Tons of AI News That You Probably Missed

For the second week in a row, we've had an insane amount of AI news drop. And as you can see by my tabs above, we got a lot to cover. So let's get right into it. At the end of last week, Meta announced a new image model called Chameleon. The model is capable of text to image and as you can see from their examples, they're kind of flexing on some of the other text to image models by generating an image of a hand and also generating an image with the text 1991 in it. But Chameleon isn't only capable of doing text to image, it can also do text guided image editing. So you can see here, they give an initial input and then they give the prompt what would she look like as a bearded man? And it generates this image. Put a pair of sunglasses, she looks 100 years old. Apply face paint. And so adding the additional text combined with the image, you can actually tweak the images with this. This model can also do text tasks. They gave it this image of a dog holding a stick and asked what is the dog carrying? And it told them a stick described the given image in very fine detail. In this image there is a dog holding a stick in its mouth. There is grass on the surface. In the background of the image there are trees. It can also do some pretty advanced stuff like look at this prompt, generate high quality IM of a room that has a sink and a mirror in it with bottle at location. And then it's actually got location information in the prompt and then you can see it actually places it based on the prompt locations that was given. It's also got segmentation to image. Here's some other examples of images that were prompted with it, which in my opinion look pretty dang impressive. Now this model doesn't work like any of the other existing text to image models out there, like diffusion models that we're used to with stable diffusion and Dolly and midjourney. And it's not using GAN models either, which was sort of the popular way to do before diffusion models came on the scene. This uses what's called an autoregressive model. So if you think of a diffusion model like you get a solid image of noise and pixels and you sort of remove all the pixels until it sort of becomes the image that you want. Then you think of GANs, the General Adversarial Networks, where essentially you have two AIs that are kind of competing against each other. One's trying to create the image, one's trying to tell if the image is fake or not, and they go back and forth and back and forth and back and forth until the one that's trying to detect whether it's fake or not can no longer determine that the image is not a real image. That's what GAN models in an oversimplified way do, where an autoregressive model you can almost think of as like when you print something on your printer, it prints line by line by line, while this prints pixel by pixel by pixel and adds more and more pixels until it comes to the image that you're trying to generate. Now again, super oversimplification of this. Just know that it's a completely different sort of model than what we are already seeing with tools like Stable Diffusion, Mid Journey and Dall E. Meta also mentions that all of the training for their Chameleon model has been trained with Shutterstock data that's only using licensed images and text data. So essentially they're not using images that were just scraped from the Internet, they're using images that were licensed and they were allowed to use from Shutterstock's data that they were given permission to use. There's nothing in their announcement or in their paper that mentions when US consumers might get access to this, but it looks to be some pretty promising technology that might actually fix a lot of the shortfalls of the existing text to image models. Also last week while I was flying home from Colorado, Elon Musk did a Twitter space where he gave a little bit more information on what we can expect from his new AI startup, xai. Now, since I wasn't on the Twitter space myself, I leveraged my good friend Claude to help summarize what came out of that Twitter space. Here's a few of the notable bullet points. Elon announced that his new startup will collaborate with Tesla on silicon and software fronts. Silicon sort of referring to the processors he's going to use. XAI will use Twitter data for training. Maximally curious AI system Tesla is building custom silicon that is three to five times more capable than current hardware. The next version will be four to five times more capable. Typically, silicon refers to like GPUs. So it sounds like they're going to go head to head against companies like Nvidia and amd. Tesla's supercomputer called Dojo will be used for AI training with data from Tesla vehicles. He wants their AI to understand the physical world beyond just the Internet, using data like Tesla driving data, which is fascinating to me. So essentially tying that all together, Tesla's going to use data from Twitter. They're creating their own sort of supercomputers and silicon to train this stuff. And they're also going to use driving data from Tesla. And from what I can tell, there's no real information about when we can expect access to this or when this will actually be unveiled for people to start using. Now moving into this week, Wix, the popular website builder announced that they're adding AI into their platform where you can essentially do text to entire website. And here's a quick demo of what WIX proposes that we can expect with it. Hi, I'm Wix AI. I'm here to assist you in getting your site up and running. What are you building a website for? I want to build a website for Onyx Online Fitness. We provide dynamic online classes and in person workouts. Our trainers are certified and we offer a variety of classes. That's great. Are there any more details about your business that I should take into account? We have a variety of fitness activities. Our fitness classes are led by professional athletes. Here's several images generate site and then it just goes and generates the site for you. It just asks a handful of questions, goes back and forth and creates an entire site for you. So WIX is kind of next level here. I've