If you mean high artistic quality, who will be the judge, and what do they think of Michael Bay movies?

@GazDownright ~Any Michael Bay movie is “more or less comparable to a big-budget studio film.” I get that this market is technically inherently subjective, but in practice a lot of outcomes will be extremely obvious.

@DavidHiggs There several big-budget studio films that are of very poor quality. I think AI is already capable of making big-budget-looking absolute swill. If AI is able to create a masterpiece, however, is a more interesting question, imo.

@GazDownright I have yet to see a single example of plausibly “big-budget” looking anything from AI in terms of video, and that’s without audio.

@DavidHiggs I think current AI is approaching the level where it could make a superior version of Cats already

https://x.com/EHuanglu/status/1902765779944345666?t=B-WCy67p11fJQwgG5r5iBg&s=19

@SimoneRomeo bhahahahha yesss

Upped my NO some, thinking about "high quality". Claude plays pokemon type stuff makes me think we're pretty far off on the kind of big picture thinking that makes a 2 hour movie adequately coherent.

Question also is framed like "make me a movie" not "here's an entire screenplay, written out character information sheets, visual style guides...make it into a film"

Also I'm guessing OP leans no if people are just producing stuff like "2 hours of guy wandering through desert" Since an AI can presumably already produce certain edge cases like "a 2 hour film of a man watching grass grow"

@JoshuaPhillipsLivingReaso a specific movie AI tool could convert the prompt to a movie-script / screenplay internally and only then go text-to-video.

Otherwise I'd lean on "No"...

I think there will be an event where this market shoots up by a lot in the next 18 months

@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.

I don’t think we can meaningfully derive anything from modern day movie creation numbers. Incredible amount of noise.

@GG once you can make an outstanding 5-min video, you're 99% there. The reset is just an automation that does the first job 20 times and stitches them together, plus coherent storytelling which we already have. Not very difficult

@SimoneRomeo is making a good 5 minute tiktok video 99% of the way to making a hollywood-level movie? For humans (let alone for AI)? lol

@bens frankly I don't know: I haven't seen a truly remarkable 5-min TikTok, including storytelling, blockbuster effects, camera works and Hollywood-level actors. Could you send me a link so I may check it?

@SimoneRomeo Even if you think going from 5 minute videos to 120 minute videos is trivial, that only represents about 1/3 of the logarithmic distance we need to cover.

Remember, it’s not just about being able to make a video with AI, it’s about being able to making the reducing the human workload to the level of a one sentence prompt.

@GG my point was exactly the opposite. Going from 5 min to 120 min is trivial. You mostly need skills that are going to develop anyway meanwhile (like long-context storytelling and action taking) and indeed, I posted above a platform that already does what we're discussing about even if the quality of each clip is low for the time being

@SimoneRomeo How much work does the end user have to put in to get a five minute movie from that tool?

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

There will surely be a lot of progress, but last mile problems can be pretty significant

@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.

@GG what's your suggestion here? That actually no progress has been made?

@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.

@GG u can't calculate gradient from one point

@jim Just backprop the weights jim