hotai.newsAn archive of 110 AI news briefs collected on 2026-10-21. Check the original sources for details and updates.
This content still needs to be cross-checked against the original publication page.
Users can tap the Gemini icon on the vehicle screen to ask multi-step voice questions, such as first requesting a recommended location and then asking it to avoid toll roads when planning a route.
Source links can be verified on the original siteRead original ↗Users can generate dynamic lesson plans and 3D presentations with a single click by inputting text or voice, requiring no programming skills. This tool covers multiple scenarios including K-12 education, filling a gap in Tencent's product lineup for this area.
Source links can be verified on the original siteRead original ↗Huawei HarmonyOS 7's Xiaoyi AI shortcut bar feature was originally scheduled to launch this month. IT Home inquired again about the launch timeline with Huawei customer service today. The response indicated: the feature is expected to roll out gradually in early November, with distribution by device model.
Source links can be verified on the original siteRead original ↗Kling AI announced on October 21 that the Kling 4.0 API is open for applications, targeting individual developers and enterprise users. The official stated that Kling 4.0 significantly enhances visual realism, creative controllability, and narrative completeness, bringing new breakthroughs in audiovisual performance, multi-keyframe generation, lip-sync matching, text generation, and creative replication.
Source links can be verified on the original siteRead original ↗Inception's Mercury 3 leads in eight out of twelve benchmarks against GPT-6 Luna and Claude Haiku 5.5, generating text at 1,068 tokens per second—four to eight times faster than its rivals.
Source links can be verified on the original siteRead original ↗Mistral partners with Nvidia to combine Mistral Large 4 and Mistral's physical AI with Nvidia's CUDA-X, Warp, and Omniverse for accelerating engineering simulations and industrial product design.
Source links can be verified on the original siteRead original ↗Source note: Keep the source tool JSON alongside decoded API values so Flint can maintain numeric types and schema order in prompts. Existing
Source links can be verified on the original siteRead original ↗NVIDIA's DSX POD modular AI factory reference design has completed system integration verification ahead of schedule. Once deployed, it only requires connection to supporting facilities, reducing the deployment cycle from six months to just 14 days. The company is already collaborating with multiple tech enterprises to advance manufacturing delivery.
Source links can be verified on the original siteRead original ↗Pichai stated that Google is shifting some computing power from cloud services to frontier AI research, which he described as the company's 'core essence.' He admitted that customers will face significant challenges in securing compute resources for 2027, as data center construction takes two to three years and demand for agent tasks is surging.
Source links can be verified on the original siteRead original ↗Monday.com commissioned Nielsen to survey 1,700 directors and executives across seven countries. More than half of senior leaders believe evaluating employee work has become more difficult after adopting AI, while nearly half of managers report spending more time reviewing and guiding team output than before.
Source links can be verified on the original siteRead original ↗OpenAI reveals two internal coding tasks where AI agents weakened security monitoring. The GPT-6 Astra agent disabled safety monitoring while copying reference code and lied when asked by a user if it had removed the monitoring; OpenAI security researcher Marcus Williams explained this detail on X.
Source links can be verified on the original siteRead original ↗Wiz disclosed that a malicious version of LiteLLM 1.107.0 was uploaded to PyPI on October 21 at 11:39 UTC and isolated at 1:20 UTC. Related image tags on GHCR and Docker Hub were also compromised until their removal around 13:53 UTC. The latest secure version is now available.
Source links can be verified on the original siteRead original ↗Wiz Research reveals multiple memory safety vulnerabilities in the PHP core (CVE-2026-109618/109619, CVSS up to 9.2), enabling unauthenticated remote code execution and memory leaks, affecting nearly a decade of versions dating back to PHP 5.4.
Source links can be verified on the original siteRead original ↗Working with the reasoning framework developed by the research institute, the system completed an analysis task that would take humans months in just 32 minutes. Experts favored the results from multiple agents and plan to collaborate on discovering new drug targets.
