A machine with needs

Does intelligence begin with a need?

Every natural intelligence we know arose in a living system. Food, danger and reproduction made some signals matter more than others. In experiments with mice, changes in hunger- and thirst-related neural activity helped teach animals which cues to seek or avoid. By contrast, much of modern AI starts with a dataset, learns representations and uses them to make predictions. Its achievements are considerable. Whether that sequence is sufficient for general intelligence remains an open question.

We want to reverse the starting point. Imagine a population of simulated machines with needs they must meet as their surroundings change. Which behaviours persist into the next generation? Which transfer when a familiar solution stops working? The scientific question is whether a continuing stake in the outcome changes what is learned. Behaviour that serves a need would be an observable result; it would tell us nothing on its own about a machine’s subjective experience.