LLMs

Stanford study reveals Western bias in AI chatbot alignment, compromising global effectiveness. Examined nine languages and regional dialects, finding cultural nuances lead to misunderstandings. Researchers explore causes and solutions for more inclusive AI.

OpenAI's SearchGPT prototype combines AI models with web info to provide quick answers and clear source attribution, currently in limited testing.

OpenAI's new Rule-Based Rewards method improves AI safety without extensive human data collection. It uses simple rules to evaluate outputs, balancing helpfulness and safety in AI models.

NVIDIA launches four NeMo Retriever NIM microservices to improve LLM accuracy and performance. These tools enable efficient retrieval-augmented generation, connecting models to business data. They offer 30% fewer inaccurate answers in enterprise applications and integrate with various platforms.

Mistral AI and NVIDIA launch Mistral NeMo 12B, a 12-billion-parameter language model for enterprise use, excelling in diverse tasks and easy customisation.

Mistral AI releases Mathstral, a STEM-focused language model built on Mistral 7B. It excels in mathematical reasoning, achieving top performance on MATH and MMLU benchmarks for its size. The model is available for use and fine-tuning.