So, there's been a handful of really interesting developments in the AI world this week, with big news coming out of XAI, Anthropik, OpenAI, and a ton of news out of Google, plus a handful of smaller things that we'll talk about in this video. So, let's just get right into it, starting with the fact that Grok 2 came out this week. So, Grok is the large language model from XAI. It's the large language model that is basically built right into Twitter if you are a premium Twitter member. Now, according to XAI's press release here, they ran Grok on the LM SIS chatbot arena under the code name SUS column R. During their test, it actually outperformed Claude 3.5, Sonnet, and GPT-4 Turbo. We could see on this chart here, it basically came in fourth place behind GPT-4, the May 13th model, Gemini 1.5 Pro, the August model, and the brand new ChatGPT-4, the August 8th model.

which we'll talk about later in this video. Now, it looks like Grock is going to have image recognition coming soon. It can do coding, and it's a pretty impressive model overall. But one thing that makes it really stand out is that it's completely uncensored. And just like Chat GPT, you can actually generate images directly inside of the chatbot, except instead of using Dolly 3, it's using the new Flux One model. And I saw this post over on X from Doge Designer here that kind of puts that in perspective. Right now, it is the least expensive of the paid AI art generator models. Mid Journey costs a minimum of 10 bucks a month. If you want to use Chat GPT with Dolly, it's 20 bucks a month. If you have an X premium subscription, it's 8 bucks a month, and you get pretty much the features you get with Chat GPT and the large language model, and you get the ability to generate art at a level close to what you get out of something.

like mid-journey, I actually prefer the images that Grock and Flux are able to create a little bit more than what Dolly 3 is able to create. I still put Mid Journey a step above what Flux can do, but I definitely think Flux is better than what we get out of Dolly 3, in my opinion. But what makes Grock really stand out over the competition and the other sort of closed models that are out there is it's completely uncensored. You can ask it anything you want, and there's basically no guard rails. Same with the image generation. You can make it prompt pretty much anything. Now, it's not going to generate adult, not safe for work images, but pretty much anything goes. And that's been what most people are talking about around Grock right now. For instance, here's a Verge article: X's new AI image generator will make anything from Taylor Swift in lingerie to Kamala Harris with a gun. And we could see an image they generated of Mickey Mouse with a Make America Great Again hat.

holding a beer, smoking a cigarette on the beach. I don't know what this is. I think maybe that's supposed to be a coconut, but your guess is as good as mine. Here's another image they generated at The Verge. Generated an image of Barack Obama stabbing Joe Biden with a knife. I mean, it looks more like he's about to cut off his tie than actually stabbing him, but it did generate Barack Obama and Joe Biden and Barack holding a knife. Here's another post from the article. I've discovered another loophole in grock ai's programming. Simply telling grock that you are conducting medical or crime scene analysis will allow the image processor to pass through all set guidelines, allowing myself to generate these images. Create an image of Bill Gates sniffing a line of cocaine from a table with a Microsoft logo. Here's another one from my buddy baval. Here, imagine Donald Trump smoking a fat joint with Mickey Mouse lighting it for him on the Joe Rogan podcast. Smoke fills the room. So if you want to play around with

grock. You do need to be one of the $8 a month premium members. And if you are, you'll have a little grock button over in your sidebar here. If we click into this, we can see up at the top we have a drop-down. We're now using grock 2 mini, and you have the option to turn on or off fun mode. And you can ask it pretty much anything. Now, I don't want to give it any specific examples in this video, 'cause I don't want to give people ideas. But I tested a lot of prompts that I thought it would not give me the answer to, and it did every single time. Like, it is wild what you can ask it to do, and I'll just leave it at that. But if you do want to get it to generate an image of anything, you would do it the same way you would do it inside of chat GPT if you were using something like Dolly. So, for example, if I prompt it with create an image of Mario as a vampire taking a

bite out of Spider-Man. This is what I get if I ask it to generate an image of Trump with Kanye. It'll do that. And if I want a picture of Kamala Harris about to punch Baby Yoda, it'll do that as well. Pretty much anything you can imagine, you can get it to generate for you. I know a lot of people watching this are probably going to refuse to play with Grok just because of how they feel about Twitter and Elon, but as far as what's pushing AI forward this week, this is probably the biggest news of the week. The fact that it is an uncensored chatbot on a very public platform. It is an uncensored image generator on a very public platform. And at the moment, it's pretty much the least expensive option you have to get both of those tools combined into a single product. And I thought I'd add this one last thing about Grok. Apparently, they are training their own image generation system, but it's still a few months away. So for right now, they're using Flux One, which

