The Insane AI News Happening No One is Noticing

Let's break down all the stuff you probably missed in the world of AI this week. Although it was a bit quieter of a week in terms of big AI announcements, there was still quite a bit of notable things happening that I think are worth talking about, starting with the fact that GPT4 is actually getting faster Now. In previous weeks I talked about how people seem to believe ChatGPT is getting dumber, which has been somewhat debunked by now, but it is definitely getting faster. This company, Portkey, has been keeping an eye on it. You can see the line in red is GPT4. The line in blue is GPT 3.5. We're getting to a point where GPT4 is almost as fast as GPT 3.5. There's just a teeny tiny margin left. This week, OpenAI also announced that Dall E3 is now available in chat, GPT plus and enterprise. Now, most people actually got access to it last week, but OpenAI made it official. I'm assuming this means it's pretty much rolled out to everyone now, although the fact that we can now use Dall E3 was in last week's news video as well. Do want to use Dall E3? Just log into your ChatGPT account, make sure that you are a ChatGPT plus subscriber, hover over your GPT4 button here and select Dall E3. And now you can make some amazing art like this space wolf chasing after a taco. This week, Rowan Chung also pointed out that OpenAI quietly changed their core values, and in this change, they updated it to focus on building AGI. So here's what the core values looked like before the change. Here's what they look like now. And if you notice there's now a section that says AGI focus. We are committed to building safe, beneficial AGI that will have a massive positive impact on humanity's future. Anything that doesn't help with that is out of scope. So they've made it absolutely clear what OpenAI's real purpose is now. Now, many people have a different definition of what AGI actually means, but Rohan also shared this video clip of Sam Altman giving his definition of AGI.

AGI is basically the equivalent of a median human that you could hire as a coworker and then they could say, do anything that you'd be happy with a remote coworker doing just behind a computer, which includes, you know, learning how to go be a doctor, learning how to go be a very competent coder. There's a lot of stuff that a median human is capable of getting good at. And I think one of the skills of an AGI is not any particular milestone, but a meta skill of learning to figure things out and that it can go decide to get good at whatever you need.

Also in OpenAI news this week OpenAI decided to shut down one of its AI programs. They were working on an AI model codenamed Arrakis, named after the desert planet from the Sci fi dune. The goal was to power AI applications like Chat GPT much more cheaply. However, they announced this week that they decided to scrap it because it struggled to meet expectations around efficiency. Not much else is known about this project Arrakis, but it sounds like we're never going to see it anyway. This week Anthropic made a major update to their Claude platform, making it now available for users in 95 countries. I personally use Claude at least as much as ChatGPT, if maybe not a little bit more. It's probably the best tool available right now for summarization of long PDFs, CSV files, transcripts, things like that. Claude, hands down, is better at that than ChatGPT, just because it has a much larger context window and now 95 more countries are going to be able to access that as well. However, in some bad news for Anthropic, Universal Music is suing them for distributing song lyrics. So essentially, if you were to go to Claude and type give me the song lyrics for Roar by Katy Perry, it will actually write out all of the lyrics for you. Since song lyrics are actually copyrighted, Universal is saying that Anthropic not only distributes copyrighted material without permission, but that it used these to train its language models. They go on to say Anthropic's copyright infringement is not innovation in layman's terms, it's theft. The thing that I find interesting though, is if I just went to Google and said Roar by Katy Perry lyrics, Google gives me the lyrics right here, also for free, and I've been able to do that for like two decades now. However, if you do ask ChatGPT the same question, it will say, sorry, I can't provide the lyrics to copyrighted songs. Would you like a summary instead? Personally, I think there needs to be a total overhaul of what copyright law means, because it's just getting harder and harder. The lines are getting blurred between what's copyrightable and what's not. We're seeing this with images, we're seeing this with song lyrics, we're seeing this with people's voices pretty soon we're going to see this more and more with video generations and as AI gets more and more prolific and more and more people use these simple tools to sort of recreate