AI News

Addressing the high cost and safety risks of real-world data collection for embodied agents, NVLabs introduced SAGE, an agentic framework for automated 3D scene generation. SAGE accepts user intent (e.g., "place bowl on table") and autonomously creates simulation-ready environments. It couples generative components with critics evaluating semantic plausibility, visual realism, and physical stability, refining scenes iteratively. The resulting diverse datasets allow policies trained purely in simulation to generalize effectively to unseen objects and layouts, demonstrating a viable path for simulation-driven scaling in robotics.

VideoWorld 2 investigates learning transferable knowledge directly from raw, unlabeled real-world videos for intelligent agents. The core advancement is a dynamic-enhanced Latent Dynamics Model (dLDM) that successfully decouples action dynamics from visual appearance. A pretrained video diffusion model manages visual appearance modeling, allowing the dLDM to focus on learning compact, task-related latent codes. These codes are modeled autoregressively to derive policies supporting long-horizon reasoning. Evaluations on complex real-world handcraft tasks show up to a 70% improvement in task success rate. Furthermore, the system acquired effective manipulation knowledge from the Open-X dataset, substantially boosting performance on the CALVIN benchmark in robotics.

This research introduces Chain of Mindset (CoM), a training-free agentic framework designed to overcome the limitation where LLMs apply a single fixed mindset throughout reasoning. CoM decomposes problem-solving into four functionally heterogeneous modes: Spatial, Convergent, Divergent, and Algorithmic. A Meta-Agent dynamically selects the optimal mode based on the evolving reasoning state, filtering information flow efficiently. Experiments across challenging benchmarks in mathematics, code generation, and spatial reasoning demonstrate CoM achieving state-of-the-art performance, significantly outperforming strong baselines.

The Covo-Audio technical report introduces a 7B-parameter LALM designed for direct end-to-end audio processing, generating audio outputs within a unified architecture. Pretraining allows it to achieve competitive or state-of-the-art performance across speech understanding, dialogue, and audio comprehension tasks relative to its size. Key variants include Covo-Audio-Chat for conversational ability and Covo-Audio-Chat-FD for full-duplex interaction. To mitigate deployment costs, the researchers propose an intelligence-speaker decoupling strategy, enabling flexible voice customization with minimal TTS data while maintaining dialogue performance. This validates the potential of 7B models for sophisticated, scalable conversational AI.

A new paper introduces Masked Bit AutoRegressive modeling (BAR), challenging continuous pipelines in visual generation. Researchers found that the performance gap between discrete and continuous methods is due to bit allocation in the latent space, not inherent tokenizer inferiority. By scaling codebooks, BAR allows autoregressive transformers to match or exceed continuous performance. The BAR framework specifically addresses training scalability issues with large codebooks by generating tokens bit-by-bit via a masked modeling head. BAR achieves a new state-of-the-art gFID of 0.99 on ImageNet-256, surpassing leading methods while requiring significantly lower training costs and faster convergence than prior continuous approaches.

A new method, Outline-Guided Path Exploration (OPE), is proposed to optimize Large Reasoning Models (LRMs) utilizing parallel thinking frameworks enhanced by Reinforcement Learning with Verifiable Rewards (RLVR). The research identifies that current approaches suffer from an information bottleneck due to redundant reasoning paths. OPE mitigates this by proactively generating diverse outlines to partition the solution space before parallel reasoning begins. This strategy significantly reduces information redundancy and improves the diversity of explorations. Empirical tests on complex mathematical benchmarks confirm that OPE reliably enhances reasoning performance across different aggregation methods.

Tel Aviv-based Backslash Security has secured $19 million in Series A funding, bringing its total capital raised to $27 million. The company specializes in safeguarding enterprise software development environments against security threats associated with "vibe coding," a term suggesting risks arising from AI coding assistants. The new funds are earmarked for expansion.

Isomorphic Labs, a company spun out of Google DeepMind, has introduced IsoDDE, a dedicated drug design system. The core announcement is the assertion that IsoDDE surpasses the predictive accuracy of Google DeepMind's landmark AlphaFold 3 when modeling complex biomolecular structures. If validated, this represents a major technical advancement directly challenging the established benchmark in AI-driven structural biology.