never really thought I would be talking about Wix. I've always been much more of a WordPress and more recently a webflow guy. But Wix is really stepping up their game here with this where you can just use their bot assistant here. Tell it exactly what you want from a website. It'll give you different things, theme options. You just have conversations with their little sidebar and it's going to build out your whole dang website for you. They say show me other layouts please. It says, sure, here's some more options. They ask for a round layout, it generates a whole new layout. And then look at this page that it generates. It's even got some like parallax effects on the page. And I mean it's, it's a solid looking website. It's a fairly impressive website, all with just some text prompts. I've made a handful of videos in the past where I talked about Microsoft 365 Copilot. Well, Microsoft just announced what the price is going to be. It's going to be about 30 bucks per month per user, which is going to be really interesting to see because Google is sort of launching similar features into their Google workspace. And 30 bucks a month per user seems fairly high, especially if Google only launches it for like 10 bucks a month. It'll be interesting to see how that battle plays out. Now, what's probably the largest news in the AI world this week is that Meta and Microsoft have sort of joined forces to put out Llama 2, which is a sort of open source competitor to ChatGPT. Now, I say sort of open source because as Meta themselves say that this is the next generation of open source large language model and it's free to use for research and commercial use, there are actually some terms inside of the license agreement that many are arguing make it not actually open source. For example, they say you will not use the Llama materials or any output or result of the Llama materials to improve any other language model. So you can't really combine it with other language models or use it to improve those. If you use it for a product that ends up having greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion. And basically they're going to take away your rights to use it if you ever try to sue them. So I would say for most people, for all intents and purposes, it's pretty open source. If you have 700 million monthly active users, Facebook's probably going to be pretty willing to work with you at that point anyway. Not a ton of products reach that level. But I think the bigger question that I had when I heard this announcement was more why is Microsoft working together with Meta when Microsoft is already in partnership with OpenAI? Wouldn't this essentially cannibalize some usage of Bing Chat? Cannibalize some usage of ChatGPT, which Microsoft is heavily invested in? Well, Alex here makes some great points about this announcement. This will likely knock out a lot of competition. Mosaic, Red, Pajama, et cetera, are in major trouble. What's brilliant about it from Microsoft is that they have their Azure suite of cloud computers and now they're going to have OpenAI running all of their models on the Azure infrastructure. I mean, part of their $10 billion investment in OpenAI was actually given to them in compute credits, essentially. And now working with Meta and Llama 2, they're going to have people building their open source models also on Microsoft Azure's clouds. I always say that wrong. I don't know if it's Azure or Azure, whatever, Azure. Alex here makes the point that with this partnership, they now have exclusive partnerships with top Large language models, OpenAI and Meta. Priority access to Nvidia's GPUs and strategic assets like GitHub and Azure. So while I'll never underestimate Google, this move by Microsoft where now they've got this sort of closed source large language model under their thumb, they've got this sort of open source large language model partnered with Meta under their thumb, they're working closely with Nvidia for GPUs on their Azure cloud servers. They sort of now have the whole large language model generative AI world under their umbrella. So likely a pretty damn smart move by Microsoft. Now if you're wondering if you can actually use Llama 2 right now, well, you can. You can head on over to Hugging Face. There's a space called llama2 70B chatbot. I'll make sure it's linked up in the description below so you can go mess around with it. In most of the testing. It doesn't quite perform as well as ChatGPT or Claude or some of the other more popular models that are out there right now. But but it's open source and people will be building and iterating off of it and making it better. Why would someone use this chat model instead of Chat GPT? Safety bias Factual accuracy Ethical considerations Customization Privacy it actually cut off in my test here, but you can actually go and use llama2 yourself and play around with it. At the moment I'm still more of a cloud guy myself. I've been using Claude 2 a lot. It's almost my go to and then Chat GPT has almost become my second go to. But I am always excited about open source stuff because when we see open source stuff it means that other smart people could build off of it, iterate off of it and create something bigger and better than the original thing that was created. So I'm excited to see where people take this llama 2 and iterate it into other really cool stuff. You know, speaking of large language models, there was this paper that came out this week about how is ChatGPT's behavior changing over time and essentially the paper came to the conclusion that it's been getting worse and worse at math. It actually shows that back in March it answered this question is 17,077 a prime number? Think step by step and then answer yes or no. It was actually answering this question properly back in March 97.6% of the time and in June only answered it correctly 2.4% of the time. Confusingly, GPT 3.5, what should be a lesser model, only answered it 7.4% of the time correctly back in March, but now actually answers it correctly 86.8% of the time. Here's another math problem that was given back in March. It answered it more often accurately, and in June less accurately. And