New FREE & Open Reasoning LLM Matches Open AI o1! + RTX 5090 Unboxing! AI News

Hello everyone. Welcome Back to the Matvid Pro AI YouTube channel. There is plenty to talk about today, so let's dive in. First up, Deepseek AI has released Deepseek R1 and it performs on par with OpenAI's O1 reasoning model. But while OpenAI's O1 reasoning model is closed source, Deepseek R1 is fully open source, meaning anyone can download it for completely free. They can modify it and improve it if they're a developer. Specifically, it's under the MIT license. If you want to try the model really quickly for yourself, it's going to be at this link chat.deepseek.com and of course I always link everything down below in the description if you want to get access and try our one out for yourself. A close look at these benchmarks though is pretty incredible. All of these bars with the lines slashing through them are that big bad Deepseek R1 model. Dark gray is going to be OpenAI01. Light blue is Deepseek R1 132B. So it's a smaller version of that R1 model. It's not going to perform as well. Light gray is OpenAI01 mini. And finally in the very light blue, we have the previous DeepSeek V3 model. Taking a look at AIM 2024, we can see that DeepSeek R1 not only is about on par with 01, it's actually slightly better, ever so slightly better with a 79.8% accuracy in that benchmark, nearly 80% accuracy accuracy. What's that? A B minus in aim 2024? Pretty good. Meanwhile, the smaller DeepSeek R132B model is at 72.6, so actually not too much worse than the full Deepseek R1 and quite a bit better than OpenAI's 01 Mini, which only gets a 63.6. We're talking about a C versus a D minus if we use a typical American grading scale. And finally, if we compare this to the last Deepseek V3 model, which only scores a 39.2%. That's not even within passing realm. So huge improvements across the board, especially compared to Deep seeks previous models. But honestly, what impresses me the Most is this 32B model, very, very close to the state of the art performance of OpenAI's 01. And honestly quite a bit better than Open AI's 01 mini. If we look at code forces though, so this is more focused on coding, we can see 01 still tops the charts at 96.6. Deepseek R1 is right behind there at 96.3. The 32B model though still comes really close to just straight up closing the gap at a 90% with O1 mini being just a little bit better at 93.4. Meanwhile, the older Deepseek V3 model is handily left behind at 58.7. Let's move on to the ever difficult GPQA diamond. DeepSeek R1 comes in at 71.5% accuracy with OpenAI's 01 at 75.7. So this is going to be OpenAI's biggest lead here today, but it's only by about 4%, so not huge. We can then see that the next contender is R132B at 62%, handily taking down 01 Mini and of course their previous Deep Seek V3, which still did pretty good at 59.1, at least in comparison to that 32B model. Next up, taking a look at math, we are coming very close to perfection with a 97.3. An A from Deepseek R1 and 1% behind is OpenAI's 01 model, then followed super closely at 94% with that R132B model. So close to getting the same score as the full Big Bad 01 at only the size of 32B. That's crazy, but it makes sense because they were already on this trajectory. V3 actually ties with O1 mini. At MMLU we see O1 take a slight lead by a percentile, but R1 is right up there and to be honest, the rest of the models are also super close as well. Interestingly Enough though, DeepSeek R132B actually underperforms in comparison to DeepSeek V3 in this particular task. Finally coming down to SWE Bench Verified, we can see almost a perfect tie between DeepSeek R1 and O1 again with DeepSeek's R1 32B falling behind the previous DeepSeek V3 and Zero1 mini with 36% accuracy in comparison to about 41%. Regardless, swe bench Verified is still pretty difficult. We have yet to have access to a model that gets in the 50s, so yeah TL. Dr. This model is beastly. It matches the performance and sometimes even outperforms OpenAI's 01 model, which is said to have about an IQ of 133 in comparison to a person. Obviously that comparison isn't super direct. Even using IQ as a benchmark isn't going to yield A lot of clarity. My point is though, that this thing is super smart, probably smarter than a lot of people, and it's open source and fully accessible. So I commend Deepseek for their commitment to this. They're also releasing as a bonus a bunch of distilled models, all stemming from DeepSeek R1. These models get as small as 1.5 billion parameters, which is super tiny. And check, check out these quick numbers. This 1.5 billion parameter model handily takes out GPT4 omni, which is super cool. Now to be fair, there are some benchmarks like GPQA where the obviously larger GPT4 Omni still wins, but it's pretty close. But if you bump up to the 7B sized model, you now take out GPT4.0 in all of these benchmarks, which is super crazy. 7B is a very small sized model, not very big at all. Lightweight. We've also got 14B, obviously the 32B, and then we've got an 8B sized model in a 70B sized model as well. Interestingly enough, this 8B sized llama model versus the 7B Quen model. 