The Open Source AI Revolution continues! HUGE News & Updates

Hello everybody. I'd like to welcome you back to the Matved Pro AI YouTube channel. The AI space has been crazy lately. As I'm sure a lot of, you know, a lot of really good things have been going on. A lot of really interesting stuff has also occurred. I'm not here to waste your time. Let's get into it. A bunch of you guys have been asking for my comments on this situation. Basically, when the stock markets opened on Monday, the entire United States tech sector was kind of met with a little bit of a crash. I mean a little bit would be an understatement. It's the down at least Nvidia here is down 19% in the past five days, which is absurd. Now this is apparently in reaction to the release of The Deep Seek R1 model. This open source Chinese LLM that goes toe to toe with OpenAI's 01. I actually covered it last week, so the market was pretty late to this in my estimation. But regardless, Deepseek R1 is kind of a win for everybody here for fully open source with a great research paper replicating what OpenAI did for less money. And the model also runs at a far lower cost. But somehow to the market this equated that people are going to buy less GPUs to train new AI. And this also somehow equated that American AI technology or American technology dominance is gone forever. I really don't get it. Let's take a look at the comments from Dr. Jim Fan here. He does work for Nvidia, so obviously maybe a little bit of bias here, but he's a senior researcher, really smart guy. An obvious we are so back moment in the AI circle somehow turned into it's so over in the mainstream. First up, unbelievable short sightedness. I mean I totally agree with this. Deepseek R1 is open source. American companies absolutely will and already are benefiting from it. Not to mention that deepseek R1 which is causing all of these issues, is also a thinking model. So thinking models cost a lot of inference. Compute. We've already spent the compute to train Deepseak R1 but now to run it we actually need more compute, more inferencing time compute because it's a thinking chain of thought model meant to replicate 01. I I don't understand how that somehow equates to the fact that we need to get less GPUs. More people are using AI so we need less GPUs. Does not make sense to me. The power of 01 in the Palm of every coder's hand to study, explore and iterate upon. Yep, there it is again. R1 is only going to benefit and lead people to build more things ideas compounding. The rate of compounding accelerates faster with open source. Absolutely true. The AI space has handedly seen that for years now. We as one humanity are marching towards universal AGI much sooner which is going to cost a ton of compute. People obviously making fun of the crash here saying oh, as cars got more mileage per unit of energy, the world bought less cars. Obviously that's not what happened. Anyways, Jim Fan is not alone in this opinion. It was pretty much the general consensus of a lot of people that are deep inside of the AI circle that this is super short sighted. The market reacted this way and it just kind of felt dramatic. I'm sure things are going to pick up in the future. From a marketing perspective, I don't see how Deep Seek R1 is somehow a bad thing, at least in the long term. So yeah, those are my thoughts on the whole deep seek R1 market tanking situation earlier this week. But in other news, Deepseek kept publishing this week. They kept shipping. They dropped Janus 7B, which is MIT licensed again open source. It's a multimodal LLM capable of generating images and intaking them, which is pretty good. It's also very light at only 7b, so it can run locally on a lot of machines. Much more, by the way, very soon on this channel. In regards to local AI, I'm sure a lot of you guys already know this by now, but Nvidia graciously sent me an RTX 5090 to mess around locally with tons of different AI. It's not sponsored or anything like that. I still don't know if I'm going to do a broader review on the GPU or if I just want to show off what it can do locally with AI because it's stunning, 32 gigabytes of VRAM is a ton. Anyways, be on the lookout because video is coming very soon on local AI. With DeepSeek R1 being so much cheaper than OpenAI's O1, which you need to at least pay the $20 subscription to use right now, even though O3 is on the horizon, OpenAI's next model I think OpenAI and Sam Altman, the CEO specifically, was feeling a little bit of pressure with all of this Deep Seek success and even more ships after Monday. So Sam put a statement out Saying Deep Seek R1 is an impressive model, particularly around what they're able to deliver for the price. Yeah, price to performance, we will obviously deliver much better models and also it's legit invigorating to have a new competitor come at OpenAI considering Anthropic has kind of been slacking lately. Anthropic, where are you? What's going on? We will pull up some releases, he goes on to say. Mostly we are excited to continue to execute on our research roadmap and believe more compute is more important now than ever before to succeed at their mission. And I would absolutely agree. We we need more compute for more AI. AI at this point in time kind of seems like the next technological revolution for humanity. The world is going to want to use a lot of AI and really be amazed by the next generation of models coming. Looking forward to bringing you all ag and beyond. Overall though, I really think this tweet is showing that Deepseek R1 is putting the pressure on and I think pressure makes diamonds. Competition is good for everybody. It's great for the consumer. We're going to have cheaper models that are better because everyone's battling out to have the best and win over the consumer at the end of the day. So. Well, this whole deep sea gar one situation, on the surface it seems like a bad thing, especially with how the market reacted. I really think that overall this is going to be great for everybody at the end of the day long term. Before we take a look at some other AI news, today's video does indeed have a sponsor. But don't click off guys. I can already see you trying to skip past it. This one fits perfectly in today's AI News roundup and I would definitely be covering it as AI News if they didn't sponsor the video. Anyways, I also plan to do a full video on this one releasing very soon, so take a look guys. I think you're gonna like this one. Today's video is brought to you by Refusion. This is a brand new AI music generation website and while it's in beta, users can make unlimited generations for completely free. And that's not all. They have plenty of features and are very competitive in terms of audio generation quality. Take a listen to some of these samples from the Refusion community.

