AI News
Latest news and trends from the world of artificial intelligence
Making the case for curiosity-driven science
President Sally Kornbluth spoke about the growing challenges to the U.S. research ecosystem, particularly the strain on funding for top research universities. She emphasized the critical importance of curiosity-driven and basic science for the nation's future, highlighting how it fuels innovation like cancer immunotherapy and future AI and quantum technologies. Kornbluth expressed concern over diminishing federal funding and its impact on training the next generation of scientists, warning of long-term ramifications. She also discussed AI's role in education, stressing the need to focus on fundamental skills and view AI as an augmentation tool. The importance of international talent in maintaining U.S. competitiveness was also underscored.
Introducing Advanced Account Security
OpenAI is introducing Advanced Account Security, an opt-in feature for ChatGPT accounts designed for users at higher risk of digital attacks or those seeking the strongest available protections. This feature combines enhanced security measures, including phishing-resistant login using passkeys or security keys (disabling password-based login), more secure account recovery methods (disabling email/SMS recovery in favor of backup keys), shorter session times, and clearer session management. It also automatically excludes conversations from model training. To make phishing-resistant authentication more accessible, OpenAI is partnering with Yubico to offer preferred pricing on security key bundles. Advanced Account Security will become mandatory for individual members of "Trusted Access for Cyber" starting June 1, 2026.
OpenAI says it hit its 10 gigawatt compute goal years ahead of schedule
OpenAI has announced it has achieved its goal of 10 gigawatts of AI compute capacity in the United States several years ahead of schedule. The company secured three gigawatts in the last 90 days alone, including two from Amazon. Despite this expansion, OpenAI has scaled back or abandoned several large infrastructure projects, including a data center expansion in Texas, a project in the UK due to energy costs, and a site in Norway. The ambitious "Stargate" project, a $500 billion initiative with Oracle and SoftBank announced in early 2025, is also mentioned in the context of these infrastructure plans.
Tencent's 440 MB AI model translates 33 languages offline on your phone
Tencent has open-sourced a compact AI translation model, Hy-MT1.5-1.8B-1.25bit, which it claims outperforms Google Translate and operates entirely offline on smartphones. The model supports 33 languages, five dialects, and 1,056 translation directions. Through aggressive compression to 1.25 bits per parameter, the model is reduced to 440 MB while maintaining high translation quality comparable to much larger models and commercial services. Tencent also offers an Android demo app for offline translation across applications. This development aligns with Google's efforts to bring local AI models to smartphones, such as with Gemma 4.
White House worried about compute limits as it blocks wider access to Anthropic's Mythos
The White House has reportedly blocked Anthropic's plan to grant access to its AI model Mythos to approximately 70 additional companies due to security concerns. Mythos has the capability to identify and exploit software vulnerabilities. Currently, around 50 organizations, including critical infrastructure operators and government agencies like the NSA, have access through Project Glasswing. Government officials are concerned about limited compute capacity, suggesting Anthropic has less than competitors, and worry that expanding access could restrict the government's own use of the model, especially as new compute capacity takes time to come online. The administration is also navigating its relationship with Anthropic following a Pentagon classification of the company as a supply chain risk.
Softbank plans IPO for new AI and robotics company valued at up to $100 billion
Softbank is reportedly planning to launch and take public a new AI and robotics company named Roze in the United States, potentially as early as this year, with a target valuation of up to $100 billion. The company would focus on building data centers and could integrate the recently acquired ABB Robotics. According to the Financial Times, Softbank founder Masayoshi Son aims to use the IPO to offset significant company spending, including its approximately $30 billion investment in OpenAI. However, some Softbank executives reportedly view the valuation and timeline as overly ambitious, especially given geopolitical uncertainties and the company's existing debt constraints.
Solving the “Whac-a-mole dilemma”: A smarter way to debias AI vision models
Researchers have developed a new debiasing technique for vision-language models (VLMs) called WRING (Weighted Rotational DebiasING). This method addresses the "Whac-A-Mole dilemma" associated with traditional projection debiasing, where removing one bias can inadvertently create or amplify others. WRING works by rotating specific coordinates within the model's high-dimensional space responsible for bias, altering the representation without disrupting other learned relationships. This post-processing technique can be applied to pre-trained VLMs like CLIP models efficiently and with minimal invasiveness. The researchers found WRING significantly reduced bias in target concepts without introducing new ones, and plan to extend it to generative language models.
The MIT-IBM Computing Research Lab launches to shape the future of AI and quantum computing
Building on a long-standing MIT–IBM collaboration, the new lab will chart the convergence of AI, algorithms, and quantum computing.
Enabling privacy-preserving AI training on everyday devices
A new method could bring more accurate and efficient AI models to high-stakes applications like health care and finance, even in under-resourced settings.
Cybersecurity in the Intelligence Age
OpenAI outlines a five-part action plan for strengthening cybersecurity in the Intelligence Age, focused on democratizing AI-powered cyber defense and protecting critical systems.
Archive
About Sources
AI news are automatically downloaded from various sources and translated using AI. Updates occur twice a day.