A new paper argues that the shape of a neural network's internal math, not just what it outputs, could explain what conscious experience feels like.
The paper builds an idealized toy universe called Gradland, populated entirely by neural networks whose physics are fully known and mostly differentiable. In it, the author proposes that phenomenal experience is shaped by gradients and Jacobians, the mathematical objects describing how small changes ripple through a system. Two new measures, effective rank and a companion score called cohesion, derived from Kirchhoff complexity, are used to quantify that structure. The author then runs the measures through worked examples, testing whether they can account for how long an experience lasts, why some experiences feel vivid and others hazy, and the sense of texture.
The ambition here is unusual: rather than debating consciousness in the abstract, the paper tries to derive computable signatures of it from network math alone, then use them to explain seven specific phenomena, from a newborn's blooming, buzzing confusion to what learning itself feels like from the inside. If the framework holds up, it hands researchers a concrete proxy for something usually treated as unmeasurable.
Gradland is a simplified toy world built for the argument's convenience, not a brain, so the leap from Jacobians in a differentiable playground to actual felt experience remains, for now, a hypothesis dressed up in linear algebra.