then given that same problem with GPT 3.5, it was answering more accurately in March and less accurately in June. So the argument of this paper here is that ChatGPT is getting dumber, but it's really only looking at math. And Rob Lennon here on Twitter argues that math based questions are not a particularly accurate or useful way to test a language model. Basically because these are language models, it's built around relationships between words. So his analogy is Olympic swimmers are getting slower. Just look at these skiing challenges that we've been testing them on. See, they're getting worse at sliding down snow slopes on wooden foot sleds. Swimmers are getting worse. You're asking a large language model which is trying to predict the next word in a sequence of words to do math and get better at math, where the training behind the scenes, they're not trying to train it to get better at math necessarily. They're trying to get better at predicting the next word in the sequence to create an accurate sequence of words. OpenAI does say that they are aware of this new research and that they're monitoring the reports of the declining GPT4 capabilities. Their vice President of Product recently tweeted, no, we haven't made GPT4 dumber. Quite the opposite. We make each new version smarter than the previous one. Current hypothesis, we use it more heavily. You start noticing issues you didn't see before. So he's basically arguing that the more you use it, the more you're going to find flaws. A lot of the novelty has worn off. People are used to chatgpt now. Now they're just actively looking for flaws in it. Speaking of ChatGPT, this week ChatGPT increased the message limits for plus users. If you go over to ChatGPT down here, it says GPT4 currently has a cap of 50 messages every three hours. Prior to this week, you used to only have a cap of 25 messages every three hours. So that's an upgrade. ChatGPT also added a new feature called custom instructions for plus users. So if you go over to ChatGPT, you come down to your settings down here, click on beta features, you can actually turn on custom instructions. Once you do that, you can close out of here, come back down to your little three dots, and you'll see a new menu item that says Custom Instructions. This allows you to actually Add additional context that you want ChatGPT to know to help improve its results when you ask it future questions. So it says, what would you like ChatGPT to know about you to provide better responses? Here's some starters. Where are you based? What do you do for work? What are your hobbies and interests? What subjects can you talk about for hours? What are some goals you have? You provided this additional context, it's going to use that information to help provide better results in future chats. How would you like ChatGPT to respond? Some examples here. How formal or casual should ChatGPT be? How long or short should responses generally be? How do you want to be addressed? Should ChatGPT have opinions on topics or remain neutral? So you can actually give it all of this additional context and it will Change the way ChatGPT outputs responses for you. And not only that, this information will now be saved in there. So all future chats will have this additional context that you've previously given it. And since we're talking about ChatGPT news, might as well mention this. OpenAI is partnering with the American Journalism Project to support local news. Basically, OpenAI wants news organizations to leverage tools like ChatGPT and AI to help improve the news that they're reporting on. OpenAI is committing $5 million to the American Journalism Project to support the expansion of their work, and they're offering up to $5 million in OpenAI API credits so that the organizations can use OpenAI's tools. And since we're on the topic of news, Google is testing an AI tool that can write news articles. This was reported by the New York Times. I would read the original source article, but it's paywall. Then I don't want to pay for the New York Times. The tool internally is codename Genesis, and it can take information and then generate news copy from. However, the people that were actually pitched the tool saw it as unsettling, noting that it seemed to disregard the effort that went into producing accurate news stories. So essentially Google's out there pitching a tool to help journalists write their news articles for them. And journalists are kind of going, that's kind of scary because this feels like it might take our jobs. Do you remember a couple months ago when I talked about this generative agents paper that came out where they generated this little town with people and the people interacted and they actually we created events and then set up reservations and asked each other on dates and all showed up to this event together and it all happened autonomously. Well, this company called Fable Studios took that a step further and created an entire south park episode that was completely generated by AI. Their research was called to infinity and beyond show one and Showrunner agents in multi agent simulations. You can see in their little demo video. Here they have their showrunner system. Welcome to the Showrunner subroutine. Try to come up with a suitable title for the story you're about to simulate. Next, prompt the system with a few of your own ideas of what should happen. They give it the title Westland Chronicles. And the prompt. Researchers train an AI robot pig to write infinite TV shows, but it goes horribly wrong. Cartman is a big fan of the AI generated TV show Westland Chronicles. The Writers Guild is on strike because of AI, so it gave it just that information and then it generated an entire 22 minute episode of south park off of that. Now I'm a little bit afraid to play it because this is for research purposes only. The South Park IP is owned by somebody else and by simply pressing play and listening to the audio on this, I could get in trouble. I don't know. So I'm not going to. But basically all of the background images were generated with AI image generation models. All of the dialogue was generated with large language models. The