anything, it seems like some of these bigger companies like Universal are just going to end up over their heads in potential lawsuits for all of these creations that are being made. Now. I'm not saying that it's necessarily right that people use copyrighted material. I just think that the way copyright works needs to be rethought. I don't personally have the solution for that. I'm not claiming I do. I just think copyright needs to be rethought in this post AI world that we're moving into. Also this week, Inflection AI, the creator of the PI chatbot, announced that PI can now search the web. So any chats that you have with PIE are up to date as of today. If I go to PI AI slash talk, I can ask it super recent questions. For example, who won the boxing match between Logan Paul and Dylan Dennis? A really, really stupid viral boxing match. Logan Paul won the fight, although it was a bit of a strange one. Dylan Dennis refused to strike and Inst instead tried to grapple with Paul which resulted in him being disqualified in the final round. It was a pretty chaotic fight with a mass brawl breaking out at one point. In the end, Logan Paul emerged victorious. Did you watch the fight or are you just curious about the results? That is pretty up to date information seeing as that fight happened on October 14, I think I didn't actually watch it. I actually just learned the result of it right now by asking PI. This week Stanford put out a report here that they call the Foundation Model Transparency index. They compared 10 different chatbots to find out which ones were the most transparent. They were scored on a hundred different indicators. Things like how the data was trained, the labor that went into it, the compute power necessary, the risks, the distribution feedback, and much more. And basically what it found was that Meta and their Llama two they've been the most transparent followed by hugging face, then OpenAI, then stability all the way on down to Amazon being the least transparent with their Titan model, concluding that that developers can significantly improve transparency by adopting best practices from their competitors and that the Open foundation model developers are leading the way. I find it Interesting that OpenAI actually beat out stability AI because stability most of their models are open and available for people to play around with. And finally in other large language model news this week Baidu, a company out of China, says that it's Ernie AI chatbot is not inferior in any aspect to GPT4. They gave a keynote earlier this week and they demonstrated improvements to four different capabilities understanding, generation, reasoning and memory, basically claiming that it's as good as GPT4. Now, Ernie4 isn't publicly available yet, but even if it was, so far these various Ernie bots have only been available to use in Chinese, so I likely wouldn't be able to use it anyway. And since we're on the topic of China here, there's been a new US restriction passed, making it so that Nvidia cannot send their high powered AI chips to China. They've already been blocked from sending the H100 chips because they were too powerful. So Nvidia went ahead and made a lower spec H800 chip and was exporting those to China. However, the new restrictions just closed that loophole and now Nvidia can no longer send H1 hundreds, H8 hundreds or even A800 chips to to China, apparently all in an effort to sort of win the AI race with China. Now let's shift into talking about AI art, starting with the sponsor of today's video, Wirestock IO. If you're not familiar with Wirestock, they are a single platform where you can upload stock photography that you created or even AI generated images that you created, and it will submit it to all of the stock imagery platforms on your behalf. It will even title the images, add tags to the images, create descriptions to the images, and inform the stock photography sites that these are AI generated so that you can stay compliant. Sites like Imago, Adobe Stock 123RF, Dreamstime and Freepik all allow AI generated art so you can generate an AI image, submit it to Wirestock and it will create all the metadata and submit it to all these sites for you. It's super fast. Wirestock even has their own Discord bot. So if you're familiar with tools like Midjourney, you'll feel right at home with wirestock. It allows you to generate custom images, mix multiple images together, and even reimagine existing images so you can upload an image that you already tell it to, reimagine it and it will create more images just like that one. So if an image is selling well for you already on a stock photo site, reimagine it and submit more just like it. Wirestock also gives you the ability to create custom collections. I can add this image, call it Pixel Collection, create the collection, and add more images straight into this collection. Now I can link people directly to this collection. They can purchase the whole