He even contradicts his whole premise of "actors will always be necessary" with his example of a custom generated series episode.
Also,
>implying the non-AI content we have nowadays is highly creative and unique
*facepalm*
>implying the non-AI content we have nowadays is highly creative and unique
You just inserted this, he doesn't say or imply this - he asserted that all of the new ideas (so far) have been generated by humans. Sure lots of it is slop - but all of AI content is slop. That's his point: the new, creative stuff is still coming from humans - even if they are using "AI" tools to do it.
This parenthetical is doing a lot of work here.
Also, you make a lot of claims that just don't seem to reflect what I've seen. Many human-made movies are cookie cutter plots (ie the new Twisters), and many AI prompts return seemingly novel results. Whether or not that are actually novel isn't all that relevant, since humans aren't databases that have memorized all the source material.
Can you say more about my claims? Yes, many human creations are re-hashes of old creations. All AI creations are re-hashes. As you say - they seem "novel" to a particular human, but in aggregate they are what they are - re-hashes. There's a meme about how people with limited exposure to something will over-associate the relevance of their small exposure that kind of aligns with this[1].
[1] https://knowyourmeme.com/memes/getting-a-lot-of-boss-baby-vi...
But the meaningful creations, those that drive our culture forward, are those that build on the foundations of something old, and create something new.
Show me a movie in the last decade where the premise of it couldn't be generated by some AI prompt.
(I'll even make it easier by stating that said AI prompt can be generic and to some degree trivial; i.e. not the premise explicitly laid out in the prompt).
Parasite, The Substance, Uncut Gems, mother!, Krisha, Tangerine, and The Lighthouse are some movies from the last decade that are novel in ways beyond their premises.
No generic prompts to an LLM could create those movies, even if it could spit out a vague 2 sentence premise that matches them.
Note that I’m not saying it could never happen, but LLMs as we have them in 2024 produce regurgitated slop, and require significant creative input from the prompter to make anything artistically interesting.
[1] https://en.wikipedia.org/wiki/Portrait_of_a_Lady_on_Fire
"Oh but you can't compare a paragraph to a whole movie script!"
Sure, just apply the same procedure recursively, stop at a depth where you're satisfied, see [2].
Bear in mind I've only spent like 10 minutes and a negligible amount of money. Now imagine you have a year and a budget of 10MUSD to come up with the best movie script you can create with the help of AI.
It really is only a matter of time ...
1: https://chatgpt.com/share/673bc781-0318-8003-a555-5ea4ea1463...
2: https://chatgpt.com/share/673bc976-0a24-8003-ab8e-11c7800527...
I urge you to compare this to the translation[1] of the shooting script of Portrait of a Lady on fire:
1. STUDIO PARIS-INT-DAY
A blank page. A feminine hand draws a black line, the first
stroke of a drawing. Another blank page. And a new female
hand, which starts a new line. This action is repeated
several times, coinciding with the credits.
WOMAN'S VOICE
First my contours. My silhouette.
The silhouette takes shape as we move from frame to frame and
the sketch of the figure comes to life under the strokes of a
pencil.
A large room pierced with skylights. An artist's studio. It’s
Paris in 1780. Eight young girls between 13 and 18 years old
are sitting on stools that are nearly obscured by the
fullness of their dresses. They are bent over their drawing
boards, which serve them as support. The faces of the young
girls are focused. Their eyes oscillate between their drawing
and the horizon.
show me an AI prompt that can generate a movie premise in the last decade where the underlying model hasn't already been trained on that movie's plot.
There's a whole world of cinema that exists outside of hollywood.
He said it would be "janky and weird" and likened it to a remix.
(Also, yeah, there's a very strong interest for him in saying actors will not be replaced, but let's assume he's being frank here).
All the "AI personalities" that are already popping up here and there are already decent enough, again, it will only take a couple years until their quality is on par with meatspace video.
Then it's gg for actors (and many other professions).
Movie making is about to radically shift to an extremely lower cost model to produce. Just not quite yet.
I can already ask chatgpt to 'write me a SCP story about a couch found in an abandoned factory that is keter class' and it will pump out something mildly entertaining in that style. It is not that big of a leap of logic to say 'good enough' will win out. Even now you can kinda pull it off with a small amount of work and stitching several tools together. There is no reason to think it will not be automated together.
How could anyone think they are similar?
You don’t just pump video data into an LLM training pipeline and get video models.
A lot of work goes into both. It’s non trivial and only barely related.
It’s only recently that transformer models are being used for video.
I already see really convincing deep fakes on Facebook, etc., it's not an impossible problem anymore, it's just a matter of increasing their quality.
All the deep fakes I've seen that don't look incredibly weird at first glance essentially re-use clips (like Back To The Future[0]) and they replacement an element, like a face, but the rest of the footage already exists.
If you look closely though, particularly at the mouth of the video I linked, something isn't quite natural about it still.
Same deal with this Taylor Swift fake[1]
The corridor crew noticed something about AI generated pictures. They all seem to have similar lighting and chroma. So they have this 'look' to them (once you see it). That is because the pictures start off as a random noise picture and coalesce into a picture. So the lighting stays the same on just about every pixel.
There are always people that claim to be...
Reminds me of photographs asking people if photos were photoshopped or not and people generally guessed wrong about half the time.