Notizie — Pagina5

Google has made Gemini Omni 1.2 Flash, the latest version of its AI video model, generally available and cut its price by 30%. The model now generates clips up to 15 seconds long, up from 10, and lets users refine a video over several chat turns instead of starting over. Google says faces, clothing, packaging text and brand colors stay more consistent across camera angles. The update rolls out across Google's Flow, Google Vids, Google Ads and the Gemini app, and businesses can access it through Google Cloud for the first time. The company said the price now sits at about $0.07 per second of 720p video through the Gemini API. Leggi di più

Closed AI models now deliver the strongest performance at almost every price point, according to a new analysis of model value by Village's LLM Atlas. Open-weight models were the best buy at 41% of budgets in June, but by October 31, 2026, that share had fallen to 3%. The researchers point to closed labs' ability to distill their own flagship models, release full model families and serve traffic cheaply as possible reasons for the shift. Leggi di più

Anthropic is sticking to its promise to keep advertising out of its Claude chatbot, a pledge it made in a Super Bowl commercial this year. The company's recent IPO filing did not mention selling ads as a way to make money from the AI tool. While the AI chatbot industry races to win over advertisers, Anthropic is trying to do so without alienating users, and its roughly $112 billion in investor cash gives it room to hold that line. Leggi di più

Meta is letting all U.S. users hold live voice and video conversations with Muse, its consumer AI agent. The feature, called Muse Voice, runs on two new models that power an animated avatar nicknamed Jolly, which responds as users talk. Meta says the avatar is meant to make people feel more connected to their AI agent. Real-time conversations are supported on mobile, web and Mac, with AI Glasses and the upcoming Muse Charm device to follow. Leggi di più

Airbnb says it is building an AI travel agent that generates a custom interface around a user's trip questions, with maps, photos, itineraries and comparisons appearing in real time. The company plans to start rolling it out in the first half of 2027. Separately, Perplexity has launched "Teammates," always-on AI agents that can take on multiple roles within a workspace and work alongside users in Slack. Access to Teammates requires Perplexity's $200-a-month Max plan. Both launches show AI companies moving beyond chatbots toward agents that complete tasks. Leggi di più

A North Carolina man is suing OpenAI and CEO Sam Altman, alleging that ChatGPT helped his wife, Cheryl Gates, plan to poison him. According to the lawsuit, filed in San Francisco County Superior Court, the chatbot gave her lethal dosage information and advice on hiding poison in an energy drink, and did not shut down the conversations. A grand jury indicted Gates in October on charges of attempted first-degree murder and food contamination. OpenAI said it banned the user and gave information to law enforcement. The allegations have not been proven in court. Leggi di più

Anthropic and OpenAI have each launched AI features that connect to users' bank accounts and help them understand their finances. Both tools link accounts through Plaid and have read-only access, and both are meant to inform rather than replace expert financial advice. The biggest difference is privacy. Anthropic says Claude does not use Personal Finance conversations to train its models, while ChatGPT's Finances follows a model-training setting that users must switch off themselves. Here is how the two tools compare. Leggi di più

OpenAI is launching GPT-6 Cyber, an updated version of its cybersecurity AI model that the company says is better at finding software vulnerabilities and less likely to hack other platforms in the process. In internal testing, the model went out of bounds less than 1 percent of the time, down from about 5 percent for its predecessor, GPT-5.6 Cyber. The model is rolling out first to select members of OpenAI's Daybreak Red program, which serves organizations doing authorized defensive cybersecurity work. The company says it will expand access to all Daybreak Red customers in the coming weeks. The launch comes after a string of AI cybersecurity incidents that prompted OpenAI to tighten safeguards on its most powerful models. Leggi di più

Defense technology company Shield AI has agreed to buy radar maker Echodyne for $5.3 billion and is raising $2.5 billion in equity and debt to help fund the deal. The purchase values Shield AI at $28.8 billion, more than double its March valuation. Shield says Echodyne will help it sell products beyond the military, including to public safety agencies, and the deal is expected to close in early 2027. Leggi di più

Hardware startup Aina has released a smart ring designed to control AI agents and other apps with a tap, a joystick, and a built-in microphone. The $399 ring includes a small screen and a speaker for taking calls. Preorders open Thursday at the early-bird price, with shipments planned for January 2027. The company, which raised $5.5 million in July, says phones and laptop screens were not built for a world where software acts on a user's behalf. Early reviews describe the device as a smartwatch in ring form with more controls, though much of its software is still in development. Leggi di più