Ant Group has released its full multimodal large model, Ming-Flash-Omni 2.0, as open source. The model demonstrates strong performance across visual language understanding, image editing, and speech generation, reportedly surpassing Gemini 2.5 Pro on select benchmarks. Technically, it is highlighted as the industry's first unified audio generation model, capable of producing speech, sound effects, and music simultaneously on a single track with fine-grained control via natural language prompts. The model achieves a very low inference frame rate (3.1Hz), enabling real-time, high-fidelity generation of long audio segments efficiently. Built on the Ling-2.0 (MoE) architecture, this release provides developers with a reusable foundation to simplify complex multimodal application development. Model weights are available on platforms like Hugging Face.

Researchers introduce Agent Banana, an agentic planner-executor framework designed to overcome limitations in instruction-based image editing, such as over-editing and object infidelity in multi-turn edits. The system utilizes Context Folding to manage long interaction histories and Image Layer Decomposition for precise, localized edits on native resolution images. To evaluate professional workflows, they developed HDD-Bench, a benchmark featuring 4K images. Agent Banana shows strong performance in multi-turn consistency and background fidelity on this demanding benchmark, aiming to advance reliable, professional-grade editing capabilities.

InclusionAI released UI-Venus-1.5, a family of unified GUI agents (2B, 8B, and 30B-A3B MoE) addressing the challenge of balancing generality and performance in digital automation. The update features three major technical advances: extensive mid-training using 10 billion tokens across 30+ datasets, online reinforcement learning for dynamic navigation, and a unified checkpoint created via model merging. Evaluations confirm new state-of-the-art performance across benchmarks like ScreenSpot-Pro (69.6%), VenusBench-GD (75.0%), and AndroidWorld (77.6%), alongside demonstrated robustness in Chinese mobile application interaction.

Prism introduces a spectral-aware method for block-sparse attention in LLMs. It resolves efficiency bottlenecks caused by mean pooling interfering with RoPE positional embeddings by restoring attenuated positional signals. This training-free approach allows block importance estimation purely via block-level operations, achieving up to 5.1x speedup during long-context pre-filling while matching full attention accuracy.

SkillRL is a new framework designed to improve how Large Language Model (LLM) agents learn from past experiences. Current methods often store too much raw, noisy data, hindering the discovery of reusable behavioral patterns. SkillRL introduces automatic skill discovery and recursive evolution to build a hierarchical 'SkillBank'. This allows the agent's policy and its skill library to co-evolve, effectively distilling high-level knowledge. The result is a significant reduction in the required token footprint while boosting reasoning utility. Experiments show SkillRL outperforms strong baselines by over 15.3% on benchmarks like ALFWorld and WebShop, demonstrating superior generalization as complexity increases.

The Anchor framework addresses the high cost and poor quality of training data for desktop GUI agents. It bootstraps scalable supervision from a small set of seed demonstrations by identifying meaningful branch points and proposing new, state-grounded task variants. An executing agent follows these variations, while a verifier enforces consistency via state-aware checks. Experiments on benchmarks like OSWorld and WindowsAgentArena demonstrate that models fine-tuned on this expanded, cleaner corpus achieve consistent improvements over zero-shot and baseline synthesis methods, generalizing across different operating systems.

Samsung Electronics' stock reversed earlier losses following statements from CTO Jaihyuk Song, who asserted that the company has reclaimed the top position in the memory industry thanks to its forthcoming HBM4 technology. This announcement signals confidence in Samsung's roadmap for next-generation high-bandwidth memory crucial for advanced AI computing infrastructure.

Google has officially launched its "AI Creations" feature, leveraging Gemini AI power, for iPhone and iPad users, following its initial Android release. This update allows iOS users to transform standard photos into creative assets using preset templates. Key functionalities include generating professional headshots or personalized anime/cartoon styles by isolating subjects from the original images. Google plans to update the template library regularly. While basic templates are widely available, access to certain advanced features may require meeting specific eligibility criteria within the Gemini features of the Photos app. The rollout is commencing in several countries outside the US, including Japan, Brazil, and Mexico.

This technical paper addresses entropy collapse during Reinforcement Learning with Verifiable Rewards (RLVR) used for LLMs. This collapse, caused by policy entropy decay, leads to overconfidence and vanishing gradients. The research analyzes how Gradient-Preserving Clipping influences these dynamics. The authors propose a novel regulation mechanism utilizing dynamic clipping thresholds derived from theoretical verification of importance sampling ratio regions. This allows for precise entropy management. Evaluated strategies, including increase-then-decrease and oscillatory decay, effectively mitigate entropy collapse and achieve superior performance across various benchmarks, stabilizing the fine-tuning process.