8B seems to be a little bit worse, but that big bad 70B sized model kicks some serious tail. But the Quen 32B comes pretty close to it. Some final updates for you guys in regards to this. It's only $0.14 per million input tokens with a cache hit and $0.55 per million input tokens with a cash miss. So seriously, cheap output tokens. We're looking at $2.19 per million. Just crazy pricing figures. I mean it's super cheap. You could use this model for months for like basic general tasks and not even spend five bucks. Obviously though, this is going to be amazing for people trying to actually build things, build products utilizing this API intelligence is getting so so cheap. Before we dive into the rest announcements, Today's video has a great open source sponsor. Sponsors are the reason I can actually pursue this channel full time and I really appreciate you guys sticking through them. We all know that large language models are all the rage, but training them is expensive, complex and most of all, GPU hungry. Well, today I've got something that could save you time and money if you're trying to train a large language model. Introducing YAFSDP from Yandex. This stands for Yandex Fully Shared Data Parallel, an improved version of of the classic fsdp. This is a new open source framework designed to speed up large language model training while optimizing GPU resources. You see when training large language models, the communication between the GPUs themselves often becomes the bottleneck and it slows everything down. But that's where YA FSDP comes in. It's the fastest solution among similar frameworks. Reduces memory usage by optimizing GPU communication. By eliminating inefficiencies, it can save the resources of up to 150 GPU', which has the potential to save millions in training costs. The best part here is that YAFSDP only tweaks the model itself, which means you can retain the use of your custom optimizers. If you're training a model with a transformer architecture, you can cut down your training time by up to 25% and save up to 26% in GPU resources. And again, the best part here is that it's open source. Check the link in the description below and in the comments. Now back to your regularly scheduled content. Welcome back everybody. Next up, I actually have some rather important channel news. I think you guys will be excited about this. Our green overlords Nvidia have actually sent an RTX 5090 my way. I'll include some unboxing footage I recorded as the embargo for unboxing is today. However, I can't yet show any benchmarks or me actually using it yet. But but we do know the specs on the RTX 5090 and it has 32 gigabytes of vram which is a ton. We should be able to run a bunch of awesome local AI models on this thing and I plan to include that in a video in the future. So comment below on which particular models you'd like me to run with the RTX 5090 obviously already on the list. We are going to be trying that deepseek R1 32B model. At the very least. I want to see if I can try to get some 70B sized models running as well. But we're obviously also looking into image generation and especially video generation. So super cool. Huge thank you to Nvidia again our green overlords for sending this thing my way. I am super excited to get a good glimpse at the future of locally run AI generation. And by the way, one quick mention I do have about my experience unboxing the RTX 5090. This thing's pretty small. It's only a two slot blower style graphics card and compared to my previous RTX 4080 it's actually thinner than this thing while drawing more power and creating more heat. Hopefully Nvidia lets me hold on to the RTX 5090 and if so, maybe we'll have to do a giveaway for this RTX4080. Let me know if that's something you would be interested in, because this thing still has plenty of capability to run local AI models, among other things. Alright, let's dive into the rest of the AI news. Next up. Sam Altman puts it all to rest. Okay, you set the record straight today. No AGI deployment soon from OpenAI. You can see in this tweet here he says Twitter hype is out of control again. We are not going to deploy AGI next month nor have we built it. So yeah, everyone who was like, is 03 AGI? Does OpenAI have internal AGI? What did Ilya see? Apparently nothing because they claim to have not built AGI. He goes on to say that they have some very cool stuff, but please chill and cut your expectations and you know, fair enough Sam, fair enough. But to be honest, you kind of have been hyping us up. The reason Twitter hype is partially out of control is in a lot of ways due to some of the tweets you made earlier. Take this one for example. Near the singularity, unclear which side, another tweet, common themes, AGI agents, and a bunch of other features. I mean, to be fair, people extrapolate crazy things from Sam Altman's tweets, but man, they are cryptic and they lead to this crazy hype. But to be honest, I think Sam knows that. I think he enjoys engaging with this hype. He likes being, you know, a part of this crazy technological shift. And can you really blame him? Also, gotta say, I really like this X breakdown of the news. Gives good context here. Saying the conversation around AGI was sparked by impressive results from the new O3 reasoning models and their benchmarks. Also, I'll say this again, guys, we don't really have a solid definition for AGI. It really depends on who you ask. AI has not been mainstream for a lot of time. In the grand scheme of things, I'm sure we'll have very solid definitions for AGI in like 5 to 10 years. But as of right now, everyone's kind of spitballing what they think qualifies as AGI. So I'll propose the question back to