Explicit Foundation Model Optimization with Self-Attentive Feed-Forward Neural Units