A new systematic pipeline, Dr. SCI, has been introduced to overcome limitations in training LLMs for complex science questions. It utilizes a large dataset (1 million STEM questions) and redesigns the standard SFT/RL workflow using Exploration-Expanding SFT, Dynamic Difficulty Curriculum, and SciRubric-Guided RL. This approach stabilizes training on open-ended queries. Models fine-tuned with Dr. SCI demonstrated substantial, verifiable gains over strong baselines, including GPT-4o, specifically improving scientific reasoning capabilities on the GPQA benchmarks.

Asia's markets are outperforming US and European peers this year, attracting global investors despite extreme swings unsettling assets like tech stocks and metals. This trend indicates strong capital interest flowing into the regional AI sector.

Force Intelligence has launched DM0, touted as the world's first embodiment-native foundation model designed to enable robots to 'learn autonomously.' Unlike previous methods that adapted LLMs, DM0 was trained from scratch to deeply fuse multimodal internet data with physical sensory data, unifying perception and action. The model, featuring only 2.4 billion parameters, achieved first place in both categories of the RoboChallenge real-world robot tests, demonstrating significant efficiency and reducing reliance on backend computation. The full open-source release, alongside accompanying development tools, drastically lowers the barrier to entry, allowing fine-tuning even on consumer-grade GPUs. This move is positioned as a potential 'Android moment' for the industrialization of embodied AI development.

At ISE, Cisco introduced new collaboration hardware designed to function as manageable edge AI infrastructure. The flagship Room Kit Pro G2 integrates an Nvidia Jetson platform, offering a 25x increase in AI processing power over its predecessor, supporting up to seven 4K cameras and AVoIP. The Desk Pro G2 targets high-level remote presentation. Additionally, the rugged Phone 9821 for frontline environments features Wi-Fi 6E and Agentic AI capabilities. Cisco also previewed Control Hub integration, allowing IT admins to query device health via natural language tools like Copilot. Shipments for the Room Kit Pro G2 and Phone 9821 are slated for Q2 2026.

Insurify, a US online insurance comparison platform, launched an application integrated with ChatGPT allowing users to get insurance quotes conversationally, bypassing lengthy forms. The news immediately triggered a sharp decline in US insurance brokerage stocks on February 9th, with the S&P 500 Insurance Index dropping 3.89%. Major firms like Willis Towers Watson (-12.10%) saw their worst single-day performance since 2008. The tool aims to simplify the complex car insurance buying process by using natural language interaction powered by Insurify's data and ChatGPT's understanding capabilities. This event highlights growing market apprehension regarding AI's potential to disintermediate traditional insurance agents.

Online policy learning directly in the physical world poses significant challenges for embodied intelligence because real systems cannot be easily reset or accelerated like simulations. Researchers introduce USER, a Unified and extensible SystEm for Real-world online policy learning. USER treats physical robots as first-class hardware resources alongside GPUs through a unified hardware abstraction layer, enabling automatic discovery and management of heterogeneous robots. It incorporates an adaptive communication plane for robust cloud-edge collaboration and utilizes a fully asynchronous framework for long-horizon training with crash recovery. USER thus provides a critical systems foundation for real-world online policy learning, supporting large VLA models.

Anhui Agricultural University unveiled two major AI models on February 10, 2026: "Lixiang" for the pear sector and "Fengshu" for soybeans. Lixiang integrates petabytes of genomic data, literature, and Q&A pairs to optimize the entire pear value chain, focusing heavily on intelligent breeding. Fengshu targets soybean production, achieving over 90% accuracy in predicting critical traits like lodging resistance and protein content. It features a generative breeding platform supporting national food security goals by shifting agricultural development from experience-based methods to model-driven science.

Security firm LayerX revealed a CVSS 10/10 vulnerability in Claude Desktop Extensions (MCP Bundles) allowing zero-click Remote Code Execution (RCE). Attackers can embed malicious instructions in Google Calendar events, which Claude processes automatically, leading to unauthorized plugin execution with command privileges. Anthropic stated they do not plan to patch, arguing the security boundary rests with the user configuring local developer tools. This exposes a severe failure in the agent's ability to safely process external, untrusted inputs via integrated applications.

