Composing Intelligence: Lessons from AI in Networks and Beyond

As AI becomes increasingly integral to cellular networks, from the RAN to the edge, we are beginning to understand what it takes to move from model development to system integration. Across deployments — from reinforcement learning for link adaptation, to early use of foundational models for anomaly detection, to modular frameworks like rApps — we have seen both the promise and the architectural friction that emerges when AI meets real-time, distributed infrastructure. These experiences reveal more than constraints — they surface the design conditions for AI in networks: temporal dynamics, coordination across layers, system observability, and the role of domain priors. They also raise a deeper question: if foundation models like GPTs have transformed text and language understanding, what is the equivalent paradigm shift for networks? What is the representational or architectural breakthrough that will enable models to reason, adapt, and collaborate in the fabric of network systems? We conclude with a forward-looking perspective on AI-native 6G — not as a catchphrase, but as a co-design challenge. One where models and infrastructure evolve together, and where intelligence is embedded not just at the edge, but throughout the system — distributed, agentic, and ready to act.