AI News is Getting Out of Hand!

So I just got back from a camping trip where I was without Internet for several days. It gave me the opportunity to make really cool Nerf videos and play around with some of the AI cameras that I have. But I missed a lot of stuff while I was gone. So I want to make a news breakdown and show you all of the cool stuff that I've come across in the AI world over the last few days. Let's dig in. So some of the most exciting research that's come out recently is this paper called scaling transformers to 1 million tokens and beyond with RMT. When papers like these come out, I like to use this tool called PDF GPT IO now to ask questions of the paper and to help better understand what the paper's actually talking about, because they like to use big words and sometimes I struggle with those big words. So this is essentially a way to add more memory to tools like ChatGPT, allowing it to remember things from very long sequences of information. It also uses the transformer network, which can divide long sequences into smaller parts and and use memory to remember information from previous parts. What this allows us to do is now use up to 2,048,000 tokens. So you know how when you're using ChatGPT or something like that, and you try to paste in something that's too long and it says, this is way too long for me to process, or it starts to process something and then cuts itself off. That's because it can only use up to 32,000 tokens in the very best version of GPT4. Right now, with this new research, in the near future, we'll be able to use over 2 million tokens. Now, the downside to this is it's going to require a lot more compute power, which means probably a lot more cost to use something like this. Now, to put this into easier to understand terms, one token is roughly 0.75 words. So if we bust out our calculator here and we do the math on 2,048,000 tokens times 0.75, so that's 1.5 million words that can be used in a combination of our input text and the text that's output back to us. Now, to put that into context, the entire Harry Potter book series, all of the books combined are 1,084,170 words. So you theoretically will be able to paste the entire Harry Potter book series into one prompt and ask it questions if you wanted to. Now, this is valuable for things like code. You should be able to paste in things like the entire code for a software product and have it answer questions about that code because there'll be enough tokens now to process it all. So this is really, really big news. Now we don't have access to this yet again. The largest model we have access to allows for up to 32,000 tokens, which is about 24,000 words combined input and output. So it's a pretty big leap once this technology actually gets implemented into some of these chat systems. Now, April 25th, the day I'm actually recording this video, was a huge day in the AI world. It was just news after news after news. It was one of those days where it just seemed like something huge just came one after the other, starting with Hugging Chat. Now, if you've watched my previous videos, you've seen me talk a lot about Hugging Face, where people can build an experiment and upload their machine learning models and let other people play with them. It's a really, really cool community driven machine learning platform. Today they just released their own chat platform based on Open Assistant's latest model. Now, if you're not familiar with Open Assistant, check out this video over here called a ChatGPT alternative that's free and open source. We talk all about Open Assistant and why it's interesting and why you should care about it. But by the end of the video, we kind of come to the conclusion that it's not quite there yet. You're not going to use it instead of chatgpt or probably even something like Bard yet, because it's just, it's still very, very early days. But that model is now available over on hugging face at huggingchat. You can find it over@ huggingface.com it's completely free to use and it's also open source, but again, it's not quite that great yet. When I said who won the most recent World Series, it said, the New York Yankees. How do you think they can improve next season? That was the entire response. And the New York Yankees weren't even in the World Series this most recent year. The other thing that I find interesting is it gave the title over here, the Boston Red Sox. What now? Also on April 25, OpenAI put out an article about new ways to manage your data in ChatGPT, basically telling us that ChatGPT users can now turn off chat history, allowing you to choose which conversations can be used to train our models. This is obviously an effort for them to overcome privacy concerns that people have had about ChatGPT and what information it's actually saving on you. So now you can tell it to no longer save and hold on to information that you enter into the chat if you'd like. Now, a lot of people probably didn't even realize that when you use ChatGPT, any information you were putting into it was being saved and helping better train the model. Now you can essentially opt out of that process if you'd like. Also on April 25, Replit announced that they raised $97.4 million at a $1.16 billion valuation and also held a mini live event called Replit Developer Day. Now, if you're not familiar with Replit, it's somewhat similar to like GitHub or Hugging Face, where people can collaboratively code. And at today's event they announced that they're releasing their own large language model, specifically fine tuned for coding. It's only 2.7 billion parame compared to almost every other platform, which is a lot larger, meaning that it's going to take a lot less processing power to use, which means lower costs, but it's actually performing better than most of these other platforms that are available. I was not the live event. They did a live stream. I was excited to watch the live stream once it was over on YouTube, but they pulled it down so I haven't been able to watch it yet. I'm having to make do with some of the screenshots that they put out. But long story short, repl.it is working on their own large language model from all of the data inside of the Replit deck database to help coders code. Also on April 25th, Nvidia released open source software called Nemo Guardrails. I was actually on a private call the other night with the people at Nvidia and got a demo of this. And this slide right here is probably the best representation of an explanation of what Nemo Guardrails does. So it's sort of a tool that sits in between your AI chatbot, like ChatGPT or Llama or the Stability Language model or Open Assistant or whatever language model you're using. It sits between the user and that language model and it allows you to have a set of rules, or guardrails as they call them, so you can custom tailor a large language model for your business with these additional guardrails. For example, a user might go to your custom chatbot, ask a question that you don't want to give an answer to, and it will hit this guardrail before it actually makes it to the chatbot, and the guardrail will send a response back saying, I can't answer that. If it can't answer it, then it'll send the question all the way through to the chatbot to answer the question. It's sort of a buffer between the user and the actual large language model, and you set the criteria of what can and can't be asked before the data makes it over to the large language model. This is open source. Anybody can use it. It'll work with any large language model. And they actually did a demo on this call that I was on the other day, and they showed how easy it is to set up. It's pretty dang simple. Anybody can set it up for their own chatbot. It's really easy. Maybe we'll dive into that in a future tutorial, if that's something that interests people. Also, on April 25, Yelp rolls out AI powered search updates and the ability to add videos to reviews. Personally, I don't really care about this. This just feels like Yelp jumping on the AI hype train and saying, look, we got AI too now, so I'm not going to spend any time on this. But that was in the news for April 25th as well. What now? In other news this week, the singer Grimes commented thoughts on AI generated music. With the rise of things like the Drake and the Weeknd song that got really popular and then got removed from all the streaming services, a whole bunch of people replicating Kanye songs, and all these artists and record labels having a big fit about all of this, Grimes is taking the opposite approach. She said, I'll split 50% royalties on any successful AI generated song that uses my voice. Same deal as I would with any artist I collab with, feel free to use my voice without penalty. I have no label and no legal bindings, and I applaud Grimes for this. I think this is awesome. It's gonna be really interesting to see how the music industry plays out over the next several years. Now that anybody could essentially train any musician's voice and make music with their voice, that genie is out of the bottle. Anybody can do it now with software that's trained on their own computer. There's gonna be legal battles, but people are gonna be able to do it anyway in the future. Looking back at this, artists like Grimes are gonna be applauded for leaning into the upcoming technology as opposed to fighting against it. But I also agree don't want their voices used and don't want people replicating them. They should also have that right. And that right has sort of been taken away from them. So I have very mixed feelings about this. I think what Grimes is doing here is really, really cool and this is something that will likely put her on the map in the future. When it comes to us looking back at the history of music progression, she's most likely going to be seen as somebody that was a very progressive thinker on this topic. I'm hoping more musicians do this and sort of open source their audio so that others can train on it and we get all sorts of new remixes and things like that. Now I think a lot of people are also probably going to use her voice without giving her royalties. There is a double edged sword here. And again, very mixed feelings, very mixed emotions. I'm interested in seeing how this all plays out. This is one of those issues where I can see both sides of the debate right now. Now, speaking of AI audio, there's an open source tool called Bark that you can find on GitHub over@GitHub.com suno AI bark and this is a text to speech generator. But it also includes a lot of other nuances. You can add size and laughs and you can make it sing and you can change the voice. Here's some examples are on the page. You can see the text prompt right here. And then when I press play we'll hear this back.

