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.
1,000@qumeric: First version of Midjourney (which was really bad) launched less than 3 years ago. Now people do good quality 5-10 minutes films/animations in a week (obviously with a lot of human effort still). It feels like we are more than 50% there.
So one week of human effort, call it 40 hours, can currently produce 10 minutes of decent AI video. To resolve this market yes, a human will need to be able to create 2 hours of AI video with only 1 minute of effort. That an power multiplier of 28,800 or 14.8 doublings.
There are 37 months left until the end of Q1 2028. So to keep pace with the market, we would need a doubling in AI video power (quality-weighted AI output relative to human input) every 2.5 months. Does it feel like the AI video models are twice as good as they were in mid-December 2024?
Or put another way, if SOTA AI can currently create a decent 10 minute video with 40 hours of human work, and AI power is doubling every 2.5 months, then around mid-November this year it will be possible to make a decent 2-hour movie with only 40 hours of human work. Does that sound plausible?
"Gold Gang" came out 10 months ago, was only 2 minutes long, and I'm still not sure if I could make it with 40 hours of effort.
@GG There is no doubt there is no measure of past time where there has been sufficient progress per unit time to think it remotely likely that this market will resolve YES.
But they mainly do not care because of supposed "exponential progress"
Still, this market is likely to be down to around 30% by the end of the year, and 15% at the end of the next.
@GG Googles Veo 2 model felt like a legitimate breakthrough to me, probably 50% better than Sora
@nsokolsky Using Veo 2, Jason Zada made a 104 second short film in "a few days, a few hours here and there". That's a vague estimate, but lets handwave it to mean a 3 hours of deliberate effort, and two hours of incubation creativity (where you think about your project in the shower, in bed, etc).
That's a ratio of 173 minutes of human work to 1 minute of AI output. The goal is 1 minute of human work to 120 minutes of AI output. So Veo 2 needs to improve by a factor of 20,760 in order to cross the finish line. If you define a "breakthrough" as being 50% than the previous SOTA, we'll need 24.5 breakthroughs before the market's deadline.
@JimHays I also don't expect anyone to put in the level of resources this would require without having a human at least do some review and editing.
@qumeric that feels like 20% of the way there
They haven’t even put real dialogue in the movies, no?
@AlexanderLeCampbell All components already there or almost there. Generate a scenario, split it into scenes, for each scene generate detailed sequence, split it into prompts, generate. Object permanence will be an issue but I think we are getting there. Then voice it using elevenlabs, do lipsync where needed (good lipsync exists) and generate sound effects.
Now we just need a scaffolding which will do everything for us, most importantly watching generated scenes and judging them. It can be done
I think the main issue (apart from object permanence) is cost. It would be pretty slow and exepensive just to generate 2h high quality video but you would probably need hundreds of hours of attempts
@qumeric That is nowhere near 50% of the way to making a high quality full length film.
@DavidBolin progress is exponential
based on METR research
rn 1 minute
EOY 5 minutes
2 years ~30 minutes
3 years 1.5 hours
so, it'll be close even in the worst case
except we got 1 min gen ages ago so we're actually further along
and there's a decent chance of a breakthrough getting us there sooner (like reasoning models got us to LLM advanced problem solving ~5 years earlier than expected)
@qumeric
> except we got 1 min gen ages ago so we're actually further along
That's the opposite of how we should update.
Imagine you're taking a 100km train ride. You see the 10km landmark being crossed. You estimate it's been 10 minutes since you got on the train. Extrapolating, the other 90km will take you 90 minutes. But then you check your watch and realize, "Oops, looks like I've actually been in the train for 20 minutes." You should now estimate that the rest of the ride will take you 180 minutes.
@jim I'm just saying, the longer we've been at this approximate stage of progress, the longer we should expect the remaining stages to take.
This has probably been answered somewhere in the comments but I can't find it, when the market title says "able", does that mean "able ever" i.e. it has done it at least once even if the success rate of the resulting movie being actually high quality is <1%, or "able normally" i.e. the success rate is >50%, or "able almost always" i.e. the success rate is >99%?
@TheAllMemeingEye not OP but I’d argue even a 0.1% success rate would qualify because that’s still going to be an incredibly mind blowing achievement.
@johnwhiles Was about to comment something similar. Imagine a world where the AI asks you for a lot of money, and then directs a live-action film.
I'd imagine this is unlikely, but a fun idea nonetheless.
@TiagoChamba There is a movie where an AI directs live actors on a movie set.
The Second Act, by Quentin Dupieux (so, expect something weird)