State of the Artless

By now, the whole world has heard about Willy’s Chocolate Experience in Glasgow (archived), or as the page title has it, “Willy Choclate Experience”, which sounds like something quite different and not at all for the weans. That mangled English is in keeping with the AI-generated graphics promising “a pasadise of sweet teats”, “enigemic sounds” and “ukxepcted twits”, which is the usual xepctation in the uk these days.

The event was, as widely reported, a spectacular bust rather than a spectacular. Beleaguered actors struggled with an AI-generated script in a sparsely decorated warehouse that brought back fond memories for any Aussie of a certain age—the D-Generation nailed this phenomenon thirty years ago. It’s a joke that never gets old, although not much fun for the young ones, who left with little more than a glass of lemonade and a few jelly beans—not even a single bar of wonky willy choclate. Their parents were furious on the day, but I suspect some now feel that being part of meme history was £35 well spent.


Beyond being just another example of a poorly produced event, Willy’s Chocolate Experience reminded us yet again of the power of AI to mislead. Despite the words on the graphics being garbled AI nonsense, the pictures were worth a thousand more, drowning out alarm bells with their promise of Encherining Entertainment in the Imagnation Lab. AI-generated images have as much potential to mislead as AI-generated texts, even when the text on them is rubbish.

We’ve had other such reminders lately. We’ve known for a while that generative AI creates images full of stereotypes. Google, in an attempt to counteract the problem, sent its new Gemini image generator too far in the opposite direction, serving up racially diverse Nazis (archived) who would have had a hard time down at the Nazi bar.

I missed the chance to play with Gemini in the few days before Google took it down for retooling, but have been trying out SDXL Lightning, an image generator which is indeed lightning fast. Partly I’ve been doing it to remind myself of the joy of playing with these tools, rather than pondering their depressing implications as so many have been. (Not so joyous, though, that I’d be happy to train AI to take my job.)

Besides boasting much higher image quality than the tools we first encountered in 2022, SDXL has the advantage that entering the same prompt and seed number will bring up the same image, meaning that although the possibilities are effectively endless every result can be preserved. (One thing I noticed was that adding or leaving off the full stop at the end of a prompt returned close but slightly different images, as did adding a number, like 3 or 74.) The tool is almost too fast; I sometimes lost an image by refreshing too quickly and missing my chance to write down the random seed. But I managed to capture a few dozen interesting ones from the same prompts that I used in 2022, to produce a new gallery of the state of the artless. Click through to see them, along with my commentary.

SDXL Lightning rendition of a spaceship over Edinburgh


Before I leave the subject: for a few hours a week ago, ChatGPT went berserk and started speaking complete gibberish. Mattsi Jansky analysed the problem, and much more, in one of the best overviews of the pros and cons of large language models that I’ve read lately.

4 March 2024 · Art