This ability of DL to convert streams of raw noisy data into labeled objects seems like exactly what's needed to solve an intelligent agent's perceptual grounding problem, where an agent that's new to the world must bootstrap its perception systems, converting raw sensory input into meaningful objects with physical dynamics. Only then can the agent reason about objects and better understand them by physical interaction and exploration. This is one of the areas where symbolic AI failed hardest, but DL does best.
With some engineering, it's easy to imagine how active learning could use DL to ground robot senses - much like an infant human explores the world for the first year of life, adding new labels and understanding their dynamics as it goes.
I suspect the potential for DL's many uses will continue to grow and surprise us for at least another decade. If we've learned anything from the past decade of DL, it's that probabilistic AI is surprisingly capable.
Instead of being stuck on the fact that deep learning and the current methods seem to have hit a limit I think I am actually excited about the fact that this opens the door for experimenting other approaches that may or may not build on top of what we call AI today.