Guardian Angels: LLM Personalization for Productivity and Security - By Gwern

In this post, Gwern argues for “Guardian Angels”: deeply personalized LLMs designed not as generic chatbots, but as trusted extensions of a specific person’s values, style, preferences, and goals. It’s a wide-ranging essay about productivity, security, AI-mediated work, and what it might take for future AI systems to genuinely amplify individuals rather than merely replace or manipulate them.* 00:00:00 - Introduction* 00:06:27 - Chatbot Incentives Are Misaligned* 00:10:03 - Chatbot Problems* 00:10:44 - Mode-Collapse* 00:13:09 - Laziness* 00:13:52 - Brittle Because Fast* 00:16:41 - Too Helpful* 00:18:14 - Amnesiac* 00:20:03 - Chatbot Fixes* 00:20:07 - Cooperative RL* 00:22:47 - Continual Learning* 00:22:50 - Catastrophic Forgetting* 00:23:48 - Generalizing* 00:26:32 - Creative Writing* 00:30:03 - Over-Parameterizing* 00:30:30 - Extremely Large LMs* 00:30:57 - Active Learning* 00:33:14 - Preference Learning* 00:36:10 - Brain Imitation Learning* 00:37:05 - Personality Emulation* 00:39:01 - Guardian Angels* 00:40:48 - Principles* 00:42:07 - Anti-Principles* 00:47:06 - UX* 00:48:23 - Use-Cases: Politics & Politics* 00:51:17 - Hardware* 00:52:57 - Cost* 00:55:11 - Organization* 00:56:18 - Startup Business Model* 00:57:43 - Competition* 01:00:20 - Initial Steps* 01:01:35 - GBT* 01:04:18 - For Writing* 01:06:44 - Data Augmentationhttps://gwern.net/guardian-angel Get full access to Askwho Casts AI at askwhocastsai.substack.com/subscribe