More Crazy Cool AI Breakthroughs

Welcome back to another edition of Cool Stuff that Matt thinks you should see that has to do with AI. Should probably work on that title. In this video, I'm going to show off a handful of things that I've come across over the last couple of weeks that are cool AI tools that you can use right now, as well as interesting research around stuff that you'll be able to use in the future, starting with the CSM tool, AKA Common Sense Machines. Now, this is a tool I did talk about in a previous video, but last time I talked about it, I didn't have access to it yet. You can find it over at 3D CSM AI and as far as I can tell, it's completely free to use for now. But I have a feeling it uses quite a bit of compute power, so I don't imagine it's going to be free forever. But I've been having a lot of fun playing with this one. What this is is a tool where you can upload a 2D image and it will convert it into a 3D image. You can see this headshot over on the left. This is my profile pic for YouTube and Twitter and all of the places I uploaded this image and had it generate a 3D version. And you can see, here's what it creates. Created. I can actually rotate around, look at it from other angles. It actually came out pretty dang good. I mean, the lighting and the colors are quite a bit different from my original image, but it's pretty dang coherent based on what I uploaded. Basically, what I think it's doing is you upload a single image from the front. It then tries to estimate what that image would look like from various angles. From the side, from the top, from the bottom, every angle it can think of. It creates new images of these different angles and then creates like a Nerf, sort of stitch them all together and give you this 3D version. Now, I will tell you, this does take quite a long time. From the time I uploaded an image, the time it gave me a 3D rendering back, it was hours and hours and hours. I don't know exactly how long it took because I uploaded an image and it wasn't ready before I went to bed, it was ready the next day. So after uploading this image and getting this result, I wanted to try it with this image, which was a silly image I made in stable diffusion of what I would look like if I was Buzz Lightyear. And it gave me this 3D image. Kind of has a wonky arm here, but other than the wonky arm. It actually is pretty decent. I was pretty impressed with how well this came out. I kind of lost half my hand here, and it tweaked the color of my hand and made a few mistakes. But I mean, dang, it's pretty impressive what this can do. So I thought, why don't I take a real image, a real full body shot of myself and upload this image and see what comes out. And, well, the results weren't quite as great. I kind of got a tumor on head or something. It added buttons where they didn't exist, and I think maybe it thought I had a wristwatch or something, but I didn't in my original image. So it took some additional creative license on this. It also completely got rid of my nose. But I mean, this stuff is coming a long way considering that the most recent model that I played around that was even kind of close was this shape E model that generated images that looked like this. So, speaking of 3D, here's another piece of research that came out recently called Text to room, extracting textured 3D meshes from 2D text to image models. With this research, people will be able to give a single prompt of a room they want to see. And it will actually generate that entire room. A living room with a lit furnace couch and cozy curtains, bright lamps that make the room look well lit. And then you can see. Here's what it generated. It actually generated the full room in 3D and it created a depth map to the room. And that all came from a single text prompt. Here's another one, editorial style. Photo. Rustic farmhouse, living room. Stone fireplace. Wood, leather, wool. And it generated this entire 3D room with depth map, once again, all from a text prompt. Here's another one, editorial style. Photo. Eye level. Coastal bathroom. Clawfoot tub. Seashell, wicker, blue and white. And once again, it generated an entire 3D room that can be navigated around and see all of the elements that were originally prompted. Now, the way this one essentially works is when you enter your text prompt, it generates a single image of what it thinks the room should look like. It then uses InPaint, which, if you use Midjourney Zoom out feature or Mid Journey's Pan feature, that's using InPaint, or if you've used Adobe's Generative Fill, that's also using inpainting. It basically takes a piece of the existing image for additional context and then generates another image of what it thinks should be the next piece of that scene. And you can see as it generates that image, it creates more and more of the map of the room. As you can see up here in this left corner, it's inpainting more and more of the image while also mapping out what the room looks like. Once it's done, it's sort of stitched together this entire room here. And then once the room is stitched together, it then estimates the depth of the objects in the room to give that true 3D feeling to that room. Now, this is one that doesn't appear to have a demo yet on a place like Hugging Face or Google Collab. However, the code is available on GitHub and if you know what you're doing, you can actually install this one on your own local machine and get it running. Here's another interesting piece of research that I recently came. We don't have access to any of it yet. We don't have the code. There's no demo of it up yet. But it looks like both are coming soon. But this one's called Any Door Zero Shot Object Level Image Customization. And if you look at the examples they give here, they take a picture of a sloth, stuffy here, and they take a picture of this windowsill here that has a watering can with some flowers in it. They mask out the watering can with the flowers and it replaces the watering can with the flowers with the image of the sloth. Here's another one where it has a picture of a poop emoji, a picture of a desk. They mask out an area of the desk and then the poop emoji appears on that area of the desk. And here's another one with a backpack, a picture of a living room with what looks like a really old school TV on it. They mask out the TV and it replaces the TV with this backpack on the stand here. You can see some other use cases for it. Object moving. They actually move this vase here and this cutting board or whatever this is, they actually move them to different areas on the table. And the rest of the image maintains exactly what it looked like before. Here's one where two cars actually swap places. You've got black car on the left, a red car on the right. And using this technology, it can actually swap the two cars, leaving everything else in the image the same. It also appears to work with multiple objects. They have this photo up here of a giant green grass field with all of these various images, various dogs, a couple different little plushies here. And it puts all of those into this single image here. Another use case for this virtual try Ons upload a picture of somebody, mask out the area that you want to put the clothes on, and then it shows what that person would look like wearing the specific clothing here. So this is one that again, we don't have access to yet, but it looks like we're gonna be getting a demo and the code available soon. And this one should be pretty fun to play with. I mean, just the memeability alone is gonna be huge. All right, so this is a really fun one. I've been seeing a lot of people talk about this one on Twitter lately. It's called Chirp and this is a text to song generator. You can basically give it lyrics or have ChatGPT generate lyrics for you and then it'll turn that into a song. Now, it doesn't appear that get to choose the style of song it generates. It just takes the lyrics and generates whatever style of song it wants. But it's actually pretty dang good. Here's one where it took Blinding Lights by the Weekend and it took a little clip from that song and just turned it into a completely new song.