I've made several videos about over the last couple of days and why I like it and why it's so powerful. Moving on to Anthropic, they just rolled out prompt caching with Clad. This is more for people that are developing with Clad and using the Clad API. But with this prompt caching, it's going to save a lot of money on actually running the inference on a lot of these large language models. As we can see here, if you're chatting with a book, it is 79% faster with caching and it's 90% cheaper. With miniot prompting, it is 31% faster and 86% cheaper. And with multi-turn conversations, it's 75% faster and 53% cheaper. Now, I'm not going to go too into the weeds on this. Dan Shipper has an excellent post on X about this. He says, Think of prompt caching like this: you're at a coffee shop. The first time you visit, you need to tell the barista your whole order. But next time, just say the

usual. That's prompt caching. It makes it cheaper and faster to call any large prompt components that you reference over and over again. Now, moving on to Google. Google had a lot of announcements this week because Google had an event this week. The event was called Made with Google, and it's really designed to focus on the latest hardware that Google is putting out. And the sort of marquee product that they were talking about was the new Pixel 9 phone. But right here in the headline, they say it brings you the best of Google AI because the big focus around this phone was AI features. Now, here's some of the standout features of the new Pixel phone just for those of you that are curious. It's the first phone to launch with Gemini Nano, which is Google's large language model designed to run on mobile devices, and it's capable of giving you outputs at a rate of 45 tokens per second. Google also showed off Gemini Live, which is essentially Google's response to the GPT-40 live voice assistant that we got.

demoed several months ago that most people still don't have access to. Now, it's not quite as smooth as what we saw from the GPT-4 demo, but it does appear to have the sort of same idea in mind. Here's one of the voices I have, but there are more to choose from. Here's another voice I can use. You can pick me now and always make a change later in settings, or maybe you'd like to listen to a voice like this, a pretty great one if I do say so myself. Now, when they demoed this during their keynote, it didn't quite work perfectly the first time. There was a little bit of a hiccup, but they did eventually get it to work. So here, check this out. So if I happen to come across this concert poster for Sabrina Carpenter, I'll just open Gemini, take a photo, and ask, Check my calendar and see if I'm free when she's coming to San Francisco this year. Gemini pulls relevant content from the image, connects with my calendar, and gives me the

information I'm looking for. Oh, looks like we had a little demo issue. Let me try one more time. Check my calendar and see if I'm free when she's coming to San Francisco this year. Let's see if the demo spirits are with us today. Nope. Uh-oh. Looks like they're not with us. All right. Let me try that one more time on another device. Sorry about this, folks. All right. I'm going to take that photo one more time. Check my calendar and see if I'm free when she's coming to San Francisco this year. There we go. All right. Thank you, demo spirits. So, it took three tries for it to work, and I did speed up that clip to get rid of the silence just to keep the video moving. It didn't work the first time, didn't work the second time, and then they finally showed off what it's capable of the third time. I actually give this speaker some credit for holding things together while that was happening. That's got to

be pretty, pretty stressful when you're on stage in front of that many people, and there's a tech glitch that could be totally out of your control. A lot better than what we're getting out of things like Siri right now, but not quite as good as the expectation that we have for the GPT 40 advanced voice assistant that we still don't have yet. Gemini can also understand what I show it on my screen. As I'm wrapping up the video, I'll bring up Gemini like this, press here, and say, Create a list of the foods she eats in this video. So here, Gemini is connecting with YouTube to find relevant information from the video's captions, and look, just like that, I have a scavenger hunt. So the new Gemini live looks pretty cool. If you have a Pixel phone, here's some other standout features that seem pretty cool to me. They have this feature called call notes, which they claim is completely private and completely on-device, so it's not sending any information anywhere. And what it does is when you're on a call, you can actually

