AI Wanderings

Posted on by Tom Woodward
An AI-generated image of 4 robot faces drawn in 1950s comic book style. One clearly has "fear" in the thought bubble over its head.

As I worked with Amy on this week’s AI article (It’s not live, it’s Memorex AI) that kind of addresses more sophisticated media generation and the weird world of whether people are people online . . . I got way too much stuff and examples that were probably too odd for the current audience. And that’s why I have my own blog.

Some of this made it into the post. Some didn’t. I’m just cutting/pasting the disaster that is how I take notes as I write.

Here’s a quick demo of some of the AI voice cloners running on Hugging Face. I still find it amazing they can run this kind of stuff for free.

 

I used hotshot.co (really weird stuff is popping up there regularly) to make an animation of Mr. T selling his own bodybuilding pills because that’s what he was saying in the weird audio I was sampling to try to create his voice.

https://hotshot.co/shot/MJhI

 

And Freddy Mercury talking to a Gecko.

https://hotshot.co/shot/MJge

 

And, to deal with bad actors, captchas have become complex enough that people struggle to complete them accurately. Ironically, as the captcha-driven training improves AI performance the AIs will be better able to complete captchas. At the same time we’ve relied on some relatively simple tests to prove you’re who you claim to be in different spaces.  have been a major way we identify the fraudulent (re)use of photos on the Internet. This defense pattern will be completely defeated by pictures that are uniquely generated via AI. Another common verification pattern for photos was to photograph yourself performing a specific task. Holding up a piece of paper with a specific message was popular on Reddit. As Photoshop improved, those tasks became more complex and you can see how easily that kind of verification is defeated these days.

The theme of AI in movies

There’s a whole lot of options to pursue by going down this road. Once I started listing them, I realized it’d get out of control.

  • The Matrix
  • M3gan
  • Her (I haven’t seen this one)
  • Blade Runner
  • Terminator
  • 2001: A Space Odyssey (nothing happens in this movie, there’s like 6 minutes of plot, lots of eating though)

If you can do video, text, audio . . .

  • Virtual influencers
    • Realest looking one (BMW ad)
    • Are people liking this because it’s virtual or do they not care either way? Is it bad not to care?
    • How different are AI personas from PR-controlled personas?
  • https://girlfriend.myanima.ai/
    • got some people arguing chatbots allow you to practice and improve social skills and others arguing that chatbots will do the opposite
  • Unrealistic beauty standards – How much does this differ from air brushing? Is it worse? Would it be better if people realized the person wasn’t real and so wouldn’t be a pursuable goal?
  • worse to have real people under KPop contracts or better to just give them fully controllable personas?
  • Given the record of ad spokespeople (Dell Guy, Subway guy, various celebrities, milkshake duck, etc.), it’s clear that a talking gecko is a much safer bet for corporations.
  • With most of these things the tendency will be to compare it to a human rather than to nothing. Is AI-whatever better than the nothing that would otherwise happen? That’s a much lower bar. A depressing bar, but still.

References

Moral aspects

  • What happens when the original source (gross term) doesn’t/can’t consent
  • When the creator knows but doesn’t tell us . . .
  • Impact on people/society

Points of contention (Where’s the uncanny valley? What moves it to AI?)

  • What moves it to AI?
    • With enough audio, you wouldn’t need to AI . . . just effort and time – is it the ease?
    • JFK in Forest Gump
  • Where’s the uncanny valley?
    • When you don’t know
    • When people do things they never did, nor could have done or consented to do

Leave a Reply