DEV Community

mycute
mycute

Posted on Edited on

We put a 1.6 KB AI in a QR code, a favicon, and a WAV

#showdev #opensource #webdev #ai

We put a 1.6 KB AI in a QR code, a favicon, and a WAV

The model is a 160-word Complement Naive Bayes SMS spam classifier. It scores 93.5%
on 1,115 held-out SMS messages. The JSON is 1,579 bytes. The interesting part is not
the accuracy; it is where the model can live.

These are not eight copies of different models. They are the same weights in eight
containers, all decoded in the browser with no server.

1. A QR code — 2.2 KB URL

The QR encodes qr.html#m=<base64 JSON>. Scanning it opens the page with the model in the
fragment; the page decodes the fragment and runs the classifier. There is no weights file
and no network call after the page loads.

2. A favicon — 1.5 KB PNG

A 32×32 RGB PNG holds the model in its pixels: a 4-byte length, the 1,579-byte JSON, and
zero padding. The page reads the pixels back with a canvas. The tab icon on the page is
the model.

3. A WAV — 1.6 KB, 0.198 seconds

The same bytes are 8-bit unsigned PCM samples. The page parses the RIFF data chunk and
runs the classifier; the audio is the model, not a rendered voice. It sounds like noise
because it is noise.

4. A spreadsheet — 10.5 KB .xlsx

The weights are in a model sheet; a classify sheet scores a text with:

=SUMPRODUCT(--ISNUMBER(SEARCH(model!A2:A161,B1)), model!B2:B161)
Enter fullscreen mode Exit fullscreen mode

No macros, no Python at inference time. Formula accuracy is 90.6% on the same test set
(the Python model is 90.9%; SEARCH is a substring approximation, not the tokenizer).

5. A PDF — 3.7 KB model card

The visible page is a model card: task, data, accuracy, weight size and caveats. The
1,579-byte model is in the PDF's /Keywords metadata as base64url JSON, not drawn on the
page. The page fetches the raw PDF, parses that string and classifies. The document and
the classifier are one file.

The fixed SMS containers deploy the raw int8 dot product. The per-class s field is
quantization metadata; applying it changes the argmax and lowers the held-out accuracy, so
the stored 93.5% is the raw-int8 rule. tests/test_sms_rule.py pins it.

The same container at 130 KB

The 1,579-byte SMS model is the small end. The 500-example AG News coreset is the large
end: 132,798 bytes of training data in a 178,761-byte PDF, in /TinyAIData instead
of /Keywords. pdf-130k.html parses the data and trains the same Complement NB in the
browser; 78.3% embedded / 74.3% fresh on the 400-row test. The PDF pays about 46 KB of
overhead over the data; no weights, just examples.

6. An SVG — 4.9 KB vector card

The visible card is a vector model card. The same 1,579-byte model sits in
<metadata id="tinyai"> as base64url JSON, with a full SHA-256 in data-sha256. The page
parses the XML and classifies; a rasterizer that flattens the SVG to PNG loses the metadata.

7. A MIDI — 4.4 KB of note events

The base64url model is packed into 1,053 note-on events: each event carries two ASCII
characters in its note and velocity bytes. Text meta events carry the model card and a
SHA-256. The page parses the MIDI and classifies; rendering it to WAV throws the note bytes
away. The first 128 notes are playable.

8. A single HTML file — train your own

Tiny AI Maker takes a CSV of text,label, trains a Complement NB in the browser, shows
accuracy and artifact size, and exports a self-contained HTML file, a QR URL, a 32×32
favicon PNG, an 8-bit WAV, and a PDF model card with the weights in /Keywords. The SMS
sample is 100 rows; 160 words; 90% on its held-out split; 2.1 KB QR. The favicon and WAV viewers load the exported files back through a file
input, so each export is round-tripped in the browser. It also writes a model card with the
split, a 95% Wilson interval, an optional 5-fold CV with out-of-fold Brier/ECE, per-label
precision/recall/F1, a confusion matrix, both artifact sizes and the caveats. The 5-fold CV
is the honest part: the 200-row AG News sample scores 70.0% on one 80/20 split but 57.0%
mean over 5 folds, a 13.0-point gap. Its out-of-fold Brier is 0.523 and ECE 0.139; the
softmax is not calibrated. A nested temperature fit (median T 1.5) lowers ECE to 0.074; on
the SMS sample it lowers ECE from 0.066 to 0.032. It is measured, not applied to exports.

Why this is useful

  • Offline and private. The model is the file; inference never leaves the page.
  • Shareable. A QR, a favicon, a WAV, a PDF, an SVG, a MIDI or a spreadsheet can travel through places a model hub cannot.
  • Cheap to version. The artifact is small enough to diff, email or embed.
  • Teaching. It makes the representation ladder tangible: the same behaviour is a prompt, a coreset, a linear model, a 5 MB adapter or a 4-bit model, depending on what you are allowed to share.

What this is not

  • Not an LLM. It is a bag-of-words classifier on English SMS.
  • Not a security proof. A favicon/WAV/QR/PDF/SVG/MIDI is a lossless container, not obfuscation; anyone with the file has the model. The PDF/SVG/MIDI checksum detects a changed model; it is not a signature. Rasterizing an SVG or rendering a MIDI to audio destroys the payload.
  • Not a production spam filter. One 80/20 split, 1,115 rows, one machine.
  • The 4-bit adapter work is separate: a 5 MB LoRA on a shared Qwen base, not this 1.6 KB model.

Which container survives everyday handling?

Size is not the only cost. We ran the shipped artifacts through the transformations a
normal workflow applies and decoded the model again:

container survives fails
XLSX openpyxl re-save Google Sheets (not measured)
single HTML byte copy a CMS that strips <script>
PDF pypdf rewrite rebuild without Info / print-to-PDF proxy
QR 10% downscale, JPEG q5, blur 4 px, 20° rotation dropped #fragment
PNG PNG re-save, metadata strip JPEG q80, resize to 16×16
WAV PCM rewrite, FLAC MP3 128k, Opus 32k

The full matrix, with exact sizes and caveats, is at
https://byjameson-mini-ai.static.hf.space/survival.html

The repo

Everything is in stream-finetune: the 130 KB browser classifier, the 1.7 MB accuracy
mode, Tiny AI Maker, the QR/favicon/WAV/PDF/SVG/MIDI/spreadsheet containers, the survival
matrix, the Minimum-Bit Table and the adapter autopsy.

If you build a container we did not think of, the table has a submission form.

Top comments (0)