Perhaps we should expect applications to evolve to become AI-enabled by embedding small local models that are specialized to each application. However, such models would have enough agency to reach out to services enabled by large models, whenever the local small model needs expertise that falls outside its own. Just like how the model should know when it should perform Web browsing to fetch the most up-to-date information; or it should invoke a calculator to do arithmetic; or it should invoke Mathematica to solve differential equations analytically; or it should generate Python and execute it to perform computations; or it should use online services for payments, commerce, and logistics; similarly, it should know what online services (large models) to collaborate with to take advantage of expertise beyond its own limited capabilities.