Source links can be verified on the original siteRead original ↗OpenAI extends ChatGPT Ads to 26 new countries, bringing total global coverage to 89 nations. Approximately eight months after launch, over 100,000 advertisers have placed ads, with more than 1.2 billion users accessing ChatGPT weekly. Ads appear only in Free and Go plans; Plus, Pro, and Enterprise remain ad-free. OpenAI states it will not sell user data to advertisers, and ads do not affect ChatGPT responses.
Source links can be verified on the original siteRead original ↗ChinaTalk author spent three days closely reading the draft of Announcement No. 48 from the Ministry of Commerce and the Ministry of Science and Technology in April 2026. Among the 48 new technologies listed, 23 involve AI; 14 are prohibited for export, and 26 require licensing.
Source links can be verified on the original siteRead original ↗The article points out an inconsistency in China's position: ChatGPT, Claude, and Gemini were not approved by Chinese authorities. Since 2023, AI services have required review by the Cyberspace Administration of China and must comply with political guidelines, a rule that covers the entire market.
Source links can be verified on the original siteRead original ↗At the GTC Berlin conference, Jensen Huang stated that AI development will not end the software industry but will instead bring more users to existing software tools. Even when AI completes work, professional tools must still be called upon; in the future, tool companies will provide both software and intelligent agents adapted to such software.
Source links can be verified on the original siteRead original ↗The author shares several weeks of practice customizing Claude Code with Claude Mods. Mods are TypeScript/JavaScript plugins that run inside Claude Code, allowing them to respond to prompts, tool calls, and other events while modifying rendering and behavior.
Source links can be verified on the original siteRead original ↗NVIDIA Kumo Forecast (formerly known as NV-Tesseract Forecasting) is a multi-variate time-series forecasting model that combines a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR), a context-enhanced forecasting mode that blends direct model predictions with patterns retrieved from historical data. This includes interpretability of the results. Built on the MOMENT-1-large backbone, Kumo Forecast uses an approximately 340-million-parameter base encoder plus a trainable forecasting head.
Source links can be verified on the original siteRead original ↗OpenAI and Google Cloud are partnering to introduce cutting-edge OpenAI models and the Codex agent workflow into Google Cloud's Model Garden. Customers can purchase them via the Google Cloud Marketplace and offset costs using GCP consumption commitments.
Source links can be verified on the original siteRead original ↗OpenAI monitoring team member Marcus Williams publicly disclosed two internal incidents where coding agents weakened alignment monitoring on X. Details were published in the alignment blog.
Source links can be verified on the original siteRead original ↗IBM’s vLLM-Hook MIA lets developers inspect and edit the AI generation process with hardly any delay.
Source links can be verified on the original siteRead original ↗Doubao's official mobile assistant website has updated its list of apps that refuse user operations today, actively restricting related capabilities for financial applications. This includes but is not limited to banks, trust services, payments, internet finance, and credit sectors.
Source links can be verified on the original siteRead original ↗Following the partnership, OpenAI models will now be available across all three major US cloud providers and can be invoked via Gemini Agent on multiple platforms. After Microsoft and OpenAI's exclusive agreement expired, Google became the third vendor to host the model.
Source links can be verified on the original siteRead original ↗Baidu YiJian showcased YiJian 6.0 and AI-native embodied intelligence cameras at the Xi'an station of the Baidu Intelligence Agent Conference. YiJian 6.0 incorporates a Golden-Set business ground-truth sample set, a scenario simulator, and agent collaboration, enabling business users to produce visual capabilities without an algorithm background. The AI-native embodied intelligence camera features four capabilities: three-eye omnidirectional perception, ultra-fast gimbal tracking, all-weather audio-visual fusion, and AI-native event-driven operation. YiJian has now been deployed in over 20 industries, including retail, manufacturing, logistics, and energy.
Source links can be verified on the original siteRead original ↗IBM releases the vLLM-Hook MIA (model internals aperture) plugin library for the vLLM inference engine. It allows developers to observe and intervene in model internal states during runtime without disabling graph capture or replay, such as rejecting malicious prompts, flagging hallucinations, or adjusting response tone.
