53 articles
Google’s EmbeddingGemma 2 can search text, photos, video and audio by meaning on local devices. The small model shows how useful AI may become quieter, faster and more private.
Robots can navigate a room yet struggle to remember which object was where. MIT researchers have developed a system that connects language and physical locations, offering a possible route towards
In a modified game of Battleship, a smaller AI system became a much stronger player when researchers taught it how to ask useful questions. The result offers a lesson for
Robots are being taught to work out what people leave unsaid. In MIT tests, an arm learnt how to move a cup without hitting a nearby laptop, illustrating why ordinary
OpenAI is preparing a college planning tool for teenagers that gathers applications, deadlines and financial-aid tasks in one place. Its educational promise arrives alongside fresh independent questions about teen safety.
Amazon’s shopping assistant can track prices and automatically buy selected products at a chosen threshold. That could save time, but it also shifts attention to payment permission, product matching and
Google’s new Playground lets people describe a game and ask AI to build it, without writing code. The idea is appealing, but game creation has a phased US rollout and
OpenAI has added native audio uploads to ChatGPT for paid users. Meetings, interviews and lectures can be transcribed, summarised and turned into follow-up work inside the same conversation.
Pinterest Beauty Guides analyse hair and nail Pins, then generate salon language, step-by-step guidance and product suggestions. The feature starts with what users can see rather than what they know
Albertsons is expanding ChatGPT from a Safeway shopping experience into selected internal teams and retail workflows. The October update matters because the customer plugin itself actually launched in August.
Nvidia’s new 64GB DGX Spark starts at $4,999 and can link two boxes into a 128GB local AI cluster, giving developers another route away from cloud-only model running.
A Stanford-led audit found AI models scored differently through developer APIs and consumer chat interfaces, suggesting benchmark results may not fully describe the product people actually use.
OpenAI says GPT-6.1 Sol delivers capabilities close to GPT-6 Astra with a cheaper, faster operating profile. The bigger question is whether frontier-level reasoning can now become an everyday default.
Anthropic says Claude Sonnet 5.5 is over 30% faster and can cost less per completed task while keeping the same token prices. That makes efficiency, not just raw capability, the
A Korean research team’s quadruped robot completed a 42.195-kilometre marathon in 4 hours 19 minutes without changing its battery. The result shows how better hardware and learned movement can stretch
Google’s first Project Suncatcher orbital test is due to launch four TPU AI chips on 1 October. It is a small experiment, not a working space data centre, designed to
Synapse combines episodic and semantic memory with spreading activation to help AI agents retrieve the right past information across long, complex conversations.
MorFiC transfers a locomotion policy trained on one quadruped to different four-legged robot bodies without fine-tuning, including real Unitree Go1 and Go2 machines.
JudgmentLens helped 16 non-expert readers complete legal-reading tasks faster and feel more confident, but scored comprehension did not reliably improve.
A Nature study found concept-based explanations helped people predict a real self-driving car's behaviour, especially when the vehicle acted unexpectedly.
An analysis of ten datasets found better decisions when humans and AI answered independently, with another person resolving disagreements.
OpenAI's new Dots are always-on ChatGPT agents with their own cloud computers. They can keep working, use connected apps and run recurring tasks after the conversation ends.
Anthropic has opened a research preview of a standard that lets AI agents coordinate microscopes, liquid handlers, robotic arms and other programmable equipment.
Claude Opus 4.6 faltered as benchmark tables widened, while classical classifiers stayed stable or improved. The preprint exposes a sharp limit in raw AI spreadsheet prediction, but it does not
AI traffic lights are cutting delay and emissions in cities worldwide, but the technology raises real questions about equity, surveillance, and who gets to audit the system. Here is what
AI port automation now decides berth schedules, crane sequencing and truck gate flow at the terminals moving 90 percent of world trade. Here is what is actually deployed in 2026,
AI translation earbuds from Apple, Google and independent testers now translate live speech in your ear. Here is what actually works in 2026, what still needs a human interpreter, and
AI in sports has moved from scouting notebooks to continuous predictive models. See how MLB, Formula 1 and Wimbledon are using AI, and what it means for players, staff and
AI in retail has moved past recommendations into autonomous checkout. See how Google, Walmart and Amazon are racing to build the agentic shopping layer, and what it means for shoppers
AI in architecture has cut a 62-year research-to-practice gap down to 2.5 years. Here's what real projects, from TestFit's Austin development to Autodesk Forma in Singapore, actually show.