Since 2019 I’ve used various language models (from the large to the very tiny) to generate drawing prompts for what has become an annual October tradition. (2019, 2020, 2021, 2022, 2023, 2024, 2025, and my personal favorite is 2022, when the GPTs were small enough to be absolutely bonkers.)
After a failed attempt to get Gemma 3 to be as interesting as the early GPT3’s (I gave up after it suggested “The Autumnal Autumning” twice in one list), I decided to go very old-school.
Specifically I’m using a tiny neural net called char-rnn, originally introduced by Andrej Karpathy in 2015, and updated by my friend Dylan Hudson to work with modern libraries. I installed it on my laptop, and then for training data I turned to a dataset that I had fun with in the past, a list of 42,000 registered racehorse names from the Jockey Club. Char-rnn runs only on my cat hair-covered laptop, and it has no general internet training, only what I gave it.
Presenting this year’s lineup!
These may not all be winning bets, but they’ll hopefully be fun to draw. Mostly I just want to see people draw horsies! Or very much not horsies, it’s up to you! Or not draw – every year people interpret these prompts in all sorts of interesting media, from flash fiction to fiber art to poetry. Use the hashtag #botober2026 if you post your art so I can check it out!
There’s bonus content for AI Weirdness supporters, which this week is a list of racehorses that didn’t make the cut, for one reason or another.
