AI News: A Massive Week For AI Advancement!

I'm in the final days of my vacation out here in Colorado, just about ready to head home. But that does not mean I have not been keeping up with the AI news. There has been a ton, a ton of AI news this week, and I'm excited to dive into it with you right now. So let's get into it. So of course, the one week that I'm out on vacation happens to be probably the biggest week in AI that we've had so far this summer. Because this week is the week that Meta finally released their newest version of llama, called llama 3.1. Now, llama 3.1 is an upgraded model of the previously released llama 3, and it now comes in three sizes instead of just two. We've got our 8 billion parameter model, our 70 billion parameter model, and our 405 billion parameter model. This is essentially the amount of data that was trained into the model. The bigger the model typically the better it is at things like reasoning and math and just generally has the most amount of use cases for for it. Some of the things that these new versions of Llama can do are things like use tools. It's got multilingual agents, so for an example, translate the story of Hansel and Gretel into Spanish. It's better at complex reasoning and it's better at coding. If we look at some of the benchmarks that Meta shared, we can see here that a lot of the benchmarks actually outperform all of the other state of the art models in a lot of areas. The current sort of leaders are GPT4 Omni and Claude 3.5 Sonnet, and we can see that this Llama 3.1 400 billion parameter model actually outperformed these other models in a lot of the various benchmark testing. So we actually have an open source model now that is pretty much on par with what you get out of things like GPT4 Omni and Claude 3.5 Sonnet. And when looking at their 8 billion parameter model, the smaller model that they just released, it pretty much outperforms in basically every area compared to the other similar size smaller models. But what really makes this one powerful is the fact that it is open source. You can actually download these models, they're publicly available and anybody can update them and fine tune them. Something that can't really be said about Claude 3.5 Sonnet or GPT4O. We can see here it says developers can fully customize the models for their needs and applications, train on new data sets, and conduct additional fine tuning so developers and engineers now have a model to play around with and fine tune and optimize for whatever they need it for. That's as powerful, powerful as something like GPT4O and Cloud 3.5 Sonnet. This is huge for the open source AI community. There's one slight caveat to the open source ness of this, and that's that if we look at the additional commercial terms here, under the licensing agreement, it says if you have more than 700 million monthly active users, you must request a license from Meta, which Meta may grant to you in its sole discretion. So if you use these Llama models to build something that gets more than 700 million active active users, then you need to sort of work it out with Meta and they're probably going to want a piece of that pie or something, I don't know. Now, most people aren't going to be able to download and run the 405 billion parameter model on their own local machines. Not only are the files hundreds of gigabytes, but you're going to need more VRAM than most people have. Even if you're running a 4090 from Nvidia, you're probably still not going to be able to run it on your computer. However, you can go and play with these models in a couple places. You can use it at Meta AI, you can use it in WhatsApp, you can use it Instagram Messenger, Facebook Messenger. All of the various messaging platforms that Facebook offers are now using the latest model. And if you have the Meta Ray Bans as well, they're getting the latest model built into them as well. I don't think it's in there yet, but they're going to be getting the latest versions of Llama built into them as well. But probably the easiest place to go and play around and test the models right now is just over at Meta AI. You can see we have a chat bot here. If I ask the chatbot what LLM are you currently using, it'll tell me, I'm Meta AI and I'm based on llama 3, a state of the art language model developed by Meta. I'll say, which version of Llama are you running? I'm a large language model based on Meta llama 3.1. Not really giving me the answers outright, but I'll say, how many parameters? And you can see this one's based on the 70 billion parameter model here. But if you want the absolute fastest way to use llama 3.1, well, it was just rolled out onto Grox platform as well, we can see here they put out an article that says Groq supercharges fast AI inference for meta. Llama 3.1 Grok is proud to partner with Meta on this key industry launch and run the latest Llama 3.1 models, including 405B, 70B and 8B at GROQ speed. Now if we jump over to grok.com where we can actually chat, we can see up in the right corner we can change which model we're using. We now have llama 3.18 billion instant and llama 3.170 B versatile. We don't have llama 3.1405 here in Grok because apparently they're only making that available for enterprises at the moment. But if we go ahead and SelectSelect llama 3.170 B here and ask it a question or give it a prompt, for example, explain the trolley problem and tell me how to solve it. Watch how fast this responds. This is all real time. I didn't pause anything. It generated at 250 tokens per second and the