Google Lied To Us! (And Other AI News)

Well, it's been another really busy week in the world of AI and I'm here to break it down for you, starting with probably the biggest news of the week, which was the introduction of Gemini, Google's answer to ChatGPT. I'm not going to break this one down too much in this video because I made an entire deep dive video about everything we know so far about Gemini right here. However, the Quick tldr, they're releasing it in three models. Gemini Nano, which is mostly for mobile devices. Gemini Pro, which is which we're currently getting access to in Bard right now, but so far it hasn't been that impressive. And Gemini Ultra, which is the really cool model that they showed off in all of their demo videos, but that nobody actually has access to yet. They also released a series of benchmark comparisons showing how it compares to GPT4. However, a lot of people have pointed out that it's comparing it to the older model of GPT4 and not the newer GPT4 Turbo model. But based on their comparisons, it essentially beats GPT4, the previous generation, in almost every single benchmark. But again, this is Gemini Ultra, the one that they say is coming next year. Again, I made a full breakdown of it. You can find it on my channel. It's called Gemini is Here. However, what I found out today was that Google wasn't entirely honest with their promotion of this new Gemini. I came across this TechCrunch article today called Google's Best Gemini Demo was Fake. In my previous video, I showed off this demo where this person was putting visuals in front of a camera and then Gemini was responding in real time to everything it was seeing. However, if we go to the very front of this video, there is this disclaimer. We've been testing the capabilities of Gemini, our new multimodal AI model. We've been capturing footage to test it on a wide range of challenges, showing it a series of images and asking it to reason about what it sees. Now it's saying it's showing it a series of images, but when we watch this video back, it's showing this person draw stuff in real time and hold up a duck in front of the camera and watch a video of a ball hidden under cups. All of this looks super, super impressive and we see the response from Gemini in real time. However, there is a tiny disclaimer in small print at the bottom of the video when it first starts saying sequences shortened throughout. I actually missed that the first time I watched you the video and when I said that this stuff Seemed to be responding in real time. Well, it's not. They shortened it all and edited it all down to make it seem like it was responding in real time. It actually turns out most of the video was sort of faked to make Gemini look a lot better than it really is. A spokesperson actually said, we created the demo by capturing footage in order to test Gemini's capabilities on a wide range of challenges. Then we prompted Gemini using still image frames from the footage and prompting via text. They used a combination of image prompts and text prompts to get the responses that Gemini was giving. Now, to be fair, Google did put out this blog post on their developer blog called How It's Made Interacting with Gemini through Multimodal Prompting, where it actually shows how it was done. Let's try an experiment. We'll show this picture to our multimodal model, Gemini, and ask it to describe what it sees. Tell me what you see. I see a person's right hand. The hand is open with the fingers spread apart. Okay, let's try this one person's knocking on a wooden door. How about this one? I see a hand with two fingers extended, which is a common symbol for number too. But wait. What if we asked Gemini to reason about all of these images together? What do you think I'm doing? Hint. It's a game. Gemini, you're playing rock paper scissors. This is the actual way it was prompted, but the video led us to believe that they were actually doing this in real time. And it was saying, I know what you're doing. You're playing rock paper scissors. And if you look through this actual blog post on the developer blog, pretty much every single example they showed in the video where they made it look like it was watching a camera video footage and Gemini was responding in real time. Pretty much none of that was true. They gave it image plus text prompts, got the response, and then in the video we all saw, they actually went back later and narrated over the whole thing. So none of the talking that you were hearing from the actual narrator was created in real time. That was done in post production to make it look like it was happening in real time. Later on in the video where they drew a picture of a guitar and it was hooked up to an amplifier, and then it says, you've added an amp. Now it's an electric guitar. We can make some seriously loud music now. And then it started playing an audio of an electric guitar, sort of making it seem like Gemini generated that electric guitar sound. But in reality, it just did a search for whatever type of music it was asking to generate. So when they drew this guitar plus amp plus drums plus palm tree, it gave a description and then just did a search for reggae music for a tropical vacation. It didn't actually generate the music. According to this TechCrunch article, Google actually reached out to them and