Gwern proposes 'Guardian Angels' – personalized LLM digital twins for productivity and security

In a LessWrong post, gwern outlines a vision for Guardian Angels (GA) – personalized LLM digital twins that emulate a user's personality, values, and preferences, moving beyond the standard assistant chatbot persona. By unifying the principal (user) and the agent (AI), GAs weakly solve the principal-agent problem: the user focuses on defining what is worth doing, while the GA handles execution and security, such as screening messages for advanced attacks like synthetic media propaganda or spearphishing. This defense-in-depth strategy is designed to counter rising threats from APTs equipped with Mythos-scale attackers.

Gwern argues that current techniques like prompt programming of frozen models are insufficient for creating useful GAs due to limitations in context windows, post-training, and offline data collection. He proposes a combination of online learning (dynamic evaluation) to keep models updated, sample efficiency via pretrained preference-oriented LLMs and active learning (querying the principal for corrections using DAgger-style bounds), and a local CLI-first logging-oriented UI/UX paradigm. While an open-source community effort is possible, gwern recommends a startup approach to ensure high security, initially targeting power users like CEOs and researchers, then scaling down as the technology matures.