Meta AI’s DINOv3 is a self-supervised vision model trained on 1.7 billion images, setting new standards in image classification, object detection, and beyond. With innovations like Gram anchoring and real-world impact from monitoring deforestation to powering NASA’s Mars exploration, it marks a paradigm shift in computer vision.
The new system enables groups of robots to act as a unified team. The MultiRobot FrameWork lets robots share real-time information about their environment, positions, and tasks, mirroring the collective behavior seen in insect colonies, but powered by advanced sensors and computation.
The new GenSeg framework significantly reduces the need for expert-labeled data and achieves high-accuracy medical image segmentation with as few as 40-50 samples. By creating realistic synthetic scans paired with exact labels, it empowers the development of advanced diagnostic tools even in data-limited settings.
OpenAI's newest open-weight models gpt-oss-120b and gpt-oss-20b bring advanced reasoning and 128K-token context windows – all under Apache 2.0 license. With support for local deployment and optimization for consumer hardware, these models mark a major shift toward transparent and decentralized AI.
Skydweller is a solar-powered drone designed for long-endurance missions. With AI-powered radar, self-healing systems, and the ability to fly autonomously for up to 90 days straight, it sets a new standard for persistent aerial surveillance and data collection.
New research reveals that LLMs like GPT-4o and Gemma 3 often stick to their initial answers even when wrong – yet quickly lose confidence when challenged. This surprising mix of overconfidence and self-doubt mirrors human cognitive biases and raises concerns about AI reliability.
MIT scientists explored a critical flaw in AI language models called position bias, where models favor information at the beginning and end of text while ignoring the middle. Their research reveals this bias is rooted not only in the training data, but also in the architecture of the models themselves.
A self-powered artificial synapse can mimic human color vision with 10-nanometer resolution using dye-sensitized solar cells. This technology enables energy-efficient AI systems capable of advanced color recognition and logic processing.
ATMO is a robot that transforms mid-air from a flying drone into a ground rover. By overcoming the long-standing challenge of hybrid robots getting stuck on rough terrain, this breakthrough unlocks new possibilities for autonomous delivery, disaster response, and planetary exploration.
MIT researchers have developed CAV-MAE Sync, an AI model that learns to precisely link sounds with matching visuals in video without any labels. This technology can bring us closer to smarter AI that can see, hear, and understand the world just like humans.
The most fascinating innovations presented at Google I/O 2025 reveal how AI is transforming Search, video creation, and communication. These advances are redefining everyday technology, making it smarter, more intuitive and personalized than ever before.
The most advanced AI models from tech giants like OpenAI and DeepSeek are generating false information at unprecedented rates – and no one knows exactly why. Due to this surge in AI “hallucinations”, the reliability of AI across critical fields is being called into question.
Microsoft’s Phi-4 family is a new generation of compact language models built for complex tasks like math, coding, and planning – often outperforming larger systems. Trained with advanced techniques and curated data, they offer strong reasoning while staying efficient for low-latency use.
Why stress over every line of code when AI can "vibe" its way through development for you? Vibe coding is a fast-growing trend where developers let AI take the lead in writing code, transforming programming into a more intuitive and spontaneous process.
NVIDIA has officially made its PhysX engine and Flow SDK fully open source, including the long-awaited GPU simulation kernel code, under the permissive BSD-3 license. Now developers can customize, port, and revive advanced physics simulations across platforms – even on non-NVIDIA GPUs.
Midjourney has launched V7, its most powerful AI image model yet, featuring smarter prompts and real-time personalization. With a redesigned architecture, V7 delivers improved object coherence, enhanced texture realism, and introduces Draft Mode for rapid, cost-efficient image iteration.