AI News: What You Missed This Week

Despite it feeling to the outside world like the AI news is kind of starting to slow down, there's been a ton of AI news this week. No, there hasn't been any huge announcements like a GPT5 or Mid Journey version 6 or anything. Just absolutely insane. But there's been a lot of little advancements that when you start to really compile all of this stuff, it adds up to a lot going on in the AI space. Now before I get into it, I want to thank this week's sponsor. This news breakdown is brought to you by Shopify and I'm going to talk about them a little bit more in a minute, but let's just get right into the news. So this actually happened last week, but it didn't make last week's news video and I didn't want to miss talking about it. But Meta actually released an AI music generation model that's open source and freely available for anybody to use on Hugging Face. And in my opinion it sounds a lot better than the other models that we've seen like Rifusion or Music LM from Google for example, one of the samples here, a 90s rock song with electric guitar and heavy drums and we get something that sounds like this.

Or we can do an early 2000s pop punk anthem and see what we get. Now there are a lot of people playing with this. As you can see. It's going to take actually a few minutes to generate this. But do keep in mind that this is open source. It is available on GitHub. You can install it locally if you have a graphics card with at least 16 gigabytes of memory, which you know is mostly the higher end graphics cards. But it is possible. There's also a Google Collab notebook that you could play around with this as well. And here's our pop punk anthem.

Pretty spot on if you ask me. Now also in last week's news, but it didn't quite make my Last Week video Adobe made some updates to Adobe Express. If you're not familiar with Adobe Express, it's kind of like Adobe's version of Canva, I guess is probably the best way to describe it. Well, among the announcements about Adobe Express, they announced that Adobe Firefly Generative AI comes to Adobe Express so you can prompt images and add text effects using AI directly inside of Adobe Express now and this was announced this week, but since we're on the topic of Adobe, let's bring it up now. Adobe also added generative AI tools in Adobe Illustrator as well. The main thing that they added is called Generative Recolor, which allows you to do things like upload a black and white image and have color added to it, or upload a design like this box on the left and have it regenerated in these other color schemes like the one on the left and the right. It uses AI to to find color palettes that look good together essentially, and then recolor your additional image in these various new color variations. So if you have Adobe Creative Cloud or just Adobe Illustrator, you can update to the latest version of Adobe Illustrator and Generative Recolor will be available now. Now, up until recently, one of the biggest complaints about OpenAI and GPT 3.5 and GPT 4 is that the model has been closed. OpenAI hasn't opened up the source code for anybody to view. As the various governments start to worry about the longer term implications of AI, companies like OpenAI and DeepMind have decided that they will open up these models when the governments start asking for access to them. For example, they agreed to open up the models to the UK government. According to this article in Politico here, Google, DeepMind, OpenAI and Anthropic have all agreed to open up their AI models to the UK government for research and safety purposes. Also, many people, including myself, have noticed that AI features have started to roll into some of the Google products. In fact, now inside of my Gmail I have this little button down here that says Help me write. If I click on this button, I can click Help me write again and then give it a prompt for what I want it to write an email about. Write a promotional email telling someone to go watch my latest YouTube video. Make it persuasive. I'm writing to you today to tell you about my latest YouTube video, which I think you'll really enjoy. It's a video about the AI news. I've also noticed this feature is now available, at least for some people, inside of Google Docs as well. If I create a new blank Google Doc immediately a button pops up that says Help me write. Once again. I can add a prompt here, write an article about how AI is changing the world. And just like that I've got an article written about how AI is changing the world. I can recreate it, I can refine it by formalizing it, shortening it, elaborating on it, or rephrasing it. Let's go ahead and elaborate. And now we've got a slightly longer article about why AI is changing the world. Now, yes, we already could have done this by using things like ChatGPT or Bard or one of the existing large language models that are out there and just copying and pasting. But it's cool to see that Google is just adding it into these tools and it's saving us a few extra steps of having to do that. Also this week, Meta announced that the first AI model that's based on Yann Lecun's vision for more human like AI is being open sourced. Now, basically what this model is, it can take a piece of an image for context and and then figure out the rest of the image by applying what it already knows about the subject matter of the image. Now this is very oversimplified, but essentially you can give it a tiny piece of a dog's head based on its training data, it knows what the rest of a dog looks like and it can fill in the gaps. You can give it a tiny piece of a bird's leg and based on what it knows with its training data, it could fill in the rest of the bird. You can see, here's an example with a wolf, here's an example with a building here where most of these AI generation models try to recreate images at a pixel by pixel level. This actually thinks more like a human, where it says that looks like it's a part of a dog. I know what the rest of a dog looks like. I will take what I know about the rest of a dog and draw the rest in. Again, this is an oversimplification of the model, but that's probably the best way I know how to explain it is that it's trained like a human and that it knows the context and it can look at a piece of something small and then fill in the rest of the context based on its data set and what it knows about that smaller piece of information that it's seeing. And once again, as is the trend with Meta lately, they are open sourcing this model and it is available right now on GitHub. Also earlier this week AMD put on a live event up in the Bay Area. If you're not familiar with amd, they create chips similar to Nvidia. They're really Nvidia's biggest and kind of only competitor right now. And during this event AMD had really two big announcements that were relevant to the AI world. So the first big announcement is that AMD is partnering with Hugging Face. Now, I've talked about Hugging Face a lot in past videos. It's a place where people can upload their machine learning models and their code similar to like a GitHub but you can also test the models directly on the site and play around with them yourself and essentially stress test them. When we were playing around with Music Gen just a few minutes ago in this video, that was a Hugging face space. With a lot of people building these machine learning products and using hugging face, it's actually a pretty big deal that AMD is going to be the company that's providing the compute power behind Hugging Face instead of Nvidia in the future.