collection at a collection price. And if you're a premium member of Wirestock, you get to keep 100% of the earnings. And if you're curious, Wirestock even sent me a handful of images that are performing well on Wirestock right now. So if you want an idea of what images to generate, here's a few. We got a landscape, a tractor, a busy street, this cool sort of surreal reef image, this cool background image, and this flooded basement. These are images that performed well. You can learn more about Wirestock by heading over to Wirestock and if you use the coupon code MATT20, you'll actually get 20% off your premium membership. So check that out over at Wirestock IO. And thank you so much again to Wirestock for sponsoring this video. Now let's check in with Ali Jules on this week's Midjourney Office Hours update. They did add one new major feature inside of Mid Journey this week. If I generate an image that I really like, let's go ahead and upscale number two here. I now have the option to further upscale this image to 2x or 4x, allowing me to get much, much larger images without loss of image quality. They also mentioned that their phase one of their new website is going to be released next week. This isn't going to be a platform where you can generate images directly from the website yet, but their phase two of the website supposedly is going to come out sometime next month, which should give us the ability to generate straight from their website and finally get Midjourney out of Discord. Other Midjourney News this week Midjourney partnered with a game company, Suzigi Studios, to launch Niji Journey, which is both an iOS and Android app which allows you to generate anime style images directly within the app. I haven't personally used this one myself yet, but it appears to give us another option to generate Mid Journey art style images, not necessarily realistic photos. In the last week and A half this YouTube video has gone viral with almost 3 million views from Peter Whitten and in this video he shows how he actually trained an AI to through Pokemon. He was able to run several instances at once of an AI character running through the game and used reinforcement learning with a point based system so that when the character did things like collected Pokemon or uncovered new parts of the map, it earned points causing the AI character to play deeper and deeper into the game and collect more and more Pokemon. I highly recommend if you love AI as much as I do, watching the entire 33 minute video because he breaks down exactly how it works, the methodology behind it, the results from it. And he even created a GitHub so that you could go and download the code and run these Pokemon simulations yourself with a Pokemon emulator. Now, a few weeks ago I talked about Stack Overflow and how they were developing their own sort of AI. Well, this week it was announced that Stack Overflow laid off over 100 people as the AI coding boom continues. That's 28% of its staff. Now, if you're not familiar with Stack Overflow, it's essentially a site where people can go to get help with coding when they run into an issue. They ask questions on Stack Overflow and then helpful people on the website help them answer their code questions. Well, the problem is now instead of using Stack Overflow, a lot of developers are just going to tools like Chat, GPT and Claude to get help with their coding when they run into issues. There's also tools out there like GitHub, Copilot, and Amazon's CodeWhisperer which just help people develop the code and troubleshoot as well. The site Stack Overflow has seen a major hit in the amount of traffic that it actually gets. YouTube announced this week that it's getting a new AI powered ad feature that lets brands target special cultural moments. So essentially, if you have a product, for instance, that's related to Halloween, you can say, I want to advertise this product on any videos talking about Halloween. And YouTube's AI will try to find the videos that are about Halloween and work your ad into those videos. So instead of specifically targeting certain channels or keywords, it uses AI to find videos on maybe unrelated channels, but that happen to be talking about the specific event that you're trying to target. Also this week Descript announced some new big AI features. They've added some new AI voices to their platform.

We're introducing an all new version of our in house AI voice tool. We basically took our beat up old AI into the shop, rebuilt the engine and outfitted it with a set of all new space age capabilities like text to speech and lightning fast authorization. The old Overdub could get you where you wanted to go. But with our new AI voices, we're handing you the keys to a rocket ship. No more waiting 24 hours to train or verify a voice. And the quality is much, much better. So good that it's now legitimately possible to create an entire voiceover without recording a single word.