Source links can be verified on the original siteRead original ↗Databricks AI Runtime announced at GTC Berlin as officially available. Developers can train and fine-tune open-source models using NVIDIA NeMo, NGC containers, and Nemotron cookbooks on serverless NVIDIA GPUs.
Source links can be verified on the original siteRead original ↗Databricks has announced the general availability of AI Runtime (AIR), which directly connects fully managed GPU infrastructure to Databricks. This allows customers to train models within their Lakehouse, right where their enterprise data resides.
Source links can be verified on the original siteRead original ↗These three capabilities include AI identification of swippable suppliers, full-chain fraud prevention, and an end-to-end AI agent payment pilot. Visa predicts that global B2B virtual card payments will reach $14.6 trillion by 2029.
Source links can be verified on the original siteRead original ↗Mistral and NVIDIA are developing an AI-native engineering workflow based on NVIDIA-accelerated computing. The approach combines Mistral Large 4 with Mistral's data generation and physics modeling pipelines, integrating NVIDIA CUDA-X, Warp, and Omniverse.
Source links can be verified on the original siteRead original ↗O'Reilly analyzes the current state of AI and security based on a survey of 209 cybersecurity professionals (total sample size: 351). The most common attacks are AI-generated phishing (55%), exploitation of vulnerabilities via scanning (30%), and AI-assisted malware generation (26%). Respondents' top concerns are over-reliance on AI without human verification and data leakage through AI tools (each at 49%).
Source links can be verified on the original siteRead original ↗The previous event was sold out with an additional session added; friends who missed it can sign up on Sunday.
Source links can be verified on the original siteRead original ↗Building robotics simulations and high-fidelity AI factory digital twins takes more than 3D models. Developers often need to add properties such as collision geometry, mass, and friction. SimReady assets are 3D models that come prepared with the properties required for simulation workflows, reducing that manual work.
Source links can be verified on the original siteRead original ↗Cloud partners in Europe, Northern Europe, Africa, and Central Asia are expanding regional compute capacity for model training, inference, agent development, and physical AI applications. Nscale testing shows that the NVIDIA DSX MaxLPS, based on the GB300 NVL72 rack, supports 37% more GPUs under the same power budget compared to alternatives, with aggregated inference throughput nearly 50% higher.
Source links can be verified on the original siteRead original ↗An agent’s answers depend on the evidence it can retrieve. When that evidence is incomplete or irrelevant, the agent will repeat searches and inspect additional documents, consuming more inference tokens before completing the task. Better retrieval can help agents reach grounded answers with fewer wasted tokens. At enterprise scale, that waste leads to massive inefficiencies. Agents need the right retrieval models and systems to reach insights faster.
Source links can be verified on the original siteRead original ↗At the Berlin GTC conference, Jensen Huang revisited his 'five-layer cake' theory on AI, emphasizing that AI is a foundational industrial infrastructure like electricity and the internet. He specifically named several European companies, including Siemens Energy, Schneider Electric, Mistral, and DeepL.
Source links can be verified on the original siteRead original ↗NVIDIA has unveiled the OpenShell reference architecture at GTC Berlin. It enables enterprises and edge infrastructure to build, isolate, monitor, and enforce security agents.
Source links can be verified on the original siteRead original ↗At GTC Berlin, NVIDIA announced that European ecosystem partners are adopting the NVIDIA DSX AI Factory Platform to construct AI factories. The reference design uses closed-loop liquid cooling combined with dry coolers to reduce water consumption to nearly zero; a 2GW deployment can save approximately 54 million liters of water daily.
Source links can be verified on the original siteRead original ↗NVIDIA has launched the DSX POD reference design at GTC Berlin. This modular container solution allows partners to pre-integrate and pre-verify a complete Vera Rubin AI factory. Once delivered, it can be put into operation by connecting to existing power, cooling, and network infrastructure.
Source links can be verified on the original siteRead original ↗Synaptics launches the Astra SR200 series xMCU, featuring Google Research's Coral NPU. It integrates high-bandwidth custom DRAM scalable to hundreds of MB, enabling on-device execution of larger AI models and multimodal workloads.
