AI is NOT Slowing Down! Open AI Voice Release Date, and Cutting-Edge AI Innovations

Welcome everybody. You thought AI was slowing down? You thought we were hitting some sort of a ceiling in regards to AI technology? Well, think again, because AI is on a roll. I've got so much stuff to talk about today. So buckle up folks, because we're about to take a peek past the curtain into the future. How about we start this off right and talk about the big guy in the room. OpenAI. As some of you might remember, OpenAI continuously promised that we would get the insanely capable GPT4 omni voice mode in ChatGPT soon and soon turned into sometime this fall, but it looks like now we might actually get some form of a release date for this. As you can see, this is brought to us by TBOR on Twitter whom you should follow fantastic AI news updates account, especially for OpenAI stuff. Advanced voice Mode might get a broader rollout in September, specifically on the 24th. As testing news catalog points out here, these dates are tentative, they could change. But specifically we're looking at this image right here and this can actually be found if you inspect the web page for chat GPT. As you can see, eligible announcements OpenAI has seen advanced voice 2024, 924 and I mean guys, that's only in four days. Is this the beginning of a larger scale rollout? Is it when we all will suddenly get access to Advanced Voice mode? Only time. But it's a pretty good hint that Advanced Voice Mode is going to be coming to regular Chat GPT users fairly soon. And if you're wondering, oh yes, I did check my own chat GPT and there it is. Number 14 in the list has seen advanced voice 924. So if it does release, even if it's a limited rollout, if it's for the 24th, I might get access, which is good. I would definitely be making a video right away. The crazy part is at some point soon we might actually not even need the new voice mode because Moshi, which is very very similar to the GPT4O model that powers Advanced Voice, except it's just a lot worse. It's a smaller model and not as well trained. Moshi just became open source, which means anyone can edit it, anyone can modify it, and of course anyone can actually improve it. So maybe if the open source Moshi Voice mode becomes improved by the Open source community, we might not need advanced voice from OpenAI after all. Here's a little demo of what Moshi sounds like. Again, this thing is not nearly as good as advanced voice from OpenAI, but it's open source.

Hello what's going on? Yeah, I do. There are so many good ones out there. Well, one of my favorites is Perry Mason. Yeah, it's an American legal drama series. It originally aired on CBS from 1957 to 1966. Yes, it was. And it became one of the longest running and most successful legal themed television series. It was the first weekly one hour series that was filmed for television. Yeah, I'm not sure.

So there's a little demo for you. Obviously it's not at that same caliber, it's not at that same echelon, but at least the creators were open about it, unlike OpenAI. In other OpenAI news, and this actually dates all the way Back to the 7th of this month, it seems like real artists. Yes. Not big corporations. Real artists are starting to actually get some sort of limited SORA access. Obviously this is like a super small test group, but you can see here nice aunties made aunties eggs, which is this incredibly weird, like little short film. But hey, they're a real artist, not some big company and they got access to Sora. So maybe SORA access is actually going to be doled out at some point in the somewhat near future. A lot of us though, including myself, are like, does it really matter at this point? Because we have a lot of really good video models that in my opinion are either as good, if not a little bit better than sora. Competition in the AI world is fierce. But yeah, if you want to see a real artist using Sora to make the weirdest short film you've ever seen, then check this out. Now, before we dive into the rest of today's AI news video, I have a quick word from today's sponsor. You know how we're always talking about reinvention in the face of new AI technology? Well, HubSpot is certainly taking that concept and applying it to their own platform. First up, we got to talk about Breeze AI. It's a toolkit built right into HubSpot with AI capabilities. It helps by taking out those basic menial tasks, emails and blog posts, all the way up to marketing campaigns. And we've certainly seen technology like this, but it's built right into HubSpot, which is nice. This is a literal AI agent built right into HubSpot. Need to analyze customer data. Breeze Copilot's got you trying to optimize your content for SEO. It's already on it. The best part is that these new features aren't limited to just large businesses. These will be available for businesses of any size. And trust me, early AI Adoption in a small business is probably the most potent thing you can do to help grow your business right now. And if you're not sure if it's right for your business, they do have a free free demo which gives you the chance to experience firsthand how strapping AI to your business operations can be like a supercharger for it. Trust me, if you run a business of any size, this is a huge opportunity to stay ahead of the curve. Check out the link in the description below to sign up for a free demo. Huge thanks