EG "make me a 120 minute Star Trek / Star Wars crossover". It should be more or less comparable to a big-budget studio film, although it doesn't have to pass a full Turing Test as long as it's pretty good. The AI doesn't have to be available to the public, as long as it's confirmed to exist.
1,000Don’t mind me I gotta pay rent 😔🙏 https://manifold.markets/NeoMalthusian/sp-500-drops-8-by-april-2025-postli
I think 10 million dollars of inference compute in early 2028 can produce a movie that would make more than 10 million dollars. But.. this is not going to be true if there are many movies that are produced and released like this at the same time.
And for that amount of money you may as well include human authorship at some level of the process. So if someone spends 10 million dollars on a "fully automated" movie around that time, they're probably doing it as a stunt more than as a good business bet
10 million dollars of inference compute in early 2028 with SOTA models is going to be something very capable. I think a lot of people are not taking into account the plausibility of compute scaling in the entertainment industry, as weird as that sounds right now.
Most of the people commenting on this question seem to be thinking in terms of the same amount of inference compute used to generate a 10 second Kling video in 2025. The max dollar amount spent on video generation of single videos is going to go up by a lot in the next three years.
Like what is the maximum amount of inference compute in dollar money someone has spent on generating a single video right now with AI? 5000 dollars? 10k?
What is the doubling time for that number? 5 months? 3 months? 1 month?
Because if it's doubling every 3 months then that graph goes way past 10 million dollars by Jan 2028
@MalachiteEagle ChatGPT says 80-90% of a TV commercial's budget is on purchasing air time, not production of the commercial itself. So a free AI would still need to be 80-90% as good as humans (as measured in conversions/airtime) in order to be economically viable. If the AI commercial costs half as much to produce, it needs to 90-95% as good.
I don't know if we'll get there in 18 months, but we're certainly not there now.
The 3-month doubling model starting from 10k dollars suggests someone has spent 1 million dollars on a single AI-generated video in November 2026. I think this is highly plausible.
Bear in mind this video was from November 2024: https://www.youtube.com/watch?v=4RSTupbfGog
@MalachiteEagle Your model suggests that an AI video cost $3,900 in November 2024. The Real Magic Coca-Cola ad, with a typical production budget, would have $500,000 to $2 million.
So either
A) Coca-cola spent a huge amount of money on an AI ad, 2 OOMs more than your model suggests.
B) The vast majority of the Real Magic production budget was spent on humans and traditional video editing expenses.
We haven't seen another big budget AI commercial. I suspect the Real Magic commercial was done for the novelty.
@GG I suspect they just did multiple calls to whatever their "Real Magic AI" thing was and selected the best versions. I think 4k dollars of inference compute sounds pretty close. The models they're using are too low-quality to motivate spending more than that.
Remember I am talking about inference compute. It's likely that Coca-cola sent much more money than that to Bain & Company but the actual expenditure on inference compute was tiny.
These markets from @SamuelKnoche should get more attention IMO, it seems to me that the question of script generation and the generation of the movie from a script factors this question pretty cleanly.
/SamuelKnoche/in-early-2028-will-an-ai-be-able-to-lezeyhikb7
/SamuelKnoche/in-early-2028-will-an-ai-be-able-to-2x8ud6ld4f
Staying at no through today's news
Seems way more lilely we get to the "someone can make a movie in a matter of days using only an AI and skill at prompt crafting" though.
How high quality?
People already can make 2 minute or more films via prompting.
Shouldn't be that hard to make a decent movie by orchestrating current tools. Probably costs a few thousand dollars to make so there's not a large market, but by 2028 the costs should maybe be 10s of dollars and similar quality as 4o image gen but for a movie.
@ChinmayTheMathGuy says it right there in the description:
"It should be more or less comparable to a big-budget studio film, although it doesn't have to pass a full Turing Test as long as it's pretty good."
I don't think any AI system can make even a 2 minute clip that is at the level of a 2 minute clip of a big-budget studio film.
No AI system can currently even make a 2 minute clip with matching audio, video, dialogue, consistent characters, etc, from a single prompt, let alone one at the quality of a big-budget studio film.
In any case, feel free to bet YES if you think so.
@bens I mean, image generation took a big leap with the OpenAI release yesterday, and still I wouldn't even say a single IMAGE is anywhere near the level of a "professional studio photograph" or "National Geographic Cover" or something. There are levels to this stuff.
I wouldn't even say a single IMAGE is anywhere near the level of a "professional studio photograph" or "National Geographic Cover"
I would.
Try to lock in what you think and average NatGeo cover looks like, and then go look at the 12 actual covers from last year.
https://ngsingleissues.nationalgeographic.com/national-geographic/2024
Same for professional studio photograph, you probably imagine some of the best portraiture of all time. Instead search your city's Yelp for photo studios, and look at the portfolio of the #3 result.
I'm not trying to move the goal posts, I'm trying to lock them in. Big budget studio films include a lot of losers, not just the best-picture nominees you might imagine. The example given is even a Star Wars/Trek crossover; a human-made big budget studio version would have some fun visuals, but it would piss off fans from both franchises and get ripped to shreds by film critics.
@robm there are probably about 100 big budget Hollywood films per year.
Do you think current image generators can make an image that is one of the 100 or so best images from a given year? And not from the sheer novelty of being AI-generated, that is, but from actually being a top 100 image in quality? No. Absolutely not.
@JimHays c'mon... let's be serious here. This question is not whether if you saw a movie on silent in the back of a waiting room whether you'd notice it's AI. The question is whether AI can generate a Hollywood-level movie! The level of the movie is not assessed by someone glancing at it while in a waiting room.
An analogous question for image generation would be something like "would Nat Geo select this image for their cover on its own merits" or something, not "would a person notice the image being weird if they glanced at it briefly".