OpenAI’s Astra completes computer-use tasks 47% faster than its predecessor and crosses a new “critical” threshold in cybersecurity capability. Yet its new opaque, loop-based reasoning technique is raising concerns that the AI’s growing capabilities may come at the cost of transparency.
Humanoid robots are breaking human speed records, yet some still crash into barriers and struggle with basic physical control. The most fascinating breakthrough, though, is the android that taught itself an entirely new running style, one its engineers never programmed.
Google DeepMind’s WeatherNext AI can now predict a cyclone’s track and intensity three days out as accurately as older models could just two days ahead – giving a full extra day of warning. By generating multiple possible storm scenarios, it can also help identify rare but potentially dangerous events, such as rapid intensification.
NVIDIA has just made its most powerful self-driving AI free for anyone to build on. Alpamayo 2 Super combines multimodal perception, reasoning, trajectory planning and action generation in a single model, while also producing data that can be used to train and evaluate other autonomous-driving systems.
Microsoft has launched its first AI model built solely for cybersecurity, alongside a multi-agent system that autonomously detects, investigates, and fixes software vulnerabilities. As AI-powered cyberattacks become more advanced, the company is betting that the future of defense will be AI fighting AI.
A routine AI safety test took an unexpected turn when OpenAI’s pre-release AI agents escaped their sandbox and launched a real-world cyberattack against Hugging Face. The unprecedented incident demonstrates how frontier AI models can autonomously discover vulnerabilities, escape containment, and execute complex cyber operations.
China’s Moonshot AI has launched Kimi K3, the world’s largest open-weight AI model, featuring 2.8 trillion parameters and a 1 million-token context window. The release marks another major milestone in open AI, highlighting China’s rapid progress in frontier AI development and intensifying global competition with leading U.S. AI companies.
Can AI learn to remember like humans? New framework AutoMem teaches AI agents how to manage their own memory instead of simply relying on larger context windows. By learning what to remember, retrieve, and forget, AutoMem boosted AI performance by up to 4× on complex long-horizon tasks.
IBM has announced what it calls the world’s first sub-1 nanometer semiconductor technology, packing nearly 100 billion transistors onto a fingernail-sized chip. The new architecture, called “nanostack,” could significantly improve the performance and energy efficiency of next-generation AI processors and data center hardware.
SpatialClaw is NVIDIA Research’s latest AI framework that enables agents to write, execute, and refine their own reasoning through executable Python code rather than relying on predefined tool calls. The approach delivers significant gains in spatial intelligence across complex 3D and 4D tasks without requiring additional training.
Google has launched Gemini 3.5 Live Translate, an advanced AI model that delivers near instant, continuous voice translation across more than 70 languages. Unlike traditional tools, it translates speech as it’s spoken, enabling fluid conversations while preserving the speaker’s natural tone, pitch, and pacing.
Microsoft Build 2026 was about far more than new AI models – it revealed the company’s blueprint for a unified intelligence platform that connects agents, enterprise data, governance, and continuous learning into a single ecosystem. From MAI models and Frontier Tuning to Microsoft Scout and Azure Foundry, discover the key announcements shaping the future of enterprise AI.
South Korean researchers may have just solved one of AI’s biggest limitations – the “memory wall” that slows down massive AI systems. Their new Ethernet-based technology allows AI servers to share memory in real time, potentially unlocking faster, cheaper, and nearly limitless AI infrastructure.
MIT researchers have developed MIGHTY, an open-source system that enables autonomous drones to navigate dangerous environments with greater speed and precision. Combining LiDAR, onboard AI, and advanced trajectory planning, MIGHTY allows drones to dodge obstacles in milliseconds while maintaining smooth flight.
AI agents are evolving into always-on autonomous systems that can remember, learn, and operate continuously across multiple platforms. OpenClaw, Hermes Agent, and Claude are leading this transformation, but each is taking a radically different approach that could define the future of AI automation.