AI News: The Best Open Source Model EVER

All right, so things are heating up in the AI world and a lot of news once again happened this week and it is my goal on this channel to keep you looped in with everything that I think you should know. There is so much AI news happening every single day that I am doing my best now to kind of just filter it down to just the things that I think you'll find interesting or that I just personally find really fun and interesting. Dive right in. Starting with the biggest announcement of the week, which is the fact that meta just released llama 3. If you're not familiar with llama, llama 2 was the open source large language model put out by Meta that almost all of the open source large language models that are out there now were initially based on. Now, despite what other content creators on YouTube may be telling you, this news did not shock nor stun the industry because Meta has been flat out telling us that this is coming and they even told us it was happening within the next couple weeks. So everybody in the industry has known that this day was coming. But here's some snippets from Zuck about this monumental day in the world of AI.

Today we're upgrading Meta AI with llama 3, our new state of the art AI model that we're open sourcing. The bottom line is that we believe that Meta AI is now the most intelligent AI assistant that you can freely use to make Meta AI even smarter. We've also integrated real time knowledge from Google and Bing right into the answers. We're also releasing a bunch of unique creation features. Meta AI now creates animations and it now creates high quality images so fast that it actually generates and updates the images for you in real time as you're typing. So today we're open sourcing the first set of our Llama 3 models at 8 billion and 70 billion parameters. They have best in class performance for their scale. And we've also got a lot more releases coming soon that are going to bring multimodality and bigger context windows. We're also still training a larger, dense model with more than 400 billion parameters.

