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
I’d be really interested in a question asking what year people do think this will be possible. I think unless there’s a slowdown in ai progress I’d be surprised if something technically passible can’t be done by 2034. Quality I think would be harder though because there’s so many different things like subplots, pacing, having a soundtrack that’s coherent and not just generic, some kind of unique to the film visual language that isn’t just a generic version of most films of a genre, set up and pay of, a very deep understanding of physics and time etc that would be required to make something “good” unless you’re specifically valuing the novelty and strangeness of what ai can make in which case I think you could already make something interesting and feature length.
@Ataraxia Taking @LoganZoellner estimate that completing making a 7-minute AI video currently takes them 2 hours, at least, and extrapolating a transformers-double-in-power-every-seven-months trend, I calculated it will take one minute for humans to generate a 120 minute film in January 2032.
On a meta level though, I don't expect the 7-month-doubling trend to continue indefinitely, and I expect we'll reach the singularity sometime before 2030. If you told me that in January 2028, an AI is able to generate a 120 minute movie with less than 1 minute of human labor, I'd be more surprised if you told me that AI was just plugging along normally than if you told me 2028 was a post-scarcity singularity.
@GG yeah, if ai improvements start to accelerate ai improvements then timelines for everything get much much shorter and as task like making a movie could become trivial. I’m personally skeptical that we’re less than three years away from that or what most ppl would count as a singularity. Now that I’m thinking about it though, if an ai can generate short videos well and then has a scratch pad to remember important details and larger scale structure, I think it would actually become a lot easier. Part of what’s pushing me to be skeptical also is that if even just one or two things are off in a two hour film, it can really tank the quality. Like if a characters voice doesn’t quiet work or the third act doesn’t really follow the second etc, there’s a really wide surface area for failure. I guess this question is kind a proxy for if there will be an intelligence explosion or not for me lol
@Ataraxia
My (only semi-sarcastic) guess is that we will have systems that questionably resolve this positive in 2028, similar to when Scott Alexander prematurely declared victory on his image-generation bet and an unambiguously good movie generation system a year or two after that.
In particular, the low quality version of this question will resolve positive and then we will endlessly debate the meaning of the words "more or less comparable to a big-budget studio film" until an AI movie achieves some huge milestone (box office, academy award, high quality turning test).
>I guess this question is kind a proxy for if there will be an intelligence explosion or not for me lol
Much like "an AI can never win at chess", "an ai can never win at go" "an ai could never pass as a human undetected on reddit" and "an ai could never generate images undistinguishable from real photographs" I very much believe that "an AI can never generate a film" is one of those milestones that we will fly past and people will still be saying "it's not really thinking"
@LoganZoellner you're right. Idk why I wrote what I did but it feels super obvious to me you don't need to be super intelligent to be able to generate quality movies.
I made a related market to see if we're on track for this: https://manifold.markets/JoshSnider/in-mid-2026-will-an-ai-be-able-to-g
LLMs are doubling in power every 7 months. Granted, video AIs like Sora and Veo are not LLMs, but they still rely on transformer architecture, like LLMs. If you have alternative trendline you think better fits video AIs, please share it!
Extrapolating from LLM trendlines, we have about ~4 more doublings in power before January 2028. So if AI can make 120-minute movies in January 2028, it should be able to do something 1/16th as impressive in August 2025. And I don't think they can.
Currently, there is no AI that can make even a 30-second, "Family Guy" style cutaway gag without significant work by a human curating, editing and instructing. (There are AIs that broadcast their ability to generate 30-60 second clips, but these tend to be slow, panning scenes that would be cut to 5 or 6 seconds in a real movie). If we were really on track to have feature length films in 2028, I'd expect to see AI created short-films, 4-8 minutes in length by now, without a human in the loop.
@GG
I agree that in general we should expect AI video to proceed at the same rate as transformers (and current AI video models are not producing 4-8 minute videos 'equivalent to a full high quality movie'). However there is a lot of room for upside here.
Video is seriously under-optimized compared to text. Veo3 is not much better than Wan2.2, a model that has only 28 billion parameters and cost a few million dollars to train. By contrast, the best LLMs have >1T parameters and cost Billions of dollars to train. That means we can instantly get 2-3 OOM by either: 1) someone decides there is real money in training a better video model or [more likely] 2) video models, LLMs and robot-control models converge into an all-in-one model (we already see hints of this with Genie-3, which again is a very small model compared to what currently possible).
The METR data itself hints at this, since 'compile a 4-8 minute video out of 10 second clips' is a task a human could do in a few hours. I suspect if you were in the mood to spend a few hundred dollars, Gemini CLI with tool access to Veo3 could do this today.
