December 2023 Summaries

In 2023, the organization made significant strides in AI safety and policy research, contributing to projects like Evaluating the Social Impact of Generative AI Systems and Goodtriever, while actively engaging with experts to address the societal and economic impacts of AI. They attended the UK AI Safety Summit in London to discuss international coordination in mitigating AI risks and launched the Cohere For AI Research Grant Program to support academic research in machine learning, granting over 20 research grants to encourage innovation in areas such as natural language understanding and biomedical integration. As they look toward 2024, the organization aims to continue advancing machine learning research by fostering open collaboration and diverse perspectives, while remaining committed to fundamental research and exploring new frontiers in the field.

Cohere has introduced fine-tuning capabilities for its Chat endpoint, allowing developers to customize AI-powered chatbots to meet specific style and knowledge requirements, thus enhancing user engagement and satisfaction. This customization is particularly impactful in specialized roles, such as hotel concierges, IT customer support, and internal finance assistants, where chatbots can be tailored to exhibit domain-specific behaviors and language. An experiment with the ConvFinQA system demonstrated the effectiveness of fine-tuning by improving the accuracy of financial query responses to 68.1%, highlighting the potential for tailored chatbots to outperform generic models in specialized contexts. Developers can access these fine-tuning capabilities programmatically through an SDK or Cohere’s new interface, with additional documentation available for further guidance.

Cohere's recent blog post highlights significant advancements in generative AI research, focusing on various innovative methods and technologies aimed at improving the efficiency, transparency, and safety of AI systems. The blog covers a series of papers curated by the Cohere For AI research community, addressing issues like data transparency, efficient evaluation of large language models (LLMs), toxicity mitigation, privacy preservation, structured pruning for model optimization, and controlled decoding for aligning LLMs with specific objectives. Notable contributions include the Data Provenance Initiative for dataset licensing transparency, the GOODTRIEVER method for adaptive toxicity control, and BitNet, a 1-bit Transformer architecture for energy-efficient scaling of LLMs. Additionally, Representation Engineering (RepE) is discussed for enhancing AI transparency by focusing on high-level cognitive phenomena in neural networks, while Self-RAG is introduced as a framework for improving the factuality of LLM outputs through self-reflection. These explorations reflect the fast-evolving landscape of natural language processing (NLP) and the growing emphasis on responsible AI development, with Cohere positioning itself as a key player in pioneering these advancements.

Neil Thompson, Director of MIT FutureTech Lab, offers insights into the evolving landscape of generative AI, emphasizing the role of innovation in societal advancement. With a diverse academic background, Thompson highlights the intricate relationship between hardware advancements, specialized computing, and AI development. His co-authored paper, "The Grand Illusion," explores the challenges of software portability amid the trend of specialized chips like GPUs and TPUs, warning that this specialization might hinder future exploration and innovation. He notes the shift from general computing, driven by Moore's Law, to specialized systems, raising concerns about the potential loss of flexibility and unforeseen innovation opportunities. The conversation also delves into the implications for AI models, discussing whether the future lies in large general models or specialized ones. Thompson advises executives to consider deep learning as a continuum, the economic implications of model scale, and the strategic value of high-quality data.

Cohere has recently expanded its operations by opening a new office in Paris, which will serve as its hub for Europe, the Middle East, and Africa (EMEA). This strategic move is part of Cohere's broader initiative to enhance its global presence and cater to the growing needs of businesses in these regions. The company is also actively involved in modernizing national defense strategies through AI adoption, highlighting its commitment to advancing technology in the public sector. Additionally, Cohere has launched "Command A Translate," a secure translation service aimed at facilitating communication for global enterprises, reflecting its focus on providing innovative solutions to meet diverse business requirements.

As companies increasingly adopt generative AI technologies, securing these systems has become a critical concern due to rising threats such as cyberattacks and data breaches. Large language models (LLMs) and retrieval-augmented generation (RAG) systems, which integrate proprietary knowledge, introduce unique vulnerabilities that traditional web applications do not face. Among these is prompt injection, a type of exploit where attackers manipulate LLMs through crafted input prompts to generate sensitive or malicious outputs. Securing LLM applications requires a comprehensive approach involving traditional web security standards, machine learning security operations (MLSecOps), and new LLM-specific concerns. Mitigation strategies should include threat modeling, robust infrastructure, and secure plugin design, alongside maintaining a strong security culture and integrating security measures early in the development process. Real-world examples illustrate the potential for sensitive information disclosure, fraudulent scams, and supply chain attacks, emphasizing the need for continuous monitoring and collaboration across the AI ecosystem to address evolving threats effectively.

Cohere has updated its fine-tuning capabilities for classification models, aiming to enhance performance and expand options for multilabel and multilingual classification. This advancement allows enterprises to employ fewer data points, with a new minimum of 32 text-label pairs required for effective model training, and results in an average accuracy improvement of 30% for small datasets. The improved system also delivers significantly higher throughput, processing over 2000 examples per second, and supports multilabel classification, enabling models to predict multiple categories simultaneously. Additionally, it offers a choice between English-only and multilingual base models, enhancing flexibility and accuracy across various language tasks. These developments are designed to facilitate new use cases in document categorization, customer support, healthcare, and more, offering a more robust and efficient classification toolset for enterprises.

Cohere has recently introduced fine-tuning capabilities to its Rerank model, enhancing its performance in complex domains such as legal, medical, and technical fields, where domain-specific jargon and intricate concepts are prevalent. The Rerank model, a semantic relevance scoring system, benefits from fine-tuning tailored to specific domains, improving search result relevance significantly. This enhancement addresses challenges posed by complex terminologies and domain-specific knowledge, outperforming large generative models like GPT-4 in terms of cost, speed, and predictability. The fine-tuned Rerank model demonstrated notable performance improvements, particularly in the legal domain using the CaseHOLD benchmark, where its accuracy doubled compared to the base model. Developers can access this fine-tuning feature through Cohere's SDK or a new interface, offering a cost-effective solution without the need for extensive computational resources.

Enterprises are increasingly integrating large language models (LLMs) into their technology strategies to harness the potential economic benefits of generative AI, which is projected to contribute up to $7.9 trillion annually. However, traditional AI practices may not be suitable for LLM projects, necessitating a reevaluation of workflows, team structures, and goals to maximize benefits. LLMs offer advantages such as lower upfront costs, faster time to value, and a broader skills spectrum, but also present challenges like high operating costs, security concerns, and potential biases. Successful deployment of LLMs requires strategic planning, careful project selection, robust team building, and effective feedback mechanisms. Security is paramount, and cloud-based solutions offer secure deployment options, although additional protection measures may be necessary to safeguard sensitive data. Understanding the distinct characteristics of LLM-based projects and exploring diverse use cases can provide companies with a competitive edge and significant value.