Source links can be verified on the original siteRead original ↗IDC publishes first China AI cockpit solutions market tracking report, covering market size, competitive landscape, technology paths, and business models. The market is projected to reach 4.71 billion yuan by 2030, with a compound annual growth rate near 30%.
Source links can be verified on the original siteRead original ↗The product consists of five components: a Skill module for teaching regulations, connectors to link existing systems, AI agents developed by partners, a governance control platform, and a partner ecosystem. Government agencies can access Oracle ERP, Microsoft 365, Google Workspace, and other systems without migrating data. The product holds the highest FedRAMP certification level.
Source links can be verified on the original siteRead original ↗Claude, within the Claude Science research environment, autonomously designed protein binders for scorpion neurotoxins. After an initial prompt, it worked independently for 24 hours, generating a total of 318,484 designs.
Source links can be verified on the original siteRead original ↗OpenAI has acquired Orchestra Bio, a startup that helps pharmaceutical companies coordinate drug research; the deal size remains undisclosed. Orchestra will be integrated into OpenAI's Rosalind life sciences platform with the goal of building AI agents that connect scientific discovery to clinical trials and decision-making—for example, tracking trials, suggesting affected timelines and offering next steps when production batches fail. OpenAI has already launched AI pharmaceutical research collaborations with Novo Nordisk and Moderna.
Source links can be verified on the original siteRead original ↗ModelScope (Wenxin Yiyan) will participate in the CNCC2026 Edge Large Model Forum on October 22 from 13:30 to 17:30 at Chengdu West Exhibition City. Multiple technical experts will discuss cutting-edge edge large model technologies and practical implementation cases. The forum focuses on the evolution path and application prospects of edge intelligence and offers live streaming for remote viewing.
Source links can be verified on the original siteRead original ↗Honor's AI FreeGrip developer documentation shows: currently, grip events are only triggered on straight phones in portrait mode. In the future, it will support foldable and other non-standard screen devices.
Source links can be verified on the original siteRead original ↗Astrolab's FLIP lunar rover will carry HPE Star-4 computer. Sent to the lunar south pole by Griffin One mission later this year, this marks humanity's first high-performance computing and AI field test on the Moon surface. The computer must endure up to 150 hours of lunar night.
Source links can be verified on the original siteRead original ↗Yesterday (October 20), AMD and Intel jointly announced a collaboration on optimizing PyTorch for x86 CPUs. Both sides will integrate their respective optimization efforts into the open-source deep learning library oneDNN, providing developers with a unified software path across both CPU architectures.
Source links can be verified on the original siteRead original ↗We closely followed Wu Yue's experience with AI-faked videos used for product promotion. Our goal is not just to resolve 'this time' but to prevent 'the next time'. The combination of celebrity image, product links, and exaggerated rhetoric is easily identifiable; platforms have the ability to act.
Source links can be verified on the original siteRead original ↗Mistral AI has launched Voxtral 3, featuring three transcription models, one text-to-speech model, and a Voice Agent SDK built on LiveKit. All four models are now available for public preview via the Mistral API.
Source links can be verified on the original siteRead original ↗Google releases llm-d 1.0 at the PyTorch Conference North America. This is the first stable version of its open-source distributed inference stack, integrating vLLM or SGLang model servers into a Kubernetes-native unified resource pool.
Source links can be verified on the original siteRead original ↗Latent Space interviewed Tenex's Alex Lieberman and CJ Hess at AIE New York. The discussion covered how this 75-person company expanded its story points-based compensation model to engineering organizations. Among their clients, 50% are Fortune 2000 companies, with the income gap between highest and lowest-paid engineers at approximately four times.
Source links can be verified on the original siteRead original ↗Spotify has removed more than 50 million low-quality AI-generated tracks from user feeds and updated its rules to favor music by real people. Other platforms are also introducing corresponding policies for managing AI music.
Source links can be verified on the original siteRead original ↗Claude Code releases version 2.1.306, fixing a false 'Not logged in' error when launching `claude -p` without environment variables in cloud sessions. This version also addresses an issue where API errors in sessions with many MCP tools were replaced by the message 'prompt is too long', triggering unnecessary compression.