Was rice like Alabice Straight device paradise Watch you swim my day where do you go when the phone rings twice and the water flows Something s in my pocket can't explain what I'm seeing lately Every morning there's a meaning and my best so I'm leaving Time reveals my truth inside our Clothes with place divine.

Yeah I'm different now watch y' all move big deals and no second got something to prove your whole block feeling When I pass through you falling from the sky that's nothing new make a cow growing you can't understand private jet status different type of man Tokyo know me they all understand we when I touch down better know who I am you was thinking why I made these plays international money, different kind of wave Showing respect in different ways When I pull up feel the ground raised.

Yeah I know pretty darn good right? That is up there with the state of the art music generation. We hear from other websites, but they've also got a few other tricks up their sleeve. First of all, they have a personalization feature that actually learns the preference, the musical preference of the user generating and can actually help them establish their own sound and this can be turned on or off. They also have more advanced features like multi sound prompts where you can actually indicate how strongly each sound is going to affect the end result. This gives you a unique and creative way to control the way your songs sound and is a new feature I haven't seen on any other platform. They also have sliders that you would expect such as lyric strength, but also new ones like song weirdness, allowing you to dictate how experimental you want your song to sound at the end of the day. Not to mention a bunch of other features like the ability to share your songs in the community, a lyric generator, and more. This beta is pretty awesome and I think they've come out swinging against the competition. It is super competitive, so if you want to try it for free in beta, check the link down in the description below and also be on the lookout because I'm definitely going to have to cover this in a full review. But anyways, fol back to your regularly scheduled content. Welcome back everybody. So yeah, as I mentioned, I am doing a full video on Refusion's Fuzz, their brand new AI music generation model for now. As I just said, it is free as long as their GPUs survive. I actually did a tweet about it talking about some of the other things going on in the music space and what we can expect for the future of AI music during this year. To be honest, the AI music niche in 2024 didn't have all that much competition in total in comparison to other AI niches and markets. But in my opinion, this is all going to change this year. First up, eleven Labs, which is a leading AI audio company, they do great sound effects, great AI voices, among other things, including real time voice agents. They showed some demos off in late 2024 that were really promising for their AI music model. It sounded really, really good. And while the demos I'm sure were cherry picked, eleven Labs has a fantastic track record at releasing high quality audio models. They really wait for things to finish baking before releasing them. Next up, we obviously just saw the release of Refusion Fuzz, a brand new AI music competitor and the beta is completely free. If you want to use something like Udio or pseudo AI music, you're gonna actually have to pay for the monthly plan. I think they do have some limited free capabilities as well though. And finally we have brand new open source models such as Yue. It's a full song music generation model and it's open source and of course Chinese. So yeah, there's a little taste of what this open source music generator sounds like. It's not as great as the closed source stuff obviously, but this is fully open source, completely free, can probably be run loaded locally on a lot of machines and is definitely going to influence the AI music generation space as a whole. So buckle up folks. I think that 2025 is going to be a huge year for AI music. Lots of new competition, lots of great models. So get excited. All right guys, let's talk more about open AI and upcoming releases such as O3 mini. Again, this is the next step, the next generation of the O series models. Remember open AI's O1 thinking model that people have access to with chat, GPT, PL and Pro and then of course the deep seek R1 competitor that started everything, all that drama we talked about earlier in the video. Anyways, O3 mini looks like it's actually going to be a lot better than expected. Chubby. A fantastic account by the way, for keeping up with AI