Same as the main market but the AI must only be at the level of the worst filmmakers, rather than the best. So instead of a Marvel movie level of quality, think The Room level of quality. That's the yardstick that I will be using. Alternatively it can be any other well-known bad movie, like Birdemic or Who Killed Captain Alex. It doesn't need to have the "so bad it's good" qualities of The Room, but it needs to have a minimally coherent plot structure and cinematography. It must not be too painful for humans to watch the entire thing, and it needs to have the basics like characters say and do things that advance the plot, at least usually.
Main market:
1,000People are also trading
Question for those saying someone might try to solve this in a couple weeks:
Has anyone yet made a 30- or 60- or 90- minute one-shot AI video that’s NOT minimally coherent / is still painful to watch?
we got lots of unwatchable/painful 15-second video clips before we got coherent ones. Seems like we should expect the same progression for hour or more long clips. Are people trying this with the current tools?
@JoshuaHedlund I would say 3-5 minutes 1 shot "not physically painful to watch" is within current model capabilities. probably need someone else to confirm/deny, though, because my tolerance for AI is unreasonably high.
@logan I’m not asking about the maximum length of not-physically-painful videos. I’m curious about 60-minute videos - is anyone generating that length at all yet, regardless of quality.
@JoshuaHedlund I'm unsure what you're asking.
The video I linked was 4 minutes long and passes my bar for "must only be at the level of the worst filmmakers". I'm not claiming you can't tell its AI or that it's something I would actively seek out. But I am claiming if you forced 100 people to watch it or Rock Me Tonight and then rate their pain on a scale of 1-10 the AI video would be rated less-painful.
edit: never mind. I misread 90min as 90sec. Yes. 90 minute videos generated by AI are still physically painful every single aspect you would expect to be a problem is a problem: screenwriting, voice acting, character consistency, etc.
I would say the reason we don't see more long-form 1-shot AI generated videos is because it's both very slow and expensive, and it's just more satisfying to spend a day and make something that's 4-minutes and borderline non-painful than a week to make something that's 30 minutes and excruciatingly so.
@logan thank you, yes that’s an example of what I was looking for. Do we know if that 80 minute video example was from a single prompt with no human edits?
@JoshuaHedlund Yes, all of the videos I linked were generated 1-shot from a short (1-2 sentence) prompt. The prompt for the 80 minute video was a high fantasy film in which a elf princess is rescued from her kidnappers by a loveable rogue and they must cross a treacherous wilderness to bring her home
@JoshuaHedlund I have a failed attempt from a few months ago: https://www.lesswrong.com/posts/24RKHEkwgZ6Hm6ygY/can-an-llm-make-a-feature-length-movie-on-its-own and will be trying again with Fable 5.5 /JoshSnider/will-the-next-claude-fable-be-able
@JoshSnider thanks, this is really great information and experimentation. Sounds like it was very much not a 1-shot process, though, even to get to what you got. Looking forward to your continued attempts.
Things are getting to a point where we will probably see some attempts at resolving this in the next couple weeks/months so would like to get a better sense of where everyones head is at.
The main limitation that I see for this (and the high quality market) is that cobbling something together wouldn't clear it, having a director at all wouldn't clear it, and having a script definitely wouldn't clear it, "to a prompt" does a lot of heavy lifting. It is, in large part, a market on context rot being solved or not by then. It does require an end-to-end solution that scales to a fairly long script to clear it. I don't think video quality was likely to ever be the main limiting factor.
I understand it does not have to be able to do it consistently, and doing it once to a prompt would probably clear the threshold, but producing something that is 1h20+ in a way that is consistently somewhat coherent still seems ways away. I haven't been tracking every new releases, but it does seem like every feature length project that would come close to the quality bar did not involve writing the script and had a lot of humans in the loop. I'm curious if someone has some examples of things that they would consider to be close to warranting a resolution that goes in the direction of "to a prompt".
@LuEE I think at current capabilities, someone could probably do this with Opus 5.5 using Seedance and like $10k
@LuEE I've already seen ~1 hr long vertical microdramas that probably clear the "quality" bar here or are very close. These films have some amount of human curation but like... I think Opus 5.5 could write a better script than the incredibly bad and pointless scripts that these films rely on, so I don't think that's the bottleneck.
@bens Thats the main disconnect, I don't think current models can take on a task as long as a feature length script even with a good harness. Unlike something more modular like coding a large project, writing a feature length script does require the model to keep the whole thing in mind, at least a lot more of the context in much more granular detail. Context cluing from heavily compressed representation leads to a lot of inconsistencies so I'm not sure compacting is the way forward. And it doesn't feel like this is an axis where improvements have been quite at the pace of other capabilities, the couple long reasoning benchmarks I'm aware off basically publish with bad results across the board and are ignored entirely thereafter, doesn't feel like anyone is looking to deal with context rot or knows how too. Altough bad movies tier script might very well be within reach here
What’s surprising is that “to a prompt” seems the clear bottleneck to me although I don’t have access to the latest models. I don’t know if the trivial minimal prompt examples provided will be sufficient for an AI to do it in the following narrow sense: the AIs have access to tools now so it could scrape a bunch of tool sets and other prompts that contain the structure/instructions to produce such a movie, but this would seem to be a gray area that goes against the spirit of the market, yet on the other hand we might not even likely know that it it did such searching unless it was a locally run AI or relied on prompting where we said “don’t search the internet and don’t use the internet for any resources for this project.”
@LuEE I think this should only resolve to Yes if there’s a clear consensus that Yes, this is now possible. One offs with an unknown amount of human curation shouldn’t be sufficient.
https://www.lesswrong.com/posts/24RKHEkwgZ6Hm6ygY/can-an-llm-make-a-feature-length-movie-on-its-own
I don't know that this merits an immediate yes -- even what we see is with post-generation guidance -- but I would say this is a low-quality feature-length film. So odds looking quite good.
@TheOtherKC Now I haven't seen The Room or any of the other bad film examples to compare, but this one is pretty bad. The main character is obviously a different person in each scene, and he keeps doing things that make no sense like going in the opposite direction of what the plot suggests he should be doing.
I like to periodically revisit this. Here's what can be created using only open source tools (gemma4, LTX23, Flux-Klein-8b). Took about 9 hours to generate. Obviously you could do much better by using commercial models (Gemini-pro-3.1, NanoBanana Pro, Veo3) but would cost a few hundred dollars.
@ItsMe nope, you are not getting this right. Nothing in the markt says anything like this. This is not one of the requirements