News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
981 articles
A new comparative study evaluated three LLMs — ChatGPT, Claude, and DeepSeek — on how well they explain cardiac imaging topics to patients. The research looks at accuracy, clarity, and usefulness of AI-generated medical explanations. Source: cureus.com
Lightweight 'Flash' variants of large language models are becoming the new flagships among Chinese AI labs, favored for their low cost and strong performance. This shift suggests efficiency is increasingly outweighing raw scale in China's LLM competition. Source: pandaily.com
Researchers have developed a system that detects flaws in software code with roughly 67% accuracy. While not perfect, the tool could help developers catch potential security issues earlier in the development process. Source: quantumzeitgeist.com
Startup Leiolai has launched a new AI model built to operate locally on user devices instead of relying on cloud servers. The approach aims to improve privacy and reduce latency for AI-powered applications. Source: itbrief.asia
South Korean telecom and tech giants SKT, KT, and Kakao have been selected to lead consortiums providing a free AI service for the public. The initiative is part of a government-backed effort to widen public access to AI tools. Source: koreatimes.co.kr
Thomson Reuters is using decades of accumulated content from its media and legal databases to train its AI systems. The move aims to give its AI tools a strong knowledge foundation built on proprietary, high-quality data. Source: pymnts.com
A discussion piece examines key concepts in AI interpretability and alignment, including J-Space, chain-of-thought reasoning, AI personas, hallucinations, and the famous 'Golden Gate Bridge' Claude experiment. It highlights how these ideas help researchers understand what's happening inside modern AI models. Source: eu.36kr.com
New research suggests that cognitive biases embedded in training data or prompts can distort the diagnostic outputs of LLMs used in radiology. Source: rsna.org
Z.ai has open-sourced its 'Ox Alpha' model, now available to the public under the name GLM-5.3-Flash. Source: siliconangle.com
A new technique called Meta^n aims to unlock deeper levels of recursive reasoning within LLMs. Source: startuphub.ai
A roundup highlights the five best-performing LLMs for coding tasks expected in 2026. Source: intuit.com
Researchers propose a deep learning framework enhanced by football-inspired optimization and LLM guidance to improve desalination system performance. Source: nature.com
A new LLM-based platform generates novel synthesis recipes for materials, significantly reducing the need for lengthy trial-and-error experimentation. Source: phys.org
OWASP has published an updated version of its Top 10 list outlining the most critical security risks facing LLM-based applications. Source: scworld.com
OpenAI is said to be developing a custom inference chip internally codenamed "Jalapeño," aimed at cutting reliance on NVIDIA hardware. The chip is reportedly designed to deliver better performance for running AI models than current NVIDIA GPUs. Source: eu.36kr.com
Search interest in prompt injection attacks has jumped 141 percent, reflecting growing attention to LLM security beyond academic circles. The trend suggests attack research on language models is increasingly reaching mainstream tech and security audiences. Source: technology.org
Marketing platform CoSchedule now lets users choose between AI models from OpenAI, Anthropic, and Microsoft Copilot. The update gives customers more flexibility in selecting the AI backend for content and workflow features. Source: desmoinesregister.com
A new review examines the current evidence, practical use cases, and necessary safety measures for deploying large language models in primary care settings. It highlights both the potential benefits and the risks that need addressing before wider clinical adoption. Source: nature.com
Researchers are using large language models to dynamically define search spaces for closed-loop, autonomous materials exploration systems. The approach aims to make automated discovery of new materials faster and more efficient. Source: nature.com
STARFlow2 is a new architecture that merges language models with normalizing flows to enable unified generation across multiple data modalities. The approach seeks to bridge text and other content types within a single generative framework. Source: machinelearning.apple.com