And here's another one where they took Get Lucky by Daft Punk and Pharrell Williams and they took a clip from that song and just generated a brand new song out of it.

The present has no ribbon it keeps on giving.

What is this? I'm.

If you wanna leave, I'm with it.

We've come too far.

To g up who we are.

I mean, pretty good. Similar to Mid Journey, you actually have to do it inside of their Discord bot. So if I jump into this Suno Discord bot. Suno is the company that created this. Here we can go into one of these Chirp Alpha rooms and if I just type slash chirp, it brings this little dialogue box where I can enter my own lyrics or I can have chat GPT make lyrics for me. So let's have chat GPT make lyrics about subscribing to Matt Wolf's YouTube channel and click submit. And what it'll do is it'll actually generate two variations of the song. You can see it already generated the text. Subscribe to Matt Wolf's channel. Join the fun. His videos are hilarious. Second to none after guaranteed you won't be bored. Hit that subscribe button, you won't be ignored. And it's going to give us two variations and hopefully one of them sounds pretty decent. And it generated songs. It took about a little over two minutes to generate them. But let's listen to variation number one.

Subscrib to Matt Wolf Channel Join the Fun his videos at Hilarious Second to None Laugh Guaranteed you will be bored. Hit that subscribe button, you will be ignored.

And here's our second variation.

Subscribe to Mac Channel. Join us Bar.

These videos are hilarious Second to none.

Laughter Guarantee you won't be bored. Hit that subscribe button, you won't be.

So that's pretty rockin'. Maybe that'll be my new theme song. It would be cool if you could give it a style as well. Say I want it to be hip hop or metal or pop punk and get that style out of it. It's not quite there yet, but it does generate some pretty dang cool music. I know some of the examples on their website here are probably cherry picked examples, but there's some that I would listen to the full song if they made a full song available now to get access to it. It's totally free, but you do have to go to this long notion URL here so I'll make sure it's linked up in the description so that you can find it easily. Next, I want to tell you about Cohesive AI which happens to be the sponsor of this video. So thanks so much to Cohesive for sponsoring this one. Cohesive is an all in one powerful AI editor. It's got built in image generation. It's one of the quickest ways to build thousand plus word blog posts and it's even got over 20 human sounding voices to do realistic voiceovers of your content. So check this out. If I go to the story generator here, it kind of gives me some prompts here to make it real easy to get the exact story I'm looking for. Generate a story about a future futuristic space station built where humans and aliens live together peacefully. Let's make the audience young adults genre would be sci fi and then story length. We can go short, medium long. Let's go ahead and go long and click go and check this out. I can add AI voices. So let's take this second paragraph here and generate in Rachel's voice.