Today we have 15,000 companies using our software and they have shared over half a million open models, data sets and demos like some you might have heard of, like Stable Diffusion, Falcon, Bloom Stock Coder, Music Gen that has just been released by Meta a few days ago. We will optimize all of that for AMD platforms. The goal is to really have the best combo between hardware and software. Hopefully this collaboration will be one step, a great step to democratize AI even further and improve everyone's life.

Now the other big announcement that they made at this event was that they are starting to build hardware specifically tailored for AI. And they even called out Nvidia saying that they're trying to make chips that are more powerful than Nvidia's chips.

We call this Mi 300X.

Now they do use a lot of acronyms and techy language. That's just totally over my head. But really the point being, AMD is trying to go head to head with Nvidia and create chips and that are great for training large language models.

To address the larger memory requirements of large language models, we actually added an additional 64 gigabytes of HBM3 memory. I am super excited to show you for the very first time mi 300x we truly designed this product for generative AI. It combines CNA3 with an industry leading 192 gigabytes of HBM3. I love this chip. By the way, when you compare Mi 300X to the competition, Mi 300X offers 2.4 times more memory and 1.6 times more memory bandwidth.

So you'll notice that in their screenshot they mentioned the competition. But on this screenshot they're saying compared to Nvidia's H1 hundreds, it's 2.4 times the HBM density and 1.6 times the HBM bandwidth. Basically, they're saying they're better chips. Now, these chips likely aren't going to be chips that you would get in a consumer PC, at least not anytime real soon. Seeing as they're comparing them to the H100 which could cost $30,000 or more. And when you compare this chip to Nvidia's H100 that has 120gb of memory, while this one has 192 gigabytes of memory, again, a lot of technical stuff, most of it doesn't have huge implications on us, the consumer yet. But it basically means that this technology is getting better and better and Nvidia and AMD are sort of going head to head to create these GPUs so that large language models can get trained faster, bigger, better, and eventually build things like the next GPT5 or the next model of Google's Palm. Now, speaking of GPT, OpenAI made an announcement this week about a handful of updates that they've made, including the new 16,000 context version of GPT 3.5. That's the standard version of ChatGPT. Before it had 4,000. The old version allowed you to use roughly 3,000 words between input and output of what you plugged into GPT and what you got out of ChatGPT. This new 16,000 context model is four times larger, so you should expect to get about 12,000 words between your input and your output for ChatGPT. They also announced that it was going to be cheaper to use GPT 3.5 with the API. So the developers that are building with the GPT 3.5 API in the back end, their costs are actually going to go down. Whether or not those companies actually pass the costs along to the end consumers who are using these tools, that's yet to be seen. It didn't seem to happen last time. The APIs got cheaper, most of the companies kept their prices the same and just pocketed more profit. But the companies that are developing with GPT 3.5 API, they are going to get even lower costs now. Also in OpenAI news this week, it came out that Maybe Microsoft and OpenAI, although they seem to have an amazing partnership, maybe they don't always see eye to eye. The news that's been coming out recently is that OpenAI reportedly warned Microsoft to move slowly on integrating GPT4 into Bing Search engine to avoid the inaccurate and unpredictable responses that it launched with. However, Microsoft went ahead despite warnings that it might take time to minimize the inaccurate and strange responses. Now Microsoft, to their credit, did kind of fix this quickly and fix a lot of the weirdness that was coming out of the chatbot early on. But the Wall street reported that there's tensions between the two companies as they work together and compete on AI features. And you know, in the past I made videos about why would somebody purchase Chat GPT if they can just use Bing for free and it has GPT4 built in. And it sounds like that's also kind of a concern that's going on inside of Microsoft and OpenAI right now. You can see in this article right here it says Microsoft and OpenAI have a rather unique partnership that has led to some conflict behind the scenes as the two companies simultaneously support and compete with each other. That's because they both have chat models. You've got ChatGPT plus and you've got Bing Chat, and both of them have GPT4 built into them. Anyway, moving right along, it was announced this week that Sir Paul McCartney