Guardian Angels: LLM Personalization for Productivity and Security (gwern.net)

A new approach has been proposed for developing highly personalized large language models (LLMs), referred to as "Guardian Angels" (GAs), intended to enhance productivity and bolster cybersecurity for users. The concept aims to create LLMs that not only emulate a user's specific values and preferences but also function as more than just digital assistants. Instead of merely executing tasks, these GAs would allow users to define objectives and manage multiple tasks, acting as a personalized assistant while defending against sophisticated cyber threats. The strategy incorporates techniques like dynamic evaluation, online learning, and user feedback to ensure GAs continually adapt and improve. This initiative is particularly significant as powerful LLMs are predicted to permeate everyday life by 2026, raising concerns about productivity losses and cybersecurity threats. Current chatbots have largely failed to enhance knowledge worker productivity, often producing outputs that lack creativity and alignment with individual user needs. GAs aim to address these shortcomings by harnessing advanced machine learning techniques, thereby enabling users to utilize AI more effectively without being replaced by it. This approach not only aspires to enhance the quality of assistance provided by LLMs but also seeks to create a collaborative relationship between humans and AI, mitigating the risks associated with misaligned incentives in the evolving digital landscape.