Short answer: Sam Nadler says Sora 2's face fidelity and social app impress, but physics in active clips like skateboarding were wonky next to Veo 3. Jordan Metzner says generations take three to five minutes and have cost him nothing so far.
Watch this answer from Episode 15 (starts at 10:19) · Episode 15 notes and transcript
From Episode 15: Sora 2 Took Over AI Video + Our New Parking Data Tool, starting at 10:19.
Sam Nadler [10:18]: Super cool. Yeah. Let's transition. Yesterday, Sora two came out. You and I have both been playing with it pretty much all day. I would say it there's a few things that make it really impressive, and then I personally, I think it trails a little bit in to veo three. But, for me, what I think is really, really impressive is the fidelity of, the face and who's generating it. I also think, like, the social app launch included with Sora two was a really cool way to get users, posting videos and sharing what they're making. But I feel, and this is, one day of use, the physics. There's a couple kind of active videos I made, whether it's skateboarding or playing soccer. Didn't you know, it was pretty wonky compared to Veo three. But, yeah, I would love to hear your opinion, what you've built. Maybe let's build one live right now and then take it from there.
Jordan Metzner [11:23]: Yeah. So I have done a ton of Sora two videos. As you see here, this is kind of the Sora two homepage on web. Lots of Pikachu, lots of SpongeBob, lots of Martin Luther King and John Kennedy, dogs and cats driving cars. So we've seen a lot of really fun stuff. This is like security footage, seen a lot of Sam Altman kind of stuff. Here are the things. Here's Hamilton. So it's been great overall. I think it's kind of like a really fun tool to play with. Sometimes the physics are great. I think in this case of this jumping on the trampoline, it looks pretty good. You can see the trampoline bounce. The you can see it kind of inflect and inflect back up again. But to your point, sometimes the physics have been a bit weird. I was doing some clicking one and I had like three fingers or something, so it looked a bit strange. But why don't we just do a video here? I have two dogs. One of them is a black lab. So, let's do a black Labrador.
Jordan Metzner [12:24]: And then what should we have her doing? Maybe we could have her cooking, maybe sushi in a high end restaurant. How's that sound? With a chef's hat on. Okay. Cool. So, it does seem that the generation time is somewhere between, like, three to five minutes. Meaning that like, to run this model is probably incredibly expensive for them. I've generated at least like 10 or 15 plus videos so far at a cost of zero. And if you recall like when we were using Veo three, the costs were like 6 to $8 a video and then driving downward. So, I don't know how much it's costing ChatGPT to produce these videos. I presume it's incredibly expensive. Right now access to Sora is limited through invitation only to plus members. And so, they've done some mechanisms to restrict the amount of volume. I even was trying yesterday and kinda got denied because there wasn't enough available servers.
