It's like self-driving cars. They use almost very effective statistical models, certainly better than our previous models, but they never seem to shake off that "almost" and become truly effective.
That's a bold prediction. Why do you think that?
The first thing I thought was the exact opposite. This isn't very good, but it's only version 1. Motion pictures are less than 150 years old. In another 150 years I bet virtual filmmaking will progress a lot.
For the record, I'm actually rather bullish on self-driving cars. There's nothing physically impossible about solving the problem, but I'm not surprised it's harder than it sounds. But I don't see humans being fundamentally prevented from solving the problem in the same way that humans are fundamentally prevented from ever engaging in everyday space travel.
It's not hard to imagine that this kind of thing could end up doing a lot of the heavy lifting for things like background scenes in the future, opening up the kind of stuff we saw in The Mandalorian, Game of Thrones, and the LOTR film trilogy to increasingly lower and lower budget productions.
I think in a lot of modern stuff they go too far. They now use CGI for things that could easily be practical effects, but they go with CGI because it's simply cheaper or because they want to A/B test different colors of wall paneling behind a character in post production, or who knows what. The end result is apparently good enough to make a billion dollars, but I can't stand it. Movies don't feel authentic anymore. It's hard to describe rationally, but the word 'soulless' sums up how movies made the new way make me feel. Even the scenes which are wholly practical/real get degraded; the excessive CGI and compositing used in the rest of the movie cast a miasma of unrealness across the entire movie.
https://64.media.tumblr.com/248d25e2185a58bf827d329490480fb9...
I remember listening to the DVD commentary track for X2 (from 2003), and loving that scene near the beginning where Nightcrawler is hiding out in a cathedral but losing control of his teleportation powers because he's being remotely controlled/possessed by the villain of the story. Bryan Singer talks on the track about how great it was to fling the character (can't remember if it it was actually Alan Cumming or a stuntman) from the rafters, and how much better and more weighty it looked to have a "real" shot and just have to remove wires/harnessing vs filming it as a blank canvas and having to insert the character digitally.
This video does a good job of explaining why, I think: https://www.youtube.com/watch?v=DY-zg8Oo8p4
For real though, I think the best CGI has always been when it's a light touch, providing only slight enhancements to an otherwise mostly-practical scene. And that's also when it's most invisible— so while superhero movies are obvious CGI-fests and can be clearly said to have been "ruined" by it, I think the most interesting modern CGI use is in lower-budget productions like TV shows, enabling the insertion of fantasy- or period-themed backdrops that would never be possible if you had to actually come up with all those props and extras in real life.
And I think to the extent that this is already happening, it's a lot harder to track because it's so much more subtle than the bombastic in-your-face effects of a Marvel movie showdown.
It would be interesting to see a revival of a show like Drive [0] as that was pretty ambitious and expensive for its time, but might be a lot more possible to do now.
Laypeople tend to think everything is linear, technologists tend to think everything is exponential; more often than not, reality tends to be sigmoidal. Technologies have exponential takeoffs followed by logarithmic plateaus. We're clearly well into the exponential phase of deep-learning ML, but it's only a matter of time before this approach hits its logarithmic phase.
Of course, the hard part of sigmoidal prediction is determining where in the curve we are. Does the current paradigm have an even steeper part ahead of it? Maybe. And yet, we could just as easily be right in the middle of the function, with a leveling off coming as the state of the art gives way to incremental improvements.
The groups supporting such absurd claims largely boil down to:
- Money-driven researchers in the applied AI field. Those are the people that spam popular ML conferences with barely novel contributions other than some minor tweaks on code from their previous papers.
- People unable to critically think and evaluate the significant limitations of SOTA methods occasionally marketed as AGI breakthroughs.
The first category is much more troubling however, since they can significantly influence research directions due to the broken citation system in academia (more citations --> higher quality contributions).
I agree with your deeper criticism, though preferential attachment/ranking is very much How Humans Do Things. You could do a much improved citation system by expanding the time dimension and looking at papers that were unpopular at first and then attracted wide interest later.
Of course, academics also have a tendency to over-cite (because they don't want to be rejected for inadequate literature review), so there are incentives to cite a bunch of research whose premises or conclusions you hope to overturn.
I don't think AI will be able to create a movie anytime soon, but I think it will become "good enough" to serve as inspiration for creatives, or to replace simple stock footage (Much like SD and DALLE-2 is now).
What you’re going to see is a race to the bottom, the same as with claymation films and 3d.
It will suddenly require a lot less (expensive, highly trained) people to make the same films.
You’ll still need expensive highly trained people to do it, but with different skills and a lot less of them can do a lot more a lot more quickly.
…and that means that some studios will make bad, low grade films… and some studios will make amazing films using hybrid techniques (like 3d printed faces for claymation).
…but overall, the people funding movies will expect to get more for less, and that will mean a downsizing of the number of people employed currently in certain roles.
Traditional film making over? Hm… it’s complicated. Is it over if the entire industry changes, but people are still making films? Or is that just the “traditional” part of it which is over?
It’s definitely going to change the industry.
People will still definitely film things.
…but, I wager, less people will be doing highly payed skilled manual work, which will replaced by a few people doing a different type of AI assisted work.
…and we’ll see some really amazing indy films, of small highly technical teams producing content with very little physical filming.
Ok? It's a nascent technology. Look at the original DALL-E blog post from last year [1]. Now compare it to DALL-E 2 and Stable Diffusion.
While these early versions are primitive, as a traditional filmmaker I think in a couple years these technologies will creatively empower visual storytellers in exciting new ways. The key will be developing interfaces which allow us to engage, direct and constrain the AI to help us achieve specific goals.
Exactly. Every step forward in creative technology is additional leverage for the artist with a vision.
The much more likely scenario IMO is that people get used to the artifacts and notice them less.
A good path forward is to fuse these image-element compositing tools with some of the 3d scene inference ones. So you start out with 'giant fish riding in a golf cart, using its tail to steer', then give that as ground truth to a modeling tool that figures out a fish and a wheeled vehicle well enough to reference some canonical examples with detailed shape and structure, the idea of weight etc. Then you build a new model with those and do some physics simulation (while maintaining a localized constraint of unreality that allows the fish to somehow stay in the seat of the golf cart).
Reading the paper, it seems to be the "right" approach (separating temporal / spatial for both convolution and attention). Thus, I am optimistic what remains is to scale it up.
It's not over for traditional moving-making. It would be decades before the software and hardware could surpass. But it will improve tremendously, just like computers do for nearly everything.
It's just the first version and seeing Stable Diffusion come out while openai's tools were coming out were something to remember and think about.
Any more specific reason why you feel this way? Curious
I thought the same way you did about speech-to-text and image search, back in the day. boy was I wrong.
who is saying that?
You'll see an excellent mostly- or all-AI feature within 5-10 years. There will be terrible ones before that, maybe 2-3 years. The first really good one will be enjoyed on its own merits, ad the artificiality of it will come to light after it has gained popularity.
*produced, assisted, or discussed on set one time during lunchbreak”
That Luddites have been successfully maligned as irrationally anti-technology crazy people is a propaganda victory by industrialist factory owners and their friends, the newspaper men.
IMHO the best argument for unfettered innovation is the impossibility of slowing innovation globally; it can be slowed in one country, but that country can't force all the rest to get with that program and will eventually be overtaken by technologically superior foes.