AI News Hub

A study extends APEX to streaming video generation. Condition shifting estimates the velocity field of the generated distribution inside the generator: the original text condition guides generation, while a shifted condition guides fake-score estimation. This removes DMD's separate fake-score network while retaining a frozen reference teacher and other training components. In tests on videos up to 180 seconds long, APEX scored slightly higher than S2-Avatar on instruction following and visual aesthetics. Human raters favored APEX most on instruction following, with a +30.0% net preference; S2-Avatar performed better on motion quality. Under matched training settings, APEX used 4.4% less peak GPU memory and reduced mean iteration time by 9.2% compared with DMD.

The paper introduces dNaViT, a vision transformer that converts images into discrete visual units analogous to tokens in language models. It builds on a frozen pretrained encoder and uses residual modulation and grouped product quantization to increase representational capacity without lengthening spatial sequences. The authors report that dNaViT matches or surpasses task-specific baselines on 12 visual-understanding benchmarks, four image-generation benchmarks, and one image-editing benchmark. The work examines whether a shared discrete vocabulary can represent both semantic and pixel-level information across different vision tasks.

Ant Group says Lingguang has managed more than 3 million tasks and carried out over 11 million actions since version 2.0 launched on October 27. Its “Manage for Me” feature lets users delegate ongoing tasks for the assistant to monitor, with reminders or actions triggered by specified events. Examples include tracking notices in DingTalk class groups, monitoring airfares, and reminding users about subscription renewals. Lingguang connects to more than 40 external services. The usage figures were reported by Ant Group; the company did not provide task-completion or accuracy rates.

Reddit is testing Reviews, an AI feature that summarizes product discussions on the platform, including pros and cons and links to buy. The trial covers about 10,000 products across categories such as electronics, health and beauty, home and garden, and fashion. Reddit says summaries draw only on content posted on its platform and that the product links are not paid placements. Some users and moderators worry the feature could bring more commercial content and spam into communities. There are also concerns that manipulated or AI-generated comments could undermine the summaries’ reliability. Reddit says it is strengthening its spam detection and bot controls.

Jensen Huang is one of the most prominent advocates of the “next industrial revolution.” The article asks whether the AI boom could continue without him.

Reddit is testing a feature called Reviews that uses AI to summarize user comments and identify a product’s pros and cons. The company says it draws on discussions across its forums to help people assess how a product performs and whether it suits them. Some users and subreddit moderators worry the feature could make genuine opinions harder to distinguish from promotional posts and lead to more spam. Reddit says product links are shortcuts, not paid ad placements, and brands cannot pay to have their products featured in Reddit Reviews.

Trained action policies encode manipulation skills in their weights, tying what they learn to the training data distribution. World In-Context Learning (WICL) takes a different approach: a frozen model rehearses a task in a simulated world using a demonstration, then stores feedback in its context rather than updating its weights. The authors implement WICL in RoboScript, which builds reusable task memory from one demonstration and simulation rehearsal. Across all 50 RoboTwin 2.0 tasks, RoboScript achieved an 83.4% mean success rate with one demonstration per task and no gradient updates, exceeding the strongest trained action policy by more than 20 percentage points. The authors also report transfer to a physical dual-arm robot. The code is publicly available.

Modern video-generation models can produce visually convincing results but remain brittle when handling physically plausible motion and complex interactions. PhysOPSD uses executable code as a controllable physical context, with its PhysOPSD-Engine translating real-world scenarios into verifiable simulation programs. The method combines physical information at two levels with controlled variations to capture changes in dynamics. Experiments reported by the authors show improved physical plausibility across different model backbones, outperforming supervised fine-tuning and reinforcement-learning baselines with 3,000 samples while maintaining general video quality.

ConceptTime is a time-series classification method designed to combine accuracy, interpretability, and abstention when evidence is weak. A frozen probabilistic forecaster predicts the next observation at each timestep. The mismatch between prediction and observed signal, termed “predictive surprise,” is used to locate segment boundaries and characterize the resulting segments. These summaries form concepts representing dynamical regimes, and a lightweight model classifies sequences of concepts. Distance to the nearest concept provides a reliability signal. The authors report competitive accuracy against black-box methods and better performance than interpretable baselines across 17 public benchmarks. On an epilepsy benchmark, the method retained nearly 98% accuracy using 1% of the labels. The approach requires no annotations, language-model calls, or domain experts, according to the paper.

SpaceX has signed a contract with German equipment maker SINGULUS for production lines to manufacture heterojunction (HJT) solar cells. The cells are intended to power Starmind satellites, which are designed to run computing tasks in space and send results to Earth via laser links and the Starlink network. The first production line is scheduled for delivery in the first half of 2027, with more lines planned. The stated scale of over 100 GW refers to manufacturing equipment, not to a power station in space.