Hello, my name is Suno and and I like pizza but I also have other interests such as playing tic tac.

Toe so you can hear there's a lot more nuance in the voice and there's some laughing and there's some pauses. Here's another one where it actually switches languages. In the first part it's speaking Spanish and in the second part it's speaking English. But it maintains a Spanish accent. So listen to this one.

Buenos dias Miguel. Tu colega piem sequeto ale mines extremada mentemalo. But I suppose your English see it.

Actually maintained that accent when it went to English. Now you can install this on your own computer and run it. My friend AI entrepreneur put out a really good tutorial on how to install it. He does a great job of walking you through how to put it on your computer. So I'll link to his video in the description below so you can see how to install Bark on your own computer. If you don't want to install on your own computer and you just want to run it in the cloud, there is a Google collab that you can use right now. It's real easy to set up and run. Just press the Play button here under install. Once you get this Green check mark. You can minimize this area. Then click this play button under Basics here to do a few more install setups here. All right, so now all of the models are installed here and we can go ahead and generate something. You've got the basic model here, which was that one that we listened to a minute ago. We can change it to whatever we want. I could say, my name is Matt and I like tacos. Press play here to have it actually generate and process real quick. Takes a few seconds to do that. And now we can listen back.

Hello, my name is Matt and. And I like tacos.

And if I like the audio that it came out with, I can click on these three dots and download it. Now, there's some advanced examples like the one that said Buenos dias, Miguel in it. You can change this however you want. You've got one where it can be a man or and a woman, and it can actually change the voice back and forth.

I would like an omelette latte, please.