So a couple things to break down about this announcement real quick. They released two versions of llama 3, an 8 billion parameter model and a 70 billion parameter model. And it's pretty comparable to the AI models that are available for free right now. If we look at the various benchmark tests here, it performs roughly about as well as Claude 3 Sonnet, the free version of Claude and Gemini Pro 1.5 when it comes to the existing open source models. The benchmarks of the 8 billion parameter model outperform some of the best existing open source models. So if you're looking for these models to really outperform Something like Claude Opus or the current version of GPT4, they these aren't quite there yet. However, what I think is going to be the bigger news from llama3 is when they finally release the 400 billion parameter model as an open source model. This 400 billion parameter model when it comes out is going to have so much better capabilities. Things like multimodality, the ability to converse in multiple languages, much larger context windows, and stronger overall capabilities. This is the llama 3 model that everybody's waiting for. This is the one that is expected to compete with current models of GPT4 and Claude 3 Opus. They're even showing some of their early benchmark scores. And this is before the model is even done training. And these benchmarks are kind of on par right now with what you get from the closed models like from Claude and from OpenAI. And of course with the announcement of Llama 3, Nvidia wants to remind you that they were trained on Nvidia GPUs and Groq. Grok, the one that really, really speeds up inference from large language models, announced that Llama 3 is going to be on Grok pretty soon. Now if you want to actually use llama 3, there's a couple ways to do it. Right now it is available over on Hugging Face. If you don't want to use it on Meta's platform, you can use it via the API over on Hugging Face. So simply go to huggingface co chat over under current model. You can see that I'm currently set on the Meta Llama 370B. But if you're not, you click the little settings icon and you can select whatever model you want, including Meta's Llama 370B model. But the cooler way to use it right now is the new website that Meta released over at Meta AI. This is Meta's new front end user interface that just feels like a chat GPT style chatbot. One thing they did mention is that this new model actually will search the web when you ask it questions. So let me quickly ask it a super relevant question about something that happened yesterday. I'll say who won the Padres vs Brewers game yesterday and what was the score? And it correctly answered that the Milwaukee brewers won against the San Diego Padres yesterday with a score of one to nothing. And then it cited its sources. Now, while I love Claude, I especially love Claude 3 opus. This is something that Claude still can't do natively in their own chat app yet. I think one of the coolest features of this Meta AI website though is under the Imagine tab, if you click over here, they have a really cool AI image generator in here that will actually generate images in real time as you type. So I am not going to speed this up at all. I am just going to type a cat on the beach eating a taco at night with glowing eyes. And as I typed it was literally just shifting and changing that image around. When I like the image and what it processed, I could go ahead and submit it and you can see up here it gave me the image that it generated and now it's generating three additional images for me a couple seconds later. I have a few variations. Let's go ahead and click on the initial one that I made and and now they have an animate feature. This is something we can't even do over on Dall E3 inside of ChatGPT yet. I can click this animate button and it's gonna take my little cat picture here and turn it into a little short animation. I mean there's not much movement going on. But I just animated my cat picture and this whole thing took me a couple seconds. Let's do a wolf howling at the moon in a realistic style. I How about in the style of the Simpsons, in the style of a eight bit video game. And again, as I type it just changes it. It's really cool, really fun to play with and totally free to play with. Over at Meta AI, one of my favorite new features that meta just rolled out again. I'm going to do a deeper dive on llama 3, but I'm kind of waiting for that 400 billion parameter model to come out because I think that's going to be the showstopper that's going to be the one that stuns the entire industry. So expect more chat about llama3 from me in the future. If you're running an online business, you really should be able to Support your customers 24. 7 with this new age of AI that we're in. That's why for this video I partnered with GPT trainer.com because it is the first and only no code slash low code framework that allows you to build multi agent chat GPT like chatbots that have function calling capabilities and can use your own data. And since human still love humans in the loop, your customers can escalate the chat and actually talk to a real human. The GPT Trainer chatbots have things like frustration detection that can better steer your chatbot's responses when a customer is getting frustrated, repeating a question, using foul language, or doing other things that indicate they're not happy with how the conversation's going. With the function calling capability, you can send customers where they need to go. For example, if they need to book a call, it can use APIs and webhooks to book the call to straight through the chatbot. If you need the bot to process a return on your company's behalf, it can do that as well. With this function calling capability, GPT Trainer can also be a great lead collection tool for your website, so when potential customers have a chat, you can collect things like their name and email and save them to a database for future marketing opportunities. They have some amazing documentation with video tutorials and step by step walkthroughs to make sure that you understand exactly how to use the platform to its full, fullest potential. You can learn more about GPT trainer over@GPT-trainer.com and if you use the coupon code Matt Wolf20, they'll hook you up with 20 off of any subscription plan on the website. Thank you so much to GPT trainer.com for sponsoring this video. And again, if you have an online business, it's definitely a way to stand out among the competition by being able to offer 24. 