You may have seen DALL·E mini posts appearing on social media for a little while now – it’s been viral for a couple of weeks, according to Know Your Meme. It’s an artificial intelligence model for producing images, operating as an open-source project mimicking the DALL·E system from company OpenAI but trained on a smaller dataset. Actually, since I had a play with this yesterday it’s renamed itself at the request of OpenAI and is now called craiyon. Since the requests all take between 1-3 mins to generate, I’m not going to re-generate all the images in this post using craiyon so that’s why they have the old ‘DALL·E mini’ branding.
AI image generation is a massively impressive technical achievement, of course. craiyon doesn’t create as stunning images as DALL·E 2, but still it can create some ‘wow’s.
What’s interesting, sometimes, is how it interprets a prompt. The data craiyon is trained on is “unfiltered data from the Internet, limited to pictures with English descriptions” according to the project’s statement on bias, and this can lead to problems including that the images may “reinforce or exacerbate societal biases”.
To see that in action, we can take a look at how the model manifests cultural expression around mathematics. When I gave it the simple prompt ‘mathematics’, it produced this.