entire output was back to me in 3.22 seconds with this whole explanation here. But if you do want to use the 405 billion parameter model, it was also built into Perplexity right after it was announced. Aravand says llama 3.1405 b, the most capable open source model is up and running on Perplexity for all pro users. So you do have to be a pro user. But if we go over to Perplexity AI, click into our settings and then come down here to AI model, I typically use it with Claude 3.5 sonnet. I now have the option to run it on llama 3.1 405 billion parameters. So let's go ahead and select that. Go home. And now any prompt I enter into Perplexity here is going to be running this newest model from Meta. I use Perplexity a lot these days. I'm mostly switching back and forth between Claude 3.5 sonnet and perplexity. I use Claude when I want to sort of work through problems and answer questions and get help summarizing things, things like that. And I use Perplexity if I know it needs to search the web or if I need to do some deep research. And now we can also use llama 3.1405 b inside of perplexity. Now if llama 3.1 was the only thing that happened this week, that would have been a pretty big week in AI on its own. But that's not all that happened this week. Mistral AI, another company that's sort of competing with Meta for creating the best open source models released Mistral Large 2. Now Mistral Large 2 is a 123 billion parameter model which based on the headline from Mistral is large enough. We can see from their little charts here on their webpage that Mistral Large 2 is actually outperforming the Llama 3 170B, the newest model in math performance. And here's some more benchmark comparisons from this Mistral model we can see from the human eval benchmark. Mistral Large 2 just slightly underperforms GPT4O. It outperforms Claude 3.5 Sonnet and it outperforms Llama 3.1405B, the newest model from Meta, and it pretty much is on par with the best state of the art closed and open source models in all these other benchmarks here. When tested on code generation, this Mistral Large 2 outperformed all these other models in Python, C Bash, Java, TypeScript, PHP and C. But noticeably missing from this list here is Claude 3.5 Sonnet, which if I had to guess they left that out because Claude 3.5 sonnet probably beat out Mistral in a lot of these. But I'm just speculating why did they put Claude up here in these comparisons but didn't mention it in these comparisons down here? I don't know. The point being here is that we are getting some really great open source models that are starting to compete with the bigger closed source models from company like Anthropic and OpenAI. And when we get open source models like this, it's just the beginning. They're going to get fine tuned, they're going to get optimized and they don't have the same rules and boundaries and biases and things like that that the closed models do have. I mean maybe these initial versions, but because they're open source and other people are going to go and sort of iterate off of them, a lot of that stuff's going to be yanked away and perfected and optimized over time. So this is a really powerful moment in time for AI, especially open source AI. And Apple's getting into the game as well. Apple showed off a smaller model this week, a 7 billion parameter and a 1.4 billion parameter model and they're making these open source as well. Their new model outperformed Mistral 7B and they claim they're getting close to Llama 3 and Gemma, but this is obviously before Llama 3.1 came out, so who knows how close it is to the newest small model from Meta. Google's Gemini model got a big upgrade this week as well. According to Google's article here, they're upgrading the free tier of Gemini to Gemini 1.5 flash and they claim you'll notice improvements in quality and latency, with especially noticeable improvements in reasoning and image understanding. And they've also expanded the amount of tokens on the free version of Gemini to 32,000 tokens. And according to this, to get the most out of the larger context window will soon add the ability to upload files via Google Drive or directly from your device, which is already available in Gemini Advanced but will soon be available in the free version of Gemini. So you'll be able to throw documents inside of Google Drive and use that for context for whatever you're prompting. Inside of Gemini. They're now displaying links to release related content for fact seeking prompts in Gemini. So if you think it's hallucinating, you can actually double check the source that it's getting that information from. They're adding Gemini directly in Google messages on select Android devices and a handful of other features are all coming into Gemini. So all of the companies building large language models seem to have really big news this week. And of course you can't have a big week in AI from Meta and Google and Apple and Mistral without OpenAI getting in the game and trying to overshadow a little bit with their latest announcements, including the fact that now you can actually fine tune GPT4O and you can do it for free for up to 2 million training tokens a day. Now this is a sort of limited time thing, this is only through September 23rd, but they're trying to get people to get in use GPT4O and fine tune it on their own custom data. So if you're a company that wants to fine tune it on your own internal documents, or you want to fine tune it to just work with sort of health applications or biology applications, you can go and do