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.
1,000People are also trading
@0xseraphim First 15 seconds: pedestrian walk / don't walk sign is facing the sidewalk instead of the street, trash can vanishes (on the left, next to tree), parked car appears out of thin air, trash can reappears in different position (teleport).
Length 1min06sec but all shots are 1-10 seconds long.
Nothing new here.
@pietrokc https://matrix.fandom.com/wiki/The_Matrix/Mistakes
Continuity: When Neo is being bugged the progression of the bug is inconsistent between shots.
Continuity: As Morpheus introduces Neo to the Construct, his glasses are completely dark. The following shot they suddenly turn highly reflective. This was because, for that shot, the glasses were painted black to avoid the cameraman from being reflected in them.
Continuity: After Neo is brought into the real world, Tank comes to introduce himself; while they're talking, the strap on Tank's right shoulder repeatedly jumps from the middle of his shirt sleeve to the edge between shots.
Continuity: When Neo gives Agent Smith the finger, the position of his hand changes between shots.
Continuity: Trinity is wearing a minidress over her tank top and vinyl pants that disappears when the group is first shown inside the wet-wall.
@pietrokc I think a feature film doesn't really need any shots longer than 30 seconds. Maybe less. And I would say the sign facing the sidewalk etc mostly fall under "doesn't have to pass a full Turing test".
The path to this is so straightforward I am shocked the odds are this low. Isn't seeddance already 'pretty good'? And LLMs can write ok screenplays. And there's about 1.6 years to go.
The functionality will probably exist, it kind of exists now in a form not meeting the quality requirement. But it might not make enough sense as a product to be packaged exactly in the form of prompt in, 120 minute movie out.
$0.1 / second for current seedance means $720 for a 2 hour movie. If it stays like that, people probably aren't going to be issuing $720 prompts without significant steering.
@JimHays I would not have thought writing would be the hardest part of video generation. But here we are.
@JimHays I asked Claude Opus 4.6 to write a full-length script for a 120 minute romantic comedy featuring hot air balloons. It does indeed look okay.
@comicstosteal and remember, seedance is what a Chinese lab has done with worse data, less compute, and less experience than Google DeepMind. Whatever GDM releases next will blow it entirely out of the water
If it stays like that, people probably aren't going to be issuing $720 prompts without significant steering.
Only needs to happen once for this to resolve YES. Even if none of us want the mana enough to blow $750, seems like an obvious marketing play for any of the 500,000 AI startups with millions in seed funding
@comicstosteal that $720 is probably a 3x-10x underestimate, since the winning solution will almost certainly be some kind of agentic ai with ability to retry shots and pick the best one.
@LoganZoellner I’d say 3x-1000x (wide error bars) underestimate more likely, but yeah, highly agentic AI (or medium-fast takeoff) is the only way this market is resolving yes and the amount of tokens going into a full blown good movie would be insane.
It wouldn’t be just generate more than 2 hrs of frames/shots to pick better than average ones, there’d need to be a lot of editing, some redoing of whole scenes or arcs of the movie, etc. And you’d need acting quality voice performances, probably VFX added on top of base video generation, musical scores, sound mixing, etc.
The main problem of course will be that the capability might not exist due to voice acting, narrative cohesion, overall long-run film direction skills and other things falling under the general category of “taste.” Agentic time horizons on visual-media skills and audio/video realism + continuity may well have extended “far enough,” but likely not to the point where reliability and quality make a single-shot cohesive high quality movie happen, even at exorbitant (for AI inference) prices.
Remember that things like computer use, reasoning based on images, creating visual-media content, etc., has continued to lag significantly if not far behind coding, math, spreadsheets, etc.
@musteval I'm sorry, that is a crucial misunderstanding of this market.
For this market to resolve YES, a model must be able to generate "good" movies in response to a prompt, on command, not just once. Otherwise you can just train a model to memorize Citizen Kane and output that in response to a prompt that asks for <plot summary of Citizen Kane>.
https://www.youtube.com/watch?v=CNbmoVdirxw
This is absolutely high quality. Expand it to 90 minutes and I would pay to watch. (and yes I know it probably took tons of prompting for each shot)
@ErickBall I think the bar is much lower for a trailer - it can just mash cool scenes together without worrying about narrative coherence.
@TimothyJohnson5c16 Is narrative coherence still hard? I think coming up with clever/original/funny thoughts is the hardest part. This trailer has several, but they're probably still all human-generated.
@ErickBall No one wants to watch a movie that's all 2-second cuts. Sora 1 has been able to give us 1-10 second cuts for 2+ years now. The challenge is longer scenes maintaining physical plausibility.
If OpenAI or DeepMind could currently generate a 1-minute long scene with complex interaction between many objects, don't you think they would be advertising it everywhere? Yet all we see is 10-second clips.
I think that kind of highlights how this market is going to end poorly.
AI, Dawkins and crypto fans are pretty likely to love this but a bunch of others are going to point out it's pretty incoherent nonsense. It really falls off the longer it goes on for. Still, it's a pretty subjective call with priors like your bets in this market easily biasing an assessment. Plus the tools list makes it look like there was a lot of cut scenes thrown together and I'm not sure how you make sure that wasn't stealthily done.
@pietrokc tbc I wasn't claiming this meets the "high quality" bar. I just thought it was fun. But I do think a yes resolution would most likely come from a dumb flashy action movie. Transformers, not Citizen Kane.
@ErickBall As I've said elsewhere, that's a misunderstanding of this market. The market requires the AI to succeed for a wide variety of prompts, if not most or all. It is not sufficient for it to produce one "good" movie. Otherwise I could make a "model" today that just memorizes Fast and Furious X and returns it when asked for <plot summary of Fast and Furious X>.
@pietrokc I mean. I agree it has to work on lots of different prompts. But nothing in the resolution criteria says it has to work in every genre.