1841 articles
Researchers from Peking University and Tsinghua University introduced WaveFormer, a new vision architecture that models image features as wave-like signals, achieving 1.6 times higher throughput with 30% fewer computations than standard Vision Transformers.
A January 2026 governance framework establishes principles for managing risks and accountability in autonomous AI systems as their adoption grows across industries.
Memory chip shortages are squeezing the AI industry hard, forcing major tech companies to cut production while prices skyrocket across the board.
Nvidia's CEO praised Anthropic's Claude as a breakthrough in coding and reasoning capabilities, revealing widespread internal use at Nvidia and calling it essential for software companies.
A new breakdown of agentic AI reveals that autonomy comes from system architecture, not just model power. The framework maps AI evolution across five distinct layers—from basic machine learning to fully governed autonomous systems.
Researchers developed Think3D, a new framework that lets vision-language models understand and reason in three-dimensional space through scene reconstruction, delivering measurable accuracy improvements without requiring model retraining.
Resemble AI released Chatterbox-Turbo, an open-source text-to-speech model delivering real-time voice generation on a single GPU with sub-200ms latency and significantly lower compute costs.
ByteDance just dropped a diffusion-based code generation model that's turning heads with benchmark scores above 83 on HumanEval. The real kicker? It's blazing fast and comes with an MIT license for maximum accessibility.
A newly released Beacon Object File enables command execution inside Windows Subsystem for Linux without launching the wsl.exe process. The technique operates entirely in memory, reducing visibility for endpoint detection tools.
Tesla is getting ready to train its Optimus humanoid robots at the Austin manufacturing plant. The initiative builds on training efforts already running in California.
Qwen3-TTS has been released as an open-source, real-time text-to-speech system built on the Qwen3 language model. Benchmark results show strong multilingual performance and high speaker similarity across multiple languages.
AI data centers are driving massive spikes in electricity and water consumption alongside trillion-dollar investments. The scale now rivals traditional heavy industry, raising critical questions about grid capacity and resource sustainability.
GPT-5.2 Pro just scored 31% on the brutal FrontierMath Tier 4 test—a 12-point jump over anything we've seen before. It's the clearest sign yet that AI is getting genuinely better at advanced math, not just memorizing patterns.
Scientists from UC Berkeley, UMD, and the University of Toronto unveiled MomaGraph, a breakthrough vision language model that helps robots understand their surroundings and plan tasks more effectively. The system beat all open-source competitors on a specialized benchmark test.
AI agents are increasingly operating beyond the visibility of traditional identity and access management systems, creating new security challenges where valid access gets misused rather than stolen.
A recent overview examines eight Vision-Language-Action models driving embodied AI forward, featuring ChatVLA-2's Mixture-of-Experts architecture and Microsoft's newly launched Rho-alpha model.
Columbia Engineering researchers have revealed a humanoid robot capable of highly realistic lip-syncing, featured on the January cover of Science Robotics. The system learns facial movements by observing humans, aiming to reduce the long-standing "uncanny valley" effect.
Ollama's latest update brings a unified launch command for multiple AI coding models and optimizes GLM 4.7 Flash to handle 64,000+ token contexts with reduced memory consumption.
Researchers at Switzerland's EPFL have created a robotic hand that can detach from its arm, crawl independently to reach objects, and reconnect afterward—expanding what robots can do beyond traditional designs.
ModelScope's DASD-4B-Thinking model topped multiple reasoning benchmarks with 88.5% accuracy on AIME24, 83.3% on AIME25, 69.3% on LiveCodeBench v5, and 68.4% on GPQA, marking significant progress in open-source reasoning capabilities.
We use cookies to improve your experience on our site and to show you relevant advertising. To find our more, read our privacy policy and cookie policy
Usman Salis
Saad Ullah
Eseandre Mordi
Sergey Diakov
Alex Dudov
Victoria Bazir
Marina Lyubimova