7 support along with all of the amazing AI features that GPT trainer comes along with. Hey AI Matt here. One last thing I forgot to mention. GPT Trainer also has a commercial license program. You can partner with them to sell chatbots or white label the entire GPT Trainer platform. Make money with GPT Trainer as well. Check it out. Moving right along. This was actually news that came out last Friday, but as you probably know, I record these videos on Thursdays, so Friday news usually makes the following week. But at the end of last week XAI announced Grok 1.5 with vision, and the benchmarks show that it's pretty on par with the other models that also have vision in them. We can see some examples that they shared over on the XAI website. It can write code from a diagram where they gave it a whiteboard with a diagram and then that whiteboard was turned into code. There's a handful of other examples here that you can take a peek at. I just checked my own GROK account because I do have access to GROK over on X. This feature hasn't been rolled out. From what I can tell, I had no ability to upload images but once I have access, I'll do some deeper testing on what this is actually capable of if you use the PO chatbot at all. They just released a new feature called Multi Bot Chat on PO where basically you can ask questions on PO and it will basically pick which model to use based on which model is best for the question that you asked it. You can also summon any bot that you specifically prefer by mentioning that bot bot. We can see in their little screenshot here, they asked help me understand inflation. Claude3opus responded. Then they agged web search and then asked about the current interest rates. It did a web search, answered those questions and then you can see there was also options for Gemini 1.5 Pro, Mistral Large and GPT4 so that you can change chatbots depending on what you need. Now, I sort of predict this is where a lot of the large language models are going to go. I think we're going to find that specific large language models are better than others at certain tasks. Some models might be better at coding, some might be better at writing, some might be better at medical analysis. We're going to have a chatbot that we interact with and that chatbot is either going to pick the best model to use to ask the question, or you're going to be able to tag a model and specifically ask it that question. I think PO is a little bit ahead of the game here with what they're doing because this to me is the future of large language models. I don't necessarily see a future where you're either a ChatGPT person or a CLAUDE person or a Gemini person. You use a single front end user interface, ask it questions, chat with it, and it's going to decide the best model to respond to that question. I'm not always right with my predictions, but I really think this is where it's all kind of headed right now. A couple weeks ago I talked about how Microsoft and OpenAI are teaming up together to build a hundred billion dollar data center so that they can really increase compute and sort of push their way closer to AGI. Well, Google answered back. And the head of DeepMind said over the next several years, Google is also spending at least $100 billion to build the same kind of infrastructure. We have Both Microsoft and OpenAI over here working on a hundred billion dollar data center for more GPU computer. And we have Google over here saying they're doing the same thing. We're building a hundred billion dollar plus data center over the years so that we can really scale up our AI. Efforts and we want to be the ones to hit AGI first. They didn't actually say that. I'm just kind of inferring it from like the whole storyline that's playing out here in the AI art world. There was some pretty big news here as well with the release of stable diffusion 3. Now there is a bit of an issue with stable diffusion 3 and that we don't really have any sort of front end user interface where we can use stable diffusion 3 yet. They released the API, they made it available so that people that are building software products can integrate stable diffusion 3 into their software. But because stability AI doesn't own clip Drop anymore, we don't have a free website to play around and test it yet. But this main image here was actually generated with stable diffusion 3. Even the text in this image was all stable diffusion 3. There's some more examples. We can see that it's pretty good with text in the images. I wouldn't necessarily say the image quality of the images themselves are like insanely far ahead of what we get out of Dall E3, and honestly probably not even quite as good as what we get out of Mid Journey, but it does do text really well. You can see another example here. They say it's not good to think in here. That's quite a big chunk of text that it managed to get into the image and follow the prompt pretty well. So although we don't have access to stable diffusion3 yet, in a easy user interface, it will probably roll out into a lot of AI image apps soon. One of it that I'm fully anticipating that it's going to roll out into soon is Leonardo AI. They've been way ahead of the game so far. One of the first tools I remember to have SDXL available in it. So in all likelihood we'll see Stable Diffusion three in Leonardo soon. Before we do, I was recently talking with somebody over at Leonardo and they told me about a new feature that they are dropping, like really soon, possibly even by the time you're seeing this video. They said that this weekend they're going to release a style transfer feature. And the example that they gave me was they uploaded this image as the style reference and then had Leonardo generate a handful of images but use the same style reference. And here's a few of the images they generated of like this person skydiving in that same style, somebody wearing a helmet and a futuristic cyberpunk world. Here's another kind of cyberpunky image in that same style, and another one and here's even another one. They didn't tell me exactly what prompts were used. They just told me that this was the initial style reference to get images like this and this and this. And again, I was told this feature is rolling out this weekend. So within the next few days, possibly by the time you've seen this video, that feature is in Leonardo. This week Microsoft also released some really cool research called Vasa1. Basically what it is is you can upload an image of a hedge, upload an audio clip and then it generates a talking video combining that headshot and the audio clip. Now we've seen stuff like this. Tools like Sad Talker did Heygen can do stuff like this right now. However, the big difference is the amount of emotion in the face, the way they move, the way they blink, the way they raise and lower their eyebrows and move their head and body around in the videos. Really, really makes it stand out. Like here's one of the examples.

