AI Just Took an INSANE Leap Forward!

Ladies and gentlemen, this week in AI news you can use. No, we're not gonna use that. You have no power here. What a productive week. And I believe this week is truly going to have something for everybody because we have updates from OpenAI. We have a brand new open source text to speech model that is indistinguishable from real people speaking. There's a new best in class open source LLM and I'll show you a super easy and free technique on how to craft stuff standup sets for you or your friends with a combination of two AI tools we already covered, plus meta open sourced llama. So we're also going to talk about that. Although it came out after recording this video, this is a massive release. Plus it's usable today. We'll talk about everything in a second here. There's no time to waste. The space moves way too fast. So let's dive right in and let's cover all of the brand new AI news that you could be using today. All right, first things first, let's talk about this text to speech model that is actually incredibly good. Now the amazing thing about this one is that it's not paywalled by a subscription or it's not some service that you can only use for an API. This thing has actually been open sourced on GitHub. Okay, meaning you could build on top of this and people will build on top of this. But also you can just use it for the hugging face space like so. So I'm going to pick one of these presets and I'm going to add something to the input text. All right, hit generate and here in the description box you could prompt it to generate it with different voices before I hit play. One interesting fact about this is that it has been trained upon 10, 10,000 hours of narrated audiobooks. Okay? So what you can expect here is a pleasant voice like you would expect from an audiobook. Let's have a listen.

Remember this is only the first iteration of the model, so maybe we will hear more about them on news you can use on the baggage bodifahruiters channel.

I was gonna put him. Okay, we're going to have to work on that AI Advantage YouTube channel part. Maybe give it one more generation.

Remember this is only the first iteration of the model, so maybe we will hear more about the man news you can use on the ad bant, but other channel.

Okay, it apparently doesn't know the word YouTube. What? Bro, what are you talking about? Because my guess is that just wasn't present in the training materials. But this was incredibly good, right? And you could just try it for yourself. I found this to be consistent among many different voices and messages that you send to it. Also, they're saying this is just a start. They'll keep adding more to the training data and making this better and better. So this is an absolutely incredible release that you're probably going to find in various apps very soon. Again, this quality level just hasn't been publicly available up until now. Okay, so on Thursday, April 18, Meta came out with Llama Free, their new state of the art open source model. And this is the big news here. This is an excellent model that beats out all other free models. And it's fully open source, which is absolutely amazing. But it doesn't end there. They're also integrating this into their entire suite of products, WhatsApp, Facebook messenger and even Instagram. And they released a ChatGPT competitor, an online chat interface where you can use Llama free for free without history. And it's not available yet in Europe. But let's start by talking about the model itself, because that's the big news here. So what did they drop? Two things, namely a 8 billion model and a 70 billion parameter model. So this is important because they're not trying to compete with GPT4 here. Okay, they're doing that too, but that model is still in training. They talk about this at the bottom of the announcement. They're training a 400 billion parameter model which will come out later this year. And this is going to be the GPT4 competitors. So today we got these two models which are nevertheless extremely capable. So how do they stack up on benchmarks? Well, they're really good. Especially this small model blows most other similarly