news. Shows off some tweets from OpenAI employees who seem to be very excited for the release of O3 mini. Dylan Hun here says O3 mini is so good, exceptional intelligence and blazing speed. I haven't been this excited for a model since GPT4, which says a lot. GPT4 was a huge release and a massive step up from GPT3. He really says it's worth the hype. Victor here in the comments thinks that he may be right. GPT03 and 03 Mini might be the models that perfectly merge what we love about GPT4 Omni and the GPT01 series, but seamlessly together. And to be Honest, after using ChatGPT 4.0 and 01, I kind of want the best of both worlds. I understand what Victor is talking about here because O1 right now doesn't have access to a lot of the capabilities such as voice or the ability to search the web. There's a lot of different use cases and different features that are available on certain models and then not available in other ones. Combining them together in a much more fast and intelligent model overall is definitely going to be something we all want now, guys. There's also a ton of rumors floating around that O3 mini is going to be released today. In fact, it might actually be released right now. These rumors are definitely not unfounded. OpenAI is most certainly gearing up to release O3 mini, especially after all of this deep seak R1 stuff. So I would expect it either this week or next week. And if you want to know if it has released I will definitely drop a link down in a pinned comment below. So be on the lookout for that if O3 mini does release, and I'll be doing a fuller deeper video on O3 mini pretty soon if that's the case. Anyways, in other OpenAI related news, it appears that GPT4 Omni was also upgraded by OpenAI and from Mark's initial testing here, the model shows a pretty substantial improvement in image recognition. Not just a subtle upgrade. The model can understand fine details and images a little bit better, including faces. Here are some patch notes if you guys want to take a closer look. You can pause to read these, but it definitely mentions better image understanding, smarter overall, uses more emojis for some reason. Again, in my mind this is a reaction from the immense pressure that R1 has seemingly put on everybody. Keep shipping, keep updating. I like to see it. Good work OpenAI. Alright, this next piece isn't necessarily news. I just came across it and thought it was so cool I had to share it with you guys. Edwin here asked 01 to help him code the Wii Menu and it built a react app that renders in the Chat GPT Canvas. This is a new feature from about a week ago I think. I did mention it briefly in a recent AI News roundup. But yeah, it can display your code. It can actually build an app right in the Chat GPT Canvas interface. Edwin here fed it a screenshot and it basically one shotted, according to Edwin, the basic layout of the Wii menu, the classic Wii menu, even the striped background. He then kept prompting it to add animations and more features and what he was left with was pretty impressive. So again I just had to share this. I think A lot of people are starting to forget how far along large language models have gotten to the fact where they can code something this good. I mean it is a pretty decent Wii menu clone. It opens up, you know, the weather app for example. You can go back to the Wii menu scrolling text just like you see from the classic Wii clicking on the disc channel. We've got the rotating spinning discs. I mean that just gives me literal nostalgia from this recreation alone. But yeah, that's super cool. It's just a great example of how powerful large language models have gotten. Keep in mind deepseek R1 is competitive with GPT OH1 and apparently O3 mini is going to be even better while also being cheaper and faster. It's. It's just crazy. All right. AI is booming. We've got even more OpenAI news. Recently we saw OpenAI Dev Day. They introduced the Real Time Voice API finally and they have a nice demo of an interactive solar system that you can actually navigate with your voice. And they've actually open sourced something for once. They open sourced this little demo to show you how you can use the Real Time API to build an interactive solar system demo. So I, I guess good job OpenAI for finally open sourcing something. I just wish it was actually an AI model and not just a project. But I'll show you the demo because it's pretty cool. The technology that is powering this OpenAI real time voice is pretty exclusive. Not a lot of other AI companies have produced something quite like this. It's natively multimodal. So this is going to be a cool demo.

