AI is Shifting Gears! Exploring GPT‑5, Grok 3 & Open‑Source Innovations

Hey everybody. Welcome back to another Matved Pro AI news roundup. First off, AI text to video generators are still getting better. New ones are still releasing and being open sourced. So AI creativity is becoming more and more democratized. Sam Altman has revealed the roadmap for future GPT series models like GPT 4.5 and 5. There's quite a few comments about them. I thought we were done with the GPT series of models, but apparently not. Seems overall the giants in the AI space are shifting straddles strategy as new challengers enter the arena of AI competition. And finally, guys, Elon Musk is claiming he's built the smartest and most capable AI in the world. And that of course comes from his company Xai. And Grok3 is the model. And actually Grok3 apparently is being demoed live later today, so maybe expect a Grock 3 video tomorrow. Without further ado, let's take a look at our first piece. First up, brand new open source AI video generator. Another Chinese one from Step Fun AI. Never heard of it, but this is Step Video AI. They've got a demo website here that I'm going to show off to you guys so you can get a good idea of how this video generator operates and what the quality levels you can expect are from it. It's pretty darn good, especially for an open source model. This is competitive little open source pocket Hollywood. You can think of little AI generating cinema quality videos in a lot of circumstances. Seems to be really great at doing things that are also realistic. Like that looks like a TV broadcast for example, or you know, live video footage of a running sport event. We can also see this thing has the ability to do text as well. Year of the snake 2025. And a little animated snake comes across the screen. Pretty cool. So it can do abstract stuff as well. Here's a kid in a superhero costume. Here's a kid apparently in the middle of a galaxy wearing a crown. Oh, he's, he's turning into a princess apparently. If you ever wanted to see a Lego Harry Potter movie, maybe this is the AI for you. Even Steve Jobs is making a little cameo here, but it also shows in the background. Again, it can do text. Step Video is coming. We've also got firearms in here. We've got a skiing pov, you know, all kinds of really cool little demos. It's definitely a pretty diverse model and they're here to compete. I think that's what this demo website is really trying to convey. Not now. As you also can see, there is a Description and a generate button on this so we can try it out for ourselves. What are these inspirations? You know, we got something normal. Train passing through golden wheat fields. Fireworks bloom in the night sky. Zhang Yu committed. I can't say that word on YouTube. Tragically by the Wu River. What? Why is this a recommendation? What? Did the AIs come up with this recommendation or did someone punch that in themselves? This is not giving me good inspiration, just to let you guys know. Anyways, let's, let's try, you know, Hot air balloon rising. It does seem to put the prompt immediately in Chinese, so this might not accept English prompts just yet, but you could use ChatGPT to translate them quite easily. Let's see if it'll. No, you need to log in. Wait, can I sign in with a US phone number? Okay, even though you can select like USA in this, it just keeps it at plus 86, which. Which is China. So I assume you need a Chinese phone number to actually use it on this website. But as I said, it's open source, so that's not stopping you from running yourself. So yeah, you'll be happy to see that this thing is right on. Hugging face. Step. Fun. AI step. Video. Text to video literally. Released today, February 17th. And you can download the model weights and the inference code. So it's. It's fully open source, ready to roll. I would love to see this on Pinocchio pretty soon. So Cocktail Peanut, could you get on that? They've also got a technical report as well, but the basics for you guys, it's a 30 billion parameter model. It can generate up to 8 seconds video right now, which isn't super awesome. Love to see up to 20 seconds. I think that's where the industry should be aiming for right now. Still though, it's open source and it's usable and it could be modified by the community because it's open source to generate longer videos in the future. The compression here is also pretty impressive with a 16 by 16 spatial compression ratio and 8x temporal compression ratio. They also applied direct preference optimization to further enhance the visual quality overall. And according to their own benchmarks, it's state of the art text to video quality compared to both open and closed source models. Since this just released today, the community hasn't really got their hands on it yet to make it ready for commercial hardware. Even an RTX 5090 isn't going to be enough. So while it is state of the art, it still uses a whopping 80 gigabytes of VRAM for the Best generation quality. That's the recommendation. So yeah, open source community, could you guys get on this and help us achieve this on consumer grade hardware? So if you're a developer, definitely get on this one because the rest of the community would appreciate it if we could actually run this model and get really high quality video generation like this that isn't messing around on consumer grade hardware a little bit older. On February 11th, magic one for one was also announced. And apparently this model can generate one minute long video clips within one minute. So essentially in one second it's supposed to be able to generate a second worth of video, hence the name. Magic One For One seems like a really cool idea. And this one also seems to be leaning towards open as well. Even the test clips that we're taking a look at are decently impressive. We've got a dragon blowing fire, the tomato cutting test classically, which it does a pretty good job on, some typing, some animated graphics around. A documentary seems pretty solid. Now this one does indeed have a paper and Connor here in the comments says they released the model weights, but if you actually take a trip over to the GitHub, you'll see that