OpenAI STUNNED As Sora 2.0 LEAKS!

What would OpenAI be without its drama? Recently, the SORA model was leaked on Hugging Face from a group of early testers in OpenAI's program. So in this video I'm going to dive into exactly what happened, because there's a lot of stuff that happened here and I think this situation is genuinely rather strange. So of course, today it was announced all over Twitter that essentially individuals from OpenAI's Early Access Research program that were getting access to the new video generation model, sora, essentially leaked access to this model by providing a hugging face link where you could essentially use the model to generate short clips. Now, this was pretty crazy because currently SORA is private use, meaning that it's not for public use just yet, like ChatGPT. But this was something that a lot of people didn't expect because we haven't received any specific date for sora. Now you can see here that on this link where they shared Sora, you can see that they posted a really long paragraph containing a few messages that essentially explain their grievances with OpenAI. Now I'm going to read to you exactly some of their grievances and then I'm going to talk to you guys about what I think the real issue is here. So this is where it starts. It says, dear corporate AI overlords, we've received access to SORA with the promise to be early testers, red teamers and creative partners. However, we believe instead we are being lured into artwashing to tell the world that SORA is a useful tool for artists. And then goes on to say, artists are not your unpaid research and design. We are not your free bug testers. PR puppets, training data validation tokens. Hundreds of artists provide unpaid labor through bug testing, feedback and experimental work for the program for a $150 billion valued company. While hundreds contribute for free, a select few will be chosen through a competition to have their SORA created films screened, offering minimal compensation, which pales in comparison to the substantial PR and marketing OpenAI receives. So essentially the gist of this article here is that these artists are quite upset. Now, I'll dive into why. I don't necessarily agree with everything they're saying, but I'm going to continue reading this so you guys can get an understanding of exactly what was said. So they say here that it says, denormalize billion dollar brands exploiting artists for unpaid research and design and pr. And it says, furthermore, every output needs to be approved by the OpenAI team before sharing. This early access program appears to be less about creative expression and critique and more about PR and advertisement. And it says here that we are releasing this tool to give everyone an opportunity to experiment with what 300 artists were offered a free and unlimited access to this tool. It says we are not against the use of AI technology as a tool for the arts because if they were, they, they wouldn't have been invited to this program. What we don't agree with is how this artist program has been rolled out and how the tool is shaping up ahead of a possible public release. We are sharing this to the world in hopes that OpenAI becomes more open, more artist friendly, and supports the arts beyond PR stunts. And this is where they basically leave a link to some open source tools. They show tools like Cog Video, X Maki One, LTX Video and Pyramid Flow. And they state that some open source tools are available. They allow artists to experiment without gatekeeping commercial interests or serving PR to any corporation. Now, I do want to give my two cents on this, and I'm not sure it's going to be the popular opinion, but I just want to state that why is it that people who are in the early access program are upset that they're in the early access program? I mean, if you've been invited to this early access program, it's quite likely that you weren't forced to do this. So if you weren't exactly happy with this early access program, I'm guessing you could have easily just left and said, you know what, this is not something I'm comfortable doing. And stating that you were lured into artwashing, I'm not sure if that seems entirely genuine now. I can understand that. Look, stating that your artists aren't unpaid labor, like we aren't the guys that, you know, do all the bug testing and stuff like that. I can get that. But part of being an early access program is that you do provide feedback and of course, relevant information regarding the bugs that you do experience. I've been a part of a few early access programs and this is really normal. It might feel like work, but that's what you get in return for being someone who gets access to that content for free. Now, I'm not sure on the entire details of this SORA competition. It says that while hundreds contribute for free, a select few will be chosen through a competition to have their SORA created film screened. Basically, what they're stating here is that we as artists are going to create something amazing and then OpenAI gets to show off the best one. And it puts their company in the best light. I guess I Can kind of understand that, like an artist goes through a lot of hard work and effort and OpenAI gets the marketing efforts because they can state that this was created by Asura. But I also think that if the artist does it, it allows them to get the recognition from working with a big brand such as OpenAI. Now, also them stating that, you know, every output needs to be approved by the OpenAI team before sharing. This actually does make sense if you're going to share something about a company. Companies all the time are really strict about what they share because it's their brand. If you publish something that is coming from a company, of course the brand would need to approve what you share. Considering the fact that this tool isn't available for public release, as if these outputs are pretty bad. We all know that outputs that are bad could definitely damage the reputation of this company because the video generation model is quite fierce. And if people are allowed to just produce absolutely any videos, then you could literally post videos that are quite bad in terms of the physics. And a lot of people might just think that the, the tool isn't that good. Now, when they state that they're releasing the tool for everyone to use, this isn't the kind of release that you do think. First of all, it's not the model weights, because they don't have access to the model weights. It's essentially