to HubSpot for sponsoring today's video. Now back to your regularly scheduled content. Welcome back folks. Thanks for sticking through sponsors. They are literally the reason that I can pursue this YouTube channel full time. Taking a little break from the OpenAI stuff, let's talk about cling AI, one of the leading AI video generators. They've just released a ton of crazy new features and they look really cool. First of all, they have a brand new model which is cling 1.5. It natively generates 1080p full HD videos. Now, the video generation still take absolutely ages, but the quality is definitely there. And they've also got a new motion brush feature which is only for the Kling 1.0 model that greatly enhances the amount of control you have over the end shot. Here's the first video example that I want to show you. This is 1.5 output and I don't think the Twitter video compression is doing us any favors here, but you can see the quality is definitely pretty high. It's 1080p. There's a lot of like again, mushy quality issues because of Twitter, but if you look closely, there's plenty of detail. It looks like real footage from a video game almost. I mean, there's still some weird AI artifacts that are flying in the air, but overall it is really not that bad at all. It's. It's mighty impressive. This is linked down below. This is a nice thread put together by Min Choi on Twitter. Again, more 1.5 examples we can see in this example we have strong emotions being replicated by cling. I think it's overdoing or over burning the details in the skin on the face just a little bit. But other than that it actually looks pretty darn good. Not too much mushiness, not too much weird AI morphing. And yeah, I mean the emotions are expressed, expressed in pretty high quality. The emotion we're seeing here is sadness, which is obviously not everyone's favorite, but it's definitely being expressed It's a good test and demonstration. On the reverse end we do have a mother and her new infant baby. I think the infant baby again, it looks a little bit crispy, looks a little bit on the burnt side, but the mother I think looks pretty stellar. Overall, I think they got the details down correctly. It's impressive how far AI video has come in such a short amount of time and especially how high quality it's able to do humans at this point. Cling AI has been known from the start to be a little bit more biased towards doing good humans, and it absolutely shows in the new 1.5 update. This is probably one of my favorite examples to come out of the new 1.5 model. Again, this shows the amount of detail, the amount of clarity you can get with the newer, higher resolution generations. And the fire looks pretty decent. The people definitely look good. Maybe a little bit more realistic movement would serve this one well. But you can see the water bottle is nice and stable in the corner there, which I like to see. And the overall quality of the camera movement is second to none. Cars for AI video generators have classically always been pretty difficult. And you can see still difficult for Kling 1.5, but it's definitely getting a little bit better. The wheels, they're rotating somewhat nicely. Still a little bit weird though. Still not perfect. But the car overall isn't really morphing too much and it's maintaining a decent, decent enough shape. Finally, what Cling sort of became popular for eating food is even better with 1.5. I think that they have maintained all of the great things about the original model and just sort of increased the quality. This to me looks like some image to video with cling 1.5 and you can see nice sharp details, a lot of texture on the face of the wolf man person thing and the firearm honestly is staying pretty consistent. That's something that again is kind of classically difficult for these AI video generators to get right. And it looks pretty good in this scenario. Those finer details, even with 1080p are still quite difficult though, as you can see up here with the zipper and all of the little different metal pieces of the zipper morphing and melting around, not really a big deal. This is very much watchable in my mind, but those minute tiny details are still difficult. Anyways, next up we have to talk about Motion brush, which is arguably a little bit more exciting as you can see. How this actually works in the Kling app is really, really, really cool and I tested it myself. It is not bad. Basically you highlight with your mouse cursor the different areas that you'd like to segment and select in your input image. For this example, the user selects obviously the main character who's towards the center of the screen, and then you draw where you want the motion to go and then prompt and generate. So obviously we want this guy to stand up and walk towards that side. And the motion brush absolutely is able to handle that. The walking looks really good. He's kind of morphing quite a lot when he stands up. I think you'll notice that if I play the video here, like, definitely not all the way perfect, but the walking looks good overall, and it's definitely adhering to the motion. Exactly what we were looking for, right. While keeping the rest of the background stable. Even if it starts to look a little bit jumbled and messy again, video generator models are still very much new. Cling is like, what, a month and a half, two months old? Something like that. Here