says artificial intelligence has enabled a final Beatles song. McCartney told BBC radio that technology had been used to extricate John Lennon's voice from old demos so he can complete the song. We just finished it up and it'll be released this year, he explained. We don't know the name of this song or necessarily when it's going to be released. We just know that we might get a new Beatles song pretty soon thanks to AI. Also this week Google announced a new generative AI model for virtual try ons. Users can generate try on images with AI. So for instance you can be selling a top online and upload a picture of yourself and see what you would look like wearing that top. It uses diffusion models so similar to what we get out of stable diffusion Leonardo Midjourney tools like that to add the top onto person wearing it. This isn't some existing model that they just took and applied it to this use case. They actually created a new model using Google's Shopping Graph which has all sorts of images of people wearing different types of clothes and trained it on that imagery. They say that starting today you can use virtual try on for apparel on women's tops from brands across Google Shopping Graph, including Anthropologie, Loft, H and M and Everlane. And over time it will become more precise and expand to more brands. Now speaking of the aforementioned Google Shopping graphics, the Google Shopping Graph actually works with today's sponsor, which of course is Shopify. Now I know I talk a lot about tech and AI and all of this stuff that may sound very complex, but when it comes to setting up an online store, you really don't need to be a tech genius. Literally anybody can set one up and I know a lot of the people that watch these videos have their own online presence and are content creators yourself and a great way to earn some extra income as a content creator Whether you're using AI or not is to sell some of your own merch. Now Shopify, as many of you probably already know, is a commerce platform where anybody can set up their own store to sell physical products. They're making the complexities of running a business simpler so that anyone, anywhere can now become an entrepreneur. Now, next week I have a call with a team who's going to help me set up a line of merchant for the Future Tools brand so that we can start selling shirts and hats and cool AI themed goodies. And of course, when I set that up, I'm turning to Shopify to set up the store. Because while I love tech and I love learning about all the latest AI technology, I still want things to be easy to set up and easy to build. And Shopify just makes setting up these stores easy. That's the power that Shopify provides. They're democratizing technology for entrepreneurs and they're helping build tomorrow's economy today. And the best part, if you go to shopify.com Matt Wolf, you can get a free trial and start setting up your own online store. Now, I'm going to get back to the AI news. But remember, you don't need to be some sort of tech engineer to create an online store. Sometimes all you need is something really simple and easy to use, like Shopify. So thanks again to Shopify for sponsoring this video and allow me to live my dream by keeping up to date with all the AI research and sharing it with you for a living. Now, if you are going to set up an E commerce store at some point, a really cool way to do it and drive people to it could be using this new trend that has been absolutely flooding Twitter lately, which is the trend of using Control Net and Stable Diffusion with QR codes to make really, really amazing images. And when I say it's a trend, I mean literally everybody is talking about it. This thread from Rowan chung here got 9.6 million views. My good buddy Linus Ekinstam has been talking about this trend. I came across this blog post over on stable diffusionart.com which breaks down how to set this up. Copy Sutra made a three step tutorial on how to do it here. Michael Gao created a tutorial on how to do this here. AK on Twitter showed off that there's actually a model that you can download directly from Hugging face and do it yourself inside of Stable Diffusion. And most recently Rohan Chung again made a tutorial with 1.3 million views and he published this tutorial on the day that I'm recording this video. Now this is a really hot trend and the images that are coming out of it look really, really cool. But if I'm being totally honest with you, I've tried to generate these myself and it seems like right now you can either get something that looks really cool or a QR code that works. So far I've had a really, really hard time actually getting something that looks this good. And the QR code actually works when I scan it with my camera. So that's a nut that still needs to be cracked. But I think somebody's going to figure it out because some of them did work. It seemed like when I was playing with it earlier, I did get a few of them to