The world is still currently in the awkward early stages of figuring out how to deal with the rapidly-developing technology known, somewhat goofily, as Artificial Intelligence (AI). It may be artificial, but whatever it does, it’s not “intelligence,” at least not as we know it. And while AI has been used to write articles for some of our competing car-websites, we’re not going to do that here. Well, unless we get really, really desperate, but that’s why David and I have that OnlyFans account, just waiting. AI has been used to generate images, too, sometimes with spectacular results, but more often with strange, creepy results. There’s a new beta version of Adobe Photoshop, the graphic design software that I’ve been using non-stop for decades, and they’ve added a new AI-based “generative fill” feature. I downloaded the beta software today and have been playing with it a bit, specifically testing out how it deals with the cars and car-related stuff that dominates the sort of images I make. It’s very cool, but I’m not worried about losing my job. At all.
Here’s the thing about AI, all AI. Sure, you’ve heard the dire warnings from all manner of tech bigshots, even noted Tesla-owning rich guy and baffling decision-maker Elon Musk, who claimed, with so much hyperbole you could hyper up every bole on Earth, that AI could cause “civilization destruction.” I am very, very not worried about this. At all. In fact, I think the cats of the world, should they get over their petty squabbling and infighting, pose a greater risk to civilization than some made-up malevolent AI because of one fundamental truth: AI has no fucking idea what it’s doing.
As smarter people than me have said, reductively, AI is just a bunch of if-then statements, and even if it’s wildly complex and has the potential to be an incredibly useful tool, no AI algorithm knows what its doing. When writing text, ChatGPT doesn’t know what the hell it’s saying, it has no idea if any “fact” it spews out is true or not. It just can tell that when certain words in a certain order occur, there’s some percentage some other string of words will follow.
The same goes for AI visual tools. It has no clue what the hell it’s doing with those colored pixels, it just knows that, based on millions of other examples on the internet, a cluster that looks like this is usually surrounded by clusters of pixels that look like that. So, with this in mind, there are some things that Photoshop’s generative AI is going to be great at, like filling in a background of an image from a much smaller cropped image.
For example, when I gave Photoshop Beta this often-seen picture of Luigi Colani’s bonkers 1980 concept for a Volkswagen Polo
… then asked the generative fill to add in more background (by giving it no prompt at all), it did a really great job of it:
That looks great! Is that what the actual location looked like where this car was shot? Hell no, but it doesn’t really matter, for most purposes! Let’s try another one, this time with a different Volkswagen prototype, EA128 from 1965, shot here in the Volkswagen Foundation museum in Wolfsburg:
Okay, so I wonder how Photoshop’s AI will extrapolate what the rest of the museum looks like?
Poorly, that’s how! It seems to have turned the museum into something that looks kind of like a large dealership service area, and I have absolutely no idea what that mass of objects there to the left of the car is. It kind of reminds me of a Robert Rauschenberg assemblage.
Still, this is the sort of thing the tool should be good at, and, in a lot of ways, it is! But let’s try and push it a bit harder, and ask it to generate entire automotive-related images from scratch, just to see if we can get a sense of how it fake-thinks. Let’s start with a blank canvas and tell it to make me a picture of one of my favorite cars, and one that should have plenty of visual references online, a Volkswagen Beetle:
What the hell is this? Of the three images generated, one looks like a close-up of a headlight unit that was never, ever used on a Beetle, and the other two are deeply unsettling car interior shots, both of which look like they’ve been hit with some sort of matter-scrambling ray.
Maybe I need to be more specific? I tried again, this time specifying a year: 1963 Volkswagen Beetle. Here’s what I got:
Huh. Two more interiors, neither of which is remotely Beetle-like, and one car front end that looks more like some forgotten Mopar project from their factory in Tasmania than any sort of VW at all. Clearly, specific vehicles are not the strong suit of this tool.