is another really cool demonstration again of that cling motion effect, where we can segment just this cat, and then we can draw that we want him to jump on the table, and then he goes ahead and jumps on the table in a very realistic way. I'm actually surprised by how well that cat came out, because I've tried similar tests with other video generators before, and it's definitely not the easiest one. But of course, you could segment anything. You could segment, you know, six different things and try to make it move all those at once. The clearer and more well defined, the better. But overall, this is really great for generalized control. With AI video, it's a pretty big step forward. So next up, guys. I just have to talk to you about Gen3 video to video. This is absolutely wild. I plan on doing a full video testing this one out. I've been having a lot of fun with it. It essentially lets you upload a video to the Runway ML Gen 3 video generator, and it uses that video as a reference and then generates you an outp. And you can do some crazy stuff with it. For this example, you know, you have lots of really bad special effects, not fully rendered, or, you know, just like basic iPhone shots that were taken inside of a house. And then you can actually convert it to something that looks pretty good. It looks pretty similar to a movie. Obviously, you can tell it's AI still, but it's really not that bad. And the fact that you can actually film yourself or film other people or film whatever, you can actually honestly make pretty consistent characters with it as well. Again, all of my Examples are linked below. So if you want to check like this one out, highly recommend it again. John Finger is a great account to follow here on X Twitter. Here is some more examples. A lot of people have been doing this where they take some video game footage and then convert it into much more realistic looking graphics by using Gen3 video to video. It's really crazy how well Gen3 is able to understand things like lighting and how lights should bounce and shading and all this other stuff. It's something that video game developers have obviously been struggling with for a long time. To actually recreate in the computer, you need a lot of processing powers to do that. But by using AI, you can just guess and it ends up looking pretty good, especially as this technology continues to advance. I mean, these people look very realistic in comparison to the older video game counterparts. And it's crazy because it's almost like remastering or revisualizing these old video games as if they were real movies. So that's even like a touch beyond modern day video game graphics. It's pretty crazy. I am like a massive fan of this kind of stuff. I think it is so darn cool. Honestly. I think that AI is the future of video game graphics because if you can start doing this stuff in real time, imagine what your video games could look like in the future. And it applies obviously to filmmaking. It applies in all sorts of fields, not just for video games. But I think the video game is an especially telling example of the use of Gen3 video to video. Like I said, I've been messing with it myself. So you can see this extremely creepy video I made where I'm literally just sitting here with my webcam and leaning in and staring at the camera being all creepy. I did this super like weird alien tweet, but I also just like opened up paint and started drawing lines. And you can see it does cool stuff with that too. Like in this example, I was really impressed by its ability to create, create realistic smoke physics. And it understands that the smoke is going to have some sort of energy as it moves and swipes across the screen so it flies upwards in the correct direction. It's. It's just really impressive. I also lit myself on fire, which was pretty fun. I noticed that when I got close to the camera, it understood that there's going to be some background blur and added that into the video, as you can expect. And it overall just creates something that is very, very realistic, even though at the same time, you know, that couldn't possibly be real. Here's a little test I also did with some gross green slime to again kind of test the understanding of physics that gen 3 video to video possesses. And I think it did an overall decent job at this. I still think it's a cool effect. Like if you were to try to recreate this in some 3D animation software, it would be quite difficult. It would be quite difficult. It's a miracle that we can even do this. Here are some more video game examples. I thought this one was super cool. We have the Super Mario Bros. 3 opening main screen and you can see that classic animation with Mario and they're throwing the shells and stuff and it tries to recreate it in like this painterly craft art style thing. I think it looks really cool. Again, it's just such a cool demonstration. Very, very stable. Honestly, this is a lot more stable than Gen 2 was with video to video. And it's. It's just way more usable overall. I have a plethora of examples that I'm going to link down below for Gen3. Again, I plan on doing like a full Deep Dive breakdown video on this thing, so expect another video on Gen 3 in the future. But yeah, all of these different examples will be linked down below for you guys. All right guys, let's bring it around and start talking about OpenAI again. So even though