The "Minimal Latents" Approach to Natural Abstractions — AI Alignment Forum