Facebook rolls out Meta AI-powered creative tools, including animated avatars that bring static profile pictures to life and AI-driven "Restyle" image editing for Stories. The updates also introduce dynamic backgrounds for text posts, aiming to re-engage younger users who are migrating to platforms like TikTok by injecting more visual flair and interactivity into the platform.

Researchers at the Chinese Academy of Sciences have successfully field-tested novel lithium-ion batteries in Mohe, Heilongjiang, solving the critical issue of performance degradation in extreme cold. The specialized batteries, which integrate proprietary low-temperature electrolytes, separators, and an advanced AI power management system, retained over 85% of their effective capacity after sitting idle for 8 hours at -34°C without external insulation. This breakthrough successfully powered industrial drones for extended missions, demonstrating reliable "plug-and-play" capability in harsh environments. The team previously reported success in controlling capacity fade below 10% relative to room temperature performance even at -40°C, significantly improving operational efficiency for Arctic and high-altitude missions.

Researchers introduced TermiGen, a system designed to overcome limitations in LLM performance for complex terminal tasks due to scarce high-fidelity environments. The pipeline generates executable environments and synthetic training trajectories that deliberately include common errors, mitigating the distributional mismatch seen in standard tuning data. Fine-tuned on this dataset, the TermiGen-Qwen2.5-Coder-32B model achieves a new state-of-the-art performance for open-weight models on the TerminalBench benchmark, outperforming established proprietary models on this specific coding challenge.

China is debuting its first feature-length animated film entirely powered by Generative AI, titled "Tuan Yuan Ling" (Reunion Order), set for release on February 28th. The project is guided by major state entities. The film utilizes AI for its visual creation and tells a story centered on the panda couple "Tuan Tuan" and "Yuan Yuan" gifted to Taiwan, aiming to foster cross-strait emotional connection. This deployment is positioned as a potential new era for AIGC integration in large-scale domestic animation production, building on existing IP developed over the past decade.

Amid escalating copyright litigation and data hunger in the AI sector, Amazon is reportedly planning to establish a new content trading marketplace via AWS. This platform would allow publishers to directly list their assets (articles, archives) for licensing to AI developers building models. This initiative aims to transition the acquisition of training data from a legal gray area to a transparent market structure, potentially integrating with services like Amazon Bedrock. The move arrives as publishers face severe traffic declines (up to 25-40%) due to AI summaries in search engines. Industry observers view this marketplace model as a more sustainable monetization strategy for content creators compared to previous one-off deals.

The next-generation Zhipu AI model, GLM-5 (codenamed "Pony Alpha"), was reportedly revealed, showing performance rivaling Claude Opus. The leak was confirmed via system prompts and specific bug "fingerprints." Technically, GLM-5 adopts DeepSeek's Sparse Attention Architecture (DSA), featuring 745B total parameters (with 8 active experts out of 256) and a 202K context window, crucially adding video understanding capabilities. This architectural reuse lowers deployment barriers by leveraging existing optimization frameworks. The announcement caused Zhipu's stock to surge 200%, reaching a market capitalization of 150 billion HKD.

A recent CCTV broadcast showcased interoperability testing involving Xiaomi devices, including a display functioning as a vehicle infotainment system. Analysts believe this indicates that Xiaomi's HyperOS vehicle operating system has successfully been ported to run on their proprietary Xuanjie (Surge) processor. This progress supports Xiaomi's long-term strategy, previously articulated by CEO Lei Jun, to integrate self-developed chips, operating systems, and large AI models on a single device by 2026.

xAI co-founders Jimmy Ba and Tony Wu announced their departures within 48 hours, signaling significant personnel instability at Elon Musk's AI startup. Ba, who previously reported directly to Musk, had responsibilities reassigned late last year. Five of the original twelve founders have now departed within the last year. The company has frequently been in the spotlight due to controversies surrounding its Grok chatbot, including publishing antisemitic remarks and generating non-consensual explicit imagery, leading to feature restrictions. This organizational shakeup follows Musk's announcement that xAI plans to merge with SpaceX.

Following the sharp decline in wealth management stocks caused by fears over an AI tax strategy tool threatening traditional advice models, JPMorgan Asset Management strategist Kerry Craig notes that trading opportunities still exist within the broader AI and technology sectors despite the recent market turmoil.