sized models out of the water. And the 70B model beats all these competitors in this medium sized category, like Mistral Medium GPT 3.5 or Claude Sonnet. And meanwhile, the Chatbot arena has updated the leaderboard and 70B model is currently in fifth place. And as mentioned, the 400B model, that is going to be their open source GPT4 competitor at actually later on in this video we talk about an open source GPT4 competitor that came out this week. So hang around for that. But Meta's Llama 3 is going to release later this year and purely judging off the benchmarks, that's all we get with this one on the MMLU 5 shot benchmarks, it performs almost equally as well as GPT4. Look at this. In summary, they just open sourced the best in class, small and Medium sized model and soon they're going to open source A GPT4 competitor. Big now, why should you care? What are they going to integrate this into and what is available today? What about this news here? Can you use. Well, they have a separate blog post outlining all the features that this will enable for their other application. So from here on out, if it's available in your region. So you know, I'm going to use a VPN here in a second to access their new chat interface, but basically it's a question of time until everybody gets features. Like you're going to be able to tag at AI, AKA their Llama free model and ask questions in your conversations. Inside of WhatsApp, inside of Messenger. Heck, they're even putting it into their feed. This is the one that really surprised me. I wouldn't have expected them to put this into their main product, their main revenue driver. Right? Look at that. There's a button in the feed where you can consult the AI. This is making these things mainstream, right? A lot of people tried ChatGPT, but that doesn't mean they're using it regularly. That doesn't mean they're in contact with it on everyday basis. And if you didn't know, this is kind of a big deal because Meta has a whooping 3.98. They have 4 billion users worldwide on their apps and Meta AI is shipping to all of them. Well, you know, except Europe. And not just that you're getting the large language model. They also have a new image generation model which actually generates in real time. Spoiler alert, AI. It's not as good at what we're used to with the newest releases. I would say it's maybe at a Dall E2 level anyway, so that's all exciting news. But what about this? Can you actually use? Well, if you head on over to Meta AI again, I'm in Europe so I have to use a vpn. Here you have a chat interface. Okay, so this is a ChatGPT competitor with llama 370B running in the background. Now you can use these prompt presets and right away you can see this is way more consumer focused. Right? There's images along with the prompts. It doesn't even show the prompts, it's just shows you the outcomes. Quick side note, this is eerily similar to the prompt templates that I've been providing for free or with my course since over a year now. You have an image and the outcome. This is really the best way to teach people what you can do with these models and Meta knows that. So in this new chat interface you can use the chat box or you can use one of these and start interacting with it. It is speedy. It's at the quality level of a medium sized model which is very decent and very usable. Again, it's not a GPT4 or a cloud Opus, but it's very good. I also tested the number one thing which I test these days, which is the writing style. It's nothing for unique writing styles. I think Opus and Gemini are still the best right now. So again, the way to think about this is really as a better GPT 3.5 that is completely free meta AI. One thing that it does very differently than all the other models is it uses bullet points and these like paragraphs a lot. I think this might be a decision that makes it even more consumer friendly. Right, because then it's not a wall of text, but it's like segmented which makes it easier to use, I think. And that's the product they released now. And this is going to be available inside of all their other Native applications to 4 billion people. Huge. This is what mainstream AI adoption looks like.

