Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM
Summary
Mini-AGI presents a continually learning, byte-level language model trained from scratch on an 8 GB VRAM GPU, using a dynamic mixture-of-experts architecture and adaptive depth. The project emphasizes on-device training with weights stored on disk and paging between disk, RAM, and VRAM, promoting ownership and ongoing learning. The README outlines architecture, paging, growth/pruning, and the claim that this is a toy-level demonstration rather than frontier capabilities.