AI improvement is in 'dog years'

Last week at the Agency Hackers event "The Robots are Coming " I learned a great deal. One takeaway is the sheer speed of progress (and proliferation) of the tools available. Another takeaway is our human capacity to keep up - everybody is talking about it but resources and budget to stay on top are thin on the ground.

Back in October 2023 I played with DALL-E to see what it could do. It was fun, especially concocting crazy prompts with my 13 year old son, to see what images could be produced. "Good, but no cigar" was my conclusion and I haven't used it since. Witnessing the rapid progress of the past four months I decided to return to DALL-E and see how it has changed. I've documented my findings below and feel like 4 months has seen the kind of progress you might once have expected would take a couple of years - hence the 'dog years' reference.

One of my prompts from last year was "What does AI look like?" Very meta I thought (meta as in "showing or suggesting an explicit awareness of itself or oneself as a member of its category : cleverly self-referential"). The output was rubbish to be frank, take a look:

Just four months later I used the same prompt with the same tool. An astonishing upgrade. I would never have used one of the above images but you could imagine the one produced below to accompany a blog post, for example (just as I did with the header image of this article).

Back to October 2023... I took that original output and asked DALL-E to reimagine it in a few different artistic styles. I thought it did a good job of understanding and reflecting the art direction I gave it. Three examples below:

1. using felt tip pens
2. in the style of Damien Hirst
3. in the style of Aubrey Beardsley

For each I've given the October 2023 attempt and the version I created yesterday. See what you think...

WOW. In four short months we've gone from something fun, clever but unusable to something that is near production-ready. Certainly good enough to grace a blog post.

One final observation. Back in October 2023 I asked for an image in the style of Banksy. DALL-E told me "I'm unable to generate images in the style of Banksy due to content policy restrictions". I got the exact same response yesterday. I wonder why it felt able to copy the style of the living artist Damien Hirst, but not Banksy?