Category: World

For years, Alibaba and Tencent have dominated China’s internet landscape. Their investment strategies shaped the market, with each company vying for leadership in a highly competitive environment. Investors previously viewed them as nearly equal players in the tech ecosystem.

Recent enthusiasm for Asian chipmakers has disrupted this balance. Alibaba’s ambitious foray into semiconductors is generating renewed interest from investors. Meanwhile, Tencent struggles to capture the same level of excitement.

Alibaba’s stock has significantly outpaced Tencent’s in the wake of this shift. The company’s dedicated efforts in semiconductor technology are resonating with market players. As a result, Alibaba’s shares are climbing, reflecting stronger investor confidence.

This divergence carries serious implications for both companies. Alibaba’s growth in the semiconductor sector could solidify its market position. In contrast, Tencent may face increasing pressure to pivot, risking its long-standing influence in the tech industry.

Approval-based committee voting systems have garnered considerable interest due to their potential for proportional representation. Thiele rules, particularly Proportional Approval Voting (PAV), feature prominently in discussions because of their appealing properties like Pareto optimality. However, calculating outcomes under these rules has remained a significant hurdle due to NP-hard complexity.

A breakthrough has emerged with new findings that address the long-standing complexity issue in the voter interval (VI) domain. While earlier approaches using linear programming (LP) faced setbacks, researchers have now established that an optimal integral solution is obtainable even when the constraint matrix fails to be totally unimodular. A novel algorithm has been introduced to compute these solutions efficiently.

This newly discovered technique not only applies to the VI domain but also extends to the voter-candidate interval (VCI) and linearly consistent (LC) domains. The investigation revealed crucial insights into the relationship between VCI and LC, leading to the conclusion that LC strictly includes VCI. A fresh definition of LC has been proposed, enhancing its relevance to approval elections.

The implications of this advancement are profound. By establishing a more efficient computational method for Thiele outcomes, the research could reshape how social choice theorists approach complex voting scenarios. As these methods find application, they may facilitate more democratic and effective decision-making processes in various approval-based elections.

Recent research has unveiled a pressing gap in the capabilities of large language models (LLMs) regarding creative reasoning. While these models excel at reasoning tasks, their ability to repurpose tools creatively remains largely untested. The introduction of CreativityBench aims to address this deficiency, marking a significant shift in how AI creativity is evaluated.

CreativityBench sets out to benchmark affordance-based creativity by creating a comprehensive knowledge base. This resource features over 4,000 entities and more than 150,000 affordance annotations. The project generates 14,000 tasks that challenge LLMs to find innovative uses for objects based on their physical properties rather than their traditional applications.

Initial evaluations across ten leading LLMs indicate that while models can occasionally identify plausible objects, they struggle with pinpointing the correct parts and their associated affordances. As a result, performance in solving tasks plummets. Notably, enhancements from model scaling appear to plateau quickly, and common strategies like Chain-of-Thought yield minimal improvements.

These findings underscore a critical hurdle in advancing AI creativity, even with state-of-the-art models. The establishment of CreativityBench not only sheds light on this vital aspect of intelligence but also has significant implications for future AI development. As researchers continue to explore these challenges, the potential for more versatile and innovative agents could reshape various applications.

Machine learning has long relied on externally imposed regime switches, a limitation that has hindered the emergence of autonomous systems. The current landscape predominantly features scalar-reducible dynamics, which simplify decision-making through clear, gradient-driven processes. This conventional framework restricts the potential for true self-directed learning.

Recent research introduces a groundbreaking classification that distinguishes between scalar-reducible and scalar-irreducible dynamics. This new approach reveals that scalar-irreducible dynamics can facilitate internal regime switching. By leveraging feedback between fast-moving variables and slower structural changes, systems can adapt without relying on external schedules.

The study employs a minimal dynamical model to illustrate how these internally driven transitions occur. This mechanism allows for the sustained adaptation of systems in unpredictable environments. The findings demonstrate a significant shift toward a new paradigm that encourages autonomous behavior in learning systems.

As autonomous intelligence progresses, these insights could revolutionize machine learning frameworks. By enabling systems to govern their own dynamics, researchers could open doors to more advanced, self-sustaining learning applications. The implications for industries such as robotics and AI-driven decision support are profound, promising a future where machines learn in ways previously thought impossible.

Traditionally, representation learning has focused on creating meaningful sensory representations through unsupervised methods. This domain aims to model elements akin to human cognitive development, yet defining what constitutes a “good” representation has proven challenging. Researchers have long sought effective ways to enhance the learning process and improve model performance.

Recent work introduces a shift in approach. By utilizing Parameter Division within the framework of Group Decomposition Theory, the need for auxiliary assumptions has been eliminated. This new method analyzes transformations between input pairs more effectively by focusing on imposed constraints, thus addressing limitations seen in earlier attempts.

The study demonstrates that by splitting transformation parameters, it identifies normal subgroups with greater precision. Evaluations of the new method on image pairs subject to rotation, translation, and scale show significant advancements. Results indicate that group-decomposition constraints greatly enhance categorization accuracy and efficiency.

This innovative approach could reshape the landscape of machine learning and representation categorization. The absence of auxiliary assumptions allows for broader applications across various fields. Potential ramifications include better understanding of human-like learning mechanisms and improved performance in tasks requiring unsupervised learning.

In the realm of machine learning, standard adaptive optimizers like AdamW have long been the backbone of efficient training. These optimizers apply uniform hyperparameters across all model parameters, simplifying the tuning process. However, this approach often overlooks the unique dynamics associated with different layers and modules.