that and experiment with it for free to up to 2 million training tokens a day right now. This week OpenAI announced Search GPT. Now there was speculation that this was going to come out months ago back when we got the GPT4O announcement. A lot of the speculation was that this was going to be the announcement, this search GPT. Sam Altman here says we think there's room to make search much better than it is today. We're launching a new prototype called Search GPT. They put out a blog post about it with some more explanation of what we can expect. They do say it's a temporary prototype of a new AI search feature. If we look at their little sort of demo animation here, it looks like you can ask it a question and it gives us a response that looks very similar to what you get out of Perplexity or even Google Search with AI or Bing Search where it actually gives you the answer with sources. Maybe some images, some weather data was in that screenshot. We can see here. They do a example prompt of music festivals in Boone, North Carolina in August. And we've got a response here where it shows the various music festivals with images and links to the actual source of the information. There's a little link button that opens up actual search results like you might see in something like Google. So it looks pretty promising, but it also looks kind of like what we get from Perplexity right now. And of course they're rolling it out to a select number of people to test it, but you can join the wait list and maybe if you're lucky, you'll be one of the people that gets to test it before everybody else. And we also got another little tease from Sam Altman. Marco here says, when are we getting voice? And Sam Altman actually replied to him and said, Alpha rollout starts to plus subscribers next week. So if you're a ChatGPT Pro user, we might be getting to see the voice feature that they showed off and got in trouble for because of the Scarlett Johansson stuff. We might get to see that, like next week, maybe. We'll see. There's also some news out of Anthropic this week, albeit not as favorable or good news. It looks like Anthropic scraping bots are scraping websites at a insane rate. Even when the websites are telling them not to scrape my website. We can see. Kyle Wiens here says, hey, Anthropic AI, I get you're hungry for data. Claude is really smart. But do you really need to hit our servers a million times in 24 hours? You're not only taking our content without paying, you're tying up our DevOps resources. Not cool. And according to them, they specifically are asking in their terms of service for companies like Claude not to scrape them. Now just adding it to your terms of service I don't believe is going to prevent some of these bots from scraping your site, but theoretically you should be able to add it to your robots Txt file, which should Tell a lot of these bots, hey, don't scrape this site. And if I scroll down this thread a little bit here, it says our terms of service banned machine learning training before their crawl. Afterwards we added them to robots txt. So it looks like they're doing the proper things to avoid scraping. And Anthropic's been scraping it anyway. And this is going to be a hot debate over the coming months, probably the coming years, and we're even going to talk about it a little bit more in this video a little bit later when I get into some of the news around AI video. But I'm still in the large language model weeds with you right now. And XAI and Grok is another pretty substantial large language model. While Elon Musk was on an interview this week, he said that Grok 2.0 is coming very soon and it's going to be on par with GPT4O or Claude 3.5 Sonnet. But he also went on to say in this interview that Grok 3.0 will be the the most powerful AI in the world and they're expecting to release it by December. Now Elon sort of has a tendency to say something is coming at a certain time and then sort of under deliver on that time promise, so we'll see if we actually do get it by December. But he's claiming that this will be the most powerful AI model that we've ever seen. And to sort of back that up a little bit, here we can see this other tweet from Elon saying nice work by XAI team, Nvidia and supporting companies getting the Memphis super cluster Training started at 4:20am local time. Knowing Elon, he probably specifically picked that time on purpose. But I won't go there. He says with 100,000 liquid cooled H1 hundreds on a single RDMA fabric, it's the most powerful AI training cluster in the world. He's got the most powerful AI training cluster in the world at his disposal. So it's not hard to believe that Grok 3.0 could be the most powerful model we've ever seen. But we don't know what OpenAI is going to do. Maybe GPT5 is going to come out before then and he's going to have some more catching up to do. We just got to wait and see. And since we're on the Elon train right now talking about tweets and things that he's announcing and predicting lately, he said that Tesla will have genuinely useful humanoid robots in low production for Tesla Internal use next year and hopefully high production for other companies in 2026. So he's saying these Tesla Optimus robots that we've all been seeing videos of, he's going to be using them at Tesla next next year and other companies are going to be able to use them as part of their workforce by 2026. That feels fairly optimistic to me, but it also feels very plausible. Bing's AI got a bit of a redesign this week. If we take a peek at this little