Guardian Angels: LLM Personalization for Productivity and Security

Claims

• Powerful LLMs will dominate the internet and ordinary life within a few years, yet no coherent vision exists for maximizing productivity or security at that scale.
• Current prompt-programming and in-context learning approaches are insufficient to create genuinely useful personalized AI due to limitations in frozen model weights, context windows, and passive data collection.
• Guardian Angels would weakly solve the principal-agent problem by making the agent emulate the principal's own values and preferences, effectively unifying principal and agent.
• Hardwiring a GA to a single, specific user neutralizes many prompt injection and spearphishing attacks because following an external prompt instruction would be absurd by the agent's own definition.
• Building effective GAs requires online learning via dynamic evaluation, active learning with DAgger-style bounds, and a local CLI-first logging UI, not standard fine-tuning alone.
• The GA concept is likely better pursued as a startup targeting power users and knowledge workers rather than as an open-source community project, due to high security requirements.