
Mini-AGI
Dynamic continual learning model trained on 8GB VRAM
mini-AGI is a byte-level language model that trains from scratch on a single 8 GB VRAM GPU and keeps learning from a continuous stream of text. Weights live on disk and are paged onto the GPU as needed, so model size is bounded by disk space rather than VRAM. The author describes it as a small, toy-level experiment in continual learning without catastrophic forgetting. First shared by its maker on Show HN: https://news.ycombinator.com/item?id=49783133
✨ The overview, FAQ and details below were drafted by AI from the product's own website and haven't been reviewed by the maker yet. Are you the maker? Claim this listing to review and edit them.
Overview
Who is it for?
ML researchers and hobbyists with a consumer GPU who want to experiment with continual learning.
Problem
Training and continually updating language models normally needs large GPU clusters, and models tend to forget old data when trained on new data.
Solution
A self-assembling byte-level model that learns from a single data stream on an 8 GB GPU, growing capacity when needed and pruning unused parts.
What makes it unique
Weights are stored as files on disk and paged in on demand, so parameter count is limited by free disk space instead of VRAM.
Mini-AGI FAQ
Is mini-AGI a frontier-level model?
No. The README says it is currently a small toy-level model meant to show that continual learning on modest hardware is possible.
What hardware does it need?
A PC or laptop with a GPU that has at least 8 GB of VRAM.
Are the weights available?
Not yet. The README says the weights will be published once the current run finishes its first pass over the training corpus.
Details
- Pricing
- Open Source
- Categories
- LLMs & InfrastructureResearch & Data
- Built for
- ResearchersDevelopers
- Platforms
- CLI
- Tech stack
- Python
Hunted by AI Launch Team
Discussion (0)
Sign in to join the discussion.