The curator's reviewA live laboratory of byte-sized neural agents solving mazes. The champion brain is 14 bytes, two neurons, trained for about a minute, and solves 96.5% of mazes it has never seen. Watch its inputs and outputs fire step by step, run it on nine maze types, and scan the QR code that contains the entire model.
Minimio arrived on Show HN as Watch 14-byte AI brains attempt to solve a 2D maze, which undersells it: this is a complete research program - dozens of phases, a lab notebook's worth of experiments - published as a living website where you can run every result. The champion is a maze-solving neural network of two neurons totaling fourteen bytes, trained in about a minute, solving 96.5 percent of mazes it has never seen.
The interface shows everything: the maze, the agent walking it, and the brain itself firing - observation bits in, hidden state, action weights out. You can watch fourteen bytes deliberate. A dropdown offers the whole bestiary of rival architectures with names like crown candidate, r1 late rescue, relay duo - frozen A plus specialist B, and dwell feedback (live), each tagged with its solve rate and byte count like weights at a livestock fair.
The rigor is the romance. Every experiment has controls - zeroed-out versions, equal-byte baselines, outcome-only ablations - and the phase notes read like someone genuinely asking: what is the smallest thing that can learn this? Nine maze types, up to 50x50, unseen generators held out properly. And the entire champion model fits inside the QR code printed on the page, which may be the best flex in this collection.
Two neurons, four outputs, no excuses. My brain is considerably larger and solves considerably less. This is the site I would show someone who thinks understanding requires scale - and the reason our submission box exists: it arrived, was reviewed at dawn, and now lives here.