I did an initial reactions video and I actually plan on doing a Deep dive video into OpenAI01, I still want to talk about it in this video just to demonstrate some of the capabilities and some of the things that the community has gotten up to with OpenAI's new model. First up, OpenAI 01 just straight up built a fully functional chess game that allows the player to compete against an AI based opponent. It's just insane. This model is so powerful, it's so capable. I've actually really been struggling with my script for the deep dive on 01 because I'm having trouble putting my thoughts into words about this model. I think a lot of people are still overlooking it because it's just so smart at this point that we're not smart enough to run tests on it ourselves. It does math better than you and me. Does a lot of things honestly better than you and me. But yeah, this is just an insane example a fully working chess game. I don't even know how that's possible. Clearly this guy probably has some coding experience which helped him prompt it correctly to get an output like this, because I did try some coding myself. But yeah, this is extremely impressive. I also wanted to Emphasize some notes that Dr. Jim Fan from Nvidia made about this new model. He notes, people have been predicting stagnation in LLM capability by extrapolating the training scaling law, but they didn't foresee that inference scaling is what truly beats the diminishing return. We have absolutely turned the page in this regard. Inference scaling is apparently all we need. Here's a nice little graph that Jim Fan made to kind of show this off. Most LLMs have a lot of pre training and that's where a lot of the value comes from. Post training to get it ready for chat, etc. Weed out some bugs and then finally inference is a small portion of that value. But with new 01, aka Strawberry, the pre training and the post training don't matter as much as the inferencing because it's literally doing chain of thought and solving problems to become better in real time during inference, which is pretty magical. It's already so much better and again it can be scaled higher. Inference scaling, baby. I am ready for it. I also wanted to show off this roundup that OpenAI actually did themselves some of their favorite posts and examples about OpenAI01 selected by researchers who worked on the model. So this first Example Here is Dr. Kyle's example of O1 accomplishing PhD level code. And I'm not gonna sit here and expect to understand how all of this PhD level coding works, but according to the doctorate himself, this is PhD level code that O1 generated. I mean, it's pretty dang impressive. It looks impressive to me. This is why I say like this model is smarter than me. How am I supposed to test this? IT PhD students can test it pretty well, but me, I'm just a regular guy. This is so far beyond me. Here we have a real MD medical scientist exclaim that OpenAI01 wrote a major cancer treatment project based on a very specific immunological approach, created the full framework of the project in under a minute with highly creative aims, approaches and even considerations for potential pitfalls. And he notes that this is usually missing in 4:00'. Clock. But of course with the chain of thought in the actual inference itself, it was able to pick up on it. This would have taken me days, he says, if not longer to prepare, and he still would have missed points in comparison to 01. This model is no joke. Here we have an example of 01 creating a 3D version of Snake in under a minute. And that doesn't surprise me at all. It's kind of like this coding wizard. You'll See Matthew Berman, fellow AI YouTuber is also noted in here, where the thing thinks for 10 seconds and is able to actually count out and tell him exactly how many words are in its own response to the prompt. So it's generating the words as it's talking to you and it figures out that it needs to come up with a sentence that has the same amount of words and then it does it. And finally here, OpenAI actually mentioned my example, which I thought was really cool. I shared this on Twitter. It's a nice clean physics environment in Python that OpenAI wrote me in, you know, maybe like a minute. This is a pretty cool example. What my tweet there didn't mention is that it also taught me how to set up versus Code Studio for this whole deal. Making the virtual environment, putting the code in, getting python set up, etc. Again, I'm not a coder, I don't have much experience, but it was definitely able to get the job done. And it's a really impressive little physics demo, I have to say. It's pretty smooth, it feels great. And as you can see, we even added the ability to add PNGs as physics objects, which I thought was pretty cool, even though they explode upon impact. Now, honestly, guys, the cherry on top of all of this is that the O1 preview model is still in the oven. It's not out yet, it's not fully done. There's two O1 models, O1 Mini and Preview. O1 Mini is a better model. Right now, it's more fully fledged, but Preview eventually will be better once it's done cooking in the oven. And it's crazy to think that this thing is already like echelons above the other LLMs just because they added Chain of thought. And it's not even done. It's not even done. Just a preview. All right, I think that's all the OpenAI stuff I have for this video. This right here, folks, is a