Guardian Angels — LLM Personalization for Productivity and Security
Source: HN, gwern.net (Gwern Branwen, 2025-12-01 / updated 2026-06-05) HN Discussion: https://news.ycombinator.com/item?id=48906041 Original: https://gwern.net/guardian-angel
Summary
Gwern proposes Guardian Angels (GA) — personalized LLM digital twins that emulate a single user’s personality, values, and preferences. This solves the principal-agent problem by unifying principal and agent. The user becomes “CEO/board” of an “AI corporation.”
Technical Components
- Online learning via dynamic evaluation — realtime weight updates matching user preferences
- Active learning (DAgger) — query principal for corrections, obtaining low-regret bounds
- Preference learning — learn user’s values from pairwise comparisons
- CLI-first logging UX — local, transparent, auditable interface
- Inner-monologue search + data augmentation — production-quality outputs
Core Critique of Current LLMs
- Tool AIs want to be agent AIs → economic incentive to replace users, not amplify them
- Chatbot personas are misaligned with users, aligned with their owners
- Amdahl’s law means human-as-bottleneck gets optimized away
- Current LLMs produce “creatively-stunted, endlessly repetitive prose”
Relevance to Hermes Agent
- Hermes’ user-profiling skill and gbrain memory system are early steps toward a GA-like architecture
- The “local CLI-first UX” matches Hermes’ design philosophy
- Personalization through memory (scope recall, gbrain) rather than fine-tuning
- Gwern’s critique of “passive offline data collection” validates Hermes’ active learning approach
- The GA security model (defending against synthetic media, spearphishing) maps to Hermes’ safe-agent skill pack