Practices for Governing Agentic AI Systems

In large language models (LLMs), the challenge of keeping information up-to-date is significant. As knowledge evolves, these models must adapt to include the latest information. However, updating LLMs traditionally involves retraining, which is resource-intensive. An alternative approach, model editing, offers a way to update the knowledge within these models more efficiently. This approach has garnered…

The field of research focuses on integrating machine learning (ML) in healthcare for personalized treatment. This innovative approach aims to revolutionize how we understand and apply medical treatments, shifting from one-size-fits-all solutions derived from traditional clinical trials to more nuanced, individualized care. The essence of this research lies in predicting treatment outcomes tailored to individual…

This week the AI buildout stopped being about models and became about balance sheets. Anthropic committed $19 billion to a single data center, SK Hynix raised $28 billion, and SoftBank kept pouring in $87 billion, even as a leaked Treasury draft warned the whole thing rhymes with the dot-com crash. Governments moved too: Illinois made…

Privacy concerns have become a significant issue in AI research, particularly in the context of Large Language Models (LLMs). The SAFR AI Lab at Harvard Business School was surveyed to explore the intricate landscape of privacy issues associated with LLMs. The researchers focused on red-teaming models to highlight privacy risks, integrate privacy into the training…

In human-computer interaction, the need to create ways for users to communicate with 3D environments has become increasingly important. This field of open-ended language queries in 3D has attracted researchers due to its various applications in robotic navigation and manipulation, 3D semantic understanding, and editing. However, current approaches have limitations of slow processing speeds and…

Self-supervised learning (SSL) has proven to be an indispensable technique in AI, particularly in pretraining representations on vast, unlabeled datasets. This significantly reduces the dependency on labeled data, often a major bottleneck in machine learning. Despite the merits, a major challenge in SSL, particularly in Joint Embedding (JE) architectures, is evaluating the quality of learned…