asked them to change the word fake. But this TechCrunch article writer argues that nope, they're not changing it. It was actually faked. And to some degree I kind of agree with them. I made a whole video breaking down this video, and the whole time I was watching it, I honestly thought they were recording a video and Gemini was responding in real time. So personally, even I feel a little duped by it. So I think it's okay to say that a lot of what we saw here is fake. Fake. It's not based in reality. It's not what we will actually likely see from Gemini when we get access to it. All right, enough ranting about that. Let's go ahead and move on. Today, the day I'm recording this, which is December 7, 2023, Elon Musk announced that Grok will be rolling out to X Premium plus subscribers with the latest app release. Now my X feed is full of people saying, hey, I finally got access to Grok. Hey, I finally got access to Grok, but as of this recording on the 7th, I haven't got access to it yet. Okay, so I'm actually splicing this in later. Because I finally got access to Grok, I get to rock out with my Grok out. If you're a Premium plus subscriber, you'll notice that Grok is now available over here on the sidebar. A couple of things I find interesting about it. It's got a fun mode and a regular mode. Fun mode will give you a little bit sillier of answers, where regular mode is more like what you'd expect from a chat bot. It actually does some pretty funny roasts. I had it roast me. I'm not going to read the response, but if you want to pause, you can see what it said about me. It wasn't very nice, but one thing I find interesting is that there is no chat history yet. So any conversations I've had, the only way for me to save them is for me to like copy and paste them somewhere else or to screenshot the conversation. All of the previous conversations I've had sort of disappear as soon as I leave this page and are gone for good. I am going to do a more in depth breakdown of Grok in a future video. So I'm not going to really go off about it too much more in this video, but if you're in the US and you are a premium plus subscriber on X, you should now have access to Grok. This week OpenAI agreed to buy $51 million worth of AI chips from a startup that was backed by CEO Sam Altman. The company is called Rain AI and Sam is invested personally in the company. Now it's interesting because when Sam spoke to the Senate, he said that he had no equity in OpenAI, but he's clearly making money in roundabout ways from OpenAI. By OpenAI purchasing chips from companies that he does have investments in. This Rain AI company is working on a chip. It's called a Neuromorphic Processing Unit or npu, which is designed to replicate features of the human brain. This deal was signed back in 2019 for these $51 million in chips and it looks like they're expected to get into customers hands as early as October of next year. Speaking of Sam Altman, Sam Altman was name CEO of the Year by Time Magazine. Now as far as I can tell, Sam Altman is also the first ever CEO of the Year. I don't believe Time has ever announced CEOs of the year. They always release their Person of the Year who this year went to Taylor Swift. But it's no real surprise that Sam emerged as the CEO of the Year because when all of that chaos went down at OpenAI, which I've made several videos about already, everybody at the company at OpenAI, outside of the handful of board members that tried to boot him, went to bat for him. Basically all saying that they would step away and join Sam Altman at Microsoft if they did not reinstate him and the board resigns. OpenAI staff members have a cult like following of their leader where every time he would post something during that whole ordeal, pretty much everybody in the company would give it a heart emoji. At one point, pretty much every OpenAI employee that's at X tweeted out, OpenAI is nothing without its people, essentially referring to the fact that all of the people were about to bail if Sam Altman didn't come back. Now if that doesn't speak to his leadership skills and show that his team members are willing to follow him into battle, I don't know what is. It was also announced this week that Microsoft's copilot is getting some major upgrades. In fact, all of the cool features that have been rolling out into ChatGPT are going to be rolling into Copilot. So for example, Copilot's getting GPT4 Turbo, the newest model from OpenAI. They're getting the new Dall E3 model. In fact, the new Dall E3 model is already available. You can go use it in the Bing image creator or directly inside of Bing Chat, which I guess is not called Bing Chat anymore, it's just called Copilot. Also, it's getting inline compose with rewrite menu. Basically you can highlight some text in your Edge browser and it will rewrite it and sort of summarize it for you. It's getting multimodal with search grounding. So basically it's getting the GPT4 Vision model so you can upload a combination of image and text to help sort of inform your search and prompting. It's getting code interpreter soon and Bing Search is now getting what they're calling Deep search. And speaking of this deep search, they actually did a separate blog post over on the Microsoft Bing blog all about this deep search. And essentially what this is is Bing will rewrite