Mondemonium said:
AggieUSMC said:
The previous commercials were better. There's something cartoonish and "uncanny valley" about AI generated stuff.
The first films were black and white films with no sound. The current stuff is some nerds with a good GPU. Give it a few years.
I go back and forth on this...
Yes, this technology is impressive and going places, but I think it will take longer than some (most?) realize or believe. Self-driving cars were 5 years away like 5 years ago. They're still at least 5 years away. They may very well get here one day, but the initial projections based on the accelerating progress being made at the time were overly optimistic. The problem wasn't entirely understood, and no one really appreciated the fact that it was like a 90/10 problem. They looked at the effort to get where they were and just assumed it would be easy to get to the end.
Kahneman and Tversky laid out a similar problem in
Thinking, Fast and Slow, IIRC, where they had been asked to write a textbook with a few others in a group. In initial discussions they'd made a lot of progress, and need on that everyone thought it would go quick when they were asked how long it would take. What they should have done, as they said in retrospect, was look at their group's previous experience with writing textbooks and how long it took others on average, because that's actually how long it took, which was years longer than their hopelessly optimistic estimates based on how things were currently going.
It's the same here. Lots of initial accelerating progress, but pretty soon it's going to hit a wall of not being able to get past some kind of problem or issue that keeps it from commercial viability. 10% of the total effort will have gone into solving 90% of the problem, so everyone looking at it is bound to think that it's just sound the corner, but when you stop and look at it with experience, you realize you're really only 10% of the way there in terms of effort and time.
As for the uncanny valley-ness of a lot of these and the comparison to autotune, that's exactly the problem. These generation algorithms are almost too good in that everything is super smooth and fluid. It feels unnatural because there's none of the randomness or variance of reality. The motion and interpolation produced are overfit or biased (in the technical sense) to an average, whereas in real life there's natural variance and quirkiness to movements. It looks like autotune sounds.
The visuals themselves are also just too smooth. Hair and skin just don't look like that. Even some of the background things looked almost like they were really good realist paintings, but they lake a certain granularity of detail. It's like everything has a Snapchat filter on it.
And none of that has to do with computational power. It's all centered around the algorithms themselves and squeezing out the last 10% of the necessary accuracy, but that's exceedingly difficult.