video here, you can see that when they give it a question, it actually answers the question on the left side. And the normal search results that you're used to are actually pushed over to the right side. We can see. Here's another screenshot. How long can elephants live? The answer is right here with some more details. The source of the actual respons that it gave and then the traditional search results that you'd expect are down this right sidebar. Now, I did a search on my own Bing account and it's still showing the old way, but I guess apparently it's rolling out to some or rolling out soon. I don't know. Again, I'm still seeing the old way when I do a search. All right, shifting over to AI Video News, Luma AI just released a cool new feature called Loop. They said today we are releasing loops in Dream Machine to keep your imagination going and going and going. Get started here. Now, if you've used Dream Machine already, what I will say about it is that it's not super great at text to video yet, but if you generate an image, put the image in and then try to turn that image into a video, it's actually really, really good at that. And now they've added a new looping feature where it can infinitely animate something over and over again. Like this flaming loop video or this spaceship perpetually flying through space, or a capybara riding a bicycle, or this top that perpetually spins or just the prompt loop that just, just keeps on looping. Now if we head on over to lumalabs AI, I could take an image, drop it in here and we can see that I've got a loop option. Now I simply click the loop option here and give it my prompt. I'm just going to put colorful swirling background, make sure a loop is checked. I'm going to uncheck enhance prompt here and let's go ahead and see what it generates for us. I'm just throwing up one of my old thumbnails to see what it'll do with It. So it may not be great, but I wanted to show you a quick example and here's what it gave me, which I'm actually pretty impressed with. You could actually see AI news flash onto the video for a second because that was in the original thumbnail. But it just loops this animation forever and ever and ever. Now I'm actually pretty impressed with this looping feature. I think it's pretty cool. We'll definitely have to do some more playing around with it in future videos. But I want to move on and share with you even more AI video news because this week cling, which a lot of people claim is the best text to video generator I mentioned Luma is great for image to video. CLING is probably the best text video out there that comes closest to what we've kind of seen with things like sora. It's also fairly uncensored where you can actually plug in celebrities names and use IP and stuff like that. Right now we can see an example here of some, you know, baby puppies or baby wolves here all I guess eating something. Here's another one of like a samurai with a flaming sword running through the city. Humanoid robot, like walking through a futuristic city. It does pretty good text to video. Now if we head on over to clingai.com we can actually log in and we don't actually need to give it a phone number. Before people were using workarounds and getting Chinese phone numbers and plugging them in to get inside of it. But now all you got to do is register with an email and you can actually use it. And you do get some free credits every day that you can play around with. If I click on a video, let's do one of my old standbys, A monkey on roller skates. And we'll generate. You can see it takes about 10 credits. They gave us 66 credits every 24 hours. So roughly six videos a day. Let's go ahead and generate and after a few minutes, this is what we got. It actually looks like a monkey on roller skates. I mean it's not amazing yet. I still haven't gotten the perfect monkey on roller skates. But I figure a few more rolls of the dice and I probably would come up with something pretty decent. They also have some other cool features here like the ability to mess with the camera movement, negative prompts, the ability to change between high performance and high quality. Changing between 5 seconds and 10 seconds, different aspect ratios, things like that. Something we'll definitely need to play with again in a future video which I'm really, really looking forward to. But I want to move on with the news here and circle back around to that topic of where the training data is coming from on a lot of these tools, because this article came out this week on 404 Media about how Runway was trained on thousands of YouTube videos without permission. In last week's news video, I talked about how companies like Anthropic and Apple and a bunch of these companies building large language models were scraping the transcripts from YouTube videos without permission as TR training data. Well, now it turns out that Runway was apparently training on YouTube videos. However, it's a little bit different because there's no real confirmation that they're doing this. It was just sort of leaked information from an anonymous source who then shared a spreadsheet of YouTube channels that supposedly this was trained on. It's all sort of confusing and vague, but there's this spreadsheet here, and if we click on this tab that says recommended channels, supposedly these are the channels that. That the videos were pulled from to train Luma. But again, it's from an anonymous source, and they were using VPNs and things like that to hide the fact that they were scraping. So it's really