Charlot Lab · CORE · phase 1 spike (internal)

CORE-nano: one sparse latent, doing two jobs

The whole SOTA-PAI bet rides on one identity: a network's sparse, positive activation space can be both the readable feature space (BDH's interpretable synapses) and the world-model space (JEPA's latent). This is the smallest test of it. A random encoder turns each frame of a moving world into a sparse-positive code; a Hebbian fast-weight matrix writes the transitions between codes while it watches, no gradient, no training step. Then it dreams the world back from those synapses alone. If the dream matches, the substrate holds.

world→encoder→sparse-positive latent x→Hebbian memory Z += x′·xᵌ→read Z → predict next

the world it watches (a dot on a loop)

sparse-positive latent x · —

the synaptic memory Z · written by a Hebbian rule at inference

neurons—

latent sparsity—

all activations ≥ 0—

next-frame prediction—

dream matches world—

Press Watch: the dot runs its loop, and each transition writes one Hebbian outer product into Z, the bright off-diagonal blocks are the dynamics, stored in synapses, readable. Then Dream it back: starting from one frame, the core reads Z to predict the next sparse code, re-sparsifies it (the ReLU + top-k that makes it a BDH-style state), and rolls forward, reproducing the world from memory alone. That is a world model living in fast weights, written at inference. Phase 2 adds the JEPA head that trains this prediction and lets it plan; phase 1 only has to show the shared sparse latent carries it.