About Haiku-Matic

Haiku-Matic, hAIku Generator Toy. Imagine, generate, explore. A robot writes on a scroll: Why A.I. Haiku? Another Useless Project, by Paul Hodara.

One Billion hAIku

by Paul Hodara

Copyright © 2026 Paul Hodara. All rights reserved.

Haiku-Matic is a free AI haiku generator by Paul Hodara. A compact neural network runs in your browser to create original 5-7-5 haiku, with Space, Spring, Summer, Autumn, and Winter themes and Free, One-thing, Two-part, and Mixed styles.

VERSION

V8.31 · Changes:

Corrections from three rounds of reader review, with the same V8 model.

These rules apply to Free, One-thing and Two-part generation. The earlier spring styles and seasonal corrections remain.

ABOUT THESE HAIKU

Every haiku here is one of over a billion that Haiku-Matic can generate. Its home is haiku-matic.com.

How it was made. Every haiku here is written by a tiny language model, 28.9 million parameters, first trained on children's stories and small enough to run on a microcontroller, and here in your browser. On its own it wrote broken, story-like English. So it went back to school. A much larger model, Qwen, wrote 5,000 example haiku under strict rules, and later about 5,000 more, written by Claude and checked by hand, became its training set. Training works by backpropagation: each time the model guesses the next word wrongly, the error is traced back through the network and every weight is nudged so the guess is a little less wrong. The training was also quantization-aware: it simulated rounding the weights to 4 bits as it went, so the model learned to work at the tiny size it is shipped in. Even so, the model only suggests the next word. Hand-made rules decide which words are allowed, so every haiku is exactly 5-7-5, in whole sentences. Many of those rules came from volunteer reviewers in the Haiku Lab, who rated and annotated poems line by line. Each recurring complaint, such as “sounds too much like astronomy class”, became a new rule, tested on thousands of poems before it was released.

ModelPLE / TinyStories
InferenceINT4 weights · INT8 math
StylesFree · One-thing · Two-part · Mixed (Space, Spring, Summer, Autumn, Winter)

INSTRUCTIONS

Menu bar
The bar at the top hides after a few seconds. Move the mouse, tap or scroll to bring it back.
Theme
Space, Spring, Summer, Autumn or Winter, or All themes to take turns. Changing it starts the crawl again with the new theme.
Style
Free: each line is chosen on its own. One-thing: every line stays with one idea, such as a black hole, falling snow, falling leaves, blossoms or bees in a garden. Two-part: two images set side by side, such as wind in the pines and a still pond. Mixed: one of the three at random for each poem. The styles are written for one theme at a time, so choosing one with All themes sets the theme to Space. In Spring, One-thing and Two-part poems have no written first line: the generator writes it, naming the poem's idea (blossoms, buds, rain, streams or birds).
Link
The address keeps the theme and style, so a copied link opens the page the same way: ?theme=spring&style=two-part at the end of the address, for example. The theme is all, space, spring, summer, autumn or winter, and the style is free, one-thing, two-part or mixed.
Reading view / Play crawl
Reading view stops the crawl and lists the poems so you can scroll through them, with Write another to add one more. Play crawl starts the crawl again. Esc also switches to the reading view.
Pause / Resume
Stops and restarts the crawl and the music.
Speed
How fast the crawl moves, from 0.5× to 4×.
Music
Shakuhachi music. It is off until you turn it on here.
Stars
The star field on or off.

CREDITS AND LICENSES

Invited Haiku Lab reviewers. Much of the rule development came from poems rated and annotated in the Haiku Lab by volunteer reviewers.
Thank you – Paul Hodara

Third-party components are used under their own licenses, listed below. They are not covered by the reservation of rights above.

Base language model and tokenizer. esp32-ai-tinystories, Copyright © 2026 Viacheslav Sierbov. Used and modified under the MIT License.

Training data of the base model. TinyStories, Ronen Eldan and Yuanzhi Li, Microsoft Research (2023), arXiv:2305.07759. Licensed under the Community Data License Agreement – Sharing, Version 1.0 (CDLA-Sharing-1.0). The dataset is not redistributed here.

Music. Araki Kodō III, Shika no Tōne (shakuhachi), Victor 13029, via Wikimedia Commons. Public domain.

Return to Haiku-Matic