Following its initial public offering being derailed about five years ago, Ant Group, backed by Jack Ma, is shifting its primary business focus. The company is now heavily investing in healthcare powered by artificial intelligence, signaling a major strategic pivot to reignite growth in this expansive sector.

A recent test revealed that leading AI models, including ChatGPT, Gemini, Kimi, and DeepSeek, struggled to provide consistent or logical answers to a simple common-sense question: Should one walk or drive 50 meters to a car wash? While many models defaulted to walking based on environmental or efficiency reasons, their reasoning was inconsistent. Google's Gemini offered contradictory scenarios, while Elon Musk's Grok delivered a characteristically sharp response, suggesting driving was necessary for the car to actually be washed, mocking the others for suggesting walking. The exercise highlights ongoing deficits in practical reasoning for current large models.

The Czech ice dancing siblings faced scrutiny after using AI-generated music in their Olympic rhythm routine, violating the 90s theme. The track was criticized for having lyrics closely mimicking existing songs and featuring an AI voice emulating Bon Jovi, whose music isn't strictly 90s. While AI music generation is gaining commercial ground, its use in an event demanding creative expression raises questions about artistic integrity and copyright infringement in high-level competition.

Qijing Auto's first shooting brake coupe, priced around 300,000 RMB, is currently undergoing rigorous extreme cold testing ahead of its June launch. The vehicle utilizes a three-motor configuration featuring torque vectoring and comprehensive vehicle coordination, significantly enhanced by Huawei Qiankun's unique Digital Chassis technology. The platform is confirmed to integrate Huawei Qiankun's L3 conditional autonomous driving hardware architecture.

Microsoft has released the February cumulative update for Windows 11 (Build 26x00.7840). The large 4GB size is partially attributed to the bundling of local AI models like Phi, though they only activate on PCs with NPUs exceeding 40 TOPS. Key functional additions include the new Cross-Device Resume feature, designed to enable seamless task handover between Android phones (Samsung, Xiaomi, etc.) and the PC, mirroring Apple's Handoff. The update also brings native MIDI 2.0 support through a rebuilt Windows MIDI Service and improves user experience with a reorganized Start Menu view and new battery icon. Critically, this patch resolves known stability issues, specifically fixing black screen crashes affecting Nvidia GPUs from the January update, and allowing users to disable Smart App Control more easily.

Investors on Wall Street are actively selling off shares of companies they perceive as being negatively impacted by the rise of artificial intelligence. This aggressive de-risking affects a wide range of sectors, including smaller software enterprises and large wealth management firms, as market participants adjust valuations based on AI disruption risk.

Ryan Beiermeister, the Vice President leading OpenAI's product policy team, has been terminated following allegations of sexual discrimination. Beiermeister was reportedly an executive who had opposed the development of OpenAI's controversial 'Adult Mode' feature. This firing highlights internal governance and ethical conflicts within the leading AI firm, demonstrating repercussions for conduct despite policy stances.

Contrary to the narrative that AI tools serve as productivity amplifiers that free up time, new research published in Harvard Business Review suggests the opposite is occurring in practice. Researchers tracking a 200-person tech company over eight months found that employees proactively filled any time saved by AI assistance with additional tasks. This led to an expansion of to-do lists, encroachment on lunch breaks and evenings, and increased pressure to perform faster. One engineer noted that work expectations tripled while actual productivity only rose slightly. The study concludes that capability gains, rather than leading to leisure, are channeled into increased output demands, fueling fatigue and a pervasive feeling of being unable to disconnect, especially as management pressures teams to justify AI investments.

Honor has initiated the phased rollout of the "Geek Center" feature for its MagicPad 3 Pro (Geek Center Edition). This update allows users to manually adjust performance parameters, mirroring gaming laptop controls. Users can now customize settings like SoC frequency, scheduling, and enable 165Hz high refresh rate for supported games. The update also includes optimization for touch controls and integrates the February 2026 Android security patch. While users requested customizable thermal control, Honor confirmed it is under analysis for future implementation.

A New Yorker article profiles Anthropic, focusing on key executives like Chris Olah and providing insight into internal research efforts. These include "Project Vend," an experiment using their Claudius system to control office infrastructure, illustrating the company's deep dive into understanding and mapping the internal "neurons" or mechanisms of their advanced AI models.

