- Author
- Posted by jbaudanza, with replies from many commenters
- Published
- 9 October 2026
- Kind
- Hacker News Show HN thread, with the author's linked blog post
- Original
- news.ycombinator.com/item?id=50015833
- Subject
- AI put to use
What it is
The poster jbaudanza shares a neural network that turns typed Korean into pen strokes. It was released for Hangul Day 2026. The linked blog post describes the Graves 2013 design it builds on. It also covers how the data was collected. Commenters test it and suggest changes.
What it says
- The post says the model generates sequences of pen strokes, not pixels.
- The post says the model has about 350,000 parameters and needs no server.
- The poster trained it on about 4,500 lines from one writer.
- The post says adding pen pressure made the model learn faster on less data.
- The post says synthetic sentences from Claude filled holes that Wikipedia text left.
- Commenters report failures such as a missing final syllable, and the poster confirms the bug.
- The post says data is open under CC BY-SA 4.0 and code is MIT.
What evidence it gives
The post offers a working browser demo, an open dataset and open code. The comments supply informal tests, with specific failing inputs and the poster's replies. There is no quantitative evaluation of handwriting quality.
Interests and limits
The model learned from one writer, so it shows one hand. The poster calls it the first public Korean model, as far as known. This was not checked here. Replies promise a new training run, but the blog says little more is likely. A separate post on changes to the Graves design was not read. The blog text came from a script bundle, and the demo was not run.