EP 255: AI News That Matters - April 22nd, 2024

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and Everyday Life.

Is Meta's new Llama 3 the best open model that we've seen? Is Microsoft's new Vesa one Too good? And is it scary? And also, are we going to be seeing a new chat GPT5 today? Well, I'm going to answer all of those questions well, as best as I can today and more on Everyday AI. What's going on, y'? All? My name is Jordan Wilson. I'm the host of Everyday AI and this is for you. We are your guide to learning and leveraging artificial intelligence. So we do this every single day, Monday through Friday. Live stream podcast, free daily newsletter. Free daily newsletter to help all of us grow with generative AI, to grow our companies and to grow our careers. So if that's you, thank you for joining us. If you're listening on the podcast, we do this usually every single Monday where we bring you the AI news that matters. Because you could spend hours, literally every single day trying to keep up worrying about what's, what's this new update? How can I use this to grow my company, grow my career? Or you can just tune in on Mondays and we'll give it to you straight. Right? So, extremely excited and a lot to cover today. But as a reminder, if you haven't already, if you're listening on the live stream or the podcast, please go to your everydayai.com and sign up for that free daily newsletter. Because, yes, we're going to be covering, you know, about a handful of some of the top stories in Generative AI. However, every single day, we bring you the most relevant AI news as, as well as a lot more today in the newsletter. So if you haven't already, make sure to check that out on our website as well. More than 250 recaps that you can go back and watch for free. All right, so let's get into it. And hey, if you are joining us live, like, Michael's joining us From Brooklyn on YouTube and Mike4G joining us as well, it's Sunny in Milwaukee for Raul and that's great. So, hey, thank you for everyone joining us live. I'd love to hear what you think of some of these news stories as we go along. But hey, without further ado, let's get started. So probably one of the biggest AI news stories in a long time, actually, Meta has introduced not just the Meta AI chatbot, but also Llama 3, which I think could be a game changer for large language models. So Meta, the parent company of Instagram, Facebook, WhatsApp, and I don't know just about everything else, they just introduced their Llama 3 model, a somewhat open source AI model used to develop their AI chatbot. And they're claiming it to be the best open model of their class, of its class. And potentially it looks like it's already surpassing some closed source models from OpenAI and Google. So we'll take a look at that here in a second. So here's what this means. Meta has unveiled, which I think is probably one of the bigger pieces that people aren't talking about, but Meta has revealed it's unveiled its standalone AI chatbot, Meta AI. So you can go check that out at Meta AI. And they've also integrated Meta AI into all of its major apps such as, you know, Instagram, Facebook and WhatsApp. So you can go and chat with the new Meta AI. So the launch of Meta AI outside of Meta's social media