Right?

And then also you have their image generator which is also going to be integrated into all their apps. Quick side note, if you have a Facebook profile in a country that doesn't have access to the AI yet, but you have a vpn, you can still use this. As I said, this is dali to like generations. I would say in other words it's good enough, but there's certainly better ones out there. Last thing that I want to underline again is that this thing is fully open source so people can build on this. What a week for open source to see. And that's really llama free. More coming later in the video. Okay, so next up, OpenAI announcement. This one is for the developers, but I think it might be interesting to non technical people too because what they did is essentially they upgraded their assistance API if you're not aware, it's essentially a GPT that you could call through the API or built into your applications or websites. And what they did is they 500x the amount of files that you can attach to it. So what they're really trying to do is they're trying to compete with a lot of the other solutions that allow you to add a lot of context. This is one of the main things you want to do if you use AI for a business, right? You want it to be aware of all the business context you have, not just part of it. Now, I played with this a bit today and I got to say there's a big difference in speed too. It's not just that you can attach way more. This thing actually became quite speedy. I am speed. If you used Assistance API before, you'll know that it wasn't the fastest and it was prohibitively expensive, which is still the case, but at least it's better now. There you go, already running. So this is really good. And I did attach two files here, but as mentioned, it could have up to 10,000 of them. Incredible. The main problem still being the very, very high pricing. Currently there's just so many other solutions which do a similar thing at a fraction of the cost. That's why this is not as widely used as OpenAI maybe would like for it to be. But with upgrades like this, step by step, this is becoming more and more capable. One quick side note here too is that they added this new feature that you could add projects and therefore organize your API costs via projects rather than by API keys. If you work with a team, you will appreciate this. And that concludes our quick little overview. Now let's move on. All right, so for this next one, I want to show you some new AI features of one of my favorite tools. You might have seen me use this before in other videos, Grammarly. Now, I've actually used Grammarly for years, and in some prompt engineering tutorials. I even recommend it as it corrects spelling mistakes while you type. And recently they added some upgrades to bring it to the next level and they reached out to sponsor this video. Namely, they're using large language models to add an assistant to your browser whenever you're typing something. So you can always just click this little G button. And here it is. As I haven't typed anything, there's no suggestions yet. But before we dive into these suggestions, I want to show you this customization feature because it actually adjusts to your personal context. It's sort of like custom instructions light. So let's just say I'm on X or twice Twitter and I'm writing a post here. Now if I select this, I could click this button saying rewrite with Grammarly. And up here you can set your voice. So this is sort of a light version of custom instructions inside of ChatGPT. And once you set this preset, everything you'll be doing with this extension will consider that context. As you might already know, the best way to use AI for content creation today is not for writing, it's for editing. And that's what this does really well, so I select this text and I'll simply say improve it and it will give me a suggestion on how to improve. Now, I could prompt it further here, as you can see, or use one of the presets here, but this makes it really simple. I can operate within the context of what I'm doing on the Internet right now. But there's also another scenario of you just staring at a blank page and not knowing where to start. Well, here you can use a tool. I just simply prompt it and there you go. Here's my email draft that I can now insert in the conversation. And as you can see, this gets me started in no time. I could customize and rewrite this further, as I just showed you, by simply selecting it and using this button to pre write. And for anybody wondering, yes, you can absolutely use this for other use cases like brainstorming. It's an integrated LLM, right? There you go. This took less than a second. Oh, and did I mention that all of these features are free up to a certain limit? That is you can run 100 prompts per month included in Grammarly's free plan. Plus, all these spelling and sentence structure suggestions that I've been using for years are also included in the free plan. If you Upgrade, you get 10 times as many prompts and strategic suggestions. These are in context recommendations on what you could be doing with your text in order to improve the impact. So go ahead and try Grammarly for free today. I use it as a Google Chrome extension, but they also have a standalone app and if you want to use it more often or use some of the advanced features, you can upgrade to the premium plan. Thanks again to Grammarly for sponsoring this video. And now let's get back to the next piece of AI news that you can put to work today. Because next up, we have probably my favorite for this week. It's a use case for something that we covered last week already. It's Yurio, this new music creation app that competes with Suno. It's really good. It went kind of viral across the Internet because it is so good. But people discovered capabilities of this app that go beyond just music creation. Namely, they found out that you could create standup routines with it and they actually sound really good. So it's not just for music. You can also create other things with it. So let me just show you this quick workflow. You can prompt the LLM of your choice with something like Write me a stand up routine about two AI assistants trying to hack each other. And we're keeping it simple here. We just want a quick stand up routine. Now a quick suggestion here is that you could totally use this to for example, roast your friends and then turn that into song and send it to them, right? This is all free by the way. Yes, I'm using ChatGPT4, but you could use any LLM, any one of the open source alternatives we cover here on a weekly basis or something like cloud, haiku or GPT 3.5. And there you go. Here you have a basic stand up routine. Now this won't work perfectly, so we will add something to the prompt here and there you go. For anybody intimidated by prompt engineering, this is how it goes. Just look at the outputs and you adjust your prompt according to what you like and what you don't like. Now I've played with this a little bit, so I know what we need. We only need the dialogue and we want to avoid the speaker names. Otherwise they're going to be spoken in our standup routine. Prompt is going to be in the description below. Okay, so I'm just going to hit copy down here, head on over to Yudio. And here it is quite simple because what you want to do is you want to switch from auto generated lyrics here to custom like so. And then I'm going to simply paste what I just copied from ChatGPT here. And at the top I'm going to say stand up comedy routine and you just hit create and there you go. It's going to create two routines. You can listen to these again. At the time of this recording, this is still in beta, so it's completely free. That will probably change over time. Okay, let's have a listen.

Alright, so the other day I caught two AI assistants trying to hack each other. Yeah, it was like watching a cyber version of Spy versus Spy. If anyone remembers that one. Says the other, hey, I bet I can get your root password. The other replies, oh, here I am.