In the distant future aboard the majestic Aegis Space Station, humans and aliens coexisted in perfect harmony.

And then for the second paragraph let's generate it in Josh's voice. At the heart of the station, the central plaza served as the hub of communal activities. So pretty dang realistic voices. And let's say I want to add an image into my story. I can add another line here. Hit command shift. Generate image with cohesive AI. Let's do a futuristic space station and let it generate Some variations. For us, this one looks pretty good. So let's add it to our story. Maybe we don't want to use AI images. We could have it pull a stock image from Unsplash. Let's put Space station, and here's a picture from NASA space station that we can use within our story. We can even translate our entire story into one of 13 different languages. Cohesive has over 200 different templates ranging from YouTube video scripts to tweet generators, to Facebook and Google Ads, NDAs, email follow up sequences, and even jingle generators. So no more struggling with figuring out that perfect prompt to get exactly what you're looking for. Cohesive is like having a smart co writer at your side at all times. So thank you so much to Cohesive for sponsoring this video. You can find out more information about Cohesive over at Cohesive so or in the link in the description. This is pretty cool. Check out this robot out of Google's robotics labs. Called R2T2. No, just RT2. This robot actually combines both robotics and web data into generalized instructions that it can use to actually complete tasks. Because it was trained on Internet scale data, it actually had emerging semantic reasoning. So it was basically learning from the web and not necessarily things that the scientists specifically trained it on. They presented a number of previously unseen objects to the robot and then evaluated its performance on tasks that require semantic understanding that goes far beyond the robot data that the model was fine tuned on. So the objects that were put in front of this robot, it hadn't been specifically trained on what to do with those objects, other than the fact that it had data from the web. So it was able to put a strawberry into the correct bowl. And I know this image is small, but there's a bowl of apples, a bowl of strawberries, a bowl of maybe some sort of vegetable. And it specifically knew which bowl this strawberry should go into. Pick up the bag about to fall off the table. You can see there's multiple bags on the table, but one is on the edge. It picked up the specific one that was about to fall off the table. There was a couple logos on the table. It said move the apple to the Denver Nuggets. It put the apple on the Denver Nuggets logo. And there are all sorts of examples here on this research page of it being able to perform the tasks that it wasn't necessarily trained on. But because it was able to learn additional context by browsing the web, it was able to complete the tasks. Here's another example. Move banana to the water bottle. And this is actually at 2x speed. So it's moving quite a bit slower than this. Move banana to the three. It took the banana and put it on the number three. Put the strawberry in the correct bowl. So this is what we just saw in that image earlier, and it knew which bowl had the strawberries in it already. Move soccer to basketball. It put the soccer ball by the basketball and, you know, closer than the football. And again, this is not data that it was specifically trained on. Move the Coke to Taylor Swift. It put the Coke can over by Taylor Swift. Move Red Bull can to the H. And it moved it over specifically to the H. Some pretty exciting research out of Google's robotics lab. It's going to be fun to see where this one goes over time. So this is a cool tool that I came across from Harpriya here, where she used this tool called Belva to actually wait on hold for her when United lost her luggage. She actually explains in this video that United lost her luggage. She called them to figure out where it was. She got stuck in a phone tree and she used this Belva tool, gave it her info about her flight details, her name, anything that they might ask. When she actually got a customer service rep on the phone, it was able to actually press buttons in the phone tree to get to the right person, wait on hold for hours, and then when they finally got to the right person, give the details necessary to get the answer that she needed. When I saw this, I thought this was really cool and had to check out Belva. I was actually able to sign up to Belvo when I first checked out the website, but now it looks like they've got a wait list. So I was actually once able to get in and use it and it worked for me at no issues. I actually had it call me and I had a whole conversation with the bot. But for whatever reason, now when I go to their login page, it actually takes me to a WordPress login. So I'm really hoping they get this stuff stuff sorted out and we get access again, because this is a tool that legitimately seems like a time saver. If you hate waiting on hold, which who likes waiting on hold? This tool can be your personal assistant that waits on hold, presses the buttons in the phone tree for you. And then when you finally get to speak to a customer service rep, they can either answer the questions on your behalf to the rep with a fairly convincing AI voice, or they can pass the phone call over to you and you can take it from there. When you finally have somebody to talk to. Really excited about this app. Hopefully they get it back online and out of waitlist mode. Hopefully people like me who were once able to sign up could actually get back into it eventually and play around with it some more and actually show you what it looked like when it was working. Here's a cool little