Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM
- ID
- 26961
- Status
- summarized
- Published
- 21 Sep 2026, 12:42 PM
- Fetched
- 23 Sep 2026, 7:13 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://github.com/volotat/mini-AGI/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 5.5
- Created
- 23 Sep 2026, 7:14 AM
- Tags
- Audience
- developersai_ml_learners
What happened
A Hacker News user posted 'mini-AGI,' a byte-level language model that trains from scratch on a single 8GB VRAM GPU using a batch-1 data stream. It pages weights from disk files onto the GPU as needed (so parameter count is bounded by disk, not VRAM), dynamically grows capacity when it runs short, prunes unused parameters, and uses the same code path for training and serving. The author explicitly calls it a 'toy-level' experiment demonstrating continual learning without catastrophic forgetting on modest hardware; weights are not yet published as the first corpus pass is still ongoing.
Why it matters
The architectural techniques here—disk-paged weights to decouple parameter count from VRAM, dynamic layer growth/pruning, and same-path train-and-serve—are concrete patterns builders could borrow for edge or low-resource ML projects, especially relevant in markets where cloud GPU access is expensive. However, since the model is toy-level with no published weights, the value is conceptual rather than deployable right now.
Discussion angle
Discuss whether disk-paged weight loading and dynamic architecture growth are practical patterns for local/edge ML in Southeast Asia where GPU cloud costs are prohibitive, versus just fine-tuning existing small open models.