mini-AGI
A continual learning language model that trains from scratch on a single 8 GB VRAM GPU, assembling its own architecture and adapting its size to the hardware and data available rather than shipping as a fixed pretrained model.
mini-AGI is a continual learning, byte-level language model written in Python and released under the MIT license. It assembles its own architecture and trains from scratch on a single 8 GB VRAM GPU, learning continuously from everything it reads rather than being trained once and then run as a fixed model.
It stores its weights as ordinary files on disk and pages them onto the GPU as it needs them, so its size is bounded by free disk space rather than by VRAM. It grows new capacity while training when it runs short, prunes capacity that nothing asks for, and serves through the same code path it trains on.
The project describes the current model as a small, toy-level experiment rather than a frontier system, meant to show that continual learning from a single stream of data without catastrophic forgetting is possible on modest hardware. Its weights are not yet published; the current training run is still on its first pass over the corpus.