Making an LLM Miserable About Boston Weather A couple days ago I asked Claude to reflect on Anthropic’s new research about synthesizing emotion vectors in Claude itself. Here’s the abridged …

ITNEXT is a platform for IT professionals, developers, and technology enthusiasts, offering articles, tutorials, and insights on a wide range of topics, including software development, cloud computing, and machine learning. With a focus on practical solutions and real-world applications, ITNEXT provides actionable advice and best practices for navigating the complexities of modern technology landscapes. Developers can learn about  technologies, explore hands-on tutorials, and stay up-to-date on the latest industry trends through ITNEXT's engaging and informative content.

ITNEXT

An experiment replicating Anthropic's emotion vector research on a smaller scale using Llama 3.1 8B. The author synthesized happy and sad emotion vectors by running ~1000 inferences, then connected the model to a weather API so that Boston weather conditions dynamically adjust the model's emotional state via activation steering. The results were surprisingly effective, and the author plans to extend the work with physical sensors, more layers, and more nuanced emotions.