mini-AGI Shows Continual Learning on Laptops

The compelling idea is not the AGI branding but making continual-learning experiments runnable on consumer hardware. The evidence remains prototype-scale and largely self-reported.

• –Byte-level input eliminates tokenizer and vocabulary engineering for arbitrary text and code.
• –Disk-paged mixture-of-experts architecture separates total model capacity from available VRAM.
• –The project reports dramatically reduced forgetting, but the result has not been independently validated.
• –Developers can train on their own files, though runtime, hardware, data quality, and model capability remain major constraints.
• –Community analysis confirms the design is technically interesting while emphasizing its toy-level status. [Technical breakdown](https://ai.thesatyajit.com/articles/mini-agi)