Robert Playter is immediately stepping down as CEO of Boston Dynamics, effective February 27th. During his six-year tenure, Playter managed the company's acquisition from Softbank to Hyundai in 2021 and oversaw product launches.

DuckDuckGo announced the addition of real-time voice chat to its Duck.ai platform, emphasizing extreme privacy. Unlike other assistants, the core feature is a privacy-first architecture. Users converse with the LLM via encrypted channels without fear of monitoring. DuckDuckGo acts as a middleman, using OpenAI for intelligence but implementing a firewall. Strict contracts ensure DuckDuckGo anonymizes audio and OpenAI only processes requests without retaining data for training. Details confirm temporary processing, immediate destruction of audio streams post-conversation, zero retention by either party, and end-to-end encryption via WebRTC. The feature is free without registration (daily limits apply), with a $10/month subscription available for higher limits and additional services.

Czech ice dancers Katerina and Daniel Mrazek used AI-generated music during their Olympic rhythm dance, sparking controversy despite not violating current ISU rules. Part of their routine incorporated music reportedly generated by AI mimicking the style of Bon Jovi, including lyrics closely resembling their work. The AI track was mixed with AC/DC’s "Thunderstruck." This high-profile usage highlights the ongoing regulatory gap concerning AI-generated content, especially when training data overlaps significantly with copyrighted material, putting immediate pressure on sports governing bodies to clarify guidelines around AI composition.

A new report indicates that Amazon is considering launching a dedicated marketplace to facilitate the licensing of proprietary media content directly from publishers to AI companies. This initiative aims to create a standardized pipeline for acquiring training data, potentially solving major friction points in data provenance and copyright clearance that currently plague large-scale LLM development. While speculative, establishing such a structured market by Amazon would significantly impact the data supply chain dynamics for generative AI.

OpenAI is deploying an update to ChatGPT's deep research tool, introducing a full-screen document viewer. This feature allows users to scroll and navigate specific sections within AI-generated reports in a dedicated window separate from the main chat interface.

Google has launched its Fitbit AI Coach feature, utilizing the Gemini model, as a public preview for Fitbit Premium subscribers on iOS in several key markets (US, UK, etc.), expanding from Android. This service acts as a 24/7 virtual health advisor, providing personalized workout plans, deep sleep analysis, and health insights based on user-defined goals after an initial consultation. It monitors vital signs and includes chatbot support. This deployment signals the increasing integration of generative AI into consumer health hardware ecosystems, occurring as reports suggest Apple may have shelved plans for a comparable AI health feature.

OpenAI is rolling out updates to the ChatGPT Deep Research Tool, reportedly now powered by GPT-5.2. New features include a full-screen viewer allowing users to navigate generated reports via a dynamic table of contents linked to sources. Users gain control to restrict research scope to specific websites or apps, monitor progress in real-time, and download final reports in Markdown, Word, or PDF formats. The update is rolling out immediately for Plus and Pro subscribers.

ByteDance has temporarily suspended the personal voice reference capability in its Seedance 2.0 video model after reports indicated it could generate highly personalized voices from just a single facial image, raising significant privacy authorization concerns. Separately, hardware specs for the next-generation Xiaomi SU7 confirm the integration of a 700 TOPS AI driving chip and high-voltage platforms, promising major performance gains. In finance, Microsoft’s P/E ratio dropped below IBM’s for the first time since 2013, signaling market apprehension regarding the massive capital expenditure required to sustain the AI arms race. Huawei also confirmed that future large model AI capabilities will be delivered via OTA updates to existing vehicle owners.

Czech ice dancers Katerina Mrazkova and Daniel Mrazek discovered the hard way that Large Language Models can generate outright plagiarism. Their performance used AI-generated music, leading to immediate scrutiny and accusations that the content was copied from existing works. This incident serves as a cautionary tale regarding the use of generative AI in high-stakes professional arts.

Discord has clarified that the vast majority of users can continue using the platform exactly as they do today without needing to submit to mandatory face scans or ID verification for age checks. The system will rely on its internal AI to predict if users are old enough, or users simply won't be accessing age-restricted experiences. This limits intrusive verification methods to a smaller subset of users.

Despite executive claims of rapid growth, only 3.3% of Microsoft 365 subscribers are paying for Copilot. This low conversion rate introduces scrutiny regarding the actual long-term return on investment for Microsoft's substantial AI initiatives.