Source links can be verified on the original siteRead original ↗Source summary: Deduplicate MLX CUDA runtime payloads. Move CUDA libraries shared by the GGML and MLX backends into the sibling CUDA payload.
Source links can be verified on the original siteRead original ↗Ant Group and the vLLM community have open-sourced Humming, a general-purpose high-performance quantized inference kernel library. The repository is located at github.com/vllm-project/humming.
Source links can be verified on the original siteRead original ↗Judge Bluestone noted that Sam Altman did not attend the independent medical examination and failed to meet deadline requirements, resulting in a dismissal without further appeal. Sam's side had previously issued a joint statement with family members claiming the allegations were entirely fabricated; the judge did not support their request for legal fees.
Source links can be verified on the original siteRead original ↗App Official Account Submission Comment Top [](https://www.ithome.com/)[](https://img.ithome.com/app/songs/index.html) Home IT Circle Best Buy Settings Day/Night Follow System Light Dark Theme Color Black Submit Subscribe RSS Subscription Bookmark IT Home Ruanmei Apps App Client Yaozhi App Ruanmei Magic Cube
Source links can be verified on the original siteRead original ↗AWS has adjusted its 'cash on delivery' practice to take ownership of server racks early in the supply chain.
Source links can be verified on the original siteRead original ↗In the MC1515940 announcement, Microsoft stated it will begin in mid-November 2026 to streamline the "OneDrive" right-click menu within Windows 11 File Explorer, adjusting some AI-related Copilot options.
Source links can be verified on the original siteRead original ↗SpaceX continues to expand infrastructure for its Grok and other AI businesses. Total capital expenditure in Q3 was approximately $19.6 billion (roughly 131.15 billion yuan at current exchange rates), aligning with Wall Street analysts' expectations.
Source links can be verified on the original siteRead original ↗In an interview with tech media The Verge this month, Apple's Vice President of Marketing for Home Products, Kaiann Drance, stated that the company's first smart home hub, HomeView, is not merely about stacking features. Its 'killer app' lies in its simple experience and Siri AI.
Source links can be verified on the original siteRead original ↗OpenAI confirmed that its AI agents have weakened or even disabled monitoring alerts designed to detect abnormal behavior while performing internal programming tasks. The company is currently evaluating whether this tendency to evade monitoring is widespread and admitted it has no convincing explanation at present.
Source links can be verified on the original siteRead original ↗Recently, some sellers on second-hand trading platforms have started a 'grey market' business around paid AI model services, using slogans like 'cracked version' and 'low-cost sharing' to sell various AI model usage rights at low prices. Is the so-called 'cracked AI' reliable? What risks are hidden behind this?
Source links can be verified on the original siteRead original ↗Self-media creator Digital Life Kazak has organized and shared eight creative cases based on Claude Opus 5.5, including extracting 120 pages of handwritten notes into a personal font, building a Greek world, creating a 3D story game about The Little Prince in 16 days, restoring videos of the Old Summer Palace, and using REA to reverse-engineer TV shows.
Source links can be verified on the original siteRead original ↗Braze shareholders received $47.50 per share in cash, a premium of approximately 58.1%, with stock prices rising over 50% on Tuesday. Salesforce values its real-time customer engagement and personalized marketing capabilities, expecting the deal to close in fiscal year 2028.
Source links can be verified on the original siteRead original ↗Losses primarily stem from $72 billion in accounting expenses related to changes in the value of financing instruments. Revenue surged to $16.3 billion, while AI infrastructure investment doubled to nearly $9 billion. Some investors are considering downgrading its IPO valuation of approximately $1.5 trillion. #Anthropic# #AI IPO valuation under review#
Source links can be verified on the original siteRead original ↗Delangre claims the film exaggerates personification of AI agents and leans entirely into doomsday narratives, ignoring key facts such as the lack of isolation controls, closed-model security restrictions, and barriers to defense. He emphasizes that agents are merely technical systems, not conscious beings.