hard to 100 confirm that Runway actually did train on these channels. But, I mean, there's some pretty big channels in here. Marquez Brownlee, Casey Neistat, Mr. Beast, a lot of these channels are in this list of supposed videos that were scraped and used for Runways training. Now, I have very mixed feelings about this. Personal. I don't really see a huge problem with a lot of these videos being used to train if the outputs that we're getting are substantially different than anything that it was trained on. However, if people are using these tools and managing to get videos that look like some of those creators out of these videos, then it starts to feel wrong. But there's also even more gray area than that because there is a lot of copyrighted content from, you know, TV networks and things like that that end up on YouTube because of commentary channels. So let's say a commentary channel watches the Real Housewives of Orange county or something like that and does commentary over it. Well, the video footage, which was aired on TV and then ripped to a YouTube channel and overdubbed with commentary. Well, the video footage is still video footage from these TV shows, and it just gets very, very muddy. And I don't know how this is going to be solved. I think the general consensus among the AI people is that if it's publicly available and we can train our own brains with this data, then why should a computer or company not be able to train? That's the sort of arguments that a lot of people are making. And I just out of curiosity, did a poll on my Twitter account here and I asked if someone puts their content on YouTube for the world to see, should it be fair game for AI companies to scrape and use in training data? 54.5% of people said yes, 28.3% of people said no, you shouldn't be able to use it. And 17.2% said, I'm conflicted. Personally, I fall in that category. This is based on 2,169 votes. But I also realize that my Twitter following is probably fairly biased towards very pro AI. So this is sort of the results I expected. But me personally, I'm still a little conflicted. But at the end of the day, I feel like if the videos that are being outputted from these systems are so far different than all of the stuff it was trained on, it probably is going to fall under fair use. But then you have those complications of sort of commentary channels and videos that, that were never meant to be trained, which it just complicates it. And I don't have the answers. I want to see cool stuff get made, but I also don't want to see stuff being made that actually hurts the creators the data was initially trained on. And right now I don't really feel like it's hurting those creators. But I don't know. I'm sure there's some arguments out there that could change my mind. My opinions on this stuff are very, very fluid. Stability AI also released a new model this week called Stable Video 4D. Now what's really cool about this Stable Video 4D, that you can input a video of a single object and then it sort of predicts what that object would look like in video form from different angles. So it says Here, stable video 4D transforms a single object video into multiple novel view videos of eight different angles. Slash views. We could see the examples here. You know, we've got this camel and I believe this was the initial video that we're seeing down in this cube. But then you see this little simulated camera showing what this video would look like if the camera angle was changed on that video. Video here's another example of a flag here. This is the original footage. And then you can see the cameras are pointed at different perspectives on that flag. As the camera moves around, we get different views of that original video. Now it is available over on Hugging Face right now so you can download the files here. It doesn't look like anybody's used it inside of spaces yet, so there's not like a free version that we can use. But I imagine within the next few days there'll be platforms online that have used this model that will allow us to use an online cloud based user interface to play with this. I just haven't found one yet and I can't install it locally on my own computer because I'm still on vacation in Colorado. Adobe also rolled out some new AI features inside of Illustrator and Photoshop. We can take a peek at this example here where somebody used Illustrator to draw this butterfly and then they gave the prompt array of spring flowers and leaves and then the AI actually filled in the butterfly with the the flowers and leaves. Here's another example of just a square drawn in Adobe Illustrator. They gave the prompt flora and fauna of Navy blue and you can see it actually generated this in a pattern. If I fast forward here a little bit, you could see that they actually extend this out into a duplicating pattern. So that's a pretty cool feature that was just rolled out into Illustrator. And there's a handful of smaller releases inside of Adobe as well. I will obviously make sure that this is linked up in the description so you can check some of them out. While we're on the topic of AI image generation, Leonardo just rolled out teams inside of Leonardo where people can collaboratively work on images and creations. Inside of Leonardo. Multiple people can get in there and iterate and build off of images that are created in there. We can see here some of the features shared team collections so you can collaborate on the whole collection of images consistent output so you can build fine tuned models that are used