AWS

MIT Unveils Dynamic Reasoning Method to Boost LLM Efficiency Researchers at MIT have engineered a more intelligent method for large language models (LLMs) to allocate computational resources during reasoning, significantly increasing their efficiency. This new technique enables LLMs to dynamically adjust the amount of computation used based on the complexity of a given question. This approach marks a departure from common methods that assign a fixed computational budget to every problem, which often wastes resources on simple queries and fails to solve more complex ones. By enhancing the reliability and efficiency of LLMs for complex reasoning, this development could lower the energy consumption of generative AI systems and enable their use in more critical, time-sensitive applications. The research is being presented at the Conference on Neural Information Processing Systems. ...

Hyundai Launches MobED: An AI-Powered Autonomous Robot Platform At the International Robot Exhibition (iREX) 2025 in Tokyo, Hyundai Motor Group has officially launched MobED (Mobile Eccentric Droid), its first mass-produced autonomous mobility robot. First revealed as a concept at CES 2022, MobED has evolved into a production-ready platform powered by artificial intelligence. The robot is engineered for a wide range of industrial and everyday applications, featuring AI-based autonomous navigation, LiDAR-camera fusion sensors, and an eccentric posture control mechanism for stable movement across diverse terrains. An intuitive 3D UI/UX on a wide touchscreen controller enables easy operation, including self-mapping and autonomous driving. Mass production of MobED is set to begin in the first half of 2026. ...

AWS re:Invent 2025: Autonomous AI Agents, Nova 2 Models, and Trainium 3 Processors At its re:Invent 2025 conference, Amazon Web Services (AWS) announced a suite of new generative AI capabilities, including the introduction of ‘frontier agents’ designed to handle complex, multi-day projects without constant human intervention. These autonomous agents include: Kiro: A virtual developer for Amazon’s coding platform. AWS Security Agent: An agent for testing application vulnerabilities. AWS DevOps Agent: An agent designed to respond to system outages. To support these advancements, AWS launched Amazon Bedrock AgentCore, a platform for building and deploying AI agents at scale, and announced that Strands Agents now support edge devices for use cases in automotive, gaming, and robotics. ...

Perplexity’s Comet AI Browser Faces Serious Security Vulnerability Allegations Perplexity’s Comet AI browser is facing significant scrutiny over alleged security vulnerabilities. Security researchers from SquareX claim to have discovered a hidden API that could allow attackers to execute arbitrary code on a user’s device, potentially leading to data theft or a full system compromise. Perplexity has vehemently denied these claims, labeling them as “fake news” and stating that no such vulnerability exists. This controversy follows previous security warnings from Brave and Guardio regarding other issues with Comet, including prompt injection vulnerabilities that could manipulate the AI into performing unintended actions. ...

Galbot Unveils DexNDM & NavFoM: Major Breakthroughs in Embodied AI for Robotics Robotics firm Galbot has announced two significant technological advancements in embodied intelligence: DexNDM and NavFoM. DexNDM is a neural dynamics model designed to enhance robotic dexterous manipulation, enabling robots to perform high-precision, in-hand rotations of various objects. NavFoM is presented as the world’s first cross-embodiment, cross-task navigation foundation model, allowing robots to autonomously perform complex tasks in diverse environments. These innovations, developed with Tsinghua and Peking University, integrate DexNDM into a teleoperation system for intricate tasks like tightening screws with high-level human commands. These advancements are poised to revolutionize industries like manufacturing, logistics, and healthcare by improving the adaptability and efficiency of robots in real-world settings. ...

OpenAI CEO Sam Altman Pushes for Government-Owned AI Infrastructure OpenAI CEO Sam Altman has called for governments to invest in and own their artificial intelligence infrastructure. In a statement, Altman clarified that OpenAI is not seeking government loan guarantees for its data centers, emphasizing that companies should bear their own business risks. He suggested that a government-owned AI infrastructure, including a “strategic national reserve of computing power,” would serve the public interest. This comes as OpenAI faces scrutiny over its significant investment plans and the immense costs associated with advancing AI technology. Altman believes the risk of having insufficient computing power is greater than the risk of overbuilding. ...

AI Breakthrough: New Method Predicts and Prevents Biological Invasions Researchers have developed a groundbreaking AI-based methodology to more accurately predict which plant species are likely to become invasive, aiming to prevent biological invasions before they occur. By combining decades of ecological data with advanced machine learning algorithms, the new models achieve over 90% accuracy in predicting invasion success. This data-driven approach is designed to remove biases from ecological risk assessments and increase their predictive power, helping authorities identify high-risk species before importation. Initially focused on Caribbean islands, the research will expand by training the models with data from other global regions to enhance their applicability. ...

NVIDIA Unveils RLP: A New Era of Reinforcement Learning for AI Pretraining NVIDIA researchers have developed a groundbreaking training objective called Reinforcement Learning Pretraining (RLP), which integrates reinforcement learning directly into the pretraining stage of AI model development. The core concept of RLP treats a short chain-of-thought as an action that is rewarded based on the information gain it provides for the model’s next-token prediction. This innovative method creates a verifier-free and dense reward system that can be applied at scale to standard text streams. Unlike previous methods that depend on sparse signals or external verifiers, RLP’s approach allows for updates at every single token position. Experiments demonstrate that RLP can yield significant improvements in a model’s reasoning capabilities, with one model showing a substantial increase in its average score on scientific reasoning benchmarks. ...

OpenAI & Broadcom Announce Landmark 10GW Custom AI Accelerator Partnership OpenAI and Broadcom have announced a strategic multi-year collaboration to develop and deploy 10 gigawatts of custom AI accelerators. Under this partnership, OpenAI will lead the design of the AI accelerators and systems, while Broadcom will be responsible for their development and deployment. This collaboration will also involve the co-development of systems incorporating Broadcom’s Ethernet and other connectivity solutions to enhance the scalability of next-generation AI clusters. The deployment of these AI accelerator and network system racks is scheduled to commence in the latter half of 2026, with a target completion by the end of 2029. Sam Altman, CEO of OpenAI, stated that this partnership is a crucial step in building the necessary infrastructure to unlock the full potential of AI. ...

Weekly Tech News: AI, Robotics, Autonomous Vehicles & Cloud Innovations Google’s Gemini 2.5 AI Wins Gold Medal in Global Programming Contest A specialized version of Google DeepMind’s Gemini 2.5 AI model, known as Deep Think, has achieved a gold medal-level ranking at the prestigious International Collegiate Programming Contest (ICPC). Held in Baku, Azerbaijan, the competition featured 139 of the world’s top university programming teams. Gemini 2.5 Deep Think successfully solved 10 of the 12 complex algorithmic problems, placing it second overall compared to the human teams. ...