Have you ever had maybe you're in that place right now where you want to turn your life around and you know somewhere deep in your soul there could be some decisions that you have to make.

And here's another example.

If you plan to go for a run and you don't have enough time to do a full run, do part of a run. If you plan to go to the gym today but you don't have the full hour that you normally work out, do some push ups, run up and down the stairs.

It's also going to work with like cartoon looking images as well, like this.

I'm not surprised. Not everything lasts. I've broken my heart so many times. I stopped keeping track now because this.

Is being put out by Microsoft and they're worried about the deep fake potential. We don't know when we're going to get access to this. They said they don't want to release it until they feel that it's safe to release. I don't know when they'll feel that threshold has past, but I don't know. I have a feeling we may not see this for some time. But I also believe that other companies not called Microsoft will probably make things very similar that we will be able to get access to even sooner. Personally, I think this would be really cool for things like podcasts. Like if I can't do a video to video podcast when I'm doing an interview and I'm just purely working with audio, I can come back and make a video, put my guest head in there, put my head in there and, and then actually make it look like we're having a real conversation back and forth. I don't think it's going to fool anybody. I just think it's a fun concept for a future video. If you can't actually record the video when you're doing the interview, things like that. I think it'll be really useful for if you're into creating 3D objects for, you know, game assets or 3D movies or things like that. We've got some new research called Instant Mesh that was released this week under an Apache 2.0 open source license where you can upload a 2D image and have it transformed into a 3D object that you can then download. And this one even has a hugging face demo that we can play with right now. For fun, let's take this image of a robot, toss it in there and see what it generates for us. I'll simply drag it as the input image. If I click generate, the first thing it's going to do is remove the background. So we'll see that happen first. Now we have our robot with no background and it's processing the 3D version. It took about 30 seconds. It generated multiple views and angles of what it predicts this robot to look like and then created this 3D object for us here that I can download as an obj or GLB file. Now, it's not perfect. You're not going to put something like this straight into a game, but you have a nice little rough draft that you can pull into, you know, Blender or something else and clean it up to get it ready for a game. Heck, I just tossed it into mixamo and got my new character to dance for me this week. The Spline tool also added text to 3D image right inside of their app. So if you head over to spline.designaigenerate, you can generate 3D images for about 24 bucks a month. Now, I do think this Spline tool is really cool, and from what I've seen, it generates pretty good 3D images. But on the other hand, we also have meshy AI, which in my opinion generates just about as good of images from what I can tell, and they have options to do it for free. Still, we also got some big news out of Adobe this week. If you weren't aware, the NAB conference was going on in Vegas. The national association of Broadcasters. It's a big event where a lot of companies show up to make announcements about what they're doing in the video world and a big talk of this year's event was AI, and this demo that Adobe showed off really blew some minds. We can see this scene here where they're looking in a briefcase on the back of a car, and there's a handful of scattered diamonds in this briefcase. They click on Add object pen selection, select this area where the diamonds are, and then type pile of large diamonds on black velvet cloth. And then click Generate. They get some samples to choose from. They select the first one, and now we have a video with a briefcase filled with diamonds. And we can see as the video camera moves around, everything kind of stays where it's supposed to stay. They also showed off the ability to mask and remove objects directly from a video scene. So they've got this sort of electric box here. They highlighted it, clicked remove object, and it was removed in the video. And you can see the video continues to play as normal. Now you can see a little bit of, like, where it used to be. Some of the lines kind of fade, but pretty dang good overall. More examples, they removed this cafeteria sign here, removed this backpack. In the shot here, this person looks at their watch and they actually change the style of the watch. They take this guy, toss a tie on him. And a feature that I will probably use a heck of a lot is this feature where you can extend a clip. So it uses AI to figure out what it thinks the rest of this clip will look like. They can click Extend. It shows it's generating. And now when they press play, the video actually continues to play longer than the actual video was. Now, one thing you'll notice, if you're looking really close, notice the complexion of her skin. This is something that the majority of the world probably won't notice. But if you are using AI tools a lot, these are the kinds of things you will look for to determine if it was, are you not? But if you look at her complexion on her face as we skim past here, when it gets to the generated area, you can actually see her skin, like, really, really smooth out. Look a little bit more kind of cartoony, but it's still pretty dang good. And the other thing they showed off during this is that there's other models you'll be able to use directly inside of Premiere. And I really, really hope DaVinci Resolve gets this, because I mostly edit with DaVinci Resolve. But this could be enough of a game changer that I switch back to Premiere. So check this out. They show a little dropdown showing that Pika is available. So if you use Pika to generate videos, well, you can use Pika in this upcoming version of Premiere, have it generate a chunk of video for you using Pika Labs. And they even showed off a version where they used Sora. So they click on OpenAI, they get a