News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
973 articles
Training large language models increasingly hits GPU memory limits before compute capacity is fully used, due to the size of model weights, gradients, and optimizer states. A new approach uses host offloading in JAX to reduce these high-bandwidth memory bottlenecks. Source: developer.nvidia.com
As large language models continue to spread worldwide, experts are reassessing which benchmarks actually capture meaningful performance. The piece takes stock of current LLM evaluation methods and their limitations. Source: cacm.acm.org
Researchers unveiled ARTEM (Agentic Retrieval with Temporal-Episodic Memory), a hybrid agent architecture that combines LLMs with a self-organizing neural memory system. The design aims to give AI agents better spatial and temporal episodic memory for more context-aware reasoning. Source: ojs.aaai.org
NASA has deployed Google's Gemma LLM in space, testing the concept amid ongoing debate over whether orbital data centers can realistically host the largest and most capable AI models. Source: aol.com
For three years, vision-language-action (VLA) research has assumed that top-tier robot AI models need a large language model at their core. TurboVLA challenges that assumption by achieving comparable results without one. Source: techtimes.com
Researchers found a fundamental weakness that makes it simple to manipulate LLMs into producing harmful outputs, including instructions for sabotaging an aircraft's navigation system. Source: technologyreview.com
Debian developers last week opened discussion on a general resolution addressing how AI large language models should be used within the project, with five different proposals now on the table. Source: phoronix.com
PolyAI Ltd. announced Dialog-RSN-1, a new voice dialog AI model designed to make conversations with AI-driven phone agents sound more natural and human. Source: siliconangle.com
JetBrains released an experimental IntelliJ IDEA plugin that lets Kotlin/JVM projects hand off runtime logic to an LLM directly from Kotlin code. Source: infoworld.com
The new governance layer gives enterprises a single controlled path for every tool call, API request, agent interaction, and LLM prompt, as companies increasingly seek to manage the tools and APIs used by AI agents. Source: securityboulevard.com
The open-sourcing of Kimi K3 suggests that AI developers' strategy of releasing highly optimized, high-performance models as open source is gaining real momentum. Source: eu.36kr.com
Backed by the GSMA, the African Trust & Safety LLM Challenge has released a benchmark of 4,216 verified, reproducible tests designed to probe AI safety across diverse African languages and cultural contexts. The project aims to fill gaps in safety evaluation for regions underrepresented in existing LLM benchmarks. Source: thefastmode.com
Researchers introduced a Dual-Population Co-Evolutionary (DPEC) framework that lets an LLM-driven population of algorithms learn from expert-designed strategies to solve complex scheduling problems. The approach was applied to agile earth observation satellite scheduling, combining adaptive large neighborhood search with LLM assistance. Source: eurekalert.org
GitHub user Slava S. built a 28.9-million-parameter language model capable of generating text on the ESP32-S3, a microcontroller that costs roughly $8. The project demonstrates how far small-scale LLMs can be squeezed onto cheap embedded hardware. Source: blog.adafruit.com
A study published in JCO Clinical Cancer Informatics shows that the Truveta Language Model can accurately pull critical cancer staging details from vast collections of unstructured clinical records. The research highlights a scalable way to turn raw medical notes into research-ready data. Source: globenewswire.com
OMODA unveiled its Super AI Cockpit in Jakarta, built on ByteDance's Seed LLM, positioning it as a next-generation in-car experience aimed at younger drivers. The announcement frames the cockpit as a key part of the brand's push into AI-driven mobility in the region. Source: manilatimes.net
The European Commission's DG Translation has released an open-source LLM trained on multilingual data, aimed at supporting EU languages that are often underrepresented in mainstream AI models. The project also includes a benchmark to evaluate model performance across the bloc's official languages. Source: slator.com
A US court has approved Anthropic's $1.5 billion settlement with a group of authors whose pirated books were used to train the Claude LLM. The case highlights growing legal risks for AI companies over how training data is sourced. Source: business-standard.com
As models grow capable of generating their own training curricula, data is shifting from being a static resource to acting as a medium that transmits intelligence between models. This shift reshapes how distillation techniques are understood in modern ML. Source: thesequence.substack.com
South Korean company DESILO announced that its THOR system was chosen as the reference implementation for encrypted AI within a new global benchmarking suite for fully homomorphic encryption (FHE). The initiative aims to set a standard for evaluating private AI and encrypted LLM inference. Source: manilatimes.net