September 2023 Summaries

Amazon Bedrock is a fully managed AWS service that facilitates the development of generative AI applications while ensuring privacy and security. It integrates Cohere’s Command model, which allows AWS customers to easily incorporate AI features like content creation, news summarization, and virtual assistance into their workflows without needing to train models or build infrastructure. The service supports secure enterprise environments with enterprise-grade security tools and facilitates scalable deployments using optimized AI workloads. Cohere's models are trained from scratch with known or public data sources and undergo rigorous testing for bias and integrity. This integration simplifies access to large language models (LLMs) and enhances productivity by enabling automation of routine tasks.

Cohere has introduced a Chat API with retrieval-augmented generation (RAG) capabilities, enabling developers to integrate dynamic conversational AI into applications. This API allows developers to enhance conversational AI products with reliable and relevant information by retrieving data from external sources like web searches or documents. By including citations from these sources, the Chat API improves the accuracy, credibility, and trustworthiness of AI-generated content, addressing common issues such as hallucinations. The introduction of modes—Document, Query-generation, and Connector—offers flexibility in how information is retrieved and used, allowing developers to create customized solutions for various applications, including market research assistants, customer support, and knowledge-based systems. The public beta of the Chat API is now accessible to developers with a Cohere account, showcasing its functionalities through the Coral Showcase without the need for coding.

A recent discussion highlighted the transformative potential of generative AI, particularly in enhancing workforce productivity across various sectors. Experts Erik Brynjolfsson and Andrew McAfee emphasized that generative AI, akin to electricity, is a general-purpose technology with the capability to revolutionize industries by automating tasks and augmenting human abilities. They discussed research illustrating significant productivity gains, such as a 35% improvement in call centers, and highlighted the importance of companies embracing this technology to avoid being left behind. The conversation also touched on challenges such as data quality, skills, and ethical concerns but underscored the necessity for companies to develop strategic plans and take action to harness AI's potential. The dialogue concluded with a call to action for businesses to prioritize generative AI initiatives to remain competitive and innovative in the rapidly evolving technological landscape.

At a White House event, Cohere joined other enterprise AI leaders like Adobe, IBM, and Nvidia in committing to the US Administration's updated voluntary guidelines for managing AI risks. This move reflects Cohere's dedication to addressing the challenges and opportunities in the enterprise AI sector, with a focus on responsible adoption and regulation. Cohere's President and COO, Martin Kon, emphasized the importance of safeguarding society's interests, including data protection. Cohere has been actively engaging with governments to develop effective regulations, such as Canada's Artificial Intelligence and Data Act, and has partnered with industry peers to establish best practices for deploying large language models (LLMs). The company views these commitments as a foundational step toward creating a regulatory environment that responsibly harnesses AI's potential.

Cohere has introduced "North," an enterprise-ready AI platform designed to enhance workplace productivity by leveraging modern AI tools. Among its offerings are Compass, an intelligent search and discovery system, and Command, a suite of scalable language models. A key focus is on T-Few finetuning, a technique that optimizes the finetuning of large language models (LLMs) by updating only a fraction of the model's weights. This approach reduces training time and resource use, allowing multiple finetunes to share GPU resources for efficient serving. T-Few's ability to stack finetuned models enhances serving scalability by enabling concurrent inference on a single GPU, thus maximizing GPU utilization. The process involves a finetuning workflow that updates model weights, making it highly portable and efficient, with the finetune weights being a small fraction of the baseline model size. This methodology is particularly beneficial for applications requiring efficient and high-performance language models, offering solutions across various industries including technology, financial services, healthcare, manufacturing, and the public sector.

Cohere, a company focused on advancing generative AI and natural language processing (NLP), highlights recent developments in the field, emphasizing collaboration and community involvement through initiatives like Cohere For AI. The company introduces ToolLLM, a framework designed to enhance open-source large language models (LLMs) for better API instruction-following capabilities, and explores the challenges and solutions associated with reinforcement learning from human feedback. Other notable research includes instruction tuning for LLMs to improve code understanding and generation, methods to align LLMs with human expectations, and innovative approaches to reduce hallucinations in smaller LLMs. These advancements aim to improve the effectiveness, reliability, and safety of AI models by addressing existing limitations and exploring new training methodologies.