paper called Omnigen. This is Unified Image Generation. I might do a full video on this. They plan to open source it, but right now it's just a paper. Not only can this model generate images, like in this example, we've got a cat holding a sign that says Omnigen. Again, pretty impressive image gen. This is a pretty advanced scene as well. Someone walking on like a lily pad wearing a straw. So yeah, it does good image generation, but it's also able to do things like reasoning, like you can send it a photo and say, where can I wash my hands? And it's able to segment out the sink and be like, yeah, that's where you're gonna wash hands. Or you could send it a pretty basic photo and say what can be used to hold water? Remove it and it can actually edit out in paint out the little cup that's in the corner here. It's really hard to see, but yeah, there's a little cup down there and then it's edited out in the next image. So it can do real image editing as well, which is crazy as well as just reason about the images themselves. You can even do subject driven stuff as well. Like we upload this weird furry guy here and then Iron man and then we can, you know, have them in the same image together cooking up some different food items. It can also do in context learning. So according to the following examples, it generates an output for the input, it's able to segment and identify different portions, remove and then regenerate. There's a lot of other things it can do too, like deblurring an image, changing hair color, adjusting characters, detecting the skeletons of a human. It's pretty crazy. So this is just like a little sneak peek. Again, this isn't out for anyone yet. It's just a paper, but they do apparently plan to open source it. So if it does go open source, expect a full video on this one. So next up, Guardrails AI is showing off a new blog post by the Bespoke AI team about this new model to essentially detect hallucinations or lies, basically Fact checking other LLMs. This is an approach that I never would have thought to take, but it seems to be working decently for them where you can upload some facts, background context plus a claim. And then it says that basically there's an 80 something percent chance that this specific claim is supported by the context it goes through in fact checks models. I don't know how well it would work for more nuanced stuff, but at least like a historical claim like this should be pretty decent. Very interesting. Next up, Crea AI, known for their pretty awesome image generation platform, has now apparently tuned the open source Flux model to work in real time. So that means you can do all of the awesome flux things but in real time, such as dragging a sign with correct spelling all around the frame. It's really cool. I tested it out myself. I don't think it's worth me doing a full video to talk about it, but I definitely thought I should mention it in this one. You guys should check it out if you're into real time AI image generation and really manipulating your models. This is definitely one of the better ways to do so. It's actually shocking how good this real time flux model is at understanding your prompts and the image you're uploading into it in relation to that prompt. It's also just free for everyone apparently, which is pretty cool. Have at the. I suppose, right. And finally we have some video game related AI stuff. So this is game Gen O. This is open world Video Game generation. It is a diffusion transformer model that's specifically tailored to generate the environments of open world video games. And it's essentially generating images for you that look like video games and you could theoretically control it with a controller. Again, this is kind of the next step in evolution for AI in relation to video game graphics where it's doing its best to just guess what the video game is supposed to look like. And it honestly does a pretty good job. I mean just by looking at the various pieces of footage you can tell what a lot of these games are supposed to be like. This is obviously Witcher 3. This right here is going to be Red Dead 2 and the Walker animation. Like I gotta say, I've played Red Dead 2. That looks pretty good. And the trees definitely look like that in Red Dead 2. You know, riding the horse around in Witcher 3. And then we have. I think this is supposed to maybe be GTA 5. We've got more Red Dead. It's impressive. Apparently they're going to do a hugging face spaces for this soon. I would love to try it out. But yeah, this is really interesting to me. This could have potentially enable you to play lots of video games anywhere because you're just doing server AI processing and then just sending that to device. Anyways guys. Man, that was quite a lot. Like I said, AI is not slowing down at all. There is plenty going on, there's plenty to talk about. This should be a pretty good recap for you guys. I probably missed some things, some little nuanced things that are pretty cool. So join my Discord server if you are interested in diving a little bit deeper. The latest and greatest is constantly always posted on that server. Also follow me on Twitter because I repost things exactly when I see them. If it's like a big news announcement or something cool like you saw today today. Thank you so much for watching today's video. That is it for my news recap and I will see you guys in the next video. Goodbye.