they're still currently working on some clearances to get that additional information out. For now, it's just a paper and this little project page on GitHub. Once they get their clearances, it looks like this should work on a single gpu, which would be pretty cool. Also has the capability to be quantized as well to optimize for model performance and use less memory. If you're trying to run, you know, a business or a research lab or something, they do have a multi GPU interface available. So this looks like it could shape up to be highly efficient and really cool. Again, these open source video generators are not messing around. They really want to put Hollywood right in your hands. Development rate is fast and these models seem to be going open source left and right. We're still waiting on this one, but they aren't the only one in the pack to be used inside of the app. Quickly. Now, the big question at hand here. Will the open source AI video generators eventually catch up to the closed source ones? Right now, as we speak, closed source still holds the Crown and it's VO2. Check out this prompt right here. I mean this is just stellar. We've got, you know, old fashioned racing scenes being projected onto a McLaren. I mean that is an insanely specific prompt and really difficult for an AI model to understand and reproduce effectively. But Google VO2 nails it with that YouTube training data and VO2 is not out for the public to just use, it's only for select few testers right now. However, Google VO2 just became available inside of the YouTube app. Now it's not the big bad Google VO2 that produced this stunning example right here as Billowal mentions and you should follow him by the way. This is likely just a distilled more optimized smaller model used for the YouTube Shorts app. It seems like they integrated it pretty effectively overall though, you can use VO2 as a green screen background or just a standalone clip. Has some basic presets, vintage anime, clay, etc. What's also really fascinating is that it does text to image first and then you pick whatever image you want then it translates that into a video. I assume they're doing it this way to give you a few different options instead of generating four videos right off the bat. So it saves a lot on compute doing it this way, generating images first and then converting those into videos. A more cost effective way to get the user closer to what they're looking for. Looks like you can do multiple clips in the same exact video as well. Only for us, Canada, Australia, New Zealand users at the moment, so pretty limited. Sorry all you folks out in the UK and other countries. No access as of right now and this little clip Bilawal shared though, you can see him messing around with this thing. Chihuahua in New York gives you a couple options, you pick one and then converts it pretty quickly into a video and there you go. Now you got your chihuahua walking around New York. It looks alright. I do have to wonder though, do you think the reason that Google VO2 is being put inside of the YouTube app as an option is to compete with open source generators from the likes of all of these Chinese companies? Or do you think it has more to do with TikTok also having AI features in an attempt to compete with them? I don't know. Let me know what your thoughts are in the comments. Also wanted to share a few more really cool little AI video related tidbits from the community. This one comes from Aberman apparently imposing a spatiotemporal weight space through Loras on dit based video models can unlock very powerful customization. It's able to capture dynamic concepts with good precision and even enable composition of multiple videos together, which is what's demonstrated right here. We give the AI two separate videos and through spatiotemporal weight via Loras we can combine them together and get a singular output that represents both. And it does so pretty coherently maintaining a lot of the details we see in the background. I wonder if this is how the likes of Cling AI or VDU AI are doing their multiple elements in the same scene. I wonder if it's similar to this method. Regardless, Aberman says that they're going to be sharing more results in a technical paper soon, so this is pretty hype. I'm excited for it. Open source, highly customizable AI video Again, think Hollywood in everyone's pockets. So crazy. Also, want to share this example from Chubby stating it's now possible to create an anime almost as good as the ones you'd find on a platform like Netflix using Luma's Rey 2 and Midjourney in combination. So I assume we're generating the anime images in Mid Journey and then pumping them right into Lumo Rey 2. It does a good job. These are good examples. I think you could share these images with someone who maybe doesn't have as trained of an eye as me or you and they might think it's a real anime at first glance. A lot of the scenes look great, Definitely hammers home the art style, the color palette and certain effects like bloom or lightning. It's pretty stellar. I think anime might actually be easier to replicate than maybe some of the other things we've been trying to replicate with AI video in the past. But think how many millions of people who are very creative have all kinds of cool ideas for their own stories to tell. This technology is democratizing that for them. You no longer need to have oodles of money and a whole team to create your own anime. Think about that could all just be done by one person. Just absolutely nuts. Alright, moving on guys. Let's talk about the roadmap that Sam Altman dumped on us for both GPT 4.5 and GPT 5. I thought we were done with all these models, but apparently Samus had a change of heart. Apparently he's channeled his inner Steve Jobs and OpenAI needs to make AI just work for consumers. We don't need 10 different models in the model picker inside of Chat GPT. We need just a few. Maybe the product lineup for OpenAI has just got way too complicated. We have GPT4O, we have O3 mini, we have O3 mini high we have 01, we have O1 pro. And don't forget about O1 mini and the legacy models like GPT 3.5 or GPT 4. Sam Altman basically said enough is