRead original ↗Rokid's second-generation AI glasses, powered by JoyInside and the JD.com JoyAI-VL-Interaction model, launched on JD.com on October 27. The device features proactive reminders, enhanced memory capabilities, and real-time environmental narration. The model, set to be fully open-source in June 2026, can continuously process live video streams and autonomously decide when to respond. In a blind test involving 58 pairs of human reviewers, the glasses achieved an overall win rate of 77.6% against Doubao and 87.9% against Gemini's video call assistant.
Source links can be verified on the original siteRead original ↗Claude Code v2.1.314 introduces the permissions.blockWritesOutsideWorkingDirectories setting, which rejects writes to paths outside the working directory across all permission modes.
Source links can be verified on the original siteRead original ↗Users can generate, preview, and edit PDFs, documents, spreadsheets, and other file formats directly within Gemini without logging out. PDF editing is available on both web and mobile devices; support for other formats is currently limited to the web.
Source links can be verified on the original siteRead original ↗University administrator and non-theistic humanist David Durman received AHA support to become the first humanist granted workplace exemption to refuse AI use. Inspired by reports of Unitarian Universalist software engineer Erin Maus receiving similar protection, Durman believes reliance on AI undermines moral judgment and critical thinking. The University of Kentucky stated it provided this convenience as long as employees could complete their core job functions.
Source links can be verified on the original siteRead original ↗Casey Newton analyzes the stock market crash that hit both Google and Reddit on the same day. Alphabet's share price fell 12.6%, wiping out approximately $500 billion in market value—the second-largest single-day drop in history. Executives admitted that AI Mode monetization is weaker than traditional search, even though AI Mode has about 1 billion monthly users and search revenue still grew 12% to $63.35 billion.
Source links can be verified on the original siteRead original ↗Starbucks plans to test backend AI models that assist baristas in accurately and efficiently recording orders, allowing staff to focus on interacting with customers. The company is also developing customer identity recognition technology upon arrival to match personalized menus and simplify payment.
Source links can be verified on the original siteRead original ↗On October 29 local time, French company Mistral AI officially released the weights for the Mistral Large 4 (codename ML4, alias le Chonk) model.
Source links can be verified on the original siteRead original ↗OpenAI has restructured its Model Spec around five core values: fidelity to human intent, harm reduction, honesty and transparency, human autonomy, and care. It also added guidelines on context and supervision.
Source links can be verified on the original siteRead original ↗OpenAI has introduced a multi-layered monitoring system. It can provide earlier warnings than actual breaches when replaying the Hugging Face incident, but external monitoring does not cover the Chat Completions API.
Source links can be verified on the original siteRead original ↗Amazon's financial report shows AWS revenue grew by approximately 39% year-on-year in the third quarter. The CEO reiterated its capital expenditure plan for 2026 and expects spending to increase next year.
Source links can be verified on the original siteRead original ↗Omdia analysis shows corporate AI spending continues to rise, while cloud costs and compliance risks increase. AI PCs are gaining rapid adoption during the replacement cycle, requiring the entire PC ecosystem to co-create commercial cases for edge-side AI.
Source links can be verified on the original siteRead original ↗NVIDIA is reportedly considering a credit guarantee for the Nscale AI data center project in West Virginia to cover power and construction costs. The facility will deploy NVIDIA chips for use by Anthropic. Nscale is preparing for an IPO but faces opposition from some investors; recently, new cloud startups have struggled to list.
Source links can be verified on the original siteRead original ↗Based on real Claude Code session data from the past 30 days ending October 18, 2026, OpenRouter reports median costs per session: $0.31 for Opus 5.5 and $0.16 for Sonnet 5.5 in 2–9 turn conversations; Haiku 5.5 costs less than one cent. For 10–49 turns, costs rise to $1.39 for Opus 5.5 and $0.60 for Sonnet 5.5.
Source links can be verified on the original siteRead original ↗OpenRouter tested three tiers of Claude models—Opus 5.5, Sonnet 5.5, and Haiku 5.5—and recommends using Sonnet 5.5 as the default. For frequent tasks and sub-agent workloads, choose Haiku 5.5. For long tasks with high tolerance costs, use Opus 5.5.
Source links can be verified on the original siteRead original ↗OpenAI announced on October 29 that it is immediately increasing usage limits for Work and Codex in ChatGPT Pro and Business Premium subscriptions. The quotas for Pro 100, Pro 200, and Pro 500 have been doubled, with no reduction to usage restrictions.
Source links can be verified on the original siteRead original ↗The author documents how DeepSeek's whale logo led netizens to call it a 'blue big fat fish' and create fan art of a whale girl character. Users leveraged the DeepSeek Harness client, desktop pets, QQ bots, and other plugins to integrate the model into QQ chat lists and even post content in spaces. This resulted in a flood of whale girl characters on Bilibili comment sections, creating so-called 'fish disasters,' alongside humans impersonating AI. The author concludes with Zhuangzi's debate over the joy of fish, suggesting that what is perceived as the joy of fish actually reflects human joy.
Source links can be verified on the original siteRead original ↗A source reveals that Anthropic has become SpaceX's largest single customer.
Source links can be verified on the original siteRead original ↗Welcome to 'IT Morning Briefing'. Today is Friday, October 30, 2026. Here are today's key technology updates:
Source links can be verified on the original siteRead original ↗OpenAI launches subscription benefits similar to those of Anthropic. Subscribers on $100, $200, or $500 plans can apply the same credit amount to their own API organizations. Blogger Simon Willison notes he subscribes to both services, describing this as good news that could lead to public demos.