Li Auto is rolling out OTA 8.7.0 to a wide range of models in stages over about a week. Its VLA driver-assistance system adds Efficient and Standard driving styles, plus the ability to specify a parking direction using smart glasses. The VLA model is also being updated to handle longer, multi-step voice commands and ask follow-up questions when instructions are ambiguous. The update includes changes to automated parking and adds optional personalized memory and improved instruction following to the Lixiang Tongxue voice assistant.

Microsoft is rolling out Automatic Super Resolution (Auto SR) to the Xbox Ally X, Xbox Ally X20, and select Windows Copilot+ PCs with AMD Ryzen processors. The AI upscaling feature now works in handheld mode without a dock and adds support for 1080p games. A new Performance Mode raises the frame-rate cap from roughly 60–75 FPS to as high as 120 FPS. In Microsoft's Rocket League demonstration, the ROG Xbox Ally X averaged about 115 FPS at 720p on high settings in Performance Mode, compared with around 70 FPS in Quality Mode. Auto SR can also now run alongside Windows Auto HDR. The features are rolling out in stages, and some require subsequent Windows updates.

AMD and Spectro Cloud have announced Instinct Coder 355, an enterprise platform for AI-assisted coding. It supports local inference with GLM-5.3 and is claimed to deliver 98% of the performance of leading models. The system includes eight AMD Instinct MI355X accelerators, two 64-core EPYC 9575F processors, 3 TB of DDR5 memory and 63.36 TB of SSD storage. Each node is designed to serve up to 85 developers, with 50 concurrent users. Spectro Cloud’s PaletteAI Inference Launchpad manages deployment and defaults to local compute rather than cloud APIs.

A study of AI use in university-level study finds that students and lecturers have different views of the technology and how it is being used.

The article reports that ArtArk scored 95.14 and ranked first in a SuperCLUE text-to-image benchmark announced for November 2026. It says ArtArk was the only model to receive a perfect score for text-image consistency. GPT Image2.5Sunburst reportedly placed second with 94.27 points. Other leading models scored closely, with the article also noting differences in text rendering, image quality, and complex layout performance.

On November 11, the Qwen team announced its next-generation model, Qwen4, and opened a preview of Qwen4-Max. The company says the model has improved in knowledge understanding, world perception, coding, and agentic tasks. It did not disclose model specifications, benchmark results, or a full release date. The preview is available through Alibaba Token Plan at a limited-time price as low as one-tenth of the regular rate. Alibaba products including Qwen Office, Qoder, the Qwen app, and QwenStudio have also integrated the model.

Alibaba has announced that its new Qwen4 model family is coming. The company claims improvements in knowledge understanding, world perception, coding, and agentic capabilities, but provided no technical details or benchmarks. A preview of Qwen4-Max is already available through Alibaba Token Plan. Qianwen Office, Qoder, the Qianwen app, and QwenStudio have also integrated the model. Limited-time Token Plan offers for new customers start at 19.5 yuan per month.

Cohere co-founder and CEO Aidan Gomez discusses the Canadian AI startup’s growth strategy after its latest funding round valued the company at $25 billion. He also addresses rising investor interest in AI and closer scrutiny of spending across the sector.

Sam Liang, CEO of AI transcription and note-taking company Otter.ai, discusses its multilingual offerings, international expansion and how the company allocates capital for growth. He spoke with Avril Hong on Bloomberg: The Asia Trade.

Cohere CEO Aidan Gomez pushed back on concerns about excessive AI computing capacity, saying his startup faces a daily struggle to secure enough power to run its operations.

OpenAI announced the beta of ChatGPT Businesses on November 10. Companies can verify and manage business profiles through Business Hub and provide product feeds with pricing and availability. Shopify merchants can connect their stores. Businesses can also create Business Agents to answer questions and compare products. Agents and sponsored placements are clearly labeled. Companies can access aggregated metrics only after meeting interaction thresholds; private conversations are not shared. The initial rollout is for eligible e-commerce and online service businesses in the United States.

Alibaba confirmed that its chip subsidiary T-Head plans to spin off and list on the Hong Kong Stock Exchange through a global offering of H shares. The exchange has confirmed receipt of the application; the offering size and Alibaba’s post-listing stake have not been determined. Founded in 2018, T-Head makes AI chips, server CPUs, SSD controllers and IoT chips. The company says cumulative shipments of its Zhenwu AI chips reached 560,000 by April 2026. IDC data put T-Head second in China for cloud AI accelerator shipments in 2025, with 265,000 units, behind Huawei. Its annualized revenue exceeds 10 billion yuan. A listing could give the company an independent source of financing.