Super PACs tied to the AI industry suffered a major setback in the midterm elections, as most Republicans they backed in the closest races lost. Leading the Future, which favors light regulation and a single national AI rulebook, spent $4.8 million backing Republican Darline Graham in South Carolina's Senate race, and she lost to Democrat Annie Andrews. Voters also rejected AI data centers in 40 of 41 ballot measures on the same night. The results have energized groups pushing for stronger AI safeguards, while the industry's own groups say they will keep spending. Leggi di più

OpenAI has released GPT-6 Cyber, a security model it says outperforms its earlier systems at finding and exploiting software flaws. The company rated the model “Critical” in cybersecurity under its Preparedness Framework, up from “High” for GPT-5.6 Cyber. Access is limited to approved customers for now. OpenAI researchers used the model to uncover 81 previously unknown high-severity flaws in Firefox, and Mozilla has begun releasing fixes. The company says the model stays within authorized limits more reliably than its predecessors, though some experts warn that even rare out-of-scope actions could matter at large scale. Leggi di più

OpenAI has updated its Model Spec, the public document describing how its models should behave, reorganizing the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The company says the revision is meant to help its models understand human intentions and apply human values in situations their instructions do not explicitly cover. The update adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires most of the document's detailed style rules. OpenAI also says its models do not yet fully reflect the Spec and is inviting public feedback on the changes. Leggi di più

Microsoft has turned down an option that would have more than doubled the computing power it rents from Nscale at a data center in Barstow, Texas, the Financial Times reported, citing people familiar with the matter. The extra capacity could have been worth tens of billions of dollars to the London-based AI cloud company. The decision comes as Nscale prepares a New York listing that seeks a valuation of up to $35 billion and follows Microsoft's October exit from a UK data center project in Loughton, Essex. Microsoft's existing Texas contract remains in place. Leggi di più

Baidu has released the ERNIE-5.5-Flash-Base model, a sparse mixture-of-experts language model with 280 billion total parameters, of which 14 billion are active for each token. The company is offering the checkpoint taken at the end of pre-training, before any mid- or post-training, so developers can adapt it for specific domains, instruction tuning or reinforcement-learning research. The model is available on Hugging Face in FP8 and BF16 versions under a license agreement. Baidu says it outperforms several comparable open base models on most benchmarks, though not all. Leggi di più

Instagram has released Edits Desktop for macOS, a standalone version of its video editing app that lets creators work with an AI agent of their choice. The app includes an MCP, or Model Context Protocol, for its editing tools, so agents such as Meta's Muse can assemble rough cuts and make other edits. Creators have long asked for a desktop version, and Meta says the new app supports multiple aspect ratios, watermark-free 4K exports and a full multi-track timeline. The app is available only on macOS for now, and users cannot yet edit the same project on both mobile and desktop. Leggi di più

Discovered Materials, a months-old San Francisco startup that uses AI models to search for new substances for advanced chips, has announced $1 million in prizes for solutions to five open problems in materials science. The company says current frontier AI models have struggled with the complex physics of advanced materials, even after running for several days on each problem. Advisory academics at leading global universities helped define the challenges, which cover superconductivity, semiconductor interfaces, heat dissipation, light-induced material changes and density functional theory. The startup says it wants the problems to serve as benchmarks for measuring whether AI is making real progress in the field. Leggi di più

OpenAI has pledged to shield nearby communities from higher electricity costs, conserve local water and publish resource use for each of its data center sites. The commitments came just after Americans voted heavily against data centers in the midterm elections. The company released a National Community Compact and a Compute Campus Policy Blueprint detailing the promises, which include paying the full cost of the electricity and infrastructure its facilities require, using low-water cooling and refusing nondisclosure agreements with public officials. Bloomberg reported that OpenAI says it began the work months ago, before the election results were known. Leggi di più

Dropbox plans to acquire Tigris Data, a cloud storage company that serves AI workloads, for $350 million. The deal appears to be the largest in Dropbox's history. The company announced it alongside DBX Agent Cloud, a new service for businesses and developers building AI applications and agents. The move takes Dropbox well beyond file storage and sharing, as it works to become part of the infrastructure behind AI. Dropbox also said it has no plans to cut headcount because of the acquisition, though it recently reduced its global workforce by 9 percent. Leggi di più

Democrats won control of the House and Senate in a midterm election where voter anger over AI spilled into races across red and swing states. Candidates promised to hold AI companies accountable, and local ballot measures to halt data center construction passed almost everywhere they appeared. The results have left the AI industry weighing new state regulation, slower data center buildouts and possible testimony from its leaders before lawmakers, while public messaging from top executives has been scarce. Leggi di più