Energy consumption places fundamental limits on the speed and accuracy of computation in both brains and machines. Remarkably, the human brain runs on only ~20 watts, orders of magnitude less than current AI models, which demand ~20 megawatts of electricity. How does the brain compute so efficiently? Can we adapt energy efficiency mechanisms drawn from natural nervous systems to inform the design of AI? We hypothesize that the brain can perform rudimentary computations within segments of dendrites, the branching tree-like structures that receive neurotransmitter signals from other neurons. By delegating some simple decisions to the dendrites, the neuron can sort through the din of inputs in a much more energy efficient manner than listening to the thousands of individual synapses throughout the dendritic tree at once. We propose to use new light-emitting reporters of to visualize in real time, with submillisecond- and sub-micron resolution, how electrical signals are initiated and processed locally in dendritic segments in response to specific patterns of neurotransmitter inputs, both in neurons in a dish and in living fly brains. These measurements will be used to test models of dendritic information processing and will reveal fundamental principles of neural computation. In doing so, we hope to reveal how evolution dramatically surpasses current engineering in energy efficiency, laying the groundwork for the next generation of AI chips to transcend thermal constraints through neuromorphic computing.