I'm curious about Earth.

Earth, sure thing. Earth is the third planet from the sun and the only one known to support life. It's packed with diverse ecosystems, abundant water, and a protective atmosphere that shields us from harmful solar radiation. Do you want to know more about Earth's characteristics?

Well, I see that there's actually a lot of water there. So can you tell me how much land is there versus water?

Absolutely. Approximately 71% of Earth's surface is covered by water, while the remaining 29% is land. Our blue marble truly lives up to its name. Let me show you this on a pie chart. Here's the breakdown of Earth's surface.

Okay, great.

Thank you. I actually have a question about Mars.

Now, Mars, the red planet, what would you like to know?

So I heard there was a lot of volcanoes there. Can you tell me which one is the biggest?

Absolutely. The biggest volcano on Mars and in the entire solar system is Olympus mons. It's about 13.6 miles. Nearly three times the height of Mount Everest. Would you like more information on Olympus Mons?

Yeah. Can you tell me like what are the three biggest volcanoes and show me their heights?

Certainly the three biggest volcanoes on Mars are are 1. Olympus Mons 2. Ascras Mons 3. Arcea Mons Around 11.2 miles.

So yeah, pretty cool demo. They're open sourcing this so if you want to build something similar, check that out. Everything's linked down below of course. I do want to make a quick comment. I don't think this UI is exactly dynamic, but I think you will see LLM powered dynamic UI crop up a lot more pretty soon. Things such as this, where I can, you know, kind of Jarvis, tell me, you know, what are the deepest oceans on Earth and then give me a comparison of all their depths in a chart and will generate all of that in real time for you with its own custom ui. That's something I think you're gonna start seeing a lot more because LLMs are getting very fast and very powerful. Next up guys. Finally we're moving out of the OpenAI news. The United States Copyright Office declared on Wednesday, so that was yesterday, that the use of AI tools to assist in the creative process does not undermine the copyright of a work. And that's a really big deal. This announcement clears the way for continued adoption of AI in post production where it has become increasingly common. And they provide a little bit of an examples, I guess the Hungarian language dialogue in the Brutalist movie. They're talking more about, you know, big name movie studios, but really this is actually better for smaller folks who are really going to use AI technology to produce creative works the likes that I think the world has never seen before. New creative ideas from people who don't have necessarily a ton of money, but AI tools are going to break down those barriers. So you don't need a ton of money to produce something. That is amazing. Check out my latest video for updates in regards to AI video. Some really cool stuff recently on that. This is a 41 page report and the Copyright Office also reiterated that human authorship is essential to copyright and that that merely entering text prompts into an AI system is not enough to claim authorship of the resulting output. So I agree honestly with all of this. When AI imagery started to first become a big deal, there was a lot of talk about this. But this kind of cements everything in stone and I really do agree with the U.S. copyright Office in regards to this final decision. Next up, I've been wanting to talk about this one For a while it kind of flew under the radar, but Netflix actually dropped some AI research. This one also has code inference, but this is Cut, Cut and Drag Go with the Flow. It's basically an AI animator and it sounds exactly as it is, where you can actually cut and then move parts of the image around and then it will generate a video animating wherever you dragged that piece around. Kind of hard to explain it, but I think this video shows it off quite well. Obviously we start here. You cut this piece of the duck out and we move it down and then up and then back to the side and then back to the center. And the Netflix AI go with the flow here. Animates it pretty well. There's still some morphing, but it's definitely very good at adhering to the movements we specified with the Cut and drag. Let's take a look. This is actually a better little demo video. That should explain it pretty well and show off some more cool examples.