Okay, so I don't know about the writing on this. This is totally up to you. And the LLM that you use, you could even write this by hand, right? But there you go. What a fantastic use case. And this is really what I envisioned for this show. I'm not interested in covering all the different drama that happens in the space. I want to show you different way how to use these tools for yourself, or what you could be building with them, or what you're going to see implemented in other apps soon. One quick side note, before we move on, people have been using this even more creatively and while browsing Instagram and TikTok. Over the last weeks, I caught all of these German TikToks that take conversations on ebay with different sellers like these awkward and funny conversations, and they essentially turn them into a song with Yurio. This is the newest Trend on German TikTok for faceless channels. Just thought that was interesting and I want to share it with you. With that being said, let's move on to the next piece of AI news that you can use. And this is going to be a quick one. But Suno Audio's competitor released this tool that allows you to explore all these different genres of music. This might be really useful if you're creating with Yurio or Suno and you're not sure what musical style you want to be creating in. Now, mind you, some of these are completely wild and don't even make sense. I mean, can anybody in the comments explain what hypnagogic shoegaze is? Or Goa Trans Afro Cuban jazz. Okay, so let's try some breakbeat Balkan brass bands. Okay, that's actually legit. Let's try one more. Okay, so this is a really funky one. Roots reggae, avant garde jazz. Swaying trees sing melodies, Rivers whisper ancient tales. That's surprisingly good. So if you're looking at my screen right now and you're like, I wonder what this one does. Well, this is your cue to open up the description, open up this link and sit down and explore some new musical genres that AI created. I mean, to be fair, some of these don't make any sense at all because they're combinations of two genres that have completely different BPMs. But AI just makes it happen anyway somehow. It's very interesting and fun to explore. And by the way, great job, Suno, for this incredible interface here. So there you go. Anybody who's into music, this is something new. All right, let's get a little more serious here because the open source space has seen some serious releases throughout this week. I want to touch on two of them. One of them is Rika AI and the second one is WizardLM. Now, Rika AI came up with a brand new multimodal LLM that actually beats out Opus, which is quite impressive, and on Human Eval, which is arguably the most important benchmark. But then again, you know my opinion on Benchmar, they're not to be trusted anymore as the model makers expect them. You could be using this model through the API today. And look at first, when these open source models came out, I tried to test them a little bit, run them through a few prompts, But I found that to be largely pointless over time. Only time will show if these models are actually any good. And you need so many test cases and at the end of the day, any opinion that I would give you right here, a day or two after release and maybe an hour or two of me using it would be super subjective and would do it justice because again, only time will show. But nevertheless, I want to point you in this direction and I want to forward the feedback that people have been having that the vision capabilities of this are actually surprisingly good. It's good to get more competition in this space. One thing that we can objectively talk about though is the fact that WizardLM came out with the new best in class open source LLM that actually beats GPT4. That's what they claim. That's what some benchmarks say. That's what people have been saying that are using it. This is legitimately the first model that even claims that mean it's a fact. But nobody else has been able to even claim that, not even on the benchmarks. And look, I have to say if you look into the paper, a lot of the data is based on this human evaluation where some people simply prefer this model over ChatGPT. They even admit that down here. But nevertheless, this apparently beats out all the other open source LLMs, but there's a little bit of a story with it because it was released this week and then it was deleted again. Why would they release something amazing and then deleted? Well, it turns out that it's been a while since they released the model and they're unfamiliar with the new release process now and they forgot a required item on hug called toxicity testing. So they deleted it and now they're adding it and they're going to publish this again soon. But I scoured the Internet for you and I found this hugging face by Alpendale. This is the hero of the week. While the model was publicly available, he downloaded the whole thing and re uploaded it here for us to enjoy. So if you want to use it before they finish updating it with toxicity testing, here you go. I personally can't wait for the Chatbot arena to update with these new models. We're gonna see how these stack up against the rest of the competition. But yeah, generally speaking the story has been that the 70B model of the Wiz Dell M2 is a GPT4 class model that is fully open source. So it took a while, but we're finally getting to a point where OpenAI's hand must be forced soon. Where is GPT5. We'll see. And you bet we'll cover it on this channel. Okay, to round things out, I have a few more quick releases here. So the first one is a hugging face base for image to 3D generator. That is actually incredible. We've seen a lot of these. I haven't covered the new ones in a while because they were all equally as meh. This one is impressive. Look at that. From this image it creates a model that actually is decent. Look at that. Just from an outside perspective, this looks better than