All published AI news and analysis.

Google DeepMind has announced the introduction of several new Gemini models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. This expansion of the Gemini family aims to offer a broader range of capabilities and optimizations for various AI applications.

OpenAI has launched the ChatGPT for Small Businesses program, an initiative designed to help entrepreneurs integrate AI into their operations. The program aims to enhance AI skills among small business owners, facilitate work automation, and support business growth through the use of ChatGPT Work.

A recent overview from Hugging Face and NVIDIA discusses the current landscape of simulation for physical AI, highlighting its critical role in developing and testing AI models for real-world robotic applications. The article emphasizes the challenges and advancements in creating realistic and scalable simulation environments.

Jargo is a new WebRTC-native, audio-first conversational AI framework for Go. It enables developers to build real-time voice agents with a streaming transcription-reasoning-speech pipeline, supporting turn-taking and barge-in capabilities.

Researchers have introduced RoboTTT, a novel robot model and training methodology that extends visuomotor context to 8,000 timesteps, a three-order-of-magnitude increase over existing state-of-the-art policies. This advancement enables new capabilities such as one-shot imitation from human demonstrations and on-the-fly policy improvement without compromising inference speed.

A new research paper introduces a method to evaluate Large Language Models (LLMs) for statistical self-consistency, particularly in persona prompting scenarios. The study found that LLMs often exhibit a 'macro fallacy,' where aggregated estimates from fine-grained subpopulation responses are more accurate than direct population-level estimates.

Researchers introduced MOJO, a training framework for spike-tokenizing neural models that combines self-supervised learning (SSL) with supervised learning (SL). This approach enables the effective use of unlabeled neural data, improving performance, especially in low-label scenarios and generalizing across different species and modalities.

This research paper introduces PAT (Pragmatic Auto-Translator), a RAG-based system designed to move large language models (LLMs) beyond sentence-level translation. PAT utilizes a comparable corpus of authentic longform texts to inform whole-document translation, aiming to produce draft translations that are contextually appropriate for the target language and culture.

Researchers have developed a multi-expert system for Optical Character Recognition (OCR) of historical Manchu documents, which exhibit diverse visual styles. The system utilizes a lightweight image classifier to route pages to specialized OCR checkpoints, improving accuracy in low-resource scenarios.

Researchers have developed Earthquaker-AI, a hybrid educational framework integrating a Retrieval-Augmented Generation (RAG) conversational AI assistant with a robotics project to teach primary school students about earthquake preparedness. The system uses RAG to provide accurate, grounded responses based on official guidelines and employs rubric-based feedback to support self-regulated learning and calmness during simulated emergencies.