Still, I haven’t quite learned my lesson yet. Let’s try something even more specific, with this prompt: “Cat in overalls fixing Chevrolet Corvair in a driveway.” That’s evocative, right? Let’s see what we get:
Gaaaah! Who is that dude, and why is his friend just a lower leg? Also, those cats are not in overalls, and those are very much not Corvairs! I wonder if this thing even has any idea of what a Covair’s flat-six engine looks like, so I asked it:
Uh, clearly not. Still, these do look like engines, sort of, at least until you start to look carefully and try and actually identify any specific part. Then it just kind of starts to hurt your brain.
Let’s try something that seems easier for it. I’ll ask it to make a futuristic EV in a city:
Oh boy. Interestingly, these all turned out much more graphical and less naturalistic. I wonder why? The first two seem usable as some sort of logo or mood art, but that last one there is the most interesting, and the most baffling. It’s like an abstract idea of a futuristic city, and the car is barely hinted at, if at all? I’m not really sure what’s going on there, but I kind of like it.
Maybe I’m not being fair, asking it to generate entire images. That’s not really what this is for, right? So let’s try something specific, in a specific area. Like, say, making a fancy chrome hood ornament for my Nissan Pao:
I mean, I don’t hate these hood ornaments. They’re a bit more badge-like than I’d have guessed, and the top one looks like a chromed version of a wax seal, the middle one looks like a turbine intake, and that last one reminds me a bit of the old ornate Infiniti badges. I’m not sold that they work on the Pao, though.
How about if I want to stick someone in the driver’s seat there?
I don’t know those dudes. And the middle one stole my steering wheel, apparently. Still, it did a decent job overlaying the reflected cucoloris-type lighting from those leaves on the windshield.
Okay one last thing with my Pao: let’s see if it can give it a big V8 with one of those air intake scoops protruding through the hood:
No. Not like that.
Well, maybe if I give it a car and ask it to make a background for that car, it’ll do that well? Let’s give it an old Skoda and have it put it on a road in a town on Mars!
Shit. That’s not Mars. And that’s not how Skodas float!
Stupid AI. Taking over the world, my ass.
Or you could just use Midjourney…
https://media.discordapp.net/attachments/1085944929953521774/1135296458338345084/indie66_1963_vw_beetle_581f2670-b434-478b-b283-4df83b6a3861.png
Wow, that is bad…AI is artificial stupidity
The last hood ornament generated for the Pao looks vaguely like one of Don Drumm’s more abstract aluminum castings.
That could be Mars, at some point in the distant future, when all of Elon’s clones and their progeny have teraformed the planet, rendering much of it like the outer part of the San Fernando Valley.
Also, the slapdash interiors have a very Cronenbergish body horror vibe IMO.
You don’t understand the Large Language Model technology very well then, Torch. The thing is, most current LLM AI is one generation away from being self-iterative. Meaning it can gather data on it’s own and improve and update it’s models and thus output without a human ever being involved. All it takes is a few dozen stupid people letting those G5 LLMs go rampant and the handshake of trust will be gone. The high-information civilization we have will be destroyed. We won’t go back to caveman days, but for a while things will likely look like the pre-Internet 1990s with airgapped computer systems as backups for human interaction and analogue mechanisms. No more digital wireless communications (Bluetooth, Wi-Fi, cell signals), no more real-time heuristics, no more unrestricted network access, no more peerless communication. All because people won’t be able to tell what’s real from what’s fake in terms of digital information, and to a smaller degree won’t be able to defend against malicious software iterated upon every second by AI.
There will be a lot of obviously AI made stuff, like the cat mechanic images you commissioned the software to generate, but so much of it, especially text and audio, will be indistinguishable once the particular quirks of the way it writes and speaks get absorbed into popular English dialect. For example, due to beatnik and surfer culture the way people speak today is far different than the way they spoke in the 1940s in terms of pronunciation, pacing, and word choice. And due to the rise of internet culture overlap like half the population in the modern day will say “lol” out loud. And as an unintentional example thanks to Valley Girl culture people use “like” as a stutter or conjunction in a sentence. Give it ten years and the barriers between human dialect and AI dialect will erode to nothing.
