News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
981 articles
As RAG pipelines increasingly adopt agentic designs that let models make decisions at every step, the number of LLM calls has grown significantly. The piece argues that reducing unnecessary calls—rather than upgrading to faster models—is the more effective way to lower latency and cost. Source: towardsdatascience.com
New research examines the conditions under which radiologists successfully collaborate with LLMs without being misled by incorrect outputs. Key factors include the radiologist's trust calibration toward the AI and their own clinical expertise. Source: radiologybusiness.com
A new training protocol uses ChatGPT to simulate high-stress family conversations for novice intensive care unit nurses. The approach aims to build skills in de-escalation and crisis communication, which are common challenges in ICU settings. Source: cureus.com
A longitudinal study tested a multimodal LLM's ability to classify disease trajectories and stratify risk in patients with oral lichen planus. The research assessed how accurately the model could follow changes over time compared to traditional diagnostic methods. Source: nature.com
The author describes switching from Perplexity, which had already replaced Google Search for internet queries, to a local LLM running directly on an Android device. The main draw is instant, offline search results without depending on a cloud connection. Source: androidpolice.com
A new practical tutorial walks through streaming, curating, and fine-tuning the SupraLabs reasoning dataset to build a small, reasoning-oriented language model. It covers the full pipeline from raw data collection to model tuning. Source: marktechpost.com
An analysis argues that large AI model developers risk higher costs and competitive disadvantages if they don't adopt distillation techniques. Distillation is described as a way to give training a fast, efficient 'cold start' toward high performance. Source: eu.36kr.com
Researchers have developed a new method called CoLA that cuts the computing power and memory required to train large language models. The approach aims to make LLM development significantly more resource-efficient. Source: anl.gov
A new study used eye-tracking data to test whether large language models truly mirror the way humans read. The comparison revealed notable differences between human reading patterns and how LLMs process text. Source: neurosciencenews.com
Tencent has expanded international availability of its Hy3 model, previously known as Hunyuan, following its official launch on July 6. The large language model is now accessible to a broader set of users worldwide. Source: telecompaper.com
Despite Beijing's push for domestic chip independence, China's top AI developers continue to rely on Nvidia hardware to train their most advanced models. Industry sources say local chips still lag behind in meeting the demands of cutting-edge LLM development. Source: amp.scmp.com
GDIT has released guidance on selecting suitable large language models for mission-critical AI applications, amid growing interest in the technologies behind generative AI. The guide touches on related components such as natural language processing and knowledge graphs. Source: news.clearancejobs.com
South Korea's UNIST has developed a smaller large language model using a Global Mixture of Experts architecture, which shares neural modules across layers. The design reduces both model size and memory usage without sacrificing performance. Source: dongascience.com
A Stanford CS229 machine learning lecture offers a concise walkthrough of how ChatGPT-like large language models are constructed. It's part of the university's broader AI curriculum available to the public. Source: mshale.com
Chinese AI firm DeepSeek has released its most advanced large language model, DeepSeek R1, as a free web app open to anyone. Early comparisons suggest it could match or even outperform OpenAI's most capable model in certain tasks. Source: mashable.com
Kimsuky, a hacking group linked to North Korea's Reconnaissance General Bureau, has reportedly set up its own locally hosted large language model to support cyberattacks. The system is said to help automate parts of the group's operations targeting South Korean targets. Source: biz.chosun.com
A report says a North Korean hacking group has developed large language model tools and gathered software designed to automate cyberattacks and analyze stolen data. The tools could help the group scale up its operations and process compromised information faster. Source: reuters.com
Researchers at UNIST developed GMoE, a mixture-of-experts architecture that reduces the parameter count of language models by 63% while keeping performance intact. The approach could make large AI models cheaper and easier to deploy. Source: en.sedaily.com
Chinese researchers have released an open-source AI framework designed to help detect and decode rare genetic diseases through DNA analysis. The team says making the tool freely available aims to ensure more equitable access to precision medicine worldwide. Source: scmp.com
AI models are increasingly beating human physicians on standardized medical tests, though full takeover of healthcare remains unlikely soon. Experts argue AI could still play a much bigger role in assisting doctors if adopted more widely. Source: reason.com