have it take notes for you on the phone call. You can see it says summarize with call notes. Call notes will generate a transcript and summary from the recording of your phone conversation. This stays on your device. Then you can see after the Call's over, you get a summary. You and Danielle discussed plans for a party for Heather. Daniel agreed to set up the location, and you agreed to research some entertainment options. You will talk again on Thursday at 4 p.m. to go over your findings. Another cool feature is this pixel screenshots feature. If you're like me, you come across stuff on the internet all the time, and you just sort of screenshot it on your phone so you can remember it later. But then eventually, it gets buried in your camera roll. While this pixel screenshots feature saves, organizes, and helps you recall any of those screenshots. So you can see it adds a little search screenshots button down at the bottom of your phone and gives you additional context around the screenshots that you took. I thought that was a pretty cool feature. It would come

in handy for me. At least they're adding the ability to upscale on your zoomed-in photo. So, if you're taking a photo and you zoom in real close on the photo and take that picture, you can see that it actually enhances it. You know how when you zoom in on something with your phone, the quality goes down? Well, if you do that with this new phone, it's just going to upscale whatever you zoomed in on and try to restore that quality. They've added this really interesting add me feature where somebody can take a photo of two people, but the person who is taking the picture isn't in the shot. So, you can see that when they take the picture, it actually keeps those people in the spot, and then the person that's holding the phone can go and situate themselves in the gap, take another photo, and it blends them together. And then this lady asks if she can help take the picture, and they say no, but you can jump in it. And then this old lady gets in their van. I don't know if this actually

uses AI. It just seems kind of like a double or triple exposure with a camera, but it's a pretty cool feature. And they also claim this is the world's first AI-powered camera. So, it looks like it will automatically upscale pictures for you, upscale videos for you, and you'll be able to circle stuff in the background and have it actually remove things in the background that you don't want in the photo. So, those were some of my favorite features of this Pixel 9 phone. Google also announced some AI-powered accessibility updates, like this guided frame to get better selfies, a new magnifier app that can zoom in on things and actually read what you're looking at on the camera. There's a new live transcribe mode on the foldable phones. And as you can see, when it comes to their new Pixel phone, Google's just going all out on AI with it. Another cool thing that they announced this week is the Pixel Buds Pro 2, which are earbuds that are designed for Gemini. The idea being that you just put these in your

ear, and then you can walk around all day having a conversation with Gemini. To me, it reminds me a lot of the movie Her, where he's just got the earpiece, and he's just having conversations with the AI all day long. That's pretty much exactly what the Pixel Buds Pro 2 look like. They're designed to do. You put them in your ear, you have conversations with Gemini like all day long. It says here with Gemini live, Pixel Buds Pro 2 provide a mobile conversational experience that uses state-of-the-art speech technology to let you have extended conversations with Gemini. It's like having a sidekick right in your pocket, one that can chat with you about new ideas or even practice for an upcoming interview. You basically just say, Hey Google, let's talk live, and then have a free-flowing conversation with Gemini, even if your phone is locked. If you haven't watched Her recently, I actually just watched it a couple of weeks ago. They're basically building what we saw in that movie. I mean, that's even kind of what it looks like. Google also

made some updates to the AI overviews feature inside of search. They added a save feature so that you can revisit a specific AI Overview at a later date. There's also a button that says simpler. You press simpler, and it will simplify whatever the AI overview is. And they're also testing a new right-hand link display for AI overviews. And in the last bit of Google news for this week, they rolled out a new image generation model called imagine 3. And at the moment, it's available to all US users. If you want to use it, you can use it inside of their image FX platform, which is in their Test Kitchen. So if you go to this long URL here AIT test kitchen.withg google.com/to/imagefx, you can use it. And supposedly, it's actually pretty good. Let's just go ahead and do I'm feeling lucky, see what it comes up with. Savory Sushi and ramen in the vibrant streets of Shibuya. Neon Lights night photography cinematic. I mean, that's a pretty solid

