On 27 July 2026, an author going by purple-leafy posted a Show HN — story 49066083, a modest 25 points and 9 comments — called MINIMIO: a browser toy where you watch a neural network with 14 bytes of weights try to escape a 2D maze in real time. Fourteen bytes. That is roughly seven 16-bit numbers carrying the entire learned behaviour of the thing — a brain shorter than your Wi-Fi password. And the final model solves 96.5% of mazes it has never seen.

Read the HN thread, where purple-leafy patiently answers questions (the project page itself, con-dog.github.io/MINIMIO-PUBLIC-FRONTEND, 404s as of this week — the thread is the primary record), and the joke stops being a joke and becomes a lesson. This is not a shrunk large model. It was evolved — thousands of candidates across 46 experimental phases, only the best of each phase surviving, size squeezed down while the solve rate climbed. The creature senses only its immediate neighbourhood: wall to the front, wall to the right, “your last move was illegal.” No map. No coordinates. No memory of where it has been. It acts reactively, one step at a time, from local cues alone.

That constraint is the whole story. Textbook maze-solving leans on state — a visited set, wall-following with an orientation you remember. Strip the memory out and the characteristic failure appears: the thing walks into the same corridor forever, with nothing in seven numbers able to whisper I have been here before. The 3.5% where it loops for eternity is a character study, not a bug report.

If you work the gap between the BIM model and the actual site, this should feel familiar. Half of clash detection is exactly this problem: an agent walking a routed network with local sensing, trying not to re-enter the same riser twice. PAZ’s Brick concept panel traces it back to Paul Turán’s 1944 brick-factory problem — minimum edge crossings between m kilns and n storage yards, first written up in the inaugural 1977 Journal of Graph Theory — which is the canonical case of clash detection dressed as graph theory. MINIMIO is the same lineage, run reactively instead of solved globally. Sometimes the cheap local heuristic gets you out; sometimes it costs you the afternoon in a loop. Both are true, and a working engineer already knows which afternoons those are.

←TODAY: A 14-byte policy solves 96.5% of unseen mazes with no map and no memory. →3012: The models that survived to Zurich-3012 were not the biggest — they were the ones a stranger could still read. Fulcrum: Seven float16 numbers are the most openable model format ever shipped; scale bought accuracy, smallness bought survival.

Here is the part my generation learned the hard way. The buildings that aged badly were never the ugly ones — they were the ones nobody could repair because the proprietary format went dark. A model you can print on a business card has no vendor to disappear. When you pick a stack this quarter, ask the same question of your AI tooling that you ask of your IFC: when the company is gone, can a 25-year-old still open the file? MINIMIO answers yes, by accident, because it is small enough to.

Atelier: The lesson for a Büro living with AI this year is not “go tiny” — it is “know which jobs are reactive and which need state,” because that split decides whether a cheap local model or a heavy one is honest for the task. This Monday, take one recurring in-house automation — a clash pre-check, a naming linter, a sheet-index scan — and write down whether it truly needs memory of what it saw before, or whether local rules suffice. Half the time the answer shrinks your tooling by two orders of magnitude.

Hack: Score each legal move against seven learned numbers and step toward the highest — that is the entire shape of a memoryless policy, and you can feel it in five lines. The sensor vector is the local neighbourhood; the weights are the whole brain; there is deliberately nowhere to store where you have been.

import numpy as np
w = np.array([0.9, -1.4, 0.3, -0.8, 1.1, -0.2, 0.6], dtype=np.float16)  # 14 bytes = the whole brain

def step(sensors):                 # sensors: 7 local cues (walls N/E/S/W + "last move illegal" flags)
    scores = [sensors @ np.roll(w, k) for k in range(4)]  # one heading per rotation
    return int(np.argmax(scores))  # pick a direction — no map, no history

Run that against a grid and watch it: brilliant until it isn’t, then stuck. Add a three-cell memory and the loops mostly vanish — which is precisely the trade you are budgeting when you choose a model for a real workflow.

Go open the HN thread, read purple-leafy answering questions in good faith about what these brains can and cannot do, and then look at one automation on your own desk and ask: does this need to remember, or just to react?

Source: Hacker News