🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,700 articles

South Central Railway upgrades Bholu AI assistant with an indigenous language model

South Central Railway has upgraded Bholu, its AI-powered passenger information assistant, with an indigenous language model. The Mini and Max versions were introduced in May at Kacheguda and Malkajgiri railway stations. Source: deccanchronicle.com

RunanywhereAI releases TrnRT, an open inference engine for Trainium LLMs

TrnRT is an open inference engine for running large language models on Trainium hardware. It bundles model execution, caching and an API server into a single system. Source: news.lavx.hu

XYBench tests whether AI can spot the flawed premise behind a question

XYBench is a new benchmark of more than 8,000 questions that checks how AI assistants handle requests built on misconceptions, such as asking how to thicken frosting. It measures whether a model addresses the underlying problem rather than just answering the question as posed. Source: huggingface.co

Anthropic: old prompting habits are holding Claude back

Anthropic has made Claude Opus 5.5 the default model on all paid plans. Alongside it, the company published a guide that retires outdated prompting habits. Source: the-decoder.com

Runway's open-source Praxis-1 turns video AI into a robot controller

Trained on video and robot data, the model takes in camera images, text instructions, and position data. From these inputs it calculates matching robot movements. Source: the-decoder.com

Hackers feed stolen LLM API keys to ARTEX AI agents to keep stealing credentials

Chinese-nexus threat actors used ARTEX, an open-source agentic penetration-testing framework, to steal LLM API credentials. The stolen keys are then fed back into the tool to power further attacks. Source: gbhackers.com

Prompting tweaks help AI seizure localization, but only for certain models

A study tested prompt debiasing and example retrieval across seven open-weight language models for seizure semiology localization. Performance gains turned out to be model-specific rather than universal. Source: bioengineer.org

Claude's Duet agent is a bit neurotic, but surprisingly useful

Duet is Anthropic's consumer-grade AI agent. It works as a helpful partner for catching things that might otherwise slip through the cracks. Source: theverge.com

Repligate shares a post on local AI as a foundation to build on

A repost from the account Repligate nods to a discussion of the ontologies of the artificial. It also includes a line arguing that local AI is a dependency developers can reliably build on. Source: digg.com

Emad Mostaque on Jensen Huang's departure and NVIDIA's succession

In a podcast discussion, Emad Mostaque weighs in on Jensen Huang's departure and what it could mean for NVIDIA. The conversation also explores broader implications for the future of AI. Source: youtube.com

Survey examines how LLM reasoning generalizes beyond training data

Large language models perform strongly on mathematical, logical and scientific reasoning tasks, but this progress is often measured within familiar data. A new survey reviews how well such reasoning generalizes outside the training distribution. Source: alphaxiv.org

Altman reveals the coolest part of leading OpenAI: holding a new AI computer

Sam Altman has shared what he considers the most exciting aspect of running the world's leading AI company. According to the coverage, it involves the thrill of holding a new AI computer in his hands. Source: youtube.com

KAIST boosts lip-reading AI accuracy using a diffusion language model

KAIST announced on the 15th that a team led by Professor Noh Yong-man of its School of Electrical Engineering has improved the accuracy of lip-reading AI with a diffusion language model. Source: biz.chosun.com

OpenAI enters cybersecurity with two Daybreak products

OpenAI has launched two security products under the Daybreak brand. One uses AI agents to automatically probe web applications for vulnerabilities. Source: the-decoder.com

Claude 3 Opus reportedly plays a real-time strategy game

Posts amplified by Repligate say Claude 3 Opus is playing an RTS game against a large field of opponents. Meanwhile, a separate commentary asks whether AI itself could make it easier to challenge Nvidia's dominance. Source: digg.com

Why Altman and Ive are building a screenless AI device

A video explainer looks at the reasoning behind Sam Altman and Jony Ive's plan for an AI gadget with no display. The same snippet also references a BBC News segment on Meta's Muse and whether it is a friendly AI agent or a 'Trojan horse'. Source: youtube.com

OpenAI rebuilds AI shopping with ChatGPT Wallet

ChatGPT Wallet lets AI agents make and pay for online purchases on a user's behalf. It launches first for Pro subscribers in the US. Source: the-decoder.com

New open-weight speech model built for real-world dictation

A newly introduced open-weight speech model aims to understand what users actually mean, from dictated email addresses to the intended formatting of their text. Source: huggingface.co

llama.cpp hits v1.0.0, cementing the foundation for local AI

llama.cpp reaches its 1.0.0 release, allowing multi-modal LLMs to run on hardware ranging from consumer machines to high-end enterprise servers. Source: huggingface.co

Google unveils Gemini 4 Hydrogen for fast, low-cost agentic coding

Google's new Gemini 4 Hydrogen model targets cost-efficient agentic coding and debugging. Developers can start building with it today in Google AI Studio. Source: blog.google