work, but the ones that I got to work were usually the ones that looked least cool. The ones that had the most detail and really looked cool didn't seem to scan. So while I love this trend and I love the art that's coming out of it, and I do think it's going to be a way to send people to places like your stores in the future, it's not quite there yet. But as of right now, it's still something that I love seeing all over Twitter. And it's also something that I had on my list of to make tutorials. Once I crack this nut and figure out how to make images that both look good and scan, most of the time I'm going to turn around and make sure I make a YouTube tutorial about that for you. But for now, I'm going to link to all of these Twitter threads that show you how to do it in various methods below in the description and you can try any one of them and see what you come up with. Also this week Google released this blog post about eight ways Google Lens can help make your life easier. And one of the things that they mentioned is it can actually help you search for skin conditions. You can literally take a picture or upload a photo through lens and you can find visual matches to inform your search. It says this feature also works if you're not sure how to describe something else on your body, like a bump on your lip, a line on your nails, or hair loss on your head. So Google Lenses computer vision technology can actually help you essentially diagnose skin conditions. Now, I wouldn't use this instead of a medical doctor, but if you're curious or worried about something unusual, it might be a good way to double check and ease your mind. Now this is some technology that I'm really, really excited about. We've all seen videos like this one that I made in stable diffusion where it's got a lot of flickering and a lot of little like extra artifacts going on. Well, this new research called Re render a video zero shot text guided video to video translation was just announced and it looks to solve that. You can see with this input video they were able to generate these videos and they do not have that same flickering effect to the videos. In fact, they have this little demo here where you can see what a typical stable diffusion video looks like with the flicker and the change and you skim across it and you can see what their model does and how clean and how clear this looks. They also have some demos down here comparing it to other video models that are out there. Like here's this input of this statue and here's the various other models that are out there and how they generate the same video to video. Here's another one where they have this girl input and they generated in three different videos here and all three of them just look excellent without the flicker. Another one here, all of them just look so amazing. You don't have that flicker. This is going to make taking real videos and turning them into animation something that anybody can do without all the flickering, without all the weirdness, without all of the coherency issues that we've seen out of other models so far, like stable diffusion and even like models like Gen1 that are out there. We don't actually know when this one's going to be available for the normal public to use, but it's one that I'm really, really excited that this research even exists because it means that we'll probably see it in places like hugging face and being added into other tools if an API is ever made available for this kind of tool or if this is open sourced and made available for anybody to use. Also, every single Wednesday midjourney does their office hours where they essentially break down everything that's going on with Mid Journey, all the news, everything that you can expect from it. And I try to catch as many of those office hours as I can. I can't catch them every week. This week I didn't catch them, but Ali Jules here on Twitter did. And she made a nice little tldr for us about what we can expect from Mid Journey after their recent June 14th office hours call. And here's her quick recap. She says that that version 5.2 of Midjourney is expected any day now. I mean it literally could be out the day that I'm dropping this video. Cause I do record these on Thursdays and release them on Fridays, so by the time you're seeing this, it might be live already. A few months ago they did say that version six they thought would be ready in six to eight weeks, but now it looks like they're estimating a release in July. It's a little bit delayed and here's what they said we can expect about version 5.2. They said it's coming in a matter of days. It's gonna be similar to version five and version five Point. It said it'll have limited Discord compatible outpainting, so you'll be able to zoom out, change the aspect ratio, or change the prompt between zooms. Now, I don't totally know what this means exactly because I wasn't on the office hours call myself, but it sounds like we're