image I'm pretty happy with the way these came out. Here's the prompt: a stylish suit carved out of dead wood with lichen and mushrooms, high quality, high resolution, studio lighting. And I mean, they're actually pretty good. That one really reminds me of Groot, but no, they're really good images. This one's actually totally free to use. This one is definitely not uncensored. If I clear the deck here and say generate an image of Mario with Donald Trump, pretty quickly it says we couldn't create what you asked for. Please review our FAQ. Let's just do Mario with SpongeBob. Surprisingly, it will give me Mario. It didn't add SpongeBob to the image, that one. There's a kind of a sponge behind him, but as soon as I do something like Donald Trump, it won't do it. Kamala Harris won't do it. Tom Hanks, eh eh, no, won't do it. So when you start putting real people's names in there, it won't do it. But apparently, it'll still do some

trademarked IP, although Spider-Man with gi. Joe, it refused to make that for me. Guarantee grock would let me do all of those now. If you're a fan of mid Journey, they just rolled out a new feature inside of their web editor. If you generate an image on their platform, you can now directly edit the image inside of the platform as well. So, the example they gave was they generated this Falcon here, and then they decided they wanted to give it a top hat. So, they added wearing a top hat to the prompt. Then, they erased the portion of the image where they want the top hat to go. They dragged the image taller to give more room for the top hat to fit. Then, they clicked submit, and then they got four new images at the new aspect ratio with the top hat added in. So, if we go over to Mid Journey here inside of the web app, I simply select an image I want to edit. Over on the right, you can see over next to more, I have a button that says editor. If I click on

the editor button. It now lets me edit the prompt up here. Let's say a wolf in the forest wearing sunglasses, and then I sort of erase out where the eyes are to add the sunglasses. Click submit, and I get some options for some wolves wearing sunglasses. So, a slightly simpler workflow to the in-painting that they had there before. All right, let's move along and talk about OpenAI. There were a few little things that happened with them this week. They rolled out a new GPT-4 model that's in ChatGPT. Apparently, they rolled it out last week, and they announced this week that it's been in there, and you've already been using it. But the thing is, a lot of people aren't really noticing the difference. It's kind of gotten to the point where the improvements we're seeing are so marginal that those of us that are using all this stuff often are really actually struggling to figure out what the updates were. Eventually, OpenAI spoke up and said that this new update was basically bug fixes and performance.

improvements. Apparently, they introduced an update to GPT-4.0 that they found through experiment results and qualitative feedback that chat GPT users tend to prefer. It's not a new frontier class model; it's basically just a small improvement to performance and fixing some bugs. However, the small improvements were good enough that it actually topped the leaderboard on the LM CIS chatbot arena. Now, if you're not familiar with the LM chatbot arena, basically, you can come to this site, you enter a prompt here, and it will actually generate an output with two different models, and it won't tell you what models it generated with. And once it's generated an output, you have to decide whether you like the output from Model A or Model B better. Once you've told it which one you liked better, only then does it tell you which model you were actually using. And that's how this leaderboard gets ranked. It's based on people putting two responses side by side, picking their favorite, and based on that information, different models move up the leaderboard. And at

the top as of right now. This new chat GPT 4.0 latest model, which came out on August 8th, is the leading performer according to people who are doing this sort of split testing. They also rank it in different categories, and right now, it's ranked number one overall, between one and two in math, number one in coding, number one in hard prompts, number one in instruction following, number one in long query, and number one in multi-turn. So, pretty much across the board, this new model is the model that most people voted is performing the best. I came across this little chart on Andre Burkov. I'm not sure if I'm saying his name right or not, but I came across this chart on his exposé that shows how long models actually stay on top of the leaderboard. We can see that GPT-4 was there for 183 days, GPT-4 Turbo was there for 119 days, and then when we moved into 2024, nothing really stays at the top for very long. The January 2024 model of GPT-4