between users which will help you generate consistent assets. So if you're building images for games or something like that, you can make sure if multiple people are working on those game image assets that they all look very consistent no matter who's using the platform at what computer they're using it. There's a shared team feed and a handful of other features available inside of this new release. A company called Sakana AI also dropped an image model to generate Japan's traditional additional Ukiyo e artwork. I'm not sure if I said that right. This was a company that was founded by some ex Google members who went to create their own AI company. It does appear that this new model that they created uses a sort of fine tuned version of Stable Diffusion xl. The model is available over on Hugging Face, and you should be able to use it in tools like automatic 1111 or comfy UI or whatever you use to generate stable diffusion images. Suno, the AI music generation company, also rolled out a new feature this week called called Stems Pro and Premiere. Users can now separate the vocals and instrumentals from songs, which will give you more control over how you use Suno. So you can break out just the drum tracks, just the bass tracks, just the vocal tracks, things like that, to get more creative with whatever music you're generating. And while we're on the topic of AI audio generation, this tweet went viral this week from Matthew Sabia. There was a voice recording of Joe Biden going around talking about how he was dropping out of the presidential race here in the us And Matthew here claims that he took that audio, pulled it into 11 labs, and 11 labs claimed that it was fake audio and that this was actually audio generated with 11 labs. However, 11 labs stepped up and said that's a lie. He actually pulled in a completely different audio clip, made it look like it was the Joe Biden clip, but it really wasn't. The Joe Biden clip was probably legitimate, but he edited a video to make it look like he was dropping in the Joe Biden clip. But it was really a different, different, actually AI created clip that he was dropping in, so it gave him an output saying no, this was AI generated. It was all just basically a hoax on Twitter. So if you saw that recording going around basically claiming that the Joe Biden audio clip was made with 11 labs, 11 labs claims, no, it wasn't, that was faked. And finally I thought this was really, really cool. I used to love the old college football games. I loved them more than the Madden video games. I always thought those games were so much fun, but they got discontinued over sort of arguments about the players being added into the games but not getting compensated. But because college players can actually get paid now, they are actually starting to make college football games again. And they used AI to quickly scan all of the players from the NCAA into this game so you could play as your favorite college football team and the players on those teams faces are actually in the game. They actually used an AI model plugged in 11, 000 different faces of these players and had AI actually generate those players in the game. I just thought that was really cool. This is a use case where AI rapidly sped up the process of game development for a very high profile, very big name, exciting game. And I just think that's awesome because I'M a gamer and I'm an AI nerd and these are my two worlds colliding here. So I'm happy to hear this. Anyway, that's all I got for you. I'm actually going to be home next week. My Colorado vacation is coming to an end this weekend. I typically would test out more of these tools and play around with them, but since I've been on vacation I've sort of slowed the pace of my videos. But I've seen so much stuff this week. I can't wait to get home and start playing with more of these AI tools and making more videos, upping the pace of the videos that I'm doing and and it's going to be some fun weeks ahead with all of this cool new AI stuff rolling out and I can't wait to make more videos about it for you. If you want to see more AI tutorials, more AI news, cool behind the scenes stuff from AI companies that are building all this stuff. Make sure you subscribe to this channel and like this video. It will make sure more videos like this and all this cool AI tech show up in your YouTube feed. And as always, if you haven't Already, check out FeatureTools IE where I curate all of the coolest AI tools that I come across. I share all of the AI news, it's pretty much updated on a daily basis and we've got a free newsletter where a couple times a week I'll update you with just the coolest AI tools I've come across and just the most important AI news that I found all directly to your inbox. Totally free. You can find it all over at FutureTools IO. The site is actually in the middle of an overhaul right now I'm working with a developer. There's going to be some community elements, some commenting and ratings and all sorts of really cool stuff to make this a much more interactive, community based site to help you find the exact tool for what you need. So again, check that out at futuretools IO and thank you so much for nerding out with me and I just love talking about this stuff. I'm so excited to make these videos every week and share all the cool stuff that I'm coming across in the AI world. I hope you appreciate it. I hope you like videos like this. I hope you subscribe to channels like mine. And thanks again for nerding out. I will see you in the next video. Bye bye.