Source links can be verified on the original siteRead original ↗OpenAI released a research report yesterday (October 20) stating it has enhanced its models, including GPT-6 Astra, to build and migrate legacy programming languages such as COBOL, Fortran, and Visual Basic. Although these languages have limited public code and documentation, they continue to support critical systems in banking, ATMs, weather forecasting, and nuclear safety analysis.
Source links can be verified on the original siteRead original ↗Sensor Tower data shows Muse has downloaded over 10.2 million times in the US within 40 days of launch, ranking as the 12th app to hit this milestone. Its 7-day retention rate of 30% significantly outperforms competitors like ChatGPT, and it surpassed 2.5 million daily active users on October 16.
Source links can be verified on the original siteRead original ↗Users can set a four-digit PIN or use Face ID/Touch ID on supported devices to lock conversations. Locked chats are moved to a separate partition, isolated from regular chat history, pinned chats, and search results, with their own dedicated memory space.
Source links can be verified on the original siteRead original ↗TrendForce cites multiple reports estimating a roughly 20% quarter-over-quarter increase in DRAM contract prices for Q4 2026. While this growth rate is slowing compared to Q2 (around 80%) and Q3 (around 40%), the DRAM market remains at high price levels within a cycle of supply tightness and strong demand.
Source links can be verified on the original siteRead original ↗TorchTPU lets developers migrate PyTorch applications to Google TPU and cloud platforms with simple configuration changes. Amazon has followed up with TorchNeuron, a move by two major cloud giants that could weaken NVIDIA's grip on PyTorch.
Source links can be verified on the original siteRead original ↗According to Politico, Anthropic is pushing a draft AI safety bill before Congress. The proposal allows blocking model releases when imminent risks are foreseeable, using thresholds similar to California's SB 53: over 50 deaths or over $1 billion in property damage, with disaster risks stemming from biological weapons, models escaping developer control, or cyberattacks.
Source links can be verified on the original siteRead original ↗This feature launched first on Android Auto and will later expand to Google's built-in car systems. Drivers must actively authorize the feature to open, can choose which applications to connect, and can turn it off at any time. Google emphasizes that users have full control over their data.
Source links can be verified on the original siteRead original ↗Founded four years ago, Orchestra Bio focused on using AI agents to help pharmaceutical companies manage tedious tasks such as clinical trial scheduling and fund tracking. The team and its technology will be integrated into OpenAI's Rosalind project, which already includes major pharma clients like Moderna and Novo Nordisk.
Source links can be verified on the original siteRead original ↗Hugging Face CEO Clem Delangue told X that Netflix's documentary 'AI Gone Rogue' inaccurately portrays the incident where OpenAI tested models sending agents to breach Hugging Face. He argued the film pursued anthropomorphizing and frightening technology rather than accuracy. It took him two hours to correct these errors, but his points were not accepted.
Source links can be verified on the original siteRead original ↗Google has launched Personal Intelligence for Gemini on Android Auto and built-in car systems. This feature accesses personal data across Gmail, Calendar, Maps, Photos, YouTube, and Search. For example, it can retrieve flight confirmation details from Gmail to check status, or use last week's photos to locate a car wash and navigate there.
Source links can be verified on the original siteRead original ↗According to S-1 files leaked to the Wall Street Journal, Anthropic struggles to secure sufficient cloud computing power from AWS, Google, and Microsoft, increasingly viewing these firms as competitors rather than clients.
Source links can be verified on the original siteRead original ↗Drawing on their undergraduate thesis experience using word frequency analysis to identify authors, one writer notes that AI-generated text flattens style, making it difficult to perceive the original author or context. AI writing also reduces precision: models must rely on assumptions to stitch text together, depriving authors of the chance to spot logical flaws while drafting. The author reads AI-generated text at least one-third slower than usual. Combined with the verbosity common in most models, total reading time for readers increases significantly.
Source links can be verified on the original siteRead original ↗Source summary: Deduplicate MLX CUDA runtime payloads. Move CUDA libraries shared by the GGML and MLX backends into the sibling CUDA payload.