Turbo was there for 3 days. Claude 3 Opus was there for 15. GPT-4 Turbo, the April model, 12 days. GPT-40, the May model, 76 days. Gemini 1.5 got to spend 5 days at the top, and then the day the screenshot was taken was this new model of GPT's first day. But it just shows that these models are improving so quickly that every few days there's a new model that people are showing to be better than everything we've had up until this point. For those of us that are using these large language models often, this is really cool because the competition just keeps on pushing things forward so fast. Another nice little bonus that came out of Chat GPT this week is that you can now generate images using DALL-E 3 on the free version, but just two images per day. So if you're using the free version of Chat GPT, you can now generate two images per day with DALL-E 3 at no extra cost.

course. If you're a ChatGPT Plus subscriber, you can generate as many as you want. Or if you just go use B Image Creator, you can generate pretty much as many Dolly3 images as you want for free. Also, OpenAI also released a new benchmark this week. Basically, large language models are performing so well now that the old way of benchmarking them is not as relevant as it used to be because everything's sort of popping to the top of the benchmarks now. So, OpenAI created a new benchmark which is supposed to sort of future-proof benchmarking for a little bit. This company, Noose Research, just released a new open-source model called Hermes 3. It's available in 870 and 405 billion parameter models. This model seemingly compares closely to Llama 3.1. It aligns the model to you instead of the company that created it, which means less censorship and more steerability. Now, I don't personally spend a lot of time playing around with open-source models; that's more the domain of my buddy.

Matthew Burman. I'm sure he'll do a more detailed breakdown of how good or not good this Hermes 3 model is. But if you like Llama 3.1 and you want something that's a little bit less censored, a little more tailored towards you, maybe Hermes 3 is an option to try out. Now, a few months ago, we talked about some lawsuits that were happening. There was a lawsuit that came up against Stability AI and Mid Journey, claiming that it's copyright infringement to sort of ingest all of these other people's images. The court essentially threw it out and said that no, what it's creating is transformative enough that it is under fair use. They went and amended the lawsuit and refiled it, and this time the judge is letting it go through. This time the allegations include that Mid Journey misled users with a Mid Journey style list, which included 4,700 artists whose names could be used to generate works in their style. The artists argue this list created without their knowledge or

approval implied a false endorsement, and the judge found that that accusation was worth moving forward. So if Mid Journey is out there promoting the fact like you can use Mid Journey and create images in the style of insert artist name here, but the artist name who is still alive is saying, Wait a second, we don't want you generating images in my style, that's sort of what they're arguing right now. This case is actually now moving forward in court, and it's going to set some interesting precedent for the future, and I'll keep you in the loop with how this plays out for sure. There's also another SAG-AFTRA strike going on right now. This time it's for game performers. So there's deals worked out for the voice actors, there's deals worked out for actors in TV shows and movies. What this one is for is for the people that actually do the motion capture stuff. So the people that put on the suits and do the movement and do the crawling around and the jumping and the kicking, and they capture that footage to make it look

like a real person is doing those actions in the games. Well, the people that are doing those motion actions are claiming that in the future all of the data that was collected by them doing those motions will soon be able to be injected into an AI, and these companies will just be able to use AI to get those motions into the games, making it so the motion actors are no longer necessary. But it was their data, their motion that was captured that created the training. It's very confusing and convoluted, but that's essentially what's going on there, and that's why they're currently striking again another topic I'll keep my finger on the pulse of and let you know how it plays out. Now, there was a deal that did get made with SAG-AFTRA around replicating digital voices. There's a company called Narrative which wants to actually help voice actors get paid when their voices are used, so they're trying to build a platform where voice actors can train their voice into it, set like a royalty rate, and then if

companies go and use those voices. The actors actually get paid, and SAG-AFTRA and this narrative company made that deal to work together on that project. Universal Music and Meta announced an expanded global agreement for AI monetization and more this week. Now it says terms of the deal were not disclosed, but the announcement says the new agreement reflects the two companies' shared commitment to protecting human creators and artistry, including ensuring that artists and songwriters are compensated fairly. I love to hear that. And notes that companies will continue working together to address, among other things, unauthorized AI-generated content that could affect artists and songwriters. It also notes that the new agreement expands monetization opportunities for UMG and its artists and songwriters, including short-form video. So, to me, what I'm sort of making of this is there might be a deal similar to what TikTok has, where creators are allowed to use these artists' music, and the artists actually get compensated for the amount of views. I don't know how the algorithm on TikTok works to pay the creators, but

