College football has never lacked people who know which play to call. Most reveal their genius after the whistle. Texas briefly gave those critics a new target. A report suggested its new AI platform, Bevo Lab, influenced a failed fourth-down decision against Florida. Athletic director Chris Del Conte clarified that he had offered the play as an example of what the system could suggest. It hadn’t made that call. The actual project deserves attention. As an example of AI in co

The cloud is a pleasant name for somebody else’s computer. Meta’s latest release offers developers a chance to bring more of the work home. Muse Spark Open is an 816-billion-parameter open-weight AI model built for coding, interpreting images, and completing tasks across multiple steps. Working with Apple, Meta optimized it to run locally on two M5 Ultra Mac Studios with a combined terabyte of memory. Meta says the setup fixed bugs, added features, and rebuilt an existing iOS

A farmer doesn’t need a machine that knows everything. He needs one that can help him understand his soil—in his language, on his phone, before the next planting. The World Bank is urging developing countries to invest in “small AI”: specialized models built for local needs that can work with limited connectivity. While much of the technology industry competes to build larger systems, the bank sees opportunity in making useful intelligence available where expertise is scarce.

A powerful tool becomes more useful when you can decide how to use it. Reflection AI Beam, an open-weight AI model introduced by Reflection AI, a company founded by former DeepMind researchers, is designed to write code, reason through problems, and carry out tasks using digital tools. Its ambition is to give organizations greater control over the AI they depend on. Reflection plans to release Beam’s model weights—the learned settings that make it work—later this month, allow

An AI model can be freely available—and still have subjects it won’t discuss. A CBS News investigation found that the Qwen model it examined avoided, denied, or reframed questions about politically sensitive subjects in China, including Tiananmen Square and the treatment of Uyghurs. Alibaba did not respond to the outlet’s request for comment. The findings raise a question that reaches beyond one company: how much of an AI’s answer reflects the evidence, and how much reflects

The face on your screen may not be human. AI is learning the small gestures that make us feel heard. A nod while you’re speaking. A pause while you think. An expression that changes with yours. Tavus Griffin is a new human interaction system designed to bring those details into live AI video conversations. Currently available through the restricted Griffin-Lite research preview, it watches, listens, speaks, and reacts simultaneously while generating a lifelike person on scree

OpenAI’s new agents can keep working after you close your laptop. At its September 29 DevDay, OpenAI introduced dots: personal AI agents powered by GPT-6 Astra, each with its own cloud computer and, according to OpenAI, access to more than 4,000 apps through its ecosystem. OpenAI dots can pursue assigned projects around the clock and can be reached through ChatGPT and, where enabled, Slack and Microsoft Teams. Think research that progresses overnight. Launch materials revised

It took less than 72 hours to reach OpenAI’s internal code repositories. The reward for finding the hole? $6,500. Security researchers at Hacktron say they chained together two vulnerabilities that allowed them to compromise OpenAI employee ChatGPT accounts and reach connected services, including an employee’s Codex access to OpenAI’s internal GitHub repositories. To prove the access without reading internal code, they instructed the employee’s Codex account to create a harml

For years, we have asked AI questions. Now it is beginning to help decide which questions are worth asking. Anthropic has created a life-sciences laboratory with an unusual division of labor: humans still perform the physical experiments, but AI increasingly helps decide what is worth looking for in the first place. One of its first assignments sounded almost absurdly broad. Search an enormous database of DNA and find something interesting. So Claude did. Roughly 950 AI agent

For half a century, computing has required a ritual. Find a screen. Sit down. Look at it. Type something. Wait. Meta thinks that ritual is beginning to disappear. At its 2026 Connect event, the company unveiled a sprawling new generation of AI glasses and demonstrated how its personal AI agent, Muse, is moving beyond the phone and computer. Muse will be able to come with you in your glasses. It will be able to act on what you’re looking at. Look at a product on a shelf and as