you guys. What do you think AGI is? I often ask this in my videos because I love to hear what your personal convictions and qualifications for this mystical, all powerful artificial general intelligence is. May the best definition win. Let's move on to AI video generation Technology right here, you're looking at a Cling AI early access feature. This is brought to us by Shiras Power AI on Twitter. As you can see, we've got three input photos at the top. Woman with brown hair and a striped yellow shirt. We've got this really high tech neon glowing room. And then finally we've got a regular can of Red Bull. Slap them all together with Kling's new character consistency and background consistency model and you get something pretty good. The woman looks, I'd say about 80 to 90% similar to the one provided in the photo. The background is also shockingly accurate. I'm really impressed by that. It's not perfect. You can see some shimmering and movement in the background. But overall it does a really good job replicating that same exact image while adding the girl in the foreground. Finally, we move on to the Red Bull can, which is pretty decent. No legible text unfortunately, but at least the model was able to understand and put everything together cohesively. The girl is standing in the room and she's holding the Red Bull can. That's what we would expect. We wouldn't expect the Red Bull can to be floating around somewhere or her to be in the background. It's intelligent and it's cohesive and dare I say, this might be the best consistent background, consistent character consisting item technology I've seen come out of the video generation space yet. Like I said though, this is currently an early access feature, so not everyone is going to be able to utilize it. But we can assume that CLING is going to deliver it to the public pretty soon. I've got some more examples. These are coming from Halim. Apparently this feature is known as elements. But regardless, it's all about consistency. And in this example we combine first our girl, who we want to be consistent. She's got blonde hair, pretty distinctive features in terms of the eyes, the nose and the face. Next up, we've got a Christmasy holiday sweater. And then finally this cute little pink creature thing. We've also got the prompt for this one. Cozy, softly lit room woman sits casually on the couch speaking warmly to the camera. In an influencer style video, she raises her hand showcasing a very small and cute cartoon monster. The cartoon monster jumps around on her hand. She's wearing a cream jacket with green patterns. Just inputting the elements and the prompt yields a result that is intelligent. It understands what that prompt means and how to combine these elements into a consistent video. That is as we expect. I am really surprised by how good the consistent character Is it looks like the same woman, relatively speaking, for the most part. I mean, the eyebrows, the eye distance, the type of nose, even down to the hairstyle with these two strands of hair at the top. Really surprising. The jacket is a little bit more complex, but I still think it does a decent job at nailing it. The patterns are pretty shockingly similar. You can see there's almost these two carrot looking red pieces at the bottom. And when I say carrot, I mean like the kind of carrot you'd see on your keyboard. And then, you know, we can see these wings that are also visible in the original image. And then finally we get down to that pink character who is jumping around on her hand. It looks pretty good. There's a little bit of shaking and weirdness going on its face as it jumps around. And I gotta say, this pink monster character isn't exactly perfect in comparison to the provided input image, but it's close enough. Clearly there is more training going on in relation to humans and their clothes than pink monsters inside of Kling. But that pink monster is an abstract idea and it still did a pretty great job, especially combining everything together. It's magic. Halim also has one more example to show us. We've got this sci fi future futuristic white ship with glowing red towards the back of the ship. And then we've also got like this futuristic city. And you can see the ship zooms out through the tunnel into pretty much the same exact city as pictured here. He also went ahead and did another test with the same ship, but this time a different city. This is a Tokyo esque cyberpunk style city, which is really cool to see and it also nails that pretty good. I think the most impressive part though is the fact that the ship is relatively consistent between these two separate generations. And again, I mean, this leads us into consistent storytelling, right? That's what it's all about. If Kling can pull this off, man, and it already looks pretty good, I think they'll have some domination in AI video generation just. Just because of the ability to do the consistency. Solid work so far, guys. Can't wait to get access to this. Continuing with the trend of AI video generation, Mincho is showing us some really great examples of how you can fine tune Huanyan AI's video generation model with Laura's. This first example is crazy. It went viral on Twitter the other day. Severance film. But John Wick as the main character instead. Take a good look at this one guys, because it is difficult to tell that John Wick here is entirely AI generated. It even changes the