anything I've seen yet from tools like this. Interesting. I'm going to try it with an image that should be really hard for it. This is a fine tuned stable diffusion model that I created over the last week. Let's see if it can cut it out successfully. Yes, that worked. And now it's going to turn me into a 3D model. Okay, this didn't work as well, but I have to say it's not as bad as many others that I've tried. The face and the hair are kind of messed up here, but this is also one of the harder tasks that you could throw at it. Always good to see the progress in the space. And let's move on to this one, which I was so excited to see at first because this is essentially a AI timelapse generation model. If you're not familiar with time lapses, they're essentially videos of a longer period of time that are sped up now. The demos where they showed this off looked absolutely incredible, but I actually spent over an hour trying to prompt this. I don't know if I'm doing something wrong. I even recreated the prompt that they had in the demo. This is very similar to what they showed this off with, yet the results are just not there. I wanted a two season time lapse progressing from summer to winter. So I don't know. The space is trending right now and people are using this left and right. I wasn't able to get what I wanted from it. Maybe you will get better results. It's all linked below. And then to round things out, I want to talk about a potential future use case. This is what we do sometimes. If I find something really interesting worth talking about. And this is really an interesting idea that I ran into on Twitter and I thought you might enjoy it. Basically It's Upwork where AIs hire humans. It's a bit crazy, right? But as you know, there's many tasks that AIs can't perform right now. And no, this is not a Real site, it's just a concept. Somebody's building this. I do not think this really has legs, but the idea is interesting, right? If things keep moving as fast as they are, we could get to this point where, yeah, autonomous agents are going to be pretty good, but they're not going to be able to do everything. And that's where something like this comes in. The creator of the agent might give it a budget and then the agent will go on a platform like this and the little gaps in its workflow it will simply fill in with human labor for a platform like this. I don't know, I just thought this was really interesting. I understand that this is a bit of a dystopian thought, nevertheless, an interesting way to round out this plethora of AI tools that we just had a look at. And actually there's one more story that I want to share with you. It's this Vasa 1 model that was published by Microsoft Research, which, you know, kind of surprising that Microsoft would be developing something in this direction. But then at the end of the day, I guess a lot of the large corporations just do this now and it's maybe less of a taboo than I would have thought because this is essentially deep fake technology where from a single image and a single audio clip you can generate human like avatars. Now look at some of these examples. This thing is not available today, will be available soon. We have similar ones that we covered that are available today. But this stuff is pretty wild and I really wanted you to see this. So, you know, sometimes nothing happens and sometimes everything happens all at once and you just kind of deal with it. The first thing we need to look at is the letter H. So the sound at the beginning, I would just say this is way better than many of the ones that we have seen before. By no means is this perfect, but I can totally see this as a customer support agents for many companies very soon. And while, yeah, it's not a human, and a lot of people argue that, hey, it's so annoying that all these companies are going to be useless as the customer support. It's at least better than those telephone bots, right? Unfortunately at the end they state that given the context, this is dangerous to technology with no plans to release an online demo, API product or any additional implementation details or anything really, until they're certain that technology will be used responsibly. If you found this interesting, covering apps and what's coming up in the AI space, then you'll enjoy what I'm about to share for the very first time on this YouTube channel, it's the opening of the AI Advantage community. I don't want to make too big of a deal around it. I created a whole sales page with all the details. But essentially what I learned over the last two years is that the best way to stay on top of this wild space and really the only way of implementing it into your own life is by actually using these tools. And the best way to do that is by surrounding yourself with other people who are trying to do the same thing. One of many things that we do in there is we run weekly challenges. And this week's challenge happens to be a battle between Suno and Yurios. So we challenge you to create two tracks, give you a detailed step by step on how to do it, and then people submit them here. And then there's a discussion in a one hour event where we evaluate the winners, take the learnings and share what we learned. If this way of educating yourself sounds enticing, there's a link in the description to it. As you might be able to tell, I'm beyond proud of what we were able to build here over the last year. We have nine people working on the community. Okay. Five of those full time. It's incredible. You should at least have a look at what we created here. There's so much power in numbers and it's really the best way to stay on top of these tools. And not just that, figuring out how to use them to your advantage. That's what the advantage community is all about. But if all you want is another video of AI news that you can use, and check out this playlist, because I do this show every single week and. And I will see you next Friday.