Hacker News front pagevolotat20 min readadvanced
Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM
Summary
mini-AGI is a byte‑level continual‑learning language model that fits on a single 8 GB GPU by paging expert weight files from disk. It uses two dense prelude blocks followed by a recurrent block applied up to 24 times per character, with adaptive depth (PonderNet halting) and per‑application top‑8 expert routing. The system maintains a RAM cache and a VRAM resident set, swaps experts based on dema…
- Continual learning without catastrophic forgetting is possible on consumer‑grade hardware by separating a small always‑resident trunk from a large disk‑backed expert pool.
- Adaptive depth via a halting head (PonderNet) lets easy characters use few compute steps while hard ones use more, keeping average compute per token low (≈8–10 rows).
- Per‑application expert routing (top‑8) and dynamic growth/pruning allow the model to allocate capacity where needed without a fixed tokenizer or vocabulary.
- Keeping Adam moments with each expert file and using hysteresis for cache replacement prevents momentum contamination and reduces churn.
If personal LLMs can be trained and continuously updated on a laptop, the barrier to owning a private, data‑specific model drops dramatically. The design shows a path to scalable, on‑device continual learning that avoids the usual trade‑off between model size and hardware limits, and it provides co…
8/10