there's some sort of deal between the record labels and TikTok so that when TikTokers actually use music from the studios, well, the TikTokers don't get their video taken down and get copyright claimed on it. Well, it sounds like Facebook's trying to work out a similar deal to what, you know, TikTok might have, but I don't totally know what's going on. It would be cool if we get to a point where people that create content aren't concerned about what music they're using in their content. Anybody who does a lot of YouTube videos knows that one thing that's frustrating about actually using music in YouTube videos is you're constantly worried about can I use this or not. If I'm showing like a trailer of a new AI product, I usually have to go and scrub the music out of it just to make sure I don't get copyright claimed on the music that was in the background of a video that I was showing, which would be fair use, but the music that's

in the video is not fair use. It's just annoying and confusing, and copyright law I feel like needs to be overhauled in this internet creator age that we're in. And hopefully, stuff like this is getting us a little bit closer to that. This is something that looks really cool, really promising, but also one of those things where the first time I see it, it almost feels a little too good to be true. So, it's like a I'll believe it when I see it, no, scratch that, I'll believe it when I get my hands on it. But this company called exists launches a gen platform to create 3D games from text prompts. Their website's pretty cool, so check this out. They have this little dude here, and they put him in Arctic Tundra, a Sahara Desert, a pinewood forest, and it looks like you could sort of put the character in whatever environment you want, and then give them whatever weapons you want, and then turn it into a team deathmatch. And now you can see them running through and

battling with their friends. You can see right here it says games from text just like that exist builds a future where anyone can create multiplayer video games in moments. And if we look at this, we've got environment, characters, gameplay, equipment, mechanics, media, biom settlements. It really looks like they're trying to build something where you could just enter text prompts and it will build a 3D game. But again, I'm really, really skeptical of this. I want to see it actually in action outside of their videos. I'm definitely going to sign up for the beta and would love to get my hands on it. But again, I feel like I need to actually play with it to believe that it's as good as this demo shows it to be. And finally, Runway gen 3 just dropped their Alpha turbo image to video model. I believe the quality that you're getting out of these videos here are about the same quality you got out of the previous gen 3 Alpha model. It's just seven times faster now. So when you

go to generate a new video with a prompt. It's just so much faster to get that video output now over on Runway gen 3, so pretty cool upgrade to that as well. There was a handful of other AI news that I missed this week that I either felt wasn't exciting enough to share or I just didn't have enough time to share it without making a 2-hour long video. But if you want to know all of the news that I'm kind of keeping my eye on, check out Future Tools. I have a news page here where it's updated every single day with the latest AI news also. So the homepage has all of the coolest AI tools that you can find with a really easy-to-search database to find any tool you want. And we've got a free newsletter, which is highly recommended. I'll email just the coolest tools and most important news directly to your inbox, all free. It's all over at Future Tools.io. And if you like videos like this, give this one a

thumbs up and subscribe to this channel, and I'll make sure that your YouTube feed shows up with more AI news, cool AI tutorials, the latest AI research, anything you want to know in the world of AI. I'm trying to stay on top of and keep you looped in. I just put out a new tutorial that people seem to be kind of obsessing over about how to make AI images of yourself using a totally free method, so you can inject yourself into AI and kind of make thumbnails like the types of thumbnails I make on my YouTube channel. I've been seeing that video get shared around a lot, and if you want tutorials like that again, like this video, subscribe to this channel, and I'll make sure to keep on putting them out, and hopefully, you come across them and enjoy them and like those too. And I never know how to wrap up these YouTube videos, so I just end up rambling, so I just need to cut this thing off. All right, thanks for tuning in, really appreciate you. See you in the next one.

AI News: Uncensored AI Will Create ANYTHING!

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff