And as long as you have to have the skilled humans you might as well let them choose how they complete their tasks, because in many cases they find that the prompt/review/prompt refinement/review process takes longer than it would to just do the thing.
i'm a creative professional so i have a searing hatred for these tools in the abstract, but even i admit they have utility. it's pretty fuckin' impressive what some of them can do. that said, they have to be used in the right way, not jammed into every angle of production because management says so
Unfortunately, the “let a person direct a photorealistic movie of a media figure or draw something well-understood in a well-understood style” are two of the things it does surprisingly well.
i'd say "reasonably well, within certain parameters" but yeah, point taken it certainly does it well enough to fire a bunch of people, so far as the C-suite's concerned
That’s true—I was mainly thinking that the deepfake stuff works so well it’s scary and it’s being used a lot for social engineering scams.
i don't think we've even seen the full power of that disinfo pipeline unleashed yet you can run a local model these days on consumer-grade hardware that can create video clips good enough to fool trained folks, let alone your average Facebook boomer next few election cycles gonna be NASTY
Yeah, and a lot of the conversations when it's pointed out go like: "It's fake. You see how this person has 8 fingers on his hand on an arm that's sticking out of his hip?" "Well, it sure sounds like something he WOULD have said" ...Instead of being outraged that someone tried to trick them.
Especially afraid when it gets to the point where an AI bot can comb someone's social media history and tailor bespoke deepfake stuff designed to specifically trigger them.
Also thinking that if it can generate something realistic from one camera angle, it can generate it from another and people will see multiple angles/edits of the same thing from different sources. It's going to fool a lot of people.
this shit is progressing incredibly fast; like, there are new consumer-grade image and video models dropping every other week and their consistency and verisimilitude only improves things like "wrong number of fingers" that were commonplace errors a few months ago are now basically solved problems
It's definitely getting better, but a lot of it still has artifacts and people don't care, even if it has a community note telling them to count Macron's legs. bsky.app/profile/rizv...