Mini-AGI trains a continual-learning model from scratch on an 8GB VRAM laptop with batch size 1

GitHub / volotat (via Hacker News, 263 points)·medium signalXBlueskyLinkedInCopy link

A Show HN posted volotat/mini-AGI, an MIT-licensed continual learning model trained from scratch on an 8GB VRAM laptop against a batch-1 stream of data, which is the opposite of the large-batch regime everything else in the field assumes. The repo was created 2026-09-19 and last pushed 2026-09-22, sitting at 450 stars, and the HN thread reached 263 points and 108 comments. The interesting claim for builders is the batch-1 streaming setup rather than any benchmark number, since continual learning without catastrophic forgetting is the thing local-model users keep asking for.