And speaking of AI voice cloning, the New York City mayor Eric Adams used AI to make Robocalls in languages that he doesn't speak. He sent thousands of calls in Spanish, more than 250 in Yiddish, more than 160 in Mandarin, 89 in Cantonese, and 23 in Haitian Creole. And a lot of people weren't happy with it, with one person saying this is deeply unethical, especially on the taxpayer's dime. Using AI to convince New York workers that he speaks languages that he doesn't is deeply Orwellian. Adams actually made a statement at a press conference saying, people stop me on the street all the time and say, I didn't know you speak Mandarin. He doesn't. This week, Meta showed off some research called towards a real time decoding of images from brain activity. Now, we've seen some of this kind of stuff before. Previous examples used MRIs requiring people to lay down in a giant machine and scan their brain as they're seeing images. This new method from Meta uses what's called a MEG or magneto encephalography. This allows the researchers to see what the brain is seeing in real time without the invasiveness of using MRIs. So this research showed people an image and then essentially read their brain waves to interpret what they were seeing. The image on the left here is what was actually shown to the person. The image on the right is actually what was decoded from the brain waves. So you can see they were shown cheese. This is what the brain saw or what was interpreted. Now, these aren't perfect replications, obviously. This is like a cheetah and it's seeing a weird, like, monkey thing. But this is getting really fascinating. We're getting to a point where AI is helping read brain waves and interpret what the brain is seeing. Here's some more examples of the viewed image of a surfer. Here's what was seen with the scan. Here's a horse. Here's what was seen with the brain scan. Plane. A plane. I mean, it's all pretty dang close to the original image. And with Meta actually being one of these companies that's trying to sell us headwear like the Meta Quest 3, I don't know how I really feel about Meta trying to read our brain waves. Is this going to go down the path where some of this technology is put inside of this wearable technology and Meta is reading our brain to show us advertising, I doubt it's going to go that far. The world as a whole would probably push back on that, but it feels like Meta is getting close to having the technology to be able to do that. All right, let's talk about Robotics for a minute. This week, Amazon said that their new AI powered robots will reduce fulfillment time by 25%, which to me is just crazy because we can already go on Amazon press buy now on something and literally have it the next morning on our doorstep. And in some cases I've ordered products and two hours later they're sitting on my doorstep and Amazon saying they're going to reduce fulfillment time by 25% with these new robots. I don't have a ton of information about what these robots actually look like or how they're going to speed up fulfillment time. I just find it absolutely fascinating that with how fast Amazon already is, they're still trying to make it faster. It's only a matter of time before we place an order on Amazon and a drone is dropping it on our doorstep 15 minutes after we order it in Dubai. They're actually getting AI powered self driving patrol cars. They're also equipped with drones. These things are designed to patrol neighborhoods. They've got 360 degree cameras on them and they, they've got facial recognition technology. They also have a drone on board where the autonomous robot could launch the drone to see into areas where the driving robot can't see. Apparently these are just essentially designed to gather information and if it sees something out of the ordinary, report it back to the police so the police could come and take care of it. This does feel a little bit like a step closer to like a robot police force that, that could be out there autonomously patrolling the streets for us. And I don't know how I feel about that yet. This is kind of creepy, honestly, but let's keep talking robots. This week this video of Figure 1 Dynamic Walking Robot was released and it shows a robot with a humanoid form that can actually walk around. You can see it here walking forward, sort of like a human, maybe a human with a bad back or something. They claim this robot is the first of its kind AI robotics company bringing a general purpose humanoid to light. At the moment, the goal of these robots are to help with undesirable jobs. They say. In the future, we believe humanoids will revolutionize a variety of industries, from corporate labor roles, to assisting individuals in the home, to caring for the elderly and to building new worlds on other planets. However, first applications will be in industries such as manufacturing, shipping and logistics, warehousing and retail, where labor shortages are the most severe. This week, Nvidia announced Eureka Extreme robot dexterity with large language models. This is an improved reinforcement learning technique that they actually figured out how to use to allow Robots to do complex things with their fingers. As one of the first tests that they use. This new reinforcement technique on text to video has been all the rage lately. We've gotten Gen 2, we've got gotten Pika Labs, we've gotten Moon Valley. There's all sorts of new AI text to video models out there. And today I came across this one called Morph Studio, which you can see some examples here. And this one claims to generate between