going to get some new toys to play with Inside of Midjourney. They also talked about their prompt analyzer, which right now you use by typing slash, describe and then uploading an image. They said they don't know if it's coming with 5.2 or not, but it is going to reduce the word barf prompts where it shows you prompts that don't seem to have a whole lot to do with the image. They are working on a web and mobile standalone version to finally get it out of Discord, but it's going slower than they'd like and they're trying to improve moderation so less words that you enter get denied and slapped down when you try to use certain prompts. Thank you so much Ali Jules, for sharing that little recap. That's super helpful for those of us that missed the midjourney call on Wednesday and then finally wrapping up the news for the week. Eleven Labs, who is known for generating ultra realistic text to speech voices. I've used them a lot in past videos. They just introduced a speech classifier, which is essentially a tool where you can upload audio into 11 labs and it will tell you whether or not it was generated with 11Labs. 11Labs has gotten some heat in the past because they're probably the easiest tool to go and take anybody else in the world's voice, plug it into their system and get that person's sort of deep faked version of the voice out of the system. But they also seem to be leading the charge in figuring out how to fight against that as well by creating technology where they can tell you whether or not that AI voice was generated with AI to sort of help cut down on the Deep faking that is going to be inevitable as we enter this world of AI being more and more prolific. Today we're thrilled to announce our authentication tool, the AI Speech Classifier. This first of its kind verification mechanism lets you upload any audio sample to identify if it contains 11 labs AI generated audio. They go on to talk about how they're taking a proactive stand against the malicious use of AI. And I for one am definitely down with that because although I love this AI stuff, the implications towards deep faking and fake information and the spread of spam and scams, that's the part that scares me the most. So I love it when these AI companies take moves in the right direction to help prevent that kind of stuff. And there you have it. That's really the breakdown of all of the news that happened in AI this week up through Thursday. There probably is more news that's coming out on Friday that I missed in this video, but anything I missed, I'll make sure it comes out in next week's weekly news breakdown. I've got some really, really excited videos in the works for you guys. There's so many videos that I want to make, including some tutorials on some really cool chat GPT prompts that use some of the plugins now, some really cool new mid journey ideas, some cool tech that's coming with Leonardo. I'm working on some AI music generation tutorials, some tutorials around Work Warp fusion. So much cool stuff. If you're not subscribed to this channel, now's a good time to do it. I've got a lot of stuff that if you're in the AI space and you want to keep up with the news, or if you want to learn how to do all of the things that you're seeing spread all over the Internet, all of the viral AI videos that are going all over TikTok and Instagram and Twitter. If you want to know how to do that, I hope my channel is a channel that can keep you in the know on how to do all of that stuff. I'm new to this whole having a popular YouTube channel and so I'm still trying to figure all this out. So sometimes my videos are hit, sometimes they're misses. I'm trying to figure out what you like to watch and make more of that. I'm also trying to figure out what I like to make and also make more of that. And I hope you guys are here for it because I'm having a blast doing this channel and I hope you're learning a lot. I hope I'm keeping you informed and also helping you understand how a lot of this stuff is working because that's really my goal. And once again, thank you again to Shopify for sponsoring this video. I really, really appreciate it. And if you haven't Already, check out FutureTools IE. I update the news on this site every single day. So if waiting till Friday to get a breakdown of the news isn't fast enough for you, I'm sharing the news in near real time over at FutureTools IO and over on my Twitter account rflo. So check out those places. Join the Newsletter I send a newsletter every Friday and yeah, lots of calls to action. Probably too many, but there's plenty of places you can follow me. I'm all over the place online these days. Days I'm done word vomiting at you. Once again, really appreciate you. Thanks so much for tuning into these videos. I hope you keep coming back, checking out more of them, subscribing, liking them, pressing the thumbs up, pressing the bell, doing all that stuff. Love you guys. See you in the next one. Bye.