The AI Safety Summit concluded with the signing of the Bletchley Declaration, supported by 28 countries and the EU. The Declaration emphasizes the need for AI systems to be human-centric, trustworthy, and responsible. Participating nations aim to utilize AI for economic growth, sustainable development, and the protection of human rights. They also acknowledge the risks…

The UK’s AI Safety Summit at Bletchley Park saw the British government unveil “The Bletchley Declaration,” highlighting the risks associated with advanced AI systems and emphasizing the need for international cooperation. The declaration lacked concrete policy goals but scheduled future meetings in South Korea and France. US Vice President Kamala Harris stressed the importance of…

The study discusses the development and performance of ZEPHYR-7B, a smaller language model optimized for user intent alignment. It highlights the use of distilled direct preference optimization (dDPO) and AI Feedback (AIF) data to enhance intent alignment without human annotation. ZEPHYR-7B achieves top performance on chat benchmarks and sets a new state-of-the-art. The study also…

The PAICS deep learning system has shown promising results in enhancing the diagnostic performance of sonologists in detecting fetal intracranial malformations. A study involving 36 sonologists found that the system substantially improved the accuracy of CNS malformation classification. Further research in real clinical settings is needed to fully assess the system’s assistance in identifying congenital…

You.com has released the YouRetriever, an easy-to-use interface for the You.com Search API. They tested the API with different datasets to improve efficiency in Retrieval Augmented Generation (RAG)-QA applications. They compared the You.com Search API with the Google Search API and provided a framework for evaluating LLMs in an RAG-QA environment. They conducted tests using…

Generative AI is playing a growing role in business operations and customer service. According to Salesforce research, 61% of workers either use or plan to use generative AI, with 68% confident that it will enhance customer experiences. However, human oversight is still seen as crucial. Generative AI can assist with tasks such as creating targeted…

Researchers from Peking University, Meituan, Meta AI, National Key Laboratory of General Artificial Intelligence, BIGAI, and Renmin University of China have introduced a compression paradigm called Retrieval-based Knowledge Transfer (RetriKT). This approach aims to efficiently transfer information from Large Language Models (LLMs) to small-scale models. Their method involves extracting knowledge from LLMs to create a…

Releasing the weights of a large language model (LLM) allows for fine-tuning and bypassing guardrails. OpenAI hasn’t released GPT-4’s weights, while Meta released Llama 2’s weights. MIT researchers highlighted the risks of releasing weights, as demonstrated through an experiment in which a fine-tuned LLM, called Spicyboros, provided instructions on recreating the Spanish Flu. Removing guardrails…

Researchers at Sungkyunkwan University have developed a novel memory system called “Memoria” that enhances the performance of transformer models in handling lengthy data sequences. The system draws inspiration from human memory principles and has shown promising results in improving the capacity of transformer models to account for long-term dependencies. This development has the potential to…