interactions between the characters were generated with this generative agents model that I was just talking about. The voices of all of the characters were generated by cloning the voices most likely using a tool like an 11 labs or something like that. And so this entire 22 minute story, with art, with voices, with dialogue, the whole thing was generated completely with AI and that one prompt that we just saw. People have been all over Twitter talking about how it's only a matter of time before we're just going to type a prompt of what type of show we want to see and it's just going to generate an entire show or entire movie for us. We're there like it's happening right now. It's not the best dialogue, it's not the best script. I don't think anybody would sit around and want to watch these episodes generated by AI yet, but it's getting a lot better. I'll make sure to link up to this GitHub page below where you can actually watch the entire episode if you want. There are actually some moments in it that are pretty dang funny. It does a pretty accurate job of getting the sort of feel of a South park story. This week it was also announced that Apple is testing a chatgpt like AI chatbot, which feels kind of late to the party. Internally. They're calling it Apple GPT. That's unlikely to be what it's actually called when it's ever released, but at the end of the day, everybody wants a new Siri, right? If you've got an iPhone and you use Siri, all Siri can really do for you these days is search the Internet and open a couple apps for you. When you've seen stuff like Chat GPT and Claude and what perplexity and tools like that can do, Siri feels really, really archaic. This TechCrunch article here says the company's chatbot is similar to Bard, ChatGPT and Bing AI as it doesn't feature any additional features that separates it from what's currently commercially available. So not a ton of news around this yet, but I'd say it's fairly likely that this technology is probably what we're going to see as the next iteration of Siri. But who knows? Now here's a cool product that was released this week. Superhuman itself isn't new. It's an email platform that helps you manage your email similar to a tool like Outlook, but it's a much faster, easier way to use email. I've used Superhuman a lot over the years. Well, Superhuman just rolled out new AI features inside of their email platform. Now you might be thinking, well, aren't Gmail and Outlook and all of these tools getting these AI features rolled into them as well? Yes, but what makes this one unique is that Superhuman actually trains itself on your past emails so it can better understand your voice and write more like you. So the first time I booted up Superhuman and turned on their AI feature, it actually had a small 30 second training period where it saw how I wrote previous emails and then figured out how to write future emails in that voice. So let's just write a test email. I'd like to propose a lunch meeting for 1pm on Saturday to share a cool new business idea. I could select this text that I just wrote, click on this AI button here, and then say rewrite in my voice. Hey everyone, I wanted to throw out the idea of getting together for a lunch meeting this Saturday at 1pm I've got this really awesome new business idea that I'm excited to share with you all. Let me know if you're interested in joining and we can figure out a great spot to grab some grub and chat. I mean, I don't know if I ever really refer to food as grubby, but other than that it sounds pretty much like the way I'd write an email. I can also simply just click on this AI button right here that says Write with AI and then outline an email brief Congratulate My Team on the latest YouTube video launch wanted to take a moment to congratulate you all on the successful launch of our latest YouTube video. It was such a fantastic effort by everyone. It just wrote up a whole dang email. Once again, congratulations and thanks for being a part of the Future Tools family. I never told that my website was called Future Tools. It just took that from the training data that it has. It says you guys sure know how to make those those AI tutorials engaging and informative. I didn't give it any context just now that I make AI tutorial videos. It just knows that from my past emails. Here's a long email thread that has multiple people in it going back and forth about a project that I'm working on. I could come in here, click the AI button, type sum and that gives me the option to summarize conversation. And look at this. Now it just bullet pointed this entire email thread for me so I don't have to skim through the whole thing again. Pretty cool time saving tool. Now Superhuman doesn't sponsor this or anything like that, but I'm constantly looking for a way to streamline my email and this tool seems like it'll be pretty handy. Last week I mentioned that Sarah Silverman and a few other authors were suing OpenAI and various other chat models for illegally scraping their books and using them as part of the training data. Well, this week, thousands of authors signed a letter urging AI makers to stop stealing books. In an open letter signed by more than 8,500 authors of fiction, nonfiction and poetry, the tech companies behind large language models like ChatGPT, Bard Llama and more are taken to task for using their writing without permission or compensation. The letter asked that companies obtain permission for use of copyrighted material and generative AI programs compensate writers fairly for the past and ongoing use of the works, and compensate writers fairly for the use of works in AI output, whether or not the outputs are infringing under current law, I think that might be a tall order, but no legal threats were made. There's a lot of legal gray area that's happening right now, and it's going to be really interesting to see how this all plays out in the corner. I mean, I do understand where a lot of these authors are coming from. It could potentially harm the sales of a lot of their books. If I can just go to ChatGPT and ask questions of this book without ever Having to pick up a copy of the book, however, figuring out a