Charlot Lab · Energy First Architecture · the fusion (internal)
EFA — the whole stack, one model
Every EFA mechanism, in a single model, running on your device. A shared sparse-positive latent is the one representation; a dendritic gate selects which of three physics-worlds it's in (learned one after another, without forgetting); a Hebbian fast-weight memory holds a goal it can no longer see; a JEPA world model predicts consequences; and it acts by planning in that learned latent. This page drives it live. Below, an honest note on the two seams that don't yet fuse for free.
obs→encoder → sparse latent→gate · which world→Hebbian Z · what's hidden→world model→plan → act
Loading the trained model…
the shared sparse-positive latent (live)
the gate · units active for this world
Hebbian memory Z · the held goal (goal hidden after step 0)
Honestly open (not dressed up): two seams don't fuse for free at nano scale. (1) Amortizing the planner into a fast reactive policy is substrate-limited in the continual/gated setting — neither policy nor world-model capacity closed it. (2) The Hebbian memory's train==live requirement is proven in a classification task (chance→99%) but breaks in this navigation loop (training on a drifting recall corrupts the policy). So this page acts by run-time planning, which is what genuinely fuses and scales. The world model, the gate, and the memory are the parts that hold.