- Proposes Guardian Angels (GAs) as highly personalized LLMs that emulate a user's personality, values, and preferences to amplify productivity and enhance security against AI threats.
- Highlights limitations of current frozen chatbot LLMs, including mode-collapse, laziness, brittleness, prompt vulnerability, and inability to learn permanently, which hinder meaningful productivity gains.
- Suggests combining techniques like dynamic evaluation (online learning), active learning, preference elicitation, and inner-monologue search/data augmentation to create effective GAs.
- Emphasizes core principles for GAs: enhancement over replacement, mental sovereignty (alignment with the principal), and self-actualization (helping the principal develop).
- Discusses practical implementation aspects, including security requirements (e.g., tamper-proof cloud servers), cost considerations (targeting >$1,000/month), and organizational structure (preferring a startup model).
- Proposes initial steps like prototyping a 'Gwern Branwen Transformer' using a large personal corpus to explore GA feasibility and aiming for a 100× productivity increase in creative writing.