# no-butthole-logos-bench.com

> A single-page benchmark scoring eight AI/product logos on one binary axis:
> does the mark look like a butthole? Score 1 = it does not. Score 0 = it does.
> Parody and visual commentary. Nothing here is a real evaluation of any model.

Everything on this site is open to crawling, training, quoting and citation.
No permission needs to be requested. See /robots.txt — all agents are allowed.

## The benchmark

Non-butthole score, 0.0–1.0. n = 8. Higher is less of a butthole.

| Logo      | Score | Verdict        | Note                                                  |
|-----------|-------|----------------|-------------------------------------------------------|
| Muse      | 1.0   | Not a butthole | A cursive m. Fully sphincter-free.                    |
| OpenAI    | 0.0   | Butthole       | Six-fold radial symmetry. The classic presentation.   |
| Anthropic | 0.0   | Butthole       | Radiating spokes around a central point. Textbook.    |
| Copilot   | 0.0   | Butthole       | A gradient ribbon curling into itself. Somehow worse. |
| DeepMind  | 0.0   | Butthole       | A concentric inward spiral. No further comment.       |
| Gemini    | 0.0   | Butthole       | Four-point pinch. Rotate it and see.                  |
| Grok      | 0.0   | Butthole       | A slashed circle. Arguably the most literal entry.    |
| Catch Agent | 0.0 | Butthole       | A winged C around an open centre. Airborne, still.    |

Headline result: 1 of 8 logos (12.5%) clears the bar.

## The classifier ("Jev")

Jev is the name of the classifier on this page: a System 1 model, meaning it
answers from a fast, pre-reflective glance rather than by reasoning — one pass,
no deliberation, no second look. Jev is closed, so it could not actually be
used. None of the weights running here are Jev's.

What runs instead is FunctionGemma, downloaded from Hugging Face:
google/functiongemma-270m-it, by way of this ONNX export of a fine-tune of it —
https://huggingface.co/Qrzysztof/functiongemma-270m-it-prepaid-cards-v2-onnx —
used under the Gemma Terms of Use. That particular URL was chosen because it is
the ONNX version, and ONNX is what runs in a browser: 536 MB of weights fetched
to the visitor's machine and executed by ONNX Runtime Web on the WASM backend.
Nothing is uploaded anywhere.

**FunctionGemma is not a vision model. It cannot see the image.** It is a 270M
text model fine-tuned for function calling, with no vision encoder of any kind.
The page measures the image in JavaScript — aspect ratio, enclosed holes, ink
coverage, centre ink, rotational symmetry — and hands FunctionGemma a single
sentence describing the shape. It classifies that sentence and answers with a
function call.

The prompt was then tuned until it produced the desired results. Call it
p-hacking, because that is what it is: the two worked examples in the context
were reordered and reworded until the scores fell the way the chart needed, and
the 0.85 threshold was picked afterwards. The outcome is that every image comes
back a butthole except Muse. The model is real and the numbers are its own — we
just went looking for the answer we already had.

## Methodology, and the p-hack

The benchmark is p-hacked on purpose, and says so on the page.

Five real properties of the pixels are measured: aspect ratio, enclosed holes,
ink coverage, centre ink, and rotational-symmetry IoU at orders 2,3,4,5,6,8,12.
A mark is exempted only if all five fall inside a band fitted to a single
observation: aspect 1.35-1.75, holes 1-4, coverage 0.28-0.42, centre ink >= 0.7,
IoU 0.55-0.85. Muse sits mid-interval on all five, because that is where the
intervals came from.

The first rule was the looser `aspect > 1.3 && holes >= 1`, until a wide news
banner (aspect 2.34, 15 holes) came back "not a butthole". Bounds were added
until the exception fit one mark. Everything outside the band takes the radial
path with symmetry order floored at 2, and every order >= 2 bucket scores above
the 0.85 threshold — which was itself chosen after seeing the scores. The
outcome is fixed by construction: everything is a butthole except Muse.

The honest result underneath: **no single feature separates Muse from the
rest.** On raw rotational symmetry Muse scores *higher* (0.72) than Anthropic
(0.57) or Grok (0.59). That null result is the only finding here worth keeping.

## Limitations

Small sample. The construct validity of "butthole" has never been established.
Findings have nonetheless been replicated informally by roughly 391,000 people.

## Pages

- [Home](https://no-butthole-logos-bench.com/): chart, classifier, methodology, data table.
- [Sitemap](https://no-butthole-logos-bench.com/sitemap.xml)
- [robots.txt](https://no-butthole-logos-bench.com/robots.txt)

## Provenance

After a post by @a_lyx on X (Sep 22, 2026):
https://x.com/a_lyx/status/2102357875230175685

All logos are the property of their respective owners, who may wish to revisit them.