just the API. So all they did was they essentially just released their API key that OpenAI gave them and then just created a hugging face space that anyone could use. Now, the crazy thing about all of this is that OpenAI is now muting people on the official Discord server. They're talking about the recent Sora leak. But that isn't the only thing about this. I think the really interesting thing about this is the fact that this Sora leak has actually probably done the opposite. Because what I've seen in the entire community is that a lot of people are saying that they're surprised at how good the quality is and how much better it is than anything out there that currently exists. So do you remember earlier this year when you had an OpenAI AMA? Basically, someone asked about Sora and Kevin Will, the OpenAI chief product officer, said that Sora was being held back by the need to perfect the model, get the safety impersonation, other things right, and to scale the compute, which basically means that they were trying to iron out certain issues. Now, if we take a look at some of the examples that were being shared on Twitter, we can see that this one was quite similar. To the original one. But it definitely looks to be a super high quality that is much higher quality than some of the other things I've definitely seen out there. I definitely feel like Sora is probably levels ahead of these other video models, despite the industry competition. Now, there's also some key details about Sora that you do want to know. For example, Arrakis AI shared this really interesting thing where it said that it seems Sora has its own mechanism for scene transitions. When looking at the generate prompt feature, that three additional prompts were generated. Aside from the initial input prompt, we can see here that of course, it initially generated this long prompt about a woman striding confidently down a vibrant Seoul street at night. And then basically the really interesting thing was that you can see how this video progresses. So it says the woman adjusts her sunglasses as she passes under a particularly bright neon sign. And it says she pauses momentarily to look at the reflection of the city lights in the puddles on the street. So it's clear that when they generate these videos, it looks like they're generating additional prompts with that to be able to ensure the video has coherence, which looks like a neat trick. Of course, I'm not entirely sure how it entirely works, but it seems like Sora has its own mechanism for these scene transitions. Now, something else that was spoken about was the fact that what we can see here is that the model that we actually have is a Turbo model and there are different models. So the model examples that you're currently seeing aren't the full version of Sora. They're essentially a lightweight version that is built for speed. So it says here from Tebo Blaho, as the information previously reported, the request payload suggests that there are multiple models mentioning the Turbo model as well as different styles. As you know, the natural style used in the hugging face space. And apparently there's also different kinds of operation modes. For example, simple compose, inpaint items and the API request can influence the number of variants and frames. And of course, the type of video generation could also hint at other capabilities, such as image generation, which was showcased previously. So basically he's referring to this, where you can see that the type here says videogen. So maybe it's going to be image generation that's also released. You can see that we've got a number of variants, we can see number of frames. And of course you can see here that it says Model Turbo. So it means that potentially there's a different model that probably has more coherent and more quality at the cost of Speed. Now, if you remember, it was reported from the information earlier this year that OpenAI is training a new version of Sora that it hopes will quickly generate clips that are higher in quality and longer than the ones it demonstrated earlier this year. This was from the Information, because of course, they were basically having issues because Sora initially took more than 10 minutes to generate a short clip, about a minute or so. So it seems that this was one of the early errors that they were trying to prevent with Sora, because of course, they were trying to ensure that they could make sure that these issues were ironed out beforehand. Now, they're also talking about the early errors here, where Sora had a hard time maintaining the same style over the duration of the clip, the person who saw a demo said. And it was even harder to make sure that the objects and characters remain consistent across different clips they generated using Sora, Cedarberg said in the interview. Which is pretty normal considering the fact that it's generative AI. Now, when we actually do take a look at some of these clips here, you can see we managed to get an AI generated Fortnite clip, which looks super interesting. I mean, I've played Fortnite quite a few times, and this thing is definitely pretty good when it comes to the kind of level of consistency you get from the character running around to the kind of character movements that we get here. Additionally, we did get a Minecraft gameplay, and we do know that this is, of course OpenAI Sora, because in the bottom hand, right, you can see that it does have the OpenAI logo. So the level of consistency there is really, really fascinating. I'm not going to lie, because it shows us something that looks really, really consistent. And I don't know of any other model that has this kind of world model that's able to correctly predict the next frame in a way that this model does, because this world model here looks to be really consistent. And even when I watched AI generated Minecraft the other day, it seemed completely different compared to, you know, every other scene. And of course we have another scene here of a horse outside the Taj Mahal with a random man just spawning in right there. But overall, let me know what you guys think about this crazy event. I think this is not that big of a deal for OpenAI, someone sharing their research preview access by giving out their API key. I think it's a little bit interesting because you could have just left the program stating that, look, I don't want to do this anymore, thanks for inviting me, and I'm sure there are other artists out there that would have loved to have access to this tool.