Artifical Id

22 Sep 2022

In science fiction, computers are typically portrayed as unemotional, detached, soulless. Their output is logically correct but cold. Thought without feeling. However, as we start to see the emergence of AI-created images, the opposite seems to be true.

Take the first image in this recent Stratechery article. It’s an image generated using Midjournery, based on a simple text prompt. It captures the feel of the concept Thompson is trying to get across, and as such works well as an illustration. However, if you look closer:

my boy on a bicycle, meanwhile, is missing several limbs, and his bike doesn’t have a handlebar

I intentionally studied the image when I first came across it (given the subject of the article, the fact that it was AI-generated wasn’t a surprise, and these images still pique curiosity), but I completely missed these logical flaws until I went back and checked. The AI had done a pretty good of evoking the emotional meaning of the prompt, while completely dropping the ball on the structural details.

This seems to be a theme across generative AI. Beyond images, GPT-3 will generate flowing, natural-seeming prose that, when you think about it, comprises entirely dead ends and wild goose chases rather than coherent reasoning. Copilot will generate code that, at a glance, appears to be just what you need, but doesn’t actually compile.

Of course, these models will get better, and the think will catch up with the feel (whether they can start to really construct a bike or an argument is an entirely different question). However, I think it’s interesting that their first successes were with the heart rather than the head.

An android dreaming of electric sheep, by Stable Diffusion

An android dreaming of electric sheep, by Stable Diffusion

This site is maintained by me, Rob Hague. The opinions here are my own, and not those of my employer or anyone else. You can mail me at rob@rho.org.uk, and I'm @robhague@mas.to on Mastodon and robhague on Twitter. The site has a full-text RSS feed if you're so inclined.

Body text is set in Georgia or the nearest equivalent. Headings and other non-body text is set in Cooper Hewitt Light (© 2014 Cooper Hewitt Smithsonian Design Museum, and used under the SIL Open Font License). BBC BASIC code is set in FontStruction “Acorn Mode 1” by “p1.mark” (licensed under a Creative Commons Attribution Share Alike license).

All content © Rob Hague 2002-2024, except where otherwise noted.