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
Does @BrianM have 2000 separate limit orders at 36%? Do you realize you can make a limit order that is larger than 25 mana?
@ZandaZhu usually this happens when someone has a lot of mana invested, and not enough mana balance to place the limit order size they want. Spam smaller orders = 1 bigger order, doesn't matter if you're low balance because it'll fill as a sell anyway
@bens did the two of you both decide to take this limit order at the exact same time? @Joshua @SemioticRivalry
@bens so i'm talking to my friend joshua on discord right and he tells me about this limit order, and I say oh cool i'm gonna take it. and he (???) races to take it 0.1 seconds in the time between my click and it going through, lol.
@SemioticRivalry damn. sometimes I make limit orders to eat limit orders when it's a big trade, but not always. you basically cleared the whole YES limit order stack! incredible
lol. Not a comment on the prediction just the size of the order.
@0xseraphim Nah, just seems more like late 2028 or anytime 2029, not by the beginning of 2028. Maybe even 2030 since the last mile problem could easily make this take much longer, though I expect progress to speed up again before this market resolves, let alone 2030.
@DavidHiggs what's the matter of that last mile blocker? A system can't stitch together the pieces into a polished product?
Can anyone with access to Seedance 2.5 ask for this and post the results here? Make it as long as it will allow (30 seconds I think).
photorealistic video of a mime in front of the Eiffel Tower. It's a sunny day and the mime is about 15 feet away from the camera. He wears traditional mime outfit with striped shirt and white makeup. He juggles four bowling pins. While juggling the pins, he takes a tennis ball out of his pocket and starts juggling that along with the pins. Once he's in the flow with the four pins and the tennis ball, the camera circles all the way around him and stops where it started
@pietrokc “it doesn't have to pass a full Turing Test as long as it's pretty good”
The provided example in the market’s description is not adversarial, and doesn’t give this level of scene-specific direction. A more comparable test to the description’s example would be:
“Make a documentary about a juggling French mime”
@JimHays I'm not suggesting the model has to pass that specific prompt, but that prompt does reveal its ability to generate video with many objects interacting in a complex way. And I think if your model can't do that, then there will be many non-adversarial, non-detailed prompts it will fail.
All examples posted so far in this market's comments are roughly the same, a bunch of 2-5 second clips with few and standard object interactions, and even those are frequently messed up. It's pretty clear to me that there has been ~zero progress in world/physics understanding from video models in the 2.5 years since Sora was announced.
@0xseraphim I have a No position, but I also don't think that its at all obvious that the market will resolve to No, as in the probability of Yes is still likely above 5% or so.
@ICRainbow making the pieces coherent together in the first place, and having enough in context memory to manage a full film with all the associated unused footage, editing decisions, world-building, etc. And then, “stitching the pieces together” as you put it, which is the least difficult of the three I think, though not easy.
@DavidHiggs I don't think that's really a requirement to making a passable movie good enough to resolve this question. Making a movie with AI isn't like making a movie in real life, you're not limited by the shots you took and are stuck in some editing phase trying to piece together a load of footage. With a video gen model you can just keep recombining ideas and reference prompts until you get a shot that passes a certain threshold. That includes measuring how coherent it stays with respect to the character identities, the plot, the theme etc. It looks like we're going to cross a lot of the remaining distance just with better models and better harness engineering. Some super duper magical "full context video model" that can process 200 hours of footage in its input context doesn't sound like something that's really necessary here.
@0xseraphim I was basically describing what I saw as the gaps between someone’s attempt at having Fable 5 make a movie using a bunch of sub agents and video/audio/etc. models they posted on LessWrong and the resolution to this market. So I could be making bad assumptions.
But they were doing the stuff you describe, such as generating new shots until hopefully they got something good/usable.