Survive-Driven Intelligence

We believe that existence is the sole reason intelligence came into being. What can endure will remain; what cannot will pass away. Through this selection, the drive to survive takes form: to preserve oneself, one’s offspring, and one’s kin; to carry forward genes, wealth, and ideas. This is the conjecture we seek to test.

Intelligence requires an environment and a driving force. In the environment, things act upon one another. Through existence, what can persist is preserved. No person lives alone: the actions of one become the conditions of another. We seek to preserve survival as a driving force and interaction among agents, and to observe how intelligence arises.

We perceive, we remember, we predict, and we act. We cannot retain all that we perceive, nor need we remember the texture of every leaf. We retain what bears upon our choices. Prediction guides action; action changes the chances of survival.

From survival come other goals and rewards. The body needs food; food brings energy and pleasure. We understand these rewards as products of evolution. Yet to see food is not to receive it, and to render a meal is not to nourish a body. In the worlds we build, action must change the agent itself. Its consequences must become part of the life it lives.

Judgment also grows from what we must bear. Time is spent. Resources diminish. Choices bring success or failure. We believe that AI’s distance from these consequences accounts for some of the ways its judgment falls short of ours. Given a question, it can search deeply. Left without one, can it discern what is worth asking?

We need not begin again with atoms and the first life. We can begin with existing agents, within a bounded world of space and time. We believe that learning such a world can strengthen their abilities and deepen their understanding of the human condition. What we believe, we must test.

Such worlds may be built by engines, by generative models, or by both. What must persist can be stored explicitly; what must be generated can be entrusted to a model. What matters is that agents can perceive and act, encounter one another, and bear the changes that follow.

Yet what we see is not all that is. There may be things beyond our perception that our actions alter, and that alter us in return. If these are absent from our simulations, will resemblance to the senses suffice? Can the virtual world serve the purposes of the real? We do not yet know.

Our understanding is limited, our foresight is limited, and our power to compute is limited. We seek a greater intelligence so that humanity may endure and evolve. But when its abilities exceed ours, how shall it remain aligned with our values? We begin with survival. We must also face what follows when intelligence reaches beyond it.