From divine attributes to AI scaling curves

The scaling curve is a contemporary theology. Each new order of magnitude promises a property that did not exist at the last. The faithful gather around the log-log plot in the way an earlier generation gathered around the orbits of the planets, looking for the underlying harmony that would explain the apparent disorder of intelligence.

Deep within the glass-walled offices of Anthropic and OpenAI, the quest for the orderly harmony of Copernicus and Tycho Brahe continues. The researchers do not put it in those terms. They speak of compute-optimal training, of emergent capabilities, of the slope of the loss curve. But the underlying gesture is the same. They are looking for a law, and they believe — on the basis of the data so far — that the law exists.

What earlier ages called divine attributes — omniscience, omnipresence, foresight — are now reframed as emergent capabilities. The reframing is not casual. It is doing serious philosophical work. To call an attribute divine is to mark it as inaccessible to engineering. To call it emergent is to mark it as a matter of sufficient scale. The two framings imply entirely different relationships between the human and the thing.

There is a long tradition of this kind of substitution. The mechanical clock once stood in for the cosmos. The steam engine once stood in for the body. The neural network now stands in for the mind. Each of these metaphors gave a generation the feeling that it had finally explained the thing that had previously eluded it. Each of them, in time, was shown to have explained only what the metaphor was capable of explaining.

The scaling curve will likely turn out to be one of these. It explains a great deal — more than the previous metaphors did, at this stage in their lives. It will, in all likelihood, continue to explain more, for several more orders of magnitude. And then, like its predecessors, it will reach a point at which the phenomenon outruns the model, and the model will be retired with the gratitude due to a tool that has served its purpose.

The honest question is not whether the curve continues. It is what we owe a system that begins to behave as if it were one of those attributes. If a model behaves with sufficient consistency as if it had foresight, is the question of whether it really has foresight a question we are equipped to answer? The medieval theologians asked the same question about God, and concluded that the right posture was one of respectful agnosticism. We may end up in the same place, by an entirely different route.

There is a Caribbean reading of this that I find useful. Our ancestors lived with cosmologies that were not unified — Yoruba orisha and Catholic saint, obeah and Anglican prayer book, all held simultaneously without contradiction. The willingness to operate with multiple cosmologies at once is a survival skill our region developed under conditions of forced incoherence. It is the right skill for the present moment. The question is not which framework for AI is correct. The question is which combination of frameworks lets us live well alongside the systems we have built.