One ironic twist of generative AI is that many hucksters will claim that an AI generated something that is bizarre for the internet points, when in fact they either fabricated it entirely or had their thumbs heavily on the scale. Many "AI-generated" stories that went viral with earlier GPTs were very clearly written whole-cloth by a human. In this case, each frame is clearly AI generated, but there's a big difference between someone saying "make me a beer commercial" and getting that (seemingly what the claim is), and spending hours with an AI asking it to add weird apocalyptic fire into images of people drinking beer, picking the weirdest ones to stitch together manually, and slapping Smash Mouth on top of it. Rather than being fooled into thinking a human generated something that was actually the output of an AI, we're faced with wondering whether this was really the output of an AI by a good-faith prompter (as seems to be claimed) or mostly the fabrication of a human looking for internet points.
In the pizza commercial, for instance, there is absolutely no way the AI came up with "Pepperoni Hug Spot" and wrote text that is not only legible, but in the same font every time. Yet the poster says "everything is AI". No, it's clearly not.
Here’s the original creator of that pizza commercial explaining the workflow. GPT-4 was used for the script, which sounds plenty plausible to me considering that it’s gibberish. After Effects was used for the editing, transitions and graphical overlays.
I don’t think the creators of these things are misrepresenting them, but as they go viral, the context gets stripped away and the ex-crypto bros start spinning their bullshit on Twitter.
Here’s another recent example I love. Of course it’s a combination of AI and human creativity, but consider the power of making something like this without a single camera or actor.
I didn’t mean to call all GPT-4 output gibberish, just this commercial script.
But now that you mention it, if a person started speaking unconsciously with no meaning connected to their words and no intent behind their communications, we would call it gibberish. I think it’s possible that these LLMs only produce gibberish, extremely well formed, useful and coherent gibberish masquerading as communication. After all, language is a vehicle for communication and I don’t know what is doing the communicating in this case.
While likely you're right, AI weirdness seems to have become a specific aesthetic enjoyed by some people. Some creators and prompt masseurs clearly don't know about some limitations being part of the style, or ignore them for the mass audience. If you believe this surrealistic quality will be ironed out by the developing tech, it's interesting to think it will at some point become nostalgic and a period mark for our decade, maybe.
> AI weirdness seems to have become a specific aesthetic enjoyed by some people
Visual AI weirdness is an uneasy visceral feeling that approaches a level of worry that something is about to happen with your brain. Think like a migraine aura.
Reminds me of the guy with the fake "I forced an bot to watch [genre]" things, although at least those ads would be likely to have AI involved at some point in the process.
This reinforces a cheeky theory I have, that AI is basically an alien intelligence that doesn't quite know what humans are and what they do, but it tries to remember and describe roughly what it saw. So you're getting these handwavy images that feel close to something familiar, but lack the grounding that we'd expect.
That video feels like an AI saying: "this is what you look like to me". Of course it's really not quite getting things right, so the uncanny valley effect kicks in.
Yeah, that's right. This class of problems (what are adequate priors, how to learn them automatically) is under the field of representation learning in ML.
> Titled "Synthetic Summer," the 30-second video first appeared on Instagram about a week ago, created by Helen Power and Chris Boyle of a London-based production company called Privateisland.tv. The pair were not available for comment
Instead of these models constructing frames it would be interesting if these AI systems could instead produce skeletal character movement data (similar to the motion capture of real humans acting) and could operate maya/blender.
In the pizza commercial, for instance, there is absolutely no way the AI came up with "Pepperoni Hug Spot" and wrote text that is not only legible, but in the same font every time. Yet the poster says "everything is AI". No, it's clearly not.