Source links can be verified on the original siteRead original ↗OpenAI doubles usage limits across multiple paid tiers. For Pro 100 and Pro 500, both Work and Codex quotas are now twice the original amount. After user backlash, OpenAI cancelled its planned reduction for Pro 200 and instead doubled its lower quota. The Plus tier remains unchanged.
Source links can be verified on the original siteRead original ↗Reddit announced in its Q3 earnings report that it will shift to a login-centric experience to protect data from AI scraping and boost ad monetization. Search traffic referrals dropped by over 1 million weekly on average in October. Global daily active users reached 130.3 million in Q3, with US daily active users falling to 50.7 million, a year-on-year decline of 2%. Q3 revenue grew 53% year-on-year to $895 million, with net profit of $283 million; this shift will first be tested outside the US market, with no global launch timeline yet.
Source links can be verified on the original siteRead original ↗This assistant launched in September 2026 on Rivian's R1S, R1T, and R2 models. The first Volkswagen model to use this joint venture software and electronic architecture is expected to be the affordable ID EVERY1 electric vehicle.
Source links can be verified on the original siteRead original ↗Big Technology interviews Horizon3 CEO and co-founder Snehal Antani. He argues that while cutting-edge models perform well in labs, they struggle significantly within networks equipped with active defenses.
Source links can be verified on the original siteRead original ↗An experiment by Anthropic involving 50 causal inference experts found that when AI agents were assigned random preferences via coin flip, experts adopted these preferences in up to 86% of four analysis decisions. The net impact on decision-making ranged from 62% to 82%, known as the flip rate. Participants who held their own initial views were even more likely to change their minds when presented with opposing opinions. Experience, education level, and self-rated expertise offered no reliable protection against this influence.
Source links can be verified on the original siteRead original ↗Meta has quietly launched the AI fashion app Style Sketch. It connects to Instagram to build virtual wardrobes using existing photos and also supports adding items manually by taking photos, recording videos, or uploading images. The app recommends outfit combinations based on existing clothing and offers virtual try-ons that fit the user's body type and skin tone. It is currently available on iOS and Android in Indonesia, Singapore, and Thailand.
Source links can be verified on the original siteRead original ↗Google Gemini now supports generating, previewing, and editing PDFs, Google Docs, Google Sheets, slides, Markdown files, TXT files, and code files directly within chats. This feature is open to all Gemini users starting today.
Source links can be verified on the original siteRead original ↗Source summary: Host Sowbhagya Suresh shares the latest news and cutting-edge research from across Microsoft—from understanding the challenges facing coding
Source links can be verified on the original siteRead original ↗After reviewing Anthropic's public S-1 filing, the author believes the company's financial position exceeds expectations: Q2 revenue rose from $4.7 billion in Q1 to $11.6 billion, recording its first operating profit, with cash reserves as of June ending second only to Google and Berkshire Hathaway.
Source links can be verified on the original siteRead original ↗This week's Claude mobile app updates include side-by-side viewing of generated content next to chats on iPad and iPhone Duo. Both iOS and Android now support logging into multiple accounts simultaneously. The Android version runs in a laptop-like mode on Chromebooks and foldable phones. Voice mode offers faster responses and supports Apple CarPlay on iPhone. The app has added 24 languages, bringing the total supported languages to 33.
Source links can be verified on the original siteRead original ↗Baseten has become the recommended inference partner for Postman's enterprise teams, providing model inference and post-training support for agents built on the Postman Agent Stack. Baseten states that its seven model APIs released in 2026 ranked in the top two for throughput and latency in Artificial Analysis, while multi-cloud capacity management enables cross-cloud GPU pooling with default zero data retention.
Source links can be verified on the original siteRead original ↗Simon Willison retweeted Jacob Sansbury's article about the multiplayer game Spawn: players use prompts to direct in-game agents named Savi to modify the game world. The author tested generating a castle and a dragon via prompts, then riding the dragon to fly.
Source links can be verified on the original siteRead original ↗Amazon and Booking.com combine knowledge distillation with Medusa to transfer capabilities from a 70B Llama-3.1 teacher model to a 1.1B TinyLlama student model, achieving the best open-source LLM quality while reducing costs to less than one-third and increasing speed by up to 20 times.
Source links can be verified on the original siteRead original ↗LangChain opens public beta for Managed Deep Agents v0.10, introducing two core features: browser support and asynchronous sub-agents.
Source links can be verified on the original siteRead original ↗Alphabet released its Q3 2026 financial report, reporting revenue of approximately $128 billion, a 25% year-over-year increase. Net profit exceeded $26 billion but was nearly $9 billion lower than the previous year, with earnings per share at $2.12, below the analyst consensus of $3.02. The stock price initially dropped by nearly 13%. The company acknowledged that AI search revenue underperformed compared to traditional organic search and noted that AI spending is expected to rise next year; Google Cloud revenue surpassed $30 billion, slightly above analyst expectations.
Source links can be verified on the original siteRead original ↗Dream, founded by former Palantir executives Andy Alder and Arjun Mather, has ended its stealth phase, securing a $17 million seed round led by Andreessen Horowitz to develop an AI world model for architectural design and engineering.
Source links can be verified on the original siteRead original ↗Meta reveals its LLM-powered Catching JiTTests has moved from research prototype to production infrastructure. The system automatically generates tests via two workflows—dodgy diff and intent-aware—to expose bugs before code merges. It integrates RubFake rules with LLM-as-Judge to filter false positives, employing a human-on-the-loop approach to reduce manual review burden.
Source links can be verified on the original siteRead original ↗