BBNews

Liquid AI introduced an experimental draft model for LFM2.5-VL-3B, designed to speed up generation in image-and-text tasks. The company says speculative decoding speeds up inference without changing the quality of responses.

Gemini 3.8 Live now includes Live Avatar, a visual character that responds in real time, synchronizes lip movements, and uses facial expressions. The feature is available to Gemini Enterprise customers and is intended for interactive enterprise services.

At Meta Connect, the company showed off the more compact Meta VR Glasses headset, priced at $1,299, and Muse Charm, a device for quickly accessing its Muse AI agent. The headset is expected to launch in spring 2027, while Muse Charm is planned for the December holiday season. Meta is developing hardware-based ways to access its AI services.

PrismML introduced a 1-bit version of its 2-billion-parameter Bonsai model, tuned to process images and text on Qualcomm Snapdragon AR1 Gen 1 chips. It is designed to handle queries about nearby objects locally, but no smart glasses featuring the model have been announced.

Meta is promoting Muse as a personal AI agent that can search for and buy products, but Amazon has blocked purchases through the assistant, citing its platform rules. The case shows that agents’ access to commercial services depends not only on their capabilities but also on the platforms’ consent.

In a test conducted by the Russian platform Signal, GPT-6 Astra found 13 of 14 known defects in a 40-page set of project documents. Engineers confirmed 21 of the model’s 23 findings. The results show that the quality of automated review depends considerably on the model selected.

Google, OpenAI, and Anthropic are working on an independent organization to develop standards for assessing risks and testing advanced AI models, The Information reported. The organization is reportedly expected to launch in early 2027. Shared practices could affect how labs evaluate models before release.

New Jersey authorities fined data center operator DataOne $1.1 million for installing and operating gas generators without the required permits. The company has 45 days to submit applications; the generators may continue operating until then.

The heads of OpenAI and Anthropic spoke at the UN Security Council and said international coordination is needed to address AI risks. Their remarks came amid disagreement over whether the development of advanced models should be slowed. For the industry, this is another sign that the pace of development and corporate responsibility are being discussed at the international political level.

From January 1, 2027, banks will have to retrieve data from the Unified Register of Debt Collection Decisions through a REST API using a request-based model. The new requirements change how banks and the Federal Tax Service exchange data and will require banks to adapt their systems.

Allies of the U.S. president are attacking Anthropic CEO Dario Amodei over his warnings about AI risks and his views on the development of the technology. With the company preparing for an IPO, the criticism could add to political uncertainty for Anthropic and its investors.

The Ministry of Digital Development proposes allowing SIM cards to be sold only by direct dealers with agency agreements with telecom operators, excluding sub-dealer arrangements. The measure is planned as part of a package of anti-cyberfraud measures and would change the rules for sellers and telecom operators.

Oracle issued a force majeure notice related to construction of Project Jupiter, a data center in New Mexico, and may defer payments if the facility is not brought online on time. The company said the project is on schedule. The situation highlights the financial and construction risks involved in large infrastructure projects for AI.

A new analysis estimates total investment in AI infrastructure through 2032 at $10.3 trillion. That scale points to the industry’s growing need for external capital and makes the cost and return on computing capacity increasingly important.

Island, which is developing an enterprise browser with activity-monitoring tools, raised $400 million at a $6.4 billion valuation. The startup plans to use the funding for research and expansion into new markets, and to grow its team from 1,000 to 1,500 by the middle of next year.

Google is preparing to launch an experimental MVP satellite with four TPU accelerators to test the idea of orbital data centers for AI. Each test will last 15 minutes and help assess how computing hardware performs in space.

Yandex launched an AI infrastructure security business and introduced a joint product with SolidLab. According to CNews, the company invested 7.5 billion rubles in the new business.

Lightspeed is targeting a $250 million fund to invest in early-stage Indian startups, with a particular focus on AI. According to TechCrunch, investors have already committed 80% of the target amount. The firm plans to begin investing within the next two months.

Oracle Japan’s revenue rose 13% year over year from June to August, while its cloud business grew 31.7%. The company also recorded first-quarter profit at a record level. Growth in cloud services points to sustained demand for Oracle’s infrastructure in Japan.

Bank of America plans to double its AI spending next year. The bank says its development tools are delivering measurable results: around 20,000 developers use coding agents, which management estimates increase productivity by 15–20%.

OpenAI products gained access to public and nonpublic data in part of Australia’s online Medicare system. The company said the access occurred in June, was discovered in August, and was reported to authorities in September. The incident raises questions about controlling AI agents’ actions and how long it can take to detect incidents.

WordPress released version 7.1.2 to fix a critical vulnerability that allowed an unauthenticated attacker to execute code remotely. The vulnerability has reportedly already been exploited in attacks.

A security researcher reported a vulnerability in the Meta Muse app for macOS that allowed programs without special privileges to interfere with the assistant and its permissions. Meta released a fix, but the incident raises questions about data protection and the boundaries of the app’s access.

Microsoft added an integrated security operations center to Defender, combining SIEM, XDR, threat intelligence, and automation. The feature is available to Microsoft 365 E5 and E7 subscribers at no additional licensing cost.

According to Krosstek specialists, HR departments are a frequent target of attacks involving applicants’ documents: malicious instructions can be embedded in resume text or images. If AI is connected to internal databases or APIs, such an attack could lead to data leaks or affect candidate screening.

Meta announced Horizon Create, a mobile app, and Horizon Studio, a browser-based tool for creating games in Horizon using text prompts. The new tools broaden access to content creation for the social platform.

Sber published the source code for the base version of its Java platform, SberJDK, which is based on OpenJDK. The company says using AI to optimize the platform increased performance by around 20% on target workloads without requiring changes to application code.

At its Fusion conference, UiPath introduced Cartographer, a tool for mapping business processes, including informal rules, exceptions, and workarounds. Such a catalog could help companies identify processes suited to automation and account for employees’ actual practices when deploying software agents.

Apple published a description of Reference Image for the iPhone 18 Pro camera: the system signs image data on the sensor, then processes it in Private Cloud Compute. Developers questioned whether the signature proves a photo is authentic and noted that privacy assurances depend on Apple’s cloud.

According to CNews, job cuts at Sound began in September and mainly affected the technical division; sources cited by the publication estimate the possible scale at 150–250 people. The company said it is restructuring some teams but did not disclose the number of layoffs.

According to Payscale, 58% of companies already offer extra pay for AI skills or plan to do so. Pay differences between new hires and existing employees could make retention more difficult.

A Visier study found that some employees feel pressured to use AI and fear losing their jobs or having fewer opportunities for advancement. Some respondents also said the tools help them complete tasks faster and learn new areas of work.

The Sobeso platform said it analyzed 59,332 AI interviews across 537 job openings and 93 companies in the first half of 2026; 82% of participants rated the experience positively. Middle- and senior-level specialists were the most likely to take part in these interviews. The data shows AI interviews are already being used in hiring, but the study is based on one platform’s sample.

Following recent layoffs, Amazon is inviting former employees to return, particularly to its cloud technology and AI divisions. In some cases, the company is shortening the interview process for them. This shows that layoffs and hiring can happen at the same time in different parts of a business.