prompt view of a city skyline in the rain at night. It generates a video of a city skyline with rain. And it's actually incorporating the Sora model into Adobe Premiere. This might be our first ever chance to use Sora. It might go into Premiere first before we get any sort of like front end user interface to it, but I'm here for it. That's really, really cool. Also looks like they plan to integrate with Runway. So whatever your preferred AI video generation model is, whether it's Pika Runway or most likely it's going to be Sora, you're going to be able to generate videos directly inside of Premiere. And as somebody who creates videos for a career, I couldn't be more excited about this. I just really wanted DaVinci because that's the platform that I'm most familiar with. But this is wild. This is going to be a huge up level for content creators. If you make videos and you have the ability to generate with Pika Runway, Sora, do inpainting, outpainting and all of that kind of stuff all within your video editing platform that you use on a daily basis. I mean, changes the game for video creators and video editors. I did mention DaVinci Resolve. They did make some announcements at the NAB event as well, including two new AI powered features, one of them being a color grading, something that I never really do on these videos, but probably should get better at doing. I like to add a lot of color in my background, but I'm not very good at color grading. So I'm looking forward to AI color grading just to give my videos like a consistent look across the board, but also adding AI powered motion tracking, which if you have a studio version of DaVinci Resolve, they already have what's called the Magic Mask, which kind of does AI motion tracking. So personally I'm not totally sure about the huge difference here, but as a DaVinci Resolve user, I'm just excited to get my hands on all of it and play around with it. These features are in DaVinci Resolve 19, which is now available in slightly scarier. More dystopian news, the US Air Force confirms the first successful AI dogfight. So basically what they did was they have a jet that can fly by AI. They actually put humans into the jet so that they can override the AI and take control if they needed to. And they had it. Dogfight with a completely human controlled jet. According to the article, this dogfight was successful and the humans that were in the AI jet never actually had to take control. But they also didn't say who won the dogfight. They didn't say if the AI was better than the human pilot or not, just that they had a successful AI dogfight in real jets. And finally, let's talk about some AI enabled gadgets because this has been a really big talk of the AI space lately, with probably the biggest news being the humane AI pin. Pretty much every review that came out about it, it's fairly negative about it. There was a bunch of controversy about whether or not Marques should have titled this the worst product I've ever reviewed for now. But that's not what I wanted to talk about. In this video today, I wanted to show off some other AI gadgets that are popping up in the AI sphere lately. We've talked about the Rabbit R1. This is a device where you can train it to do tasks and then the next time you need to do that task, it could kind of do it much more autonomously. So if you need it to help you book flights or send a specific type of email or respond to somebody on Slack, and you have a fairly consistent way of doing it, you can train the R1 to do it that way every single time. It's a sort of pocket AI agent. Well, that itself isn't new news. The news around the rabbit R1 is that it is starting to ship this week also, a few months ago I talked about the Rewind pendant. It was this little necklace that you would wear and it has a microphone in it and it would listen to your conversations all day long. And at the end of the day you can go on your computer and sort of review the conversations you had. It was supposed to be like an augmented memory device. Well, this company Rewind and their pendant just rebranded as the Limitless Pendant. And it's no longer like a necklace. It now looks like this little clip that you clip on your shirt or on a pocket or something like that. And it does the same thing. It records all day long so that you can go back to your computer later and review conversations and get transcripts, notes and summaries of what you did and talked about throughout the day. But probably the biggest question that most people have about it is the privacy thing. If I'm wearing this little pendant all day long, what happens if people don't want to be recorded. I'm just recording everything all day long, whether people like it or not. Well, that's not exactly how it works. We can see in their launch video, they give this example where they're saying, hey, Sam, it's great to catch up with you. And then somebody responds, but it just says, new speaker. New speaker is speaking, but hasn't given consent to be recorded. And the person wearing the pendant says, this meeting is really important to me, and I really want to remember your advice. Are you okay if I record and transcribe our chat? The response comes back as, sure, no problem. It puts the person's name in here and gives a little check that consent was actually received. And now, instead of blocking out everything that speaker says, we can now go back and read the transcripts and see what that speaker said. To me, this actually seems like something fairly useful. It's kind of like a brain augmentation, right? You just walk around all day. I can see using this at networking events and conferences and things like that that I go to very often. I meet a lot of people, I get a lot of new names, see a lot of new faces, and then have to remember some of the conversations I had or some of the things that I agreed to do with people like go on their podcasts and things like that. Well, if I have something like this, I can go back at the end of the day, remind myself of all the deliverables and things I said I would do and conversations and names of people that I met, and just review it all really quickly. So I'm personally really excited about this one. I can see myself using this a lot, and I like that it has that consent thing in there, so I can let people know, yes, I'm recording, but it's not recording you. It can only hear me.