There is also the possibility that RSI means the trend will not remain exponential but at some point it becomes super-exponential (but I consider this less likely prior to 2028, and balanced out by the possibility that growth instead flattens out).
@LoganZoellner
Very good comment. What follows are just nitpicks.
I wouldn't update too much on the Genie-3 showcase. Sora looked amazing but was stuck in development hell for 10 months, then turned out to be hot garbage when it finally went public.
>The METR data itself hints at this, since 'compile a 4-8 minute video out of 10 second clips' is a task a human could do in a few hours. I suspect if you were in the mood to spend a few hundred dollars, Gemini CLI with tool access to Veo3 could do this today.
You could stich 24 10-second clips together and call it a 4-minute video. But without a human curating you'll have no continuity between the clips, as if you hired a director, cloned him several times, had each clone shoot a 10 second portion of a scene, and then stitched them together. Making multi-minute AI videos, as of now, still requires a human in the loop reviewing the clips, suggesting improvements, and editing them together in a way that respects film grammar.
Remember, the example provided is "make me a 120 minute Star Trek / Star Wars crossover". That's less than a minute of work on the human's part for 120 minutes of video output. If we were four doublings away from that mark, we could make 120 films with 16 minutes of human work. Instead, humans are spending several hours editing 3-5 minute films.
@GG
> I wouldn't update too much on the Genie-3 showcase.
By "wouldn't update", do you mean you disagree that multimodal models that unify language, video and robotics are the future, or you just think they're further away than 2028?
>You could stich 24 10-second clips together and call it a 4-minute video.
No, what I mean is this is literally a thing that I have done. Using ai video generators, I have put together a few minute video in a couple of hours. A better example, though, is probably Trisha Code. Trisha puts out a video every single day which means (unless this is somehow a full-time job for them) that they are making each and every one of their (usually 1-2 minute clips) in a few hours/day.
> Instead, humans are spending several hours editing 3-5 minute films.
This is precisely my point. METR puts GPT-5 at a the ~2hr mark. Meaning that if a human can make a 5 minute clip in two hours, so can (in theory) GPT-5. Now obviously there would be some issues with this (as GPT-5 is fine-tuned for programming/math not film-directing) but that's mostly a companies haven't chosen to invest in this type of product barrier, not a fundamental limitation of current technology.
I admit this probably still leaves us a bit away from the mark. I would be hard-pressed to make a 7-minute clip in 2 hours. But we're probably closer to 5-6 doublings away than we are to 8+.
@LoganZoellner
>By "wouldn't update", do you mean you disagree that multimodal models that unify language, video and robotics are the future, or you just think they're further away than 2028?
I just mean that I don't weigh the Genie showcase very heavily when I make predictions. I don't think it tells us much, because the public hasn't had a chance to actually test Genie yet.
>No, what I mean is this is literally a thing that I have done. Using ai video generators, I have put together a few minute video in a couple of hours.
Maybe I wasn't clear. I understand that humans + AI can stitch together AI generated clips into pretty good short films. But that requires labor on the part of the human, determining which clips are good, adjusting the prompt when the AI struggles, strategic editing to remove bad parts of clips, etc.
If you told Gemini "Write a 4-minute skit, then divide that skit into 24 10-second segments, then write a Google Veo prompt for each of those segments, then stich those segments together" that would not be a good video. The quality would vary drastically between each segment, the cuts wouldn't mesh together, and film grammar (like the 180 degree rule) would be ignored. Even if you gave Gemini the ability to assess Veo's output and re-prompt it, Gemini wouldn't have the ability to evaluate each clip and see what needs improvement. You still need a human for that, for now.
>Meaning that if a human can make a 5-minute clip in two hours, so can (in theory) GPT-5.
Even if a cash dump into video AI suddenly allowed us to make 5-minute clips with a single prompt, we're still 4.6 doublings away from making 120-minute films with a single prompt. That leaves us crossing the finish line in April 2028.
I don't know why we would expect a cash dump into video AI though. Presumably, the market has already determined that this is the most profitable amount of money to dump into video generation. Claude estimates text generation is responsible for about 15-20% of US GDP, compared to video's 2-2.4%.
>I would be hard-pressed to make a 7-minute clip in 2 hours. But we're probably closer to 5-6 doublings away than we are to 8+.
A 7-minute clip in 2 hours means you are investing 17 minutes of human effort for every one minute of AI video output. This market resolves when a human can generate 120 minutes of AI video with 1 minute of effort. That's 11 doublings in power, finishing in January 2032 if the 7-months per doubling holds up.
@GG
> Even if you gave Gemini the ability to assess Veo's output and re-prompt it, Gemini wouldn't have the ability to evaluate each clip and see what needs improvement.
