A cube guessing blind across seven categories would be right about 14 percent of the time [1], so an 85 percent floor is genuine signal and the 100 percent ceiling is a small label space learned to saturation [7]. Nearly 500 shapes spread across those seven classes is roughly 70 examples per class [3], which tells you this is a consensus problem wearing the clothes of a recognition problem.
That framing matters because the interesting quantity here is how the consensus forms, not the accuracy score. Each cube carries one vector holding both its memory and its current guess, and its network does a single job: update that vector from the vectors of the cubes touching it [5]. Early on, a cube knows only something like whether it sits on an edge, and information about what lies two or ten cubes away arrives second-hand, passed along [6]. The lineage is cellular automata, Conway's Game of Life from 1970 being the familiar case [14]. The assembly settles on "chair" without any cube holding a model of the whole chair. Sabine Hauert of the University of Bristol, who was not involved in the work, called it remarkable that the bricks can infer the global shape of the collective using only local information [11].
Sebastian Risi, the senior author, puts the motivation as local self-organization "where you don't have one failure point" [3], and the damage experiments are where that gets tested. Remove 15 percent of the simulated cubes and the survivors often still name the object, work out which cubes are missing, and regenerate the neighbours [8]. Simulated systems past 18,000 cubes still guessed what they were in [9]. The physical run is where the counts drop: up to 197 blocks a few centimetres across, each with circuitry doing the computation locally, correct 100 percent of the time [10]. That is about 91 times fewer units than the largest simulation [2].
The open question is whether a real structure holds its answer when you pull bricks out of it. The damage tolerance and the regeneration are simulation results; the hardware result is a clean classification at 197 units [8][10]. Those are different experiments, and only one of them ran on silicon. The hardware also establishes that local computation on real circuits reproduces the simulated inference at that scale, not what the inference costs in latency or power as the count climbs toward five figures.
Risi's near-term use is edutainment, construction bricks that tell you where the next one goes, or a built dinosaur that roars, with modular robots that assemble themselves as the long horizon [12]. The self-healing circuits that Scientific American sketches as the eventual payoff sit further out again [13]. So the earned claim is narrow and real: a collective of independent local rules can compute a global property with no controller anywhere in it, demonstrated on 197 physical blocks and stress-tested for damage only in software.