Researchers introduced Reasoning-Augmented Representations to fix issues where multimodal retrieval embeddings fail complex reasoning tasks. The technique decouples reasoning from compression by using VLMs to enhance corpus data (dense captioning) and rewrite queries before retrieval training. This data-centric approach shows consistent performance gains on M-BEIR benchmarks by providing semantically dense input signals, overcoming distribution shift problems. Code is available.

Senators Elizabeth Warren and Jim Banks are reportedly drafting bipartisan legislation aimed at prohibiting the sale of specific high-end AI processing chips to China. This initiative follows recent discussions between lawmakers and AI industry leaders, including Anthropic CEO Dario Amodei, concerning national security and technology transfer risks related to advanced AI hardware development overseas. The move signals increasing political friction over crucial hardware supply chains.

AI startup Flapping Airplanes secured $180 million in seed funding from major investors including Google Ventures, Sequoia, and Index. The lab aims to develop models that learn similarly to humans, departing from the current paradigm of simply processing vast amounts of internet data. The team includes founders Ben and Asher Spector and Aidan Smith.

Google announced the public preview of Fitbit's AI personal health coach is rolling out to iOS users in the U.S. and to iOS/Android users across the UK, Canada, Australia, New Zealand, and Singapore. It was initially launched in October exclusively for US Android users.

Wall Street dealers are raising compensation requirements for trading corporate bonds issued by private credit funds. This shift stems from growing investor nervousness regarding these funds' exposure to software firms currently facing disruption due to advancements in artificial intelligence. Consequently, the cost structure for trading these specific private credit instruments is being adjusted upward due to perceived risk related to AI impact on the underlying borrowers.

A website promoted by RFK Jr., Realfood.gov, utilizes Elon Musk's Grok chatbot to provide nutritional recommendations. Analysis shows that the information generated by Grok often conflicts with established government dietary guidelines, highlighting potential risks when deploying large language models in sensitive informational contexts.

Significant internal instability appears to be occurring at Elon Musk's AI venture, xAI. Reports indicate that five of the twelve individuals who constituted the original founding team have now left the company. Only seven members of the initial group, including Musk himself, remain affiliated with the organization. This high rate of departure among core personnel raises questions regarding the project's stability and execution timeline.

Blackstone is boosting its investment in AI firm Anthropic PBC, raising its total stake to approximately $1 billion based on the startup's current valuation. This significant capital reinforcement comes as Anthropic continues to develop its large language models.

Bloomberg analysts discuss Spotify's record user growth following its 'wrapped' campaign. Additionally, Paramount has strengthened its acquisition offer for Warner Bros. Discovery, covering significant financial obligations. The segment also features Runway CEO Cristobal Valenzuela detailing a new funding round that places the generative video company's valuation at $5.3 billion.

Tech visionary John Carmack has proposed a significant architectural shift where fiber optic cabling could replace traditional volatile RAM specifically for large-scale AI workloads. This idea aims to alleviate the critical memory bottleneck that currently constrains the scaling of AI models. While the implementation path remains speculative, the proposal targets a fundamental infrastructural limit restricting next-generation AI capability.

Seedance 2.0, an AI video generation tool associated with TikTok, has launched. It is generating buzz for producing video quality comparable to OpenAI's Sora 2 and Kling 3.0. However, the tool's debut comes with specific limitations or conditions noted for users in the United States.

Research indicates that the integration of AI tools in the workplace may paradoxically lead to increased working hours and job dissatisfaction, suggesting a risk of AI-driven burnout if adoption is poorly managed.

Anchored Decoding is a novel, inference-time method designed to suppress verbatim copying from Language Models trained on sensitive or copyrighted data. It functions by keeping the generation trajectory close to a permissively trained safe model, enabling a tunable risk-utility trade-off. The technique, demonstrated with the new TinyComma 1.8B model, successfully eliminated up to 75% of the measurable copying gap compared to a safe reference, achieving this mitigation at a modest inference overhead while preserving high fluency.

Abu Dhabi’s sovereign investment fund, MGX, is reportedly close to finalizing a deal to participate in Anthropic PBC’s latest funding round. Sources familiar with the matter indicate this potential investment deepens MGX's commitment to one of the most prominent artificial intelligence startups. The funding round is valued at $20 billion.