The introduction of MetaAdamW marks a significant shift in this paradigm. By utilizing a self-attention mechanism, this optimizer adjusts learning rates and weight decay for distinct parameter groups dynamically. It employs a lightweight Transformer encoder to analyze various statistical features of each group, allowing for targeted and efficient optimization.

Extensive experiments across five diverse tasks confirm its potential. MetaAdamW consistently surpasses the performance of AdamW, offering reductions in training time and improvements in accuracy or perplexity. Notably, it can enhance convergence rates and alleviate issues caused by premature early stopping, all while maintaining manageable overhead.

The implications of this advancement are considerable. By tailoring optimization strategies to individual parameter groups, MetaAdamW empowers researchers and practitioners with enhanced tools for tackling complex machine learning challenges. The optimizer represents a leap forward in making more nuanced, efficient training accessible in various applications.

Machine learning has long grappled with data that resides on complex, curved spaces. Traditional methods often struggled with the distortions introduced when applying Euclidean geometry to such problems. Researchers now face challenges in scaling these techniques efficiently across diverse manifolds.

The introduction of Entropic Riemannian Neural Optimal Transport (Entropic RNOT) marks a significant shift. This new framework merges intrinsic entropic optimal transport with out-of-sample evaluation on Riemannian manifolds. By utilizing a neural pullback parameterization, the method constructs a target-side Schrödinger potential, aiming to enhance the accuracy of distance and transport calculations.

As a result, Entropic RNOT develops barycentric projections and heat-smoothed surrogates, transforming atomic target laws into continuous ones. The framework shows strong theoretical guarantees, with convergences in essential probabilistic metrics and stability in practical applications. Empirical evaluations have demonstrated its effectiveness, often surpassing benchmarks set by existing techniques.

This advancement has profound implications across various fields, including robotics and computational biology. Notably, its application in protein-ligand docking has highlighted its efficiency, adjusting poses without the need for extensive retraining. The integration of these methods signals a promising new direction for addressing complex data challenges in machine learning.

In today’s digital landscape, security operations centers (SOCs) face constant pressure to protect networks from sophisticated cyberattacks. Cyber defenders rely on traditional methods to configure endpoint detection and response policies, which often fall short under real-time adversarial conditions. The need for advancement in autonomous cyber defense systems has never been higher.

A new tool-mediated architecture has emerged, integrating large language model (LLM) agents with deterministic tools to enhance decision-making amid threats. This innovative approach employs strategies such as Stackelberg best-response and attack-graph primitives, granting SOCs improved capabilities to operate under duress. Research shows these systems provide formal guarantees that traditional methods lack, fundamentally altering how cyber defenses are managed.

Testing on 282 enterprise attack graphs demonstrated significant improvements in performance. Using the Claude Sonnet 4 controller, the approach reduced the attacker’s expected payoff by 59% compared to existing deterministic methods. Even with varied conditions, this controller maintained consistent stability across multiple trials, underscoring the effectiveness of the new architecture in real-world scenarios.

The implications of this research extend beyond enhanced defense mechanisms. By allowing LLM agents to navigate creative strategies while maintaining system stability, organizations can better adapt to the evolving landscape of cyber threats. As SOCs integrate these findings, the future of autonomous cyber defense appears not only promising but also essential for safeguarding digital environments.

Public perception of artificial intelligence has faced challenges, primarily due to its diverse and complex nature. Traditional methods often oversimplify these views by using a single dependency graph, failing to capture the varying attitudes across different demographic groups. This limitation has left researchers seeking more nuanced understanding of sentiment towards AI.

Recent advancements have sparked a breakthrough in evaluating AI attitudes through heterogeneous ordinal structure learning. A novel framework has been introduced, utilizing Bayesian nonparametric complexity discovery combined with confirmatory fixed-K estimation. This methodology allows for the identification of distinct archetypes in public attitudes, rather than relying on generalized models.

In a study conducted on the 2024 Pew American Trends Panel AI attitudes survey, researchers implemented this new framework on nearly 4,800 respondents. The results were compelling, with a 25.8% reduction in mean squared error when compared to conventional single-graph analyses. This framework not only enhanced prediction accuracy but also offered interpretable insights into the complex landscape of AI perceptions.

The implications of this research could reshape how policymakers and technologists engage with public sentiment regarding AI. By understanding the intricate nuances of attitudes through this advanced approach, stakeholders can tailor their strategies to better resonate with diverse audience segments, ultimately fostering a more inclusive dialogue on technological advancements.

Moonshot AI, the company behind the popular Kimi chatbot, has become a significant player in the tech landscape. The firm has long been recognized for its innovative artificial intelligence solutions tailored for various industries. However, recent developments have now thrust it into a new realm of valuation and investment interest.

In its latest funding round, Moonshot AI secured approximately $2 billion, backed by investors including Meituan. This influx of capital signifies a notable shift in the market, as venture capital increasingly flows toward Chinese technology startups looking to compete with established leaders in Silicon Valley. The excitement around the funding is reflective of broader investor confidence in the potential growth of China’s AI sector.

Following this funding, Moonshot AI’s valuation skyrocketed to $20 billion. This leap not only enhances the company’s financial stability but also increases its capacity for research and development. With more resources, the firm is expected to expand its product offerings and enhance its AI technologies further.

The immediate impact of this funding is evident in the heightened competition within the tech industry. As Moonshot AI bolsters its position, it may inspire similar startups to innovate aggressively. This development could reshape the landscape of AI technology, challenging existing players and ultimately benefiting consumers through improved services.