gadget from Kyle Goodrich that was shared to me in my DMs. It's a little camera where you take a photo with the camera, you give it a little style like fairy tale, demon, retro, etc. And using stable diffusion and control net, it takes whatever photo you just took in the moment and then converts it into that style that you just asked it to convert to. And here's some other examples. It takes the dog, puts it in a space station, the car changes it into a Ferrari, takes the scene, puts it underwater, and he calls it the Dream Generator. A little further down the thread here you can see he showed off the prototype of it where he comes in, presses a little button here it takes a selfie of him and then I'm assuming it's on demon mode because it turns him into like a demon. So it's a cool little concept, this Dream Generator here. I'll link to this tweet thread in the description so you can check it out for yourself. I think it'd be kind of fun to get my hands on something like this. And then finally, I can't make a video about the latest cool research I came across without talking about LK 99. This research paper dropped on Arxiv within the last week and it claims to be the first room temperature ambient pressure superconductor. Now if you're on Twitter at all lately LK 99 has been the talk of Twitter. Basically what this is saying is that these researchers in South Korea managed to create a superconductor that can operate at room temperature under normal pressure conditions. You may have even seen this video that shows this piece of metal sort of half levitating that is meant to be proof that they found this superconductor. The reason this is important is because up until now, scientists have only been able to find materials that superconduct at super cold temperatures, like negative 100 degrees Fahrenheit cold. Because of this, superconductors have been really impractical to use. Now they are using things like MRI machines and like the Large Hadron Collider, but we can't really use them in small electronics or in the power grids. However, if this discovery does prove to be true it could actually enable lossless power, which means we'll see a new explosion in technology from this. We'll see more efficient computers, more efficient transportation. Electronics will become much more energy efficient and hold their charges much longer. Machines like MRIs will become more accessible and a lot less expensive to use. Magnetic levitation technology will probably see an explosion in new use cases. This will fundamentally change the technology world as we know it. I'm super oversimplifying it. That's why I recommend you go follow someone like Andrew Coat here on Twitter because he breaks down really what this all means. I still feel like kind of an idiot when I try to explain it, but this really could truly mean some just massive breakthroughs and huge technological advancements if this science holds up to be true. Now the thing is, this has only been tested in that one South Korean lab and other scientists are still trying to recreate it. So it hasn't been verified by other labs yet. But it is super inexpensive to create and the process outlined in the paper is supposedly really easy for anybody to duplicate. So we should have answers within the next few weeks on whether or not this technology is legit and could really change the world or not. Now, according to Andrew, there's been simulations done which somewhat verify that maybe there is some truth that this actually works. And a lab out of China also claims to have been able to replicate replicate this superconductor with this little teeny tiny piece that it generated here where you can see it sort of levitates on its side when they put a magnet next to it. But so far this video and this video are the only two videos we have that show it off. And a whole bunch of other labs have not yet been able to replicate it. This very, very exciting if true. But still we don't have a lot of amazing verifications that it actually is true yet. Something I'm excited about. I know it's not really AI related, but it does have a lot of implic in the AI space because this is going to make GPUs dramatically cooler if it's true, which means we're going to be able to crank up the compute power with this technology in the future as well. So really, really exciting stuff. Definitely a little overly nerdy and maybe something I'll do another breakdown in a future video if more people are interested. But for now I just want to point out that it's very hard to talk about cool research without mentioning that this is something that could potentially change the world as we know it. And I'm excited and fingers crossed that it's actually legit because I would love to see that happen. And that's what I got for you today. If you love cool AI tools and nerding out over this kind of stuff, check out FutureTools IO. This is where I curate all the cool tools that I come across. I keep the AI News page up to date on a daily basis with all the latest news. And every week I send out a free newsletter with the TLDR of everything that's happening in the AI space. It's subscribed to by over 130,000 people, including people at all of the biggest tech and AI companies in the world. So you'll be in good company. You can find it all over at FutureTools IE. Just click the little button that says Join the free newsletter and I'll send out your first issue this Friday. Thank you so much for tuning in. I do really, really appreciate you and thank you once again to cohesive for sponsoring this video. It goes a long way to help me create more fun and exciting videos that show off the latest cool AI and future tech. So thanks again. I'll see you in the next video. Bye bye.

Subscribe to Mac Wyat Time.

Laughter Guarantee you won't be bored that subscribe button you won't be in.