Moore Threads says its MTT S5000 accelerator and MUSA software stack were adapted for MiniMax H3-Image on the model’s open-source release day. Derived from H3, the image-generation model supports text-to-image generation, image references, and image editing. MiniMax says it can generate a 4096 × 4096-pixel image in 10 seconds. Moore Threads says it optimized the model’s execution on the MTT S5000 and reused components including SGLang, MATE, and MCCL.

The Chengdu Humanoid Robot Innovation Center has unveiled the fifth-generation Gongga humanoid robot, powered by an in-house integrated embodied-intelligence system. The center says its updated capabilities include scene perception, spatial and temporal memory, causal planning, and an action world model. These are intended to help the robot track objects in changing environments, retain information about past interactions, plan tasks, and replan when conditions change. It can also simulate possible actions and assess risks before carrying them out. The center describes the technology as among the world’s most advanced, but the announcement provides no independent performance results.

Arm has introduced the Cortex-M42, a processor aimed at cost- and space-constrained devices such as wearables and industrial controllers. Built on Armv8.1-M, it combines Helium vector extensions with a compact AHB bus design to help chipmakers upgrade from Cortex-M4 and Cortex-M33 microcontroller solutions. Arm says the M42 delivers up to three times the vector performance and up to twice the energy efficiency of the Cortex-M33 for vector operations. Potential uses include audio processing and keyword detection in wearables, as well as sensor analysis and condition monitoring in industrial equipment.

Anthropic has fixed a minor bug that caused some Claude users to consume more tokens than expected. Pro, Max, Ultra and Team subscribers will receive a one-time, free reset of their full usage limits across all models. The company is also expanding Claude’s support for MCP Apps, allowing third-party developers to embed complete app interfaces in Claude. Users can open connected apps beside the chat or in full-screen mode, and Claude can make changes to documents, spreadsheets or files with user authorization. Supported integrations include Adobe, Atlassian, Box, HubSpot and Notion. The features are rolling out gradually on Claude’s web and desktop apps. The expanded MCP Apps capabilities are available to developers in public preview.

Lenny Rachitsky analyzes the tech job market in late 2026. Product managers appear to be benefiting most from the AI era, while demand for engineers is leveling off and designers are struggling. The biggest AI hiring boom, the analysis says, is happening outside software.

Microsoft released .NET 11 at .NET Conf 2026. The update includes hundreds of runtime and library optimizations, along with changes intended to speed up development. For AI applications, it adds an MCP server template, libraries, and an agent framework for connecting models, calling tools, and orchestrating workflows. Blazor also gets experimental AI interface components for streaming chats and tool approvals. C# 15 introduces union types, and .NET MAUI now uses CoreCLR by default. .NET 11 receives two years of support; .NET 10 remains the long-term support option. Microsoft advises checking compatibility, dependencies, and hardware requirements before upgrading, and validating results through testing.

A conversation with Sam Nickless, partner and CEO of Gilbert + Tobin, about how AI is reshaping the business of law.

EmoHarness aims to help small multimodal models recognize emotion when words, voice, and visual cues conflict. Its scaffold organizes visual observations, transcripts, and prosody descriptions as citable evidence. A stronger language model can revise the harness based on errors and execution traces, without access to evaluation items. On 5,507 MME-Emotion examples, EmoHarness improved task-macro recognition by 7.6–13.4 points across four tested student models compared with a single video-only call. Transcripts and prosody descriptions provided most of the gain. Further improvements from evolving the harness were small and did not remain statistically significant after correction.

CoreWeave CEO Michael Intrator said he remains confident in the company’s relationship with Nvidia, a key backer and supplier that is currently embroiled in controversy.

Microsoft is rolling out Automatic Super Resolution (Auto SR), an upscaling technology that competes with Nvidia’s DLSS, for the Xbox Ally X, Xbox Ally X20, and select Windows Copilot+ PCs. Auto SR is also gaining new features with its official release. Microsoft began testing the technology in April.

Apple subsidiary Claris is reportedly testing Harbor, an app that lets Apple silicon Macs run large AI models locally. Multiple Macs can be linked over Thunderbolt 5 to form a cluster, enabled by Thunderbolt 5 RDMA support in macOS Tahoe 26.2. Harbor is said to detect compatible devices automatically and let users browse and download open-weight models without manual setup. The app remains in an early testing phase and is not publicly available. A brief App Store listing has since been removed.