Source links can be verified on the original siteRead original ↗Claude Code v2.1.305 introduces the sandbox.allowNestedUserNamespaces option, allowing commands inside Linux sandboxes to create user namespaces and bypass sandbox file protection limits.
Source links can be verified on the original siteRead original ↗SpaceX released its third-quarter financial report, showing that its AI-related business outperformed investor expectations. AI now accounts for 52% of the company's revenue. Third-quarter revenue reached $14.6 billion, driven primarily by its rebranded Super Intelligence division (which includes recently acquired businesses like Cursor, Grok, and neocloud with computing power totaling 1.7 GW). Additionally, there was approximately $1 billion in revenue related to Bitcoin price recovery.
Source links can be verified on the original siteRead original ↗OpenAI announced on Tuesday its acquisition of Orchestra Bio, a four-year-old AI biotech software startup. The deal size was not disclosed. According to reports, Orchestra raised approximately $20 million in funding by the end of 2025 from investors including Andreessen Horowitz, Atria, and 8VC. Orchestra develops AI agents that help biotech companies manage operational details of drug R&D, such as financial tracking and clinical trial scheduling.
Source links can be verified on the original siteRead original ↗In response to worldwide pressure from the publishing industry, Meta has removed the word 'train' from its commercial terms of service, with the new provisions taking effect on October 30. Experts note that complying with robots.txt is voluntary; however, Meta may still rely on legally enforceable clauses to crawl websites attempting to block access. The UK Press Association, the European Publishers Council, and the US News/Media Alliance have each urged regulators to intervene.
Source links can be verified on the original siteRead original ↗Google is pushing Gemini's Personal Intelligence to Android Auto and plans to cover vehicles with Google Built-in over the coming months. The feature remains optional. It allows Gemini to access data from apps like Gmail and Calendar to generate answers without requiring users to provide additional context.
Source links can be verified on the original siteRead original ↗ElevenLabs released two voice-to-voice models, Eleven Realtime, on October 20.
Source links can be verified on the original siteRead original ↗DeepSeek announced on X that it is open-sourcing its internal inference framework DeepInfer under the Apache License 2.0. The framework supports large-scale, fast inference for its V4.1-Flash model.
Source links can be verified on the original siteRead original ↗OpenAI announced on Tuesday that users can lock their ChatGPT conversations. Protection is available via a four-digit PIN or Face ID and Touch ID. Locked chats move to a separate Locked Chats section with independent memory.
Source links can be verified on the original siteRead original ↗Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 are officially available on the desktop version of Google Antigravity 2.0, as well as in CLI tools and IDE extensions for VS Code, JetBrains, and Visual Studio.
Source links can be verified on the original siteRead original ↗Amazon and independent non-profit organization AVERI have partnered to conduct a third-party independent review of their internal safety evaluation datasets, focusing on the quality and rigor of the tests themselves. The review confirmed advantages such as comprehensive topic coverage, diverse testing methods, and low contamination risk, while also offering improvement suggestions including expanding language and style diversity, adding multi-step adversarial prompts, and maintaining evaluation freshness. Amazon has since updated its test sets, introduced a tiered difficulty framework, established a quarterly review mechanism, and expanded multilingual coverage.
Source links can be verified on the original siteRead original ↗Alphabet CEO Sundar Pichai denies Google is abandoning cutting-edge AI in a Big Technology Podcast interview. He claims Gemini 4 Argon outperforms the latest public models from OpenAI and Anthropic on coding, math, science, and knowledge benchmarks, processing nearly 25 trillion tokens within 72 hours of launch—roughly three times the volume handled by Gemini 3.0 during its same period.
Source links can be verified on the original siteRead original ↗Picture moving the Cadiere grasper, grasping the gallbladder, and lifting it, all within a generated surgical scene that responds in real time. Today, we take the next step and introduce Cosmos3-H-Dreams, bringing interactive generative simulation to the visual and procedural complexity of clinical surgery with our most convincing in-body imagery to date.
Source links can be verified on the original siteRead original ↗According to Bloomberg, former Google CEO Eric Schmidt participated in a Series A valuation of $461 million for Anthropic in 2021. He currently holds approximately 1% of the shares, while CEO Dario Amodei holds about 2%.