Maybe you're not familiar with Gemini? It has the ability to "see" video, and the CLI can do things like run a command, look at the output, and then re-run the command if it doesn't like it.
>Even if a cash dump into video AI suddenly allowed us to make 5-minute clips with a single prompt, we're still 4.6 doublings away from making 120-minute films with a single prompt. That leaves us crossing the finish line in April 2028.
I agree. However, if you look at the error bars on the METR, +/- 6 months is well within 1 standard deviation.
>I don't know why we would expect a cash dump into video AI though
I don't expect a $10-100 billion dollar cash-dump into video. Moreso, I expect the natural evolution of AI will be towards a general purpose system and video will be one of the modalities of that system. However I do think "first Oscar-worthy AI-generated film" is one of those vanity-metrics that Google or OpenAI would happily dump a few million dollars of engineering effort into if they thought they were close.
>A 7-minute clip in 2 hours means you are investing 17 minutes of human effort for every one minute of AI video output.
My claim (unproven) is that if OpenAI or Google invested as much into AI video generation as they did into programming or solving math Olympiads, we would have an LLM that could generate a 5 minute video with 1 minute of human effort already. I argue that this is plausible because:
1) I, a human, can generate a 5 minute video with 2 hours of effort
2) According to METR, GPT-5 can do tasks that require 2 hours of human effort
My 2nd claim (testable if someone wants to spend a few hundred dollars on Google API credits) is that Gemini CLI today with a day or two of engineering could be made into a system that generates clips significantly longer than 10 seconds based on a one-sentence prompt. I doubt (since Gemini is not fine-tuned to be a film director) that it would actually be as good as a professional working for 2 hours. But I think it would be very interesting to know what the real status-quo on video generation is.
I agree that still leaves us 4.5 doublings away from a positive resolution, and absent any other factors this doesn't happen soon enough for the market. However, my gut feeling is that the market will resolve positive anyway because:
a) 0.5 doublings is well within the statistical noise
b) RSI could very well make progress faster in the the next 2 years than the last 2 years
c) these kind of markets seem to be surprised more often by short-timelines than long ones
@LoganZoellner
>Maybe you're not familiar with Gemini? It has the ability to "see" video, and the CLI can do things like run a command, look at the output, and then re-run the command if it doesn't like it.
I showed Gemini 2.5 Pro this video and asked it to suggest improvements. Gemini seemed to think that the video was a still image, even though I uploaded the actual MP4 file. After I clarified that it was in fact a video, Gemini seemed to think that the boat was static while the camera zoomed in. This is backwards, the camera remained fixed in the video while the boat turned back and forth.
I would not trust that instance of Gemini to curate a video. Perhaps you have access to another Gemini model? Would you, for instance, be able to compile a 3-minute video, with a narrative structure (not just a clips montage), relying solely on Gemini to tell you what's in the mp4 files you're stitching together? That is say, can you make a short film without human eyes reviewing any of component clips the video is made of?
this seems to be a completely accurate description of the video.
> Would you, for instance, be able to compile a 3-minute video, with a narrative structure (not just a clips montage), relying solely on Gemini to tell you what's in the mp4 files you're stitching together?
Probably, but it would take a couple days of engineering effort and cost a few hundred dollars, which is more effort than I am currently willing to invest to win fake internet points.
@LoganZoellner In that thread I asked it if the boat engaged in any motion and it said the boat was completely still. It also said the camera was zooming in, which it wasn't.
The few NO holders who are capable of extrapolation are coming up with increasingly convoluted cope to justify their positions
https://x.com/jkbr_ai/status/1953154961988305384
Obviously resolving YES
(friendly coded)
@Bayesian
A) When do you expect AI to be able to generate a 1 minute commercial in one shot, with no human in the loop? (Or if you think we've already crossed that, when did that happen?)
B) When do you think AI will be able to generate an 11 minute short film in one shot, with no human in the loop?
Take the time between B and A, and add it to B. That's when we should expect AI to generate a 121 minute feature film. The jump from a 1 minute commercial to an 11 minute short film is the same order of magnitude as a 11 minute film to a 121 minute movie.
@GG no it's not.
Find a film crew that just completed a 20 minute short film.
Hire them to create a 90 minute feature film, or a 150 minute epic. Do they need to upskill? Do they need to go back to film school? Do they need to learn how to make 30, 45, 60 minute films first?
No, at some point length decouples from the ability to make films. If you continue thinking of this purely in terms of length, I will take all your mana.
@robm "Do they need to upskill?" If you want them to write the screenplay - Yes. Writing for a short movie is much easier than a feature film.