A photograph used to be evidence. Now it can be a search. In June, 28 law-enforcement specialists from 11 countries gathered in Tunisia for an unusual five-day counterterrorism operation coordinated by INTERPOL. Their challenge was enormous. Extremist groups had published mountains of photographs and video online. Somewhere inside that visual noise were people authorities were trying to identify. Humans could have searched it. Eventually. Instead, investigators combined human

For most of history, nobody owned intelligence. You could hire it, teach it, learn from it or compete with it. But you couldn’t turn it off. That may be changing. OpenAI announced that it intends to wind down its direct model-supply relationship with Cursor following the company’s acquisition by SpaceX. OpenAI said it cannot be confident that SpaceX will use its technology within its terms of service and argued that increasingly capable AI models require greater accountabilit

For most of human history, access to expertise was scarce. If you wanted great legal advice, scientific analysis, computer programming, tutoring, strategy or research, you needed access to a person who had spent years learning how to do it. That scarcity made expertise valuable. Something strange is happening now. Meta and Google both released powerful new AI models this week. Meta CEO Mark Zuckerberg described the performance of its newest model as “almost too cheap to meter

HOLLYWOOD WAS WORRIED ABOUT ONE AI ACTRESS. CHINA RELEASED 122,000 AI MICRODRAMAS. Last year, the entertainment industry briefly panicked over Tilly Norwood, an artificial actress designed to test whether audiences might accept a digital movie star. She attracted headlines, industry outrage and intense debate over whether synthetic performers could ever compete with real actors. It turns out AI may not need to create the next Marilyn Monroe. It merely needs to produce enough

AI MAY HAVE DISCOVERED HOW TO THINK WITHOUT TALKING TO ITSELF. Most leading generative AI systems today are built on Transformer-based or Transformer-derived architectures. Transformers power systems such as ChatGPT, Claude and Gemini. They process sequences of tokens and can solve difficult problems by generating long chains of intermediate reasoning. The harder the problem, the more computation the model may need. And often, the more tokens it produces. Researchers at AI co

AI JUST LEARNED HOW TO USE THE MACHINES. Anthropic has introduced the Anthropic Model Hardware Standard, or MHS, a new system designed to allow artificial intelligence agents to discover, understand and operate physical equipment, including microscopes, robotic arms, laboratory instruments and factory machinery. Until now, connecting AI to specialized hardware often required engineers to spend weeks or even months writing custom software for individual machines. Under Anthrop

A new kind of coworker just showed up in Slack. It does not need a salary, a desk, or an invitation to the holiday party. It just needs a prompt. Slack Code brings AI coding agents directly into shared workspaces, allowing entire teams to watch, guide, and review what those agents are doing. It currently supports Claude, Devin, GitHub Copilot, and Vercel, with ChatGPT announced as coming soon. Give an agent a task and it can create a dedicated code channel where teammates can

For years, artificial intelligence has been accused of producing answers with too much confidence and too little contact with reality. So scientists gave it something reality could grade: actual proteins. Claude was asked to design tiny protein binders capable of attaching themselves to 15 biological targets, a job that traditionally requires specialists, serious computing power, specialized design tools, and enough time to reconsider one’s career choices. It succeeded on 14.

After a hurricane, every hour matters. Roads flood, traffic signals fail and fallen trees turn familiar neighborhoods into mazes. Until now, assessing that damage has required crews to drive hundreds of miles, take photographs, complete reports and eventually return to an office where internet service still works. The storm moves in hours. The paperwork prefers a more leisurely pace. Researchers at the University of South Florida are developing AI technology to accelerate tha

Artificial intelligence has already created prompt engineers, robot trainers and digital ethicists. Now it has produced a profession no guidance counselor saw coming: the masturbation consultant. AI companion company Joi AI selected 10 participants for a 28-day “wellness” experiment, offering each $2,000 to examine whether AI-guided self-pleasure could affect stress, focus and other measures. Six ultimately completed the study. Participants tracked their mood, energy, craving