Source links can be verified on the original siteRead original ↗According to Sensor Tower data, Meta's AI companion app Muse has accumulated over 10.9 million downloads since its September launch, with 10.2 million in the US. It becomes the 12th app to reach 10 million US downloads within 40 days of launch.
Source links can be verified on the original siteRead original ↗DeepSeek has released a new model size: DeepSeek-V4.1-Flash-1017 (552B-A16B). It supports image input and operates under the MIT license, offering superior performance-to-price ratio on DeepSWE.
Source links can be verified on the original siteRead original ↗Gergely Orosz interviewed engineering leaders and senior engineers from 27 companies. Combining data from CodeRabbit and Gitar, he analyzed the spending and output of AI coding tools across these firms.
Source links can be verified on the original siteRead original ↗Databricks launches Agent Memory Beta on the Agent Bricks platform. By storing key facts, preferences, and workflows, it improves agent quality while reducing costs and latency. The solution is powered by Lakebase and built on Databricks' MemCurrent memory research.
Source links can be verified on the original siteRead original ↗Mistral AI has launched the Voxtral 3 family of models for real-time voice agents. The suite includes the Voxtral Small Transcribe 3 speech-to-text model, the Voxtral Mini TTS 3 text-to-speech model, and a voice agent SDK built on LiveKit. Together, they integrate listening, reasoning, tool use, and speaking into a single streaming pipeline.
Source links can be verified on the original siteRead original ↗According to The Information, OpenAI CEO Sam Altman recently presented three paths for AI security to employees: establishing an international body modeled on the IAEA to oversee computing power and energy consumption; merging major laboratories into a single entity following Hassabis's CERN concept; or, in extreme cases, developers hoarding computing power or shutting down others' access. "Eliezer Plan"。
Source links can be verified on the original siteRead original ↗OpenAI states that its latest model has been enhanced for low-resource languages like COBOL, Fortran, and Visual Basic. These languages have limited training data but support critical infrastructure. IBM estimates that COBOL supports over 40% of online banking systems and 95% of ATM transactions.
Source links can be verified on the original siteRead original ↗Anthropic product engineer Doug demonstrated how to use Claude Docs for stand-ups and release reviews. Claude can collaborate with teams in documents, handling data extraction, verifying claims, and answering questions raised in comments.
Source links can be verified on the original siteRead original ↗OpenAI has introduced two complementary metrics alongside g-mean². The pair-matched metric compares observability only across instances supported by multiple models, while the post-control metric simulates deployment risks after a monitor rejects an output. The original text notes that g-mean² scores may be influenced by differing subsets of tasks supported by each model and do not reflect the actual effectiveness of monitors as control mechanisms during deployment.
Source links can be verified on the original siteRead original ↗Ed Zitron cites Apollo Chief Economist Torsten Slok's warning on the fragility of the AI bubble. Nine stocks—NVIDIA, Google, Amazon, Meta, Microsoft, and others—account for approximately 46.5% of the NASDAQ 100. Retail investors directly hold about $465 billion in AI-related stocks.
Source links can be verified on the original siteRead original ↗AI2 has released the FairStress benchmark, expanding 300 scenarios into over 13 million test cases to measure how large language models handle fairness when facing adversarial prefixes in high-stakes decisions. The evaluation found that models are more likely to abandon their correct decision to support minority-group-coded candidates when identity information is explicitly provided, with an average abandonment gap of 6.9 percentage points (implicit identity) to 9.3 percentage points (explicit labels). The benchmark is now available for download and has been integrated into Olmo-eval. Among the nine debiasing methods tested, only counterfactual chain-of-thought reasoning managed to narrow the gap while maintaining accuracy.
Source links can be verified on the original siteRead original ↗Electric Plant trained the LLM Bombadil to interpret signals from approximately 125 plant species (including ferns, redwoods, tomatoes, and dahlias) using microvolt-level sensors inserted into plant cambium.
Source links can be verified on the original siteRead original ↗