hair. Wow. I mean, it's crazy. I think there's some notable differences in the facial expressions. I think the acting is a little bit more reserved in this example. But I mean, it's. It's really, really difficult to tell. It's. It's AI generated. There are some notable details that are lost. Like the plant is entirely different. These seats are a little different. This particular example, with this particular movie, these white scenes with not too much going on serves the AI generation quite well, at least in regards to making it look real. Also, by the way, the Laura is available on Civit AI John Wick. In this example, Odin Loveis actually trained a Laura on real foot of a real person, similar to the John Wick example. Apparently the training time for this took 15 hours on an A100 GPU. So we're still looking at quite a long time to be able to do natively consistent characters using Loras for Videogen. But if you have access to an A100 and you're serious about AI video generation, you could start training Loras of different characters today. However, if you don't have access to beastly hardware. Yatharth Gupta not sure if I pronounced that correctly, shows that you can actually use six to eight images to train a LoRa in just 15 minutes on fall AI, which is a fantastic website by the way. Might have to do a full video on this one. Or maybe with that GPU we can do some local training of our own. But 15 hours seems like a long time. Obviously though, the 15 minute 6 to 8 image training isn't going to yield as good of results, but still might be good enough for certain projects. I think it really is going to depend on your use case. As always, everything's linked down below. If you want to check out more of these examples. Here's another really consistent character. I mean, there is no doubt in our minds, right, that this is the same exact woman, but it's all AI generated video. None of these clips are real. And finally, to end this off, we've got Jesus inside of Squid games. So whether you're using a Laura with one Yan video or you somehow got access to the new Cling elements feature, it looks like consistent characters is the next logical step for AI video generation. I've been saying, saying this for a while now. I am excited because it's the last piece of the puzzle, at least for me, to really start to dive into storytelling using AI video models. And after this is locked down, consistent characters, consistent backgrounds, consistent storytelling we can then focus on polishing up the quality and making sure that these models can do some of the more tricky tasks. All right now I'm going to quickly run through a couple of last few pieces of news that I thought were worth including. First up we've got Synth Light. This is portrait relighting with Diffus models by learning to re render synthetic faces. You can see the examples here are all pretty incredible and do a really great job at predicting what different lighting will look like on a particular character's face without actually having to model everything in 3D. I like Victor's comment down here we're teaching AI to be a Hollywood lighting director now I believe. For now there's no GitHub for this one though. It seems to just be a paper. Regardless, it's linked down below. I think I already covered this in a previous video, but this is an awesome text to speech mod that is open source and only has 82 million parameters. It generates really good quality speech and can generate minutes of speech in just seconds because it's so lightweight. 82 million parameters. But yeah, this thing is open source and it's awesome so it'll be linked down below. It also is available to try for free in a hugging face space. Also wanted to mention this one. This is Magic Quill. It's an intelligent interactive image editing system system that uses AI to edit your images quite well. The demo that they provide showing off it running on this tablet here is pretty impressive. Adding necklaces which is typically a pretty difficult task for traditional inpainting, but not impossible. Also able to do simple brush strokes here and add color to the woman's hair flower petal headband and it really does allow you to add and edit specific things. All in all this looks super cool and it seems really fast as well, at least from this demo video I am here on the GitHub page. It seems like they have a Gradio demo. They've got a paper released as well as a full demo page. Again, this one I might have to look a little bit deeper in. Could make for a decent full video. They've also got a comfy UI custom node for this so it should be able to work locally too. Looks like minimum requirements here are going to be 8 gigabytes of vram so make sure that your GPU has requirements if you want to run this locally. Okay guys, that's going to do it for today. Obviously there's tons going on in the AI space at all times so I can't cover everything for example, there is a plethora of stuff going on in the robotics space as well as the 3D AI video generation space. Thank you so much for watching folks. If you aren't subscribed, please consider subscribing and leave a like if this video was useful. Also, check out the awesome Discord community down below if you want to stay on top of AI news as well as engage with like minded individuals who are super talented and knowledgeable about AI. By the way, it's a great spot to be. All right, I'll see you in the next one. See ya.