I recently started reading Co-Intelligence: Living and Working with AI by Ethan Mollick. I’m generally leery of AI books, for a number of reasons, but I picked this one up for my Kindle recently, when it was on sale for $2.
There’s a mention, in the chapter I’m currently reading, to an experiment where somebody asked an AI chatbot to “draw a unicorn in TikZ.” I guess the source is this paper. I decided to experiment with this idea a bit myself. The results are in this GitHub repo.
I didn’t specify TikZ, but I asked a few AI tools to write a program to generate a vector illustration of a unicorn, just to see what I’d get.
I started with Claude Code, run via CLI on my Mac, using the default Sonnet model. I was surprised by one thing there: initially, it created and ran its Python code in an ‘artifact’ rather than on my local machine. I hadn’t seen it do that before, at least when I was running from CLI. So that was interesting. I had to, at one point, explicitly tell it to save the Python script to my local folder. I probably need to learn more about that.
The actual result from Claude Code Sonnet was pretty bad. You can see it here. (This is after I asked it to make some tweaks.) And, weirdly, it took quite a lot of “work” from Claude to get this done.
Next, I asked Copilot to give it a try. This wasn’t GitHub Copilot, this was just the web browser interface at copilot.microsoft.com. That gave me an XML file that I could save as an SVG, but no meaningful program code. That SVG is here. It’s not great. I didn’t push Copilot to make it better or anything; I was just curious.
After that, I gave Kagi Assistant a try. As with Copilot, Kagi Assistant isn’t really a coding tool, so I didn’t expect much. But it spat out a Python script pretty quickly, and the resulting SVG looks pretty good. Honestly, better than the previous tests. This was with Kagi’s “quick” mode, which I think was using Kimi 2.6. As with Copilot, Kagi didn’t give me a ‘real’ meaningful program to generate the SVG. It just gave me a Python program with the SVG in there as a big old string literal, and a couple of lines to write it out.
Finally, I went back to Claude Code, but switched the model to Fable. Rob Conery wrote a post a few days ago, encouraging folks to play with Fable while it was still available. (He was specifically talking about using it against your Obsidian vault, which I also tried, and might write about later.) Fable did a really nice job of writing a good-looking Python script, which output the best unicorn SVG I got out of any of these experiments.
Does any of this mean anything? Probably not. Who knows which of these AI bots is really doing the work, vs. just regurgitating stuff that’s in its training data and/or out on the web. But I learned a few things, and got a few nice unicorn pictures out of it.