way to compensate authors for the use of these various chat models. I honestly can't totally comprehend a path where that can happen. I mean, I guess Spotify kind of has it figured out for music, but this is a little bit more complex than did somebody press play on a song or not? Because when somebody generates a response from a large language model, it's pulling from all sorts of data from all over the place to provide the response. And so I don't know how that ever totally gets rectified. And speaking of lawsuits, a US judge found flaws in the artist lawsuits against the AI companies. If you remember Stability, AI and Midjourney, and the various models that generate art have been getting sued by a lot of artists because they're trained on their artwork. The judge said that artists were unlikely to succeed on their claim that images generated by the systems based on text prompts using their names violated their copyrights. They said, I don't think the claim regarding output images is plausible at the moment because there's no substantial similarity between images created by the artists and the AI systems. So this one is actually going to be fascinating to see how it plays out in courts, because when you think about it, most artists borrow from artists that came before them. There's no law that says that I can't take a canvas and paint a picture that looks like a picture that Pablo Picasso painted. Now there's laws against me turning around and trying to sell it as if it was a Pablo Picasso, but not to actually generate that image in the first place. But now people want laws against computers generating images that are inspired by other artists. So there's this weird gray area there because I don't necessarily think there's ethical issues of generating images that look like or inspired by other artists, but I do think there are ethical issues of creating images that look like or are inspired by other artists and then making money off of those. So I don't know. There's definitely a gray area debate that I'm still personally a little torn around. This week, GitHub also made the Copilot Chat available for all organizations. It's basically a chat assistant that helps you with your code, and it's basically trained on data that's available inside of GitHub. Now, I don't personally work with a lot of code at the moment. Eventually I do want to develop. Develop some games and some apps and use AI to do creative stuff with it, but I haven't personally tried out this GitHub co pilot. But if you're a coder, this might be something that is interesting news to you. So I wanted to point it out real quick. And finally I wanted to end with something really cool that you can go test out yourself right now if you want. I know I've mentioned this company a few times in the past. Illumine AI. They just announced a new feature where you can use 2D images that have a 3D depth map and pull them into their platform. So I actually wanted to test this feature real quick in their Instaverse platform. You can see in their example here, they have this 2D to 3D mode, they click on that, they upload a 2D image, they upload a 3D depth map and now they have this 3D image that they could sort of move around and manipulate. So I actually created this little 3D world here inside of Midjourney. I just used the prompt a colorful 3D metaverse world from a video game. Now you do need a 3D depth map to be able to upload this. One of the easiest ways to get a 3D map depth map is to go to convert.leiapix.com, upload your image. So I'll go ahead and grab this and upload it. I could come down here to the share button. It gives me the option to download a depth map. So let's go ahead and download that real quick. And now if I jump over to this Illuminai site, I can click on 2D to 3D mode, upload my 2D texture here. I've got this image now, but it looks kind of funky. So let's go ahead and upload our depth map. Open this up. And now I can actually kind of sort of move around it. It gets a little warped if I move around too much, but I created this trippy psychedelic 3D depth map out of that Midjourney image. So just something fun to play around with with your 2D images. It's all completely free to use. I mean Mid Journey is not free to use, but you can use any AI image generator. Create the depth map using Leia picks real quick, pull it into Instaverse and you've got an Interesting little 3D depth mapped world. I'm sure there's some other settings I could have played around with inside of Leia Picks converter to to get my depth map a little bit better, but just wanted to show you a quick little demo of what you can do there. And that's what I got for you this week. If you want to keep your finger on the pulse of all the latest AI news, make sure you check out FutureTools IO, specifically the AI News tab. I keep this up to date every single day. It's probably the most up to date resource on AI news on the entire Internet. And also I update all the cool AI tools that I come across. So on the homepage here, I'm pretty much adding new tools every single day. And if you don't want to check in every single day, click on this button to join the free newsletter and I'll just send you one update a week every Friday with just the handful of news articles that I think you should be paying attention to, a handful of really cool tools that I think you should be paying attention to, a few YouTube videos and one interesting way to make money with AI I send it every single Friday and you can get it totally for free by going to FutureTools IE and joining the free newsletter. The last few weeks have been really, really insane with AI news. There was a little bit of a slowdown, but things have really, really picked up steam again. I'm excited for it. Gives me a lot to talk about, a lot to experiment with, a lot to make videos on for you and yeah, a lot of fun stuff in the AI space. So I'm excited to report on it, I'm excited to share it with you and if you haven't already, maybe like and subscribe and I'll make sure you keep on seeing them in your YouTube feed. Once again, thanks so much for tuning in. Really, really appreciate you. See you in the next video. Bye. By.