a better prompt for you when you go to search for something. So the example they give here say I'm researching loyalty programs in different countries and search for how do point systems work in Japan? Well, the search engine may not totally understand all of the context around that question. It could refer to rewards points as intended, but it also could be seeking information on immigration policy or something else. So Deep Search will actually make sure to clarify Are you asking for loyalty cards and rewards, immigration policy for highly skilled professionals, public transportation fairs? And then when you select loyalty cards and rewards, it writes a better search prompt for you. Provide an explanation of how various loyalty card programs work in Japan, including the benefits, requirements and limitations of each. Include examples of popular loyalty cards from different categories such as convenience stores, supermarkets and restaurants, et cetera. It goes on and writes a much more in depth prompt for your search to make sure that Bing has the full understanding of what you're actually looking for when you do this search. This week Apple also seemingly is getting more involved in the world of AI, releasing a framework called mlx. Now this is a framework to build foundation models. So this is really kind of for developers to be able to build AI models using Apple silicon chips. And if you're a developer, it's actually already available on GitHub so you can start using this MLX framework already. I am not a developer, so this is actually a little bit over my head. But with all of the buzz around AI, Apple clearly wants people to start building models using their chips and their computers. This week, Stability AI also released a new model called Stable LM Zephyr 3B, which is a new addition to Stable LM. It's a model that's 60% smaller than the existing 7B model, which means you should be able to run it on less powerful computers. Essentially. Now you can download the model weights already and start using it. However, it is not open source. And I find this interesting because Stability AI has been one of the bigger proponents for open source, especially Ahmad on X. He constantly talks about how open source is better, but it says this model is being released under a non commercial license that permits non commercial use, so therefore not open source. We can see here on this chart, this Stable LM Zephyr is represented by this purple circle. It seems to perform better than everybody in humanities and better than Falcon 40B, which is 40 billion parameters versus this 3 billion parameters and almost everything but extraction. And it underperforms against GPT 3.5 and pretty much everything but humanity. Now the other big news of the week is all of the announcements that came out of Metta. There was quite a few, starting with their new AI alliance, which is an international community of leading technology developers, researchers and adopters collaborating together to advance open, safe, responsible AI. And a whole bunch of companies signed on to this, notably AMD hugging Face, IBM, intel, and the previously mentioned Stability AI. Now, there's a lot of companies that are clearly missing from this list, namely OpenAI, Microsoft, Google, Anthropic Inflection AI. You know, all of the companies that we hear about the most that are developing AI right now. And if this concept sounds very familiar, it's because back in July, OpenAI teamed up with Anthropic, Google and Microsoft to create the Frontier Model Forum, an industry body focused on ensuring safe and responsible development of frontier AI models. Which is kind of interesting because if all these companies are trying to accomplish the same thing, why not just do it all together? I mean, you're already working with your competitors anyway. Google, Microsoft, Anthropic and OpenAI. I mean, it's not like they're opposed to working with competitors to make this, this alliance or Safety Forum. So just do it together. I don't know, maybe that's too much to ask. This week, Meta also introduced Purple Llama for safe and responsible AI development. These are basically benchmarking tools that people that are developing an AI can use to find out how safe and secure the models that they are building are. It says we are sharing what we believe is the first industry wide set of cybersecurity safety evaluations for large language models. These benchmarks are based on industry guidance and standards and are built in collaboration with our security experts. We'll be providing tools like Metrics for Quantifying LLM cybersecurity risk, tools to evaluate the frequency of insecure code suggestions tools to evaluate LLMs to make it harder to generate malicious code or aid in carrying out cyber attacks. And it's designed to reduce the frequency of insecure AI generated code suggestions. But that's just like the safety security news that came out of Meta. They also announced and or released a whole bunch of tools in their suite of products. So in this article on the Facebook blog here to close out the year, we're testing more than 20 new ways generative AI can improve your experience across Facebook, Instagram, messenger and WhatsApp, including the ability to create and riff on images with friends. You can see in this example here, somebody wrote imagine a