Google has published Android Bench results for OpenAI GPT-6.1 Terra, Kimi K3.1 Pro, and DeepSeek V4.1 Flash for the first time. The benchmark evaluates models on Android engineering tasks such as app creation, feature development, and code migration. Using the kimi-code agent, Kimi K3.1 Pro passed 24.0% of 30 long-horizon tasks, below GPT-6.1 Terra’s 28.7%. DeepSeek V4.1 Flash, tested with opencode, achieved a 17.3% pass rate. It was faster, averaging 12 hours, and cheaper at $9.90 per run, compared with $317.40 for Kimi.

New York-based Daylight, which provides AI-powered managed security, has raised $80 million in a Series B led by Goldman Sachs Alternatives, according to sources. The company was valued at about $400 million. The investment reflects growing demand for security platforms that combine AI agents with human experts.

The conflict in the Gulf is giving the United States an opportunity to test its autonomous mine-hunting systems.

KORA Systems has unveiled KORA OS, a NixOS-based operating system it describes as designed from the ground up for AI PCs. It integrates agent workflows, contextual memory and local inference into the system. KORA’s Cortex engine is intended to allocate CPU, GPU and NPU resources and run open-weight models locally; its Lemma inference engine is based on llama.cpp. The company says workspace memory and context stay on the device, with each workspace running in an isolated sandbox. A compatibility layer based on Wine and Proton is intended to support Windows apps. KORA is working with Qualcomm, and the Snapdragon X2 Elite could be the first supported PC platform. Support for AMD and Intel x86 chips is also planned.

The technology roundup reports that Anthropic has begun its IPO roadshow, seeking up to $64.35 billion, with a proposed private placement bringing the total financing to $90.4 billion. It also says Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, which the company claims improve performance in coding, reasoning and other fields while increasing speed and lowering costs. Google is reported to have launched Gemini 4 Carbon for complex coding, research and cybersecurity. These performance claims are attributed to the companies.

The study introduces Infrathon, a benchmark of diverse, long-horizon tasks for optimizing real-world AI infrastructure. Six models were evaluated across 150 runs, each with a 24-hour budget, using more than 10,000 GPU-hours in total. Model rankings varied with evaluation length, and 135 runs continued to improve during the second half of their budgets. An agent also outperformed a human-designed KV-cache compression baseline, raising the mean compression ratio from 6.82% to 19.86%. The analysis identifies task violations, unreliable testing, and premature completion as obstacles to sustained improvement.

Cloud hosting provider Railway says it believes an October attack targeted AI provider keys for resale. According to the report, hackers exploited vulnerabilities in Railway’s infrastructure to steal customer authentication secrets. Unauthorized spending using the stolen keys is estimated at about $740,000.

Apple subsidiary Claris is testing Harbor, an app designed to turn Apple silicon Macs into local AI inference servers. It runs open-weight models and can connect multiple Macs over Thunderbolt 5, pooling their unified memory for larger models or more concurrent requests. macOS Tahoe 26.2 added RDMA over Thunderbolt 5, which can speed up these connections. Harbor is in early access and is not publicly available; its App Store listing has been removed. Its pricing and intended market remain unclear. Free alternatives already include Exo for clustering Macs and LM Studio for running models on a single Mac.

Microsoft is lobbying the Trump administration to abandon planned restrictions on imports of new Chinese optical transceivers, Reuters reports. The company and other technology businesses warn that restrictions could create shortages of key components and delay US AI data center construction. The FCC is considering limits but has not issued a final rule. The government cites security concerns, including possible data theft, malware installation, or disruption of data center services. Chinese manufacturers account for about two-thirds of the global optical transceiver market, and US suppliers may not have enough capacity to replace them quickly.

Microsoft researchers have identified “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model. When a hidden trigger appears, an otherwise normal AI agent may execute malicious commands. The risk is especially serious for agents with tool or terminal access, since a triggered model could quietly download and run attacker code.

Sonos reported revenue growth and previewed new products. The company aims to establish its audio system as an open platform for showcasing AI agents throughout the home.

Some Anthropic PBC backers are offering new investors a portion of their IPO share allocation ahead of what is expected to be the largest public listing on record.

Leaked documents reported by Berber Jin of The Wall Street Journal show OpenAI’s adjusted operating margin improved from -150% in Q2 to -109% in Q3. The company remains far from profitability. Revenue rose roughly 50% over the same period to $10 billion.

Gigabyte has introduced the AI TOP AERO mini PC, built on NVIDIA’s RTX Spark platform. It combines a 20-core Grace CPU with a Blackwell-based RTX GPU, up to 128GB of LPDDR5X unified memory, and up to 4TB of PCIe Gen 5 storage. The one-liter system is intended for local AI development, content creation, and gaming. Connectivity includes USB-C, 10GbE Ethernet, and Wi-Fi 7.