Wow, that's expensive. And then in this one, you can actually change different voices. Now, I'm not exactly sure how many different speakers there are, but we can listen to this first one here.

I have a silky smooth voice, and today I will tell you about the exercise regimen of the common sloth.

Now there are other voices. Let's just go ahead and put it on like English speaker 4. Press play to have it generate again real quick. And we should get a different voice this time.

I have a silky smooth voice. And today I will tell you about the exercise regimen of the of the common sloth.

But notice how it actually added some ums and some. Some like actual sounds that people make with their mouth in the text. Pretty interesting stuff. Now, I don't think this comes anywhere close to what something like 11 labs can do. It still has a little ways to go, but this is open source. You can install it on your computer. You can use it for free. And 11 Labs is not. It's called Bark. I'll make sure I link it up in the description below. The other thing that has been popping up a lot that is really awesome is this track anything. It actually uses the segment anything from meta, but it uses it for video. Here you can see this Avengers clip where it's actually segmenting out each character in the Avengers and separating them and finding them all in the video. Here's a cool clip of Steph Curry, how it's just finding Steph Curry on the court here and then following him around as he moves. I mean, that would be really, really hard to do in something like DaVinci or Adobe Premiere or one of those kinds of tools. And this can just kind of do it effortlessly. Here's actual use case of it for video inpainting. You can see you have somebody doing parkour running through here on the left. And on the right you can actually see the video where they were just completely removed and it's following the camera along. Let's play it again. Because you can actually still see the person's shadow. It didn't segment out their shadow, it just segmented out the person. So this is something that's really cool and will probably in the future be worked into a lot of people's video editing workflow. Because this ability to segment out things is better than anything we get inside of most of the video editor tools right now. And since since this comes from Meta's open source segment anything, this itself is also open source. So people will build off of it, iterate off of it, and it will only get better over time. All right, I want to wrap up with two really cool augmented reality things. I've been really getting excited about augmented reality. I think augmented reality is the next big thing. I think we're going to have a big wave of really cool augmented reality tech and toys and software come out as the augmented reality glasses and tools get cheaper and cheaper. I think augmented reality is just going to be everywhere pretty soon. So the first thing I want to show you comes from Ian Curtis here. If you're not following his Twitter, I highly recommend you follow him. He shares a lot of cool stuff. He does a lot in the augmented reality space. I love all of the videos that he puts out. Here's one that he recently made where he actually made an augmented reality Tetris. So he's just sitting in a coffee shop here, or wherever this is, and he's playing a giant Tetris in augmented reality. He said the total development time to build this was 1 hour, 54 minutes to build the game logic using GPT 4 and 6 minutes to build the game aesthetics. And there's this cool augmented reality game that I came across from Lee Vermillion where he made this prototype of a really cool augmented reality game where you've got this tube of dripping liquid and you have to grab tools from your little augmented reality tool wrist thing and redirect the liquid until it gets into this bucket. This looks like a really fun game that I'd want to play. When I was a kid, I used to love this game called the Incredible Machine on my PC. This looks like a augmented reality futuristic version of that kind of incredible machine. Build something to get to the desired result. I really think this is cool. This looks like a game that I would play the hell out of if I had it available to me right now. But this is just a demo. This is just a proof of concept that he's building. Can't wait till I get my hands on some augmented reality goodness like this. I'm going to be sharing a lot more cool augmented reality stuff that I come across because again, I think that's the next big wave that we're going to see. I think the combo of AI and augmented reality is going to collide and we're going to see so much awesomeness come out of it. And I couldn't be more here for it. So excited to see how that all plays out. And if you love all this nerdy stuff, you love all this AI tech and augmented reality and extended reality and maybe even some virtual reality stuff in the future. Check out Future Tools IO. This is where I curate all of the cool AI tools that I come across. Right now, I'm adding new tools every single day. I'm keeping the AI News page up to date every single day with all the latest info on what's going on in the AI world. And if all of this is too much, it's all very overwhelming. Join the free newsletter and every Friday I'll give you the TLDR of the week. I'll send you just five cool tools that came out this week. I'll send you the latest AI news from the week, I'll send you a handful of YouTube videos, and I'll send you one cool way to make money with AI. It goes out every single Friday. It'll be your TLDR. All you gotta do is head over to FutureTools IO, click on the button to join the free newsletter and I'll hook you up. So thanks so much for tuning into this video. I really, really appreciate you. There's a lot happening in the AI space right now and it's coming out at a rapid pace and I'm going to try to do my best to make videos and keep you up to date with all of it that's coming out. So if you like this kind of stuff, make sure you click the like button below this video. And if you want to see more videos like this in your YouTube news feed, click the subscribe button and I'll make sure you keep on seeing more videos like this once again. Thanks so much for tuning in. I really appreciate you. I'll see you guys in the next video. Bye.