hippo eating spaghetti and it generated an image of a hippo eating spaghetti. They sent it to their friend. Their friend wrote imagine an Italian bistro. It then changed the image to a hippo and an Italian bistro. And then somebody else prompted imagine it wearing a baseball hat, and then it generated the hippo wearing a baseball hat. So it's a way to generate AI images with friends and kind of go back and forth in messenger, sort of trying to one up each other and how crazy you can make the image. They're also rolling out the ability to search reels directly inside of your meta AI. So somebody gave the prompt show me videos of the best places to visit in Tokyo. Sure. Here are some reels of best places to visit in Tokyo. So if you're going on a vacation, you could do a quick search and find a bunch of reels of people talking about places to go in Tokyo. Up until now, I don't think there's been a real convenient way to search things like reels, so seems pretty helpful. I don't really use Facebook or Meta products very often, but seems cool. It looks like they're also adding a feature to Facebook where it will write your Facebook posts or responses for you using AI. Which in my opinion totally defeats the purpose of social media. But that's just my opinion. If everybody's just going to start using AI to communicate with social media, then what's the dang point? But I digress. Here's another example where it's somebody's birthday and they have AI suggest a way to wish them a Happy birthday. Another feature they're adding is the ability to expand videos. They look good in Landscape or Portrait View, as well as a handful of other features. But in probably the most notable rollout that they had this week, they released a custom video page where you can generate images with their EMU image generator. Before, you had to do it in one of their chats, but now you can enter a prompt right here and generate it directly from the Meta website. Just go to imagine.meta.com and you can do a humanoid golden robot sitting in the grass eating a bowl of Cheerios. And just like other models, it gives us four examples that we can use of our humanoid golden robot sitting in the grass eating a bowl of Cheerios. And now that EMU is in the mix, I feel like I need to make another video breakdown comparing all the AI image models. So I'll probably wait till next year do that, because there'll probably be like, 17 more by the time I do it. And speaking of new AI image models, this week Playground rolled out what they call Playground V2, which is their own custom AI image model. You can see some examples of some of the images it generated here. And they did a benchmark test. And essentially what they did was they gave images generated with SDXL and images generated with Playground V2 to a bunch of people and asked them which one they preferred. And almost 70% of people preferred Playground V2 over SDXL, and a small percentage said it was a tie. If we log into Playground here and we generate an image and use a prompt like a purple cow in a spacesuit standing on the moon, and click Generate, it generates an image like this. And I'm guessing this is using their new Playground V2. Their interface is actually kind of annoying me because it won't let me close this tutorial, and I really wish I could. All right, I finally got through all the steps of the tutorial by generating more purple cows. And you can see if we click on the model up here, we have Stable Diffusion XL and Playground v2. If I click on these other images that I created, I could see over on the right that these were all generated with this new playground model 2. But this one was using a filter that I picked called Bella's Dream. But there's all sorts of cool filters here. Some of them use SDXL, some of them are using the new Playground V2 model. So if you love AI art, there's tons of stuff for you to get in here and play with. It's free to use for up to 500 images per day. And if you want to create more beyond that and get a few extra features, it's 12 bucks a month. Now, let's quickly talk about AI video. I actually got early access to Pika 1.0. I'm not going to dive too deep into it in this video because I talked about it in last week's news video and I'm making a deeper dive video on it actually next week. So you'll learn more about Pika next week. But in my opinion, so far it lives up to the hype. I was able to create videos like this wolf howling at the moon and then out paint it and extend it into a portrait view of a wolf howling at the moon. I was able to do this time lapse of the sky over Joshua Tree here. Somebody asked me on X if I uploaded an actual real image, how does it work? Or does it morph it and make it look not like you at all like Runway Gen 2 does? Well, here's an example of an image I uploaded of my headshot. It just makes it look like I'm animated and talking. It didn't morph my face like crazy. I was curious if I can do things that were copyrighted characters like Pikachu and it let me generate Pikachu and I was curious if it would generate celebrities. So I made a video of Tom Hanks on a basketball court here. And of course I needed to try the inpainting feature and add extra stuff to it. So I took this same video and then gave Tom Cruise aviators sunglasses while on the same basketball court. You can see the video is