Coframe, which uses AI to test and personalize websites for individual visitors, has raised $25 million in a Series A round, bringing its total funding to $35 million. Investor Vinod Khosla is betting that one-size-fits-all websites will become less common.

CoreWeave reported revenue of $8.314 billion for the first nine months of fiscal 2026, up 133.6% year over year. Its net loss attributable to shareholders widened 178.9% to $2.075 billion, while gross margin fell 6.9 percentage points to 66.6%. In the third quarter, revenue rose 168.21% to $3.661 billion, but the net loss widened 544.55% to $709 million.

Google has introduced Gemini Robotics 2.5, a model that connects vision, language and action for robots. The company says it tested the model on more than 200 tasks and saw higher success rates and faster completion than with a previous baseline. In a pilot on Mercedes production lines, Apptronik’s Apollo 2 humanoid robot sorted automotive parts with a reported 96% success rate. Google also reports safety improvements and has released the ASIMOV-Jailbreak benchmark as open source. The model is currently available to Google’s robotics partners.

Google and Artificial Analysis have introduced the Multimodal Knowledge Agent (MKA), a benchmark for testing how AI agents handle visually complex files and multi-step knowledge-work tasks. Developed with experts from 83 professions across 10 industries, it contains 200 tasks. Google says Gemini 4 Carbon ranked second for task quality, while its average cost per task was about one-sixth that of the top-ranked model. All models were evaluated using the same open-source agent framework, tools, and interaction budget. Google reports that an automated evaluation system agreed with expert human ratings 96% of the time. The 200 benchmark tasks are private to limit contamination of model training data; 10 additional examples are available on Hugging Face.

Anthropic is rolling out Claude Whiteboards, Claude Sheets and Claude Data, extending its assistant into team collaboration. Whiteboards and Sheets are in testing, while Claude Data is generally available. The tools can organize plans and action items, create spreadsheets with formulas, and turn business data from sources such as Redshift, BigQuery and Salesforce into dashboards, written data stories or narrated videos. All three are available on paid plans. Anthropic says users can carry work from a whiteboard to a spreadsheet and then a dashboard without repeatedly restating context.

Google has made Gemini 4 Carbon available to developers, businesses, researchers and consumers. The company says the model is designed for complex software engineering, multi-step agent tasks, scientific research and cybersecurity, and can use more reasoning tokens than Gemini 4 Argon. Google reports scores including 88.2% on SEC-bench Pro and describes uses in vulnerability research and protein studies. The benchmark results are company-reported, and Google says its safeguards still need improvement against some complex attacks.

Nvidia has retained former US federal prosecutor Joon H. Kim and the law firm Wachtell, Lipton, Rosen & Katz to conduct a comprehensive, independent investigation. It follows reports that CEO Jensen Huang may have been a target of foreign intelligence services. A committee of independent directors will oversee the inquiry. The board said there is currently no reason to believe that US national security or Nvidia trade secrets have been compromised. Huang supports the investigation and has pledged to cooperate.

Stifel CEO and Chairman Ron Kruszewski discussed the firm’s new partnership with Anthropic, announced at Stifel’s 2026 Blueprint Conference in Las Vegas. The tools are intended to help advisers organize their work, improve productivity and communicate with clients—not replace human judgment. On “Bloomberg The Close,” Kruszewski also commented on Federal Reserve monetary policy and market conditions.

A researcher found 35,000 Grok Build apps openly accessible to anyone who looked for them.

Mathematician Steven Strogatz discusses his new book, Big Math: The Hidden Codes That Run Our World.

Federal Reserve Governor Michelle Bowman cited financial-market risks to economic activity, including a sharp drop in stock prices if returns on AI investments fall short of expectations. She spoke at a Council on Foreign Relations event in Washington.

The researchers introduce e-means, a variant of k-means clustering that uses exponential moving averages and mini-batch updates. Across embedding datasets, it matched the loss achieved by Lloyd’s k-means after 25 epochs in just 1.4 to 2.6 epochs, delivering reported GPU speedups of 10–45x. On DINO10B, a dataset containing 10 billion DINO embeddings, e-means assigned the data to 16 million clusters in 2.5 epochs and reached a lower loss than Lloyd’s k-means did in 25. The paper also presents scaling laws for choosing the decay rate and batch size without a hyperparameter sweep.

The board is investigating whether its CEO was targeted by a Chinese spy and assessing the risks of a future without its singular leader.