News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
973 articles
China-based Maase Inc. announced a $50 million PIPE financing round intended to speed up rollout of its Star distributed intelligent computing centers and commercialize its Lingyanmiaoyu large language model. The company describes itself as an AI-focused full-scene technology provider. Source: finance.yahoo.com
Bell and Cohere have unveiled a new AI model built to help security teams investigate cyber incidents more efficiently. The launch comes as organizations worldwide accelerate AI adoption in defense systems amid rising fears that attackers are also weaponizing the technology. Source: theglobeandmail.com
Researchers have introduced LLM4MG, the first framework that adapts a large language model for multipath signal generation using the Synesthesia of Machines (SoM) concept. The approach aims to bring LLM-based reasoning into wireless channel modeling tasks. Source: nature.com
Saudi AI company Humain has developed a massive 428-billion-parameter Arabic-language model built on top of China's MiniMax architecture. The project is part of Saudi Arabia's effort to establish its own national, Arabic-first AI platform. Source: shattered.io
Researchers built a system where multiple large language model agents collaborate to assist with protein engineering. The approach aims to let scientists describe design goals in plain language, narrowing the gap between human intent and molecular design. Source: nature.com
Researchers developed a new algorithm designed to help large language models better interpret vague, jargon-filled, or context-dependent user requests. The approach focuses on human-AI collaboration to fill in gaps that machines typically misunderstand on their own. Source: bioengineer.org
An analysis published in September 2026 argues that large language models, despite their rapid rise since OpenAI's ChatGPT launch in late 2022, are unlikely to meaningfully improve decisions in national security contexts. The piece traces the technology's evolution from Sam Altman's original chatbot release to its current limitations in high-stakes policy settings. Source: globalsecurityreview.com
New Comscore figures indicate that ChatGPT, the chatbot that popularized modern AI assistants, is seeing its market share erode as rivals gain traction. Google Gemini in particular is cited as a growing competitor pulling users away. Source: tubefilter.com
Parameters in a large language model refer to the weights and biases learned during training, as defined in Google's machine learning glossary. The article explains that a higher parameter count doesn't necessarily translate into stronger real-world performance. Source: sqmagazine.co.uk
After spending an hour testing Google Gemini Live's voice conversation feature, the author found that the once-notorious lag in AI voice chats has been resolved. However, the experience revealed that awkward pauses were never actually the main problem with AI voice assistants. Source: androidpolice.com
TypeSafe AI has introduced Jev, an AI judgment model it says is 75 times faster and 170 times cheaper than OpenAI's model for classification tasks. The company positions Jev as a specialized alternative aimed at cutting operational costs for AI-driven decision-making. Source: chosun.com
Major AI companies including Anthropic, OpenAI, and Google's Gemini have each faced recent security incidents involving their large language models. The repeated breaches raise questions about why current safety defenses for LLMs keep getting bypassed. Source: eu.36kr.com
Researchers from Georgia Tech, Nvidia Research, and Stanford built BOOST, a runtime system that taps host memory and HBM at the same time to accelerate LLM inference. The approach aims to reduce bottlenecks that typically slow down large-scale model serving. Source: indexbox.io
Jev, a new AI model from TypeSafe AI, is drawing attention among developers for its distinct approach to handling decision and judgment tasks. It's being framed as a 'generalist' model that could rival established systems in speed and efficiency. Source: indianexpress.com
Anthropic has teamed up with more than a dozen industry partners to bring its Claude AI models into the workflows of registered investment advisors. The move could test how eagerly tens of thousands of financial professionals adopt AI-assisted tools in their daily practice. Source: financial-planning.com
Businesses are shifting quickly from testing generative AI to finding concrete use cases that generate measurable enterprise value. Consultancies now focus on helping organizations turn large language model capabilities into practical, revenue-driving applications. Source: tynmagazine.com
As AI adoption grows, more users are turning to offline, on-device models that keep prompts and data away from third-party servers. After testing a local LLM on a smartphone for five practical tasks, the experience proved it can be a genuinely useful privacy-focused alternative. Source: makeuseof.com
MIT researchers argue that brute-force scaling of AI models runs into fundamental mathematical constraints. Their work points to superposition effects, Zipf's law, and depth limitations as factors capping further growth of large language models. Source: c3.unu.edu
According to Stanford's 2025 AI Index, computing power used to train leading AI models doubles roughly every five months, while dataset sizes double every eight months. Electricity demand for training these models is also rising fast, doubling approximately once a year. Source: spacewar.com
China presented what it calls the country's first multilingual, fully multimodal AI-powered input method for the Tibetan language. The system was unveiled on Sunday in Xining. Source: globaltimes.cn