Percepta Introduces Spotlight: Decoupling Intelligence from Memory in Neural Networks
AI research company Percepta announced Spotlight, a novel neural architecture that decouples an intelligence computation module from an external writable memory. This design allows models to continuously expand their knowledge base and acquire new capabilities without retraining model weights or increasing per-token computation costs. By replacing traditional attention mechanisms with an arbitrarily sparse memory system, Spotlight addresses the fundamental tradeoff between model capacity and inference cost in large language models. This paradigm could enable AI models to gain unbounded knowledge over time while keeping runtime computation fixed. Unlike Mixture-of-Experts architectures that activate a fixed fraction of parameters, Spotlight uses arbitrary sparsity where each token indexes only a small, fixed number of memory cells regardless of total memory size. Because the writable memory stores both procedures and facts, the model can absorb new operational skills dynamically without modifying its core intelligence weights.
## BACKGROUND
Traditional Transformer-based large language models store knowledge directly within their static model weights during training, making knowledge updates computationally expensive. Furthermore, standard attention mechanisms require processing context lengths that scale computational costs significantly during inference. Decoupling compute logic from external writable memory allows neural networks to store continuously growing context without escalating hardware requirements.