enough. We have to simplify all of this and I agree Honestly, I think a lot of people will miss the older style of things just because that's how they're used to doing it. But if they can create an all in one model to essentially choose the right tools to use to get you whatever you're looking for and it works, that's the way to go. The next model that they're going to ship is going to be GPT 4.5. They call it Orion internally and it's going to be their very last non chain of thought model. So they're still actually going to be releasing a non chain of thought model. I don't know why they're planning on releasing it if it is going to be non chain of thought. Not sure what benefits GPT 4.5 is going to have over something like 4.0 or even 03 mini or these chain of thought models. I guess we'll have to wait. But he touts more about GPT 4.5 later, which we're going to look at, which makes me even more intrigued. Basically, at the end of the day, GPT5 is going to be a system, not necessarily one model that is integrating a lot of their technology, including stuff like O3. So O3 will no longer be available as a standalone model inside of chat GPT you'll just have GPT5, which is a system that chooses the right model for whatever task you are asking for. Even in the API though Sam, that's what my question would be. Sometimes some products you might build just need a large language model and not a system that could access these other tools. So I'm hoping that OpenAI doesn't just leave us hanging and never release O3 through the API, I feel like some people would be a little bit upset by that. Apparently. And this is big words coming from Sam, the free tier of ChatGPT will get unlimited chat access to GPT5 at standard intelligence setting. What does that mean? I feel like this is almost a way though for them to limit free users in a way that they can't exactly see very clearly. Oh yeah, this is GPT5 at standard intelligence setting, but it doesn't have access to O3 at all under the free tier. But how would we know that? Obviously us power users are probably already paying for plus at the very least, but we want to know what's going on and what the model is doing under the hood. I get that you want it to just be like magic and that might very well be useful for the free tier. But if the free tier is still barred entirely from getting access to certain models. The standard intelligence setting might not be highly regarded for GPT5, so you could be looking at some potential perspective issues from the market here. Plus, users will be able to get GPT5 at a higher level of intelligence again, whatever that means. And Pro subscribers will be able to run GPT5 at even higher level of intelligence again. What does this exactly mean? I feel like you guys still need to be very transparent, especially with the name OpenAI as your as your business name. Chubby replies here. Any eta for GPT 4.5? GPT 5 weeks months, he replies. Weeks, months, which doesn't help us at all. That means maybe we're going to see GPT 4.5 in a few weeks, hopefully not months. This answer basically to me just means probably sometimes this year. Whatever Sam, keep your secrets. At the end of the day, this is a huge shift in strategy for OpenAI. They're basically betting that they're going to be able to build something that can do it all and actually satisfy customers who are very picky. Mind you, I get that the one size fits all is supposed to make life easier, but I am worried it could become too monolithic. I don't want it to become harder to understand and I don't want it to become harder to control with less information about what model it's using or what it's actually pinging to try to get the work done. Going back and forth with the AI in an attempt to fix issues could become a roadblock. So I am really excited, but I'm also a little bit nervous. Another OpenAI related news they did release a small update to GPT4O inside of chat GPT, which they do quite frequently, trying to consistently stay for the price above the other competitors in terms of benchmarks. But this update also has a couple of other differences. The personality of Chat GPT has actually changed a little bit in my personal testing. It uses more emojis and feels a little bit more casual, a little bit more friendly. He doesn't really give us much information though. He says it's pretty good. Okay, Sam could have given us more information also saying it's soon going to get much better. The team is apparently cooking, so get excited I suppose. Again, GPT 4.5? Is that what he's kind of talking about, cooking? He goes back and forth with a few users in the comments here too, stating it's the best search product on the web. Check it out, let me know what you think. And that's in response to the CEO of Perplexity AI, which in my mind might be the best search prod product on the web right now. I don't know if ChatGPT can beat it out yet. So a little bit of Sam Sassiness right here. Also responding to McKay Wrigley here teasing that 03 pro level intelligence is on its way and apparently even better personality to go along with it. Also, finally, another tease, another little jab for us about GPT 4.5. It's much more of a feel the AGI moment among high taste testers than he expected. Also confirmed that 4.5 is not getting voice mode. OpenAI why do we do these things with our models? Understand the point of GPT 4.5. It's going to be a model that is not chain of thought, does not have access to voice mode. Hopefully it still has image recognition. But he's also touting it, saying it's a feel the AGI moment among high taste testers. How it's so mysterious. Just an interesting direction in general, but I don't think he'd be posting this if there wasn't, you know, a cat in the bag, so to speak, or something up his sleeve. So maybe it'll just be a really intelligent model. But as we know, OpenAI is not the only game in town. One of their co founders, Elon Musk, who is no longer with the company, is planning to launch a bold challenger tonight. Grok3 again touted as the most intelligent AI ever. So let's see what's what. Let's dive into this. So let's hear from Elon's own words exactly how he feels about Grok3.

