Despite all of AI's recent success, various amusing examples of AI failure have become popular on the internet, whether it is a misbehaving self-driving car or a hallucinating large language model. This is not good for the progress of the field, with various stakeholders now calling for a pause in frontier AI development until it becomes trustworthy, fair and reliable. A common thread across many of the failure cases points to an inability of current AI to handle exceptions. The so-called out-of-distribution and multi-hop problems are another manifestation of the same limitation. Purely data-driven AI, also known as Machine Learning, could be said to be incapable of handling exceptions by its very definition as improvement of generalization performance from examples. By contrast, neurosymbolic AI combines neural network learning with knowledge representation and reasoning to address the above issues of trust, fairness and reliability. In neurosymbolic AI, exceptions can be expressed in a formally-defined logical language, expert analysis of trained networks and intervention can produce compact representations satisfying logical constraints. In this talk, I will review recent progress in neurosymbolic AI and earlier theoretical results. I will argue that, to be sustainable in the broad sense of the word, AI will need to: (1) learn compressed models from fewer examples incorporating general rules and exceptions; (2) produce descriptions of what has been learned allowing validation of results and sound reasoning; (3) enable direct model user intervention without the need for reinforcement learning with human feedback. In the next five years, neurosymbolic AI is expected to scale up to enable AI that can not only answer but also ask questions, make conjectures and check its understanding towards trustworthy, reliable and safer AI.