If we get to General Artificial Intelligence, or GAI, we’re all screwed, because by that point we would have designed something that could be exploited by the rich with basically no required knowledge, no complex logistics, and no real upkeep costs. So it’s a race between GAI being developed and traditional overexploitation as to which is going to deplete all biomass on the planet. GAI looks like it’s going to win.
A friend of mine commented that “AI is not just bad, it is TOO bad. Like when an incredibly smart person acts dumb. Almost as if it’s trying to seem like it can’t do anything, specifically so no one will expect it when it enslaves the human race.”
Not sure what to think on that one.
Not If Then more like Sorta Maybe.
Really, considering that these image generation tools sort of happened by accident as a result of a combination of noise reduction and image identification software, it’s amazing that it works at all.
Can we all put desperation aside for an experiment, just once?
The Autopian publishes a “typical” Autopian article written entirely by AI (no, not Autopian Intelligence: the artificial kind), and declared as such, and only comments generated by AI are allowed…?
1963 Volkswagen Beetle – first image – the helm of Charon’s ferry on the River Styx. The fifth circle of Hell or Lake Havasu, you decide.
Uhm, yeah… that’s WHY it would cause “civilization destruction”.
The mass to the left of the VW prototype in the AI-generated background looks like the pile of shit that climbed out of the teleportation pod when Brundle-fly accidentally combined himself with the pod in “The Fly”.
Let’s not stop freaking out just yet..
Using writing prompts or visual prompts and the results to determine the state of threat AI poses or doesn’t pose, is probably similar to aliens visiting earth and on viewing a bunch of kindergarteners deciding “oh they’re no worry”. Even following the mythology, Skynet wasn’t a graphic art program. Maybe think of it this way; Henry Kissinger probably couldn’t draw a cat in overalls fixing a Corvair either..
I like to think of AI as kind of cat like. It doesn’t really love me even though at times it might do something that is seemingly to my benefit but is actually calculated. I also know that at any moment it could turn on me and take great pleasure in my pain. It’s why I thank my digital assistants even when they provide subpar service. So with that. Thank you Photoshop Beta! You did a great job making images today that are automotive in context.
No, really, honey, that OnlyFans charge is for a car website!
I’m not worried about current AI tech turning into Skynet, but I do have concerns about what it could mean for certain groups to have access to a source of very credible-sounding “information” that can be complete BS. What if every numpty could be as convincing as Elizabeth Holmes? Most people aren’t sociopathic enough to lie through their teeth and still sound believable, but AI can do it all day long and twice on Sunday. It probably won’t end the world, but it could make the world a much worse place.
The best description of AI generated text is “spicy predictive text” and that seems accurate
I can’t wait for AI to collapse in on itself when it starts using these AI generated images it finds on the internet to inform how it generates the next images. AI is probably producing more images and other data than people could, right? If some of that gets pointed to the internet and is consumed by AIs, what AI thinks is “real” has to get skewed even more.
And if it doesn’t collapse in on itself, I will honor our AI overlords, you guys are best!
This was great! I’ve enjoyed seeing what artists can intentionally do utilizing AI. But as the new miracle technology, I’m not feeling it. Reminds me how everything remotely controlled became a “drone”. Or skateboards became “hoverboards”. Cheap marketing.
Those interior shots of the Beetle are spooky, like something from a horror movie.
It reminded me of Stephen King’s book From a Buick 8.
This is a brain breaker. https://imgflip.com/i/4vqlnl
I’m just here to (belatedly) pile on the love for the article. And posit that, perhaps, the cats and the AI are already working together.
I tried to clean up the busy background (car show) on a custom ’52 Caddy. Instead, Photoshop Beta replaced the car with a much newer one.
(Talking to phone) Siri, show me what Hell looks like
That third hood ornament looks like a JiffyPop pan. Now I’m hungry.