The Relentless Arithmetic

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes a moment when a capability threshold was crossed that felt sudden but had been predicted by a line on a graph for decades. The relentless arithmetic is the mechanism behind that threshold. The machines that now write fluently and reason credibly are not the product of a conceptual breakthrough in our understanding of intelligence. They are the product of the compute Moravec said would arrive, applied to architectures that had existed for years and were waiting for the hardware to catch up.

The concept connects to Moravec’s paradox at a deep level. The order of conquest—abstract reasoning first, embodied competence last—was predicted from the difficulty gradient. But the timing of each stage was predicted from the arithmetic. The relentless arithmetic explains why the present AI moment looks like the early pages of Moravec’s prophecy: not because someone finally solved the intelligence problem, but because the hardware reached the threshold he said it would reach, on roughly the schedule he said it would reach it.

The concept also frames the present moment’s deepest uncertainty. The scaling laws that chart capability against compute on logarithmic axes are Moravec’s wager rendered as empirical science. Whether those laws will continue to hold—or whether the curve bends before reaching general intelligence, as energy and capital costs grow faster than capability—is the question on which Moravec’s entire timeline depends.

Origin

Moravec developed the argument in Mind Children (1988) and refined it in Robot: Mere Machine to Transcendent Mind (1999), where he published a graph charting the cost of computation across a century and plotted the line forward. He estimated the human brain’s effective computational capacity in engineering units—a number necessarily rough but whose direction was robust to wide uncertainty—and calculated when the cost of matching that capacity in silicon would fall within the range of a personal computer. His estimates placed human-equivalent personal machines around 2040, with machines surpassing the combined power of all human brains sometime in the second half of the century.

The argument had a specific rhetorical force that pure philosophical argument could not achieve: it made the timeline legible. Previous arguments about machine intelligence were qualitative—it is possible in principle, or it is not possible in principle—and the debate was interminable precisely because it could not be grounded. Moravec’s argument was quantitative, and it grounded the debate in a graph that anyone could inspect. Whether you agreed or disagreed with his conclusions, you had to engage with the curve.

Key Ideas

Hardware as the binding constraint. The dominant AI assumption through much of the field’s history was that intelligence required the right conceptual insight, and that progress would come from clever theoretical work. Moravec inverted this: the binding constraint was raw computational power, and the right algorithms would follow the hardware rather than precede it. The history of AI from 2012 onward has validated this claim to a remarkable degree. The deep-learning architectures that conquered speech, vision, and language had existed for decades; they became competitive when the hardware was large enough to train them.

Cross-technology continuity. The most distinctive feature of Moravec’s argument is the observation that the exponential curve runs across hardware generations rather than within them. Individual technologies exhaust themselves, but the curve inherits from the next technology at the same slope. This cross-technology continuity is what gives the arithmetic its relentlessness: it has survived five generations of underlying physics, and there is no structural reason to expect it to stop at the sixth. Gordon Moore’s law, formulated in 1965, is a single-technology expression of the same regularity Moravec had identified across a longer span.

Gordon Moore

The limits of extrapolation. Moravec acknowledged the vulnerability at the center of his method: extrapolating a hardware curve says nothing about whether the right algorithms exist to turn that hardware into general intelligence. Hardware is necessary, he rightly saw; whether it is sufficient, he assumed more than he showed. The present AI frontier has confirmed the first and left the second genuinely open. Energy costs are rising faster than capability at the frontier, and the smooth scaling curves that flatter the thesis may be approaching regions where each increment of performance costs disproportionately more. The arithmetic has been relentless so far. Whether it continues is the empirical question Moravec’s prophecy now depends on.

The lineage of the scaling hypothesis. The scaling laws that define current AI development—the empirical power-law relationships between training compute, parameters, and model capability—are Moravec’s wager in quantitative form. The field that treats scaling as the path to general intelligence is the field he predicted from a hardware graph in 1988. The researchers who discovered the scaling laws were, whether they knew it or not, confirming a specific claim he had made before most of them entered the field.