five and seven second clips, but not only that, it will actually generate them in high resolution. So if we pop into their discord, let's give it a try real quick. If I type slash video and give it the prompt a wolf walking in the snow. It took about three minutes. And here's the video I got from it. Really, really impressive. Actually I'd only generated three seconds. But let's take a peek at it in full screen. Pretty decent quality as well. So at the moment it seems pretty similar to Gen 2 or Pika Labs or Moon Valley, but pretty impressive quality. Just looking at some of the other generations that have been here. This time lapse of the Milky Way galaxy looks really cool. This happy woman here looks really good. Swim in the sky blue fish. I mean these are pretty good other than the fact that that fish had a tail face for a. But this one was actually seven seconds. So it's definitely capable of generating some longer videos from time to time. It looks like maybe you got to give it this extra dash S7 to get the longer videos though. But as of right now, it appears to be free to use. You can find it over@morphstudio.com and joining their discord. Here's another tool that I came across. This Masterpiece X company partnered with Nvidia recently to generate these 3D animated assets from a text prompt. Now, I haven't tried this out yet. If I click on generate3d now, you can see it starts us off with 250 credits. And the credit cost to generate is about 50 credits. You can generate objects, animals or humans. Let's go ahead and generate an animal. You can see the estimated time is two to eight minutes per 3D model. I'm going to click Next Step to add shape details. Let's just go ahead and do a wolf. Next step to add paint. It gives us some guidelines here. Paper r realistic 4k, gray and white. I click Advanced Settings. It gives us some negative prompt options. And let's click generate 3D models. It says it may take two to eight minutes for each model. And it now says Processing up here and we have our wolf model. You can see it took about three minutes. So let's go ahead and click on it. And I would say it's got more of a cat look to it than a wolf, but not bad. I mean, we've come a long, long way when I've used Common Sense Machine CSM AI that took a lot longer to generate. Interesting. So if I look at it from some angles, you could kind of see, like, the wolf's nose right here. Like the front straight on looks like it's supposed to be a wolf, but then when you turn it, it looks like the nose kind of got put on the neck or something. So not quite perfect, but it looks like an interesting cat. Now, this tool also claimed that it can animate whatever you generate as well. However, it looks like the animate option is only for the human models. So let's go ahead and create a human real quick. Let's generate a pirate. I'm just going to throw in some keywords here, like, like 4K and ArtStation and cartoon. And now it gives us the animation option. So we've got bones only, which looks like we can rig it for other tools or we can have this tool animate it for us. Let's do bones and animation. And let's do it punching and let's generate and see what we get with an animated human model. Well, it took about six minutes to generate this one, probably because of the animation. Let's take a peek. Okay. I mean, it's punching, I guess. Looks more like a zombie mutant than a pirate. But again, this is as bad as it's gonna get. And the amount of time it took to generate this versus something like Common Sense Machines does is, I mean, just a fraction of the amount of time generated this in six minutes. CSM would have generated this in, like 24 hours and it wouldn't have been animated. So pretty cool, pretty fun to play around with. I'm sure if you figured out the prompting and got better with the prompting, you can generate stuff much better than this. Just looking through some of the examples here, like this tiger and these people and this horse. Other people are getting some much better generations out of it than what I'm getting. So it's definitely got some promise. I just haven't spent enough time figuring out how to prompt it properly yet. And there you have it. There's the news roundup in the AI space for this week. Again, it was a bit of a slower week in terms of announcements, but still a lot of really, really cool stuff happened and some new tools and toys that we got to play with got released, so that's always a fun time. If you enjoy this video, make sure you check out FutureTools IE where I curate all of the latest AI news on a daily basis and I update the homepage with all of the coolest AI tools that I come across. I also have a weekly newsletter. If you sign up for the free newsletter, I'll send you the latest AI news and the coolest AI tools directly to your email inbox. You can find it all over at FutureTools IE, so check that out. And if you enjoyed this video and you want to stay in the loop with the latest AI news and tutorials, make sure you like this video and subscribe. Subscribe to this channel and I'll make sure more videos like this show up in your YouTube feed. Thanks so much for nerding out with me today. I really, really appreciate you. And thanks once again to Wirestock for sponsoring this video. You guys are awesome as well. I hope you learned something new and discovered some cool new AI tech or news because that's always my goal with these videos and I'm always having a blast making them and sharing them with you. So thank you once again. Really appreciate you. See you in the next video. Bye.

By.

Sam.