Grok3 has very, very powerful reasoning capabilities. So in the tests that we've done thus far, Grok3 is outperforming anything that's been released that we're aware of.

So that includes O3 mini high then, I assume. But is it going to be a chain of thought model like O3 mini high still has yet to be known.

That's. That's a, that's a good sign. Yeah, it's in fact at times I think Grok 3, it's kind of scary smart. You're like, wow, this thing's smart. It's kind of scary. Graph three is scary. It's like, wow, this thing's, you know, it comes up with solutions that you didn't even think were like, you wouldn't even anticipate, you know, not obvious solutions.

He's saying that it's coming up with novel solutions, things you wouldn't anticipate. I also see a little bit of that from models like O3 Mini High, for example, or even GPT4 was able to come up with things that I hadn't anticipated, especially with the right prompt. So I'm wondering if this is more of just his visceral reaction with the way that he's specifically using it. It will be interesting to see whether or not the community has the same reaction to elon here about Grok 3.

He was trained with the most amount of compute and I think very efficiently trained. Also notably Grok 3 was trained on a lot of synthetic data, so it's.

As to be expected, honestly.

And then it goes back and forth through the data and tries to achieve logical consistency. So if it's got data that is wrong, it'll actually reflect upon that and remove the data that is wrong. It does not concord with reality, so its base reasoning is very good. In fact, the even without fine tuning Grock 3 the base model is better than Grok 2 so that's good. That's good to hear the final stages.

Of we can't read around the bush though. This is a super bold claim, but it's Elon so it's not the first time he said things like this and then the community had a different perspective on it. I mean he's definitely got a flair for drama, calling it the smartest AI on earth before anyone's even really got to see it from the public. So it's a classic Elon hype building move. But I'll be really interested to see if Grock3 can actually back it up. Apparently we've got a live demo coming today at 8pm PT I believe. For me that's going to be like 11pm so pretty late. But I will be tweeting and probably in my discord talking about it live, so be on the lookout for that if you want to hear my first impressions and reactions to this. Now I'll go out on a limb here and make some predictions surrounding Gray Grok 3. I think it will be a really solid model. I think it'll have all the same flair that we loved with Grok 2 as well. The ability to be more uncensored than some other models and also perhaps be a little bit more comedic than other models. So there's no doubt that it'll be fun to mess around with and use. But in terms of benchmarks and raw coherency, raw intelligence, I think it's going to be really hard to top the latest models that have come out Models like, like Deep Seq, R1 models like O3 High, you know these, these recent models, they're actually using the compute time during inference to reason and then provide you an answer which is giving them a huge leg up. And especially in combination with tools like Deep Research, the intelligence level is staggering. I mean we're talking 10 page PhD level research papers flying out of deep research using O3. So that, that's a really tall order for Grok to actually come up and take down being the most intelligent AI in the world. It might turn out that Grok3 is a reasoning model. Maybe it's doing reasoning without tokens, just like that recent paper you guys may have seen. Maybe he's got a rabbit in his hat. But if I had to guess, I think it's gonna fall a little bit short. I think it'll be a really, really solid model. I don't think it's gonna be a reasoning model. And at the end of the day it especially because you're probably going to get unlimited access to this model if you already have, you know, the Twitter X subscription. So regardless, I do think this will have a positive impact on the industry as a whole, just being a powerful competitor and pushing other companies to make moves. But my personal prediction is that it's going to fall a little short. The open source community is continuously raising the bar, really pushing closed source companies to actually create models that, that compete with them. And that especially reigns true for things like censorship and pricing. Innovation is continuously speeding up and ramping up and the competition is absolutely driving AI development at warp speed. On one hand we see open models that empower the community, while on the other hand, tech giants and billionaires are running their own race. But that race is also similar to open source decided upon by the community and the people that actually use the products. And I definitely think that they cannot for get that if they want to end up as the winner of this race. As always though, I want to hear your thoughts as well. Which of these developments is the most exciting to you? Are you interested in actually trying out and using open source video generators? Do enough of us even have good enough hardware to pursue that? Do you think OpenAI's one model for everything in regards to GPT5 is a good idea or a bad one? What about Elon Musk's Grock 3? Is it just a marketing stunt or is this going to be the real deal? Smartest AI ever. Chat about it in the comments. As always, thanks so much for tuning in. Make sure to check out my Discord server and my Twitter. If you want to kind of stay in the loop with where I'm at in all of this. Until next time, folks, have a good one.