Chappie

The idea

Instead of one big model, Chappie is a population of small specialist agents that live. Every tick a scheduler ranks them by priority — relevance, plus coupling to whoever is currently active — and places only the attended few on the GPU (hot); the working set stays warm on CPU; the long tail is unloaded. They deliberate in a shared workspace and reach consensus. When energy runs low, Chappie sleeps: it replays the day's episodic memory to consolidate connections and train its long-term weights.

One moment of life

perceive → attend → schedule (GPU / CPU / cold) → deliberate
        → consensus → act → record episode → [tired?] sleep → consolidate

Regions

Memory = Complementary Learning Systems

A fast, non-parametric heap (episodic memory) is replayed during sleep to slowly improve weights (a real Burn network trained by autodiff on the GPU). Interleaved replay keeps it from forgetting. The result: a model that measurably improves over a lifetime.

Run it

git clone https://github.com/fentas/chappie && cd chappie
mise run life            # one deterministic life (pure-std, fast)
mise run life-burn       # with the Burn neural agent (long-term memory)
mise run bench           # record a git-tagged benchmark