We present Go with the Flow, a method for controlling motion and video diffusion models that uses warped noise. Let's illustrate our method with an example. Let's say we want to make the cat walk down the tree. First select the thing we want to move, which is the cat, and then define a trajectory which in this case is translated downwards and rotates counterclockwise from that trajectory. The final output of our model is the cat walking down the tree.

Kind of. Obviously the cat is growing extra legs here. More work to do in this regard. This is just a research paper after all. But it definitely moves in the correct direction, which is cool. They're doing this with warped noise, which is really interesting. You can see the crude animation versus the warped noise and how that actually translates in the output video to a smooth generation. Obviously each image or each frame is generated from noise using diffusion, so there's definitely more to get into. If you want to dive a little bit deeper into the research paper that's going to be linked down below. For now, we're going to move on. We've got another open source large language model here that surpasses the performance apparently of Deep Seq V3. And don't be confused, Deep Seq V3 is different than Deep Seq R1. It's not as good, but this massive 405 billion parameter model and Allen AI's circumstance here demonstrates that their recipe from verifiable reward learning scales all the way up to a 405 billion parameter size with performance that is about on par with GPT4 omni and surpasses prior open waves, training models. The reason I included this is not necessarily that it's a huge game breaking AI release, but the fact of the matter is that open source AI is absolutely booming right now. We're seeing a ton of great open source releases with all kinds of different research like verifiable reward learning and the fact that this method they developed does scale up. So there's so much research going on. I mean, I think that's really the point that I'm trying to get at. There's so much research into all these different methods to improve these AI models and a ton of it is going to be coming out as open source with research papers and weights. It's really, it's really the boom of a new revolution, guys. It really is. Finally, for the last news piece of today, Alibaba have quietly included video generation in their recent Release of Quen 2.5. Apparently the video generation is pretty good. Tom Lakes Robots gives us a couple of examples to take a peek at. So new video generator. Let's have a look. First generation here appears to be pretty good. This also looks like it is 1080p. I don't think Twitter compression is doing us any favors though. But you know, obviously she's got a light and she holds up her phone to take a picture. The light does mysteriously start floating when she lifts her other hand to pose for the photo. But we've got great recreation of physical body parts such as the hand, the face, slight dangling of the jewelry as she moves her head. It's all pretty crystal clear and really nice. We've also got a really nice stable video of an elephant in a river who is drinking water with his trunk. It appears he's filling his trunk up and then trying to drink all that water. So impressive. A lot of the details here, the finer details on the skin are really nice and consistent as the elephant rotates his head. And the background here is also really rock solid. Coming really close to a very true to life looking video from this model. Does it beat Google vo? Probably not, but I still like to see releases such as this. We've got another video here. A man painting with a palette here. He's just sort of doing some brush strokes. For some reason, he's using the canvas as the art palette itself. I don't know why you're usually supposed to be painting on the canvas, but bro's just having a good time. All right. We can't, we can't take that from him. More interestingly though, we've got some Mario Kart going on. It's literally a tiny little squirrel in some nice fall leaves here. He's got his acorn cart and he's rolling around driving that thing. So this is a really cool generation from the model. It shows that it does have some capability to do things you wouldn't typically see in the training data. I'm not quite sure if you can generate for free right now, but apparently if you Visit chat QuenLM AI and switch to video generation, you can sign in with Google and start to use it. I might have to give this one a shot in a larger video. Let me know if you want to see that and also let me know if you try this one out and like it. Alright, I think that's gonna do it for this video. Like I said, I've got more coming in the future. It's been super busy lately, I'm not gonna lie to you. But I'm really trying to get this content out out. Larger video coming on Refusion. Lots of local AI stuff also coming with the RTX 5090. Really exciting. And more. More down the pipeline from then on out there. I mean, there's just so much going on. Good problems though, good problems. We like to be busy in the AI space. That means things are going well, despite what the market might say. Anyways, thanks so much for watching. I'll see you in the next video and goodbye.