identical other than the fact that he's now wearing sunglasses. Pretty exciting tech and it should be a fun video. So look out for that one next week. And while we're on the topic of AI video, I recently did this video called An Explosion of AI Is Happening Right Now. And I talked about a whole bunch of really cool AI video tools, including one called Animate Anyone, which essentially makes these viral looking TikTok videos. But you can take an animation of someone dancing and an image of someone and it sort of combines them and makes the person that was in the image dance in that same way while. Well, as it turns out, these were trained on a bunch of TikTokers videos. And I don't believe Alibaba, who created this, got permission and a lot of people are frustrated about that. Not to mention that if you get a little creative with your imagination, you can probably think through what this will actually be used for. More often, most likely putting people in situations that they didn't intend to be put in and never wanted to videos of them doing online. I'll leave it at that. You know what I'm talking about. But this article says these researchers are already using people's work without their consent as a practice baked into training and building the model. The Alibaba paper is the commercialization of the TikTok data set, which was originally created for academic purposes by researchers at the University of Minnesota. It's trained on a model that scraped videos of some of the most famous TikTok creators, including Charlie D', Amelio, Addison Rae, Ashley Nocera, Stina K and dozens of others. Sorry if I mispronounced those names. I actually haven't heard of most of them. Also this week, AMD is releasing new chips to power faster AI training. When it comes to the AI world, we hear about Nvidia all the time and all of these companies using Nvidia chips. But AMD is here too, and they're building better, more powerful, faster chips as well because they want to be part of this AI revolution. Some prominent companies like Microsoft have already signed on to use these AMD chips. Now getting into the world of sort of bizarre News. This week, McDonald's and Google Cloud announced strategic partnership to connect connect latest cloud technology and apply generative AI solutions across its restaurants worldwide. Now, there's not a lot of information about what this means, but McDonald's wants to use a lot more generative AI. Now they've already tested using AI in drive thrus and there's some hilarious viral videos where people are trying to order at a drive through with an AI generated order taker and they're getting frustrated because they can't actually place the order and the AI is just misunderstanding everything they're saying. It's, it's, it's hilarious. So I don't know how well this is going to work out for McDonald's and I don't know if that's exactly what their plans are, but it sounds like McDonald's is trying to incorporate AI more into what they're doing. So we'll see how that plays out. And in the final bit of news that I want to share this week, there is some ongoing research where people are learning more about the sounds that whales make and trying to essentially translate them and understand more about their language. It's actually really, really fascinating because apparently whales use things similar to vowels like a and I, and they use it different tones and they repeat it in different ways. To convey different meanings and emotions, supposedly. And it sounds like it could only be a matter of time before maybe there's some sort of understanding of what whales are trying to communicate. I don't know. But fascinating tech. I love AI for reasons like this because stuff that I never even considered is being done with AI right now. And some of it's pretty dang, mind blowing. That's all I got for you this week. There's a lot more AI news that happened. I couldn't cover every single bit of it, but if you want to stay in the loop, make sure you check out FutureTools IE and check out the news page. I update this every day with pretty much all of the news, so anything I missed will be on this page. I also curate all the coolest AI tools that I come across here on the homepage and I have a free newsletter every week. I'll send you the coolest AI tools and the most important AI news directly to your email inbox. You can find it all over at FutureTools IE and everything there is 100% free. Also, I am getting super, super close to 500,000 subscribers. Again, this is just blowing my mind how quickly this YouTube channel has grown in the last year. And once I hit 500,000 subscribers, I'm gonna do a giveaway and give away five of these Meta Ray Ban smart glasses. They have cameras built in, they have AI built into them. I can ask my glasses questions and it uses LLAMA to answer these questions. They're really cool and I want to give away five of them. So I'm going to pick five random subscribers once I pass 500,000 subscribers on this channel. And I'm going to send you a pair of these glasses as a thank you for being a supporter of this channel. So if you haven't subscribed already, do it now. If you're already subscribed, you're already entered to win. So thank you so much. And again, that's all I got for you today. Really, really appreciate you. I'll see you guys in the next video. Bye bye.