OpenAI's Sora: Text-to-Video AI is a World Simulator?!

check out my cool lil leaderboard website!

Sora

This video is supported by the kind Patrons & YouTube Members:
🙏Andrew Lescelius, alex j, Chris LeDoux, Alex Maurice, Miguilim, Deagan, FiFaŁ, Daddy Wen, Tony Jimenez, Panther Modern, Jake Disco, Demilson Quintao, Shuhong Chen, Hongbo Men, happi nyuu nyaa, Carol Lo, Mose Sakashita, Miguel, Bandera, Gennaro Schiano, gunwoo, Ravid Freedman, Mert Seftali, Mrityunjay, Richárd Nagyfi, Timo Steiner, Henrik G Sundt, projectAnthony, Brigham Hall, Kyle Hudson, Kalila, Jef Come, Jvari Williams, Tien Tien, BIll Mangrum, owned, Janne Kytölä, SO, Richárd Nagyfi

[Music] massobeats - lucid

[madushandissanayake]

Imagine we are living in a simulation created by Advanced AI nearly 14 billion years ago using a text prompt.

[oowaz]

some of the animal videos are terrifying there's something about incorrect anatomy with that level of detail, also the turtle eating jewelry 💀

[SweetHyunho]

Those video artifacts are like free VFX

[technicalmaster]

The most interesting thing to me is that it actually keeps getting better with more compute. Imagine what a future model with much more compute will be able to simulate with this level of progress. If this continues, we might soon be simulating chemical reactions, cellular processes, rigid and soft body dynamics and so much more.

[BrianMosleyUK]

This is so reminiscent of lucid dreaming... and also the concept that we are always dreaming, it's just that our waking dreams are framed by the physical world.

Future generative models will be guided by a 3d physics engine of some sort. We're so very close!

As 2 minute papers would say... Just another couple of papers down the line... and what a time to be alive!

[gnollio]

Bling Zoo needs more footage. I've got to see what that monkey king is up to.

[yaelm]

0:45 Wow, I thought it generated nerfs and used assets, which then another AI would beatify the result.

The fact that it's that much 3D consistent (enough for SFM) is an emerging capability is insane to me.
We are going to get good 3D generated scenes in no time soon

[almundtan]

this makes us closer to having Star Trek holo deck simulators

[joelface]

Love that someone turned some of these simulations into G-Splats. SO much potential by simply prompting the model for a 3D rotation of an item/person/etc. If it can do that consistently, it can make some amazing 3D models that can then be rigged and animated, or simply viewed in holographic space, or explored in 6DOF, etc.

[YUTPIA]

I thought this would take at least a few more years.

I want to get off Mr. Bones wild ride.

[Droid]

Those failed results really make it look like dreams, where most of the time things are constantly changing and don't make any sense

[NIkolla]

one particular detail that caught my eye is them saying they are using synthetic data to train the models, this may be a clever way of solving both copyright issues and it can be used to imprint a recognizable aesthetic on generated content.

[sneedtube]

By far the best video I saw on the subject

[itssoaztek]

Always a pleasure to hear your opinion together with some good explanations of important technical intricacies. Thank you!

[Mobay]

I just hope somebody trains this on the most abundant and fastests growing data source for videos, that involve human interactions. ;-)

[bjdrknn]

As a cinematographer i was shitting my pant's seeing Will Smith eating spaghetti. Year later, at this point i really don't care. The industry is doomed and we won't do anything about it. I think at the moment it starts to collapse, there will be more things collapsing, so that would be the least of our problems

[MemesnShet]

Nobody is talking about the leapfrog SORA has made over Dalle3 for single image generation:SORA IMAGES ARE INDISTINGUISHABLE FROM REAL ONES

Can't wait to try it

[keenheat]

this current trend of synethic data remind me of an issue that was brought up during tesla FSD development. They were asked why don't use more simulated driving data like waymo or cruise. Tesla respond that you only push the problem from solving self driving to solving perfect simulation of reality. Which is a lot harder problem. Then they show a collection of weird road condition you wouldnt think it would exist. IE: an old man "shepherd" a group of washing machines chain together on a highway, weird shadow pattern that made the road looks like it split into two roads, intersection with 50 plus traffic lights, etc. Reality is a lot weirder than simulation by order of magnitude. So to capture edge case, you still have go observe reality.

I get the feeling synthetic data will have the same issue of "how close is the synthetic data to reality". These data probably don't capture reality too well. But I guess that's okay since image generation is lot less mission critical than self driving.

[Allplussomeminus]

5:02... That looks like something out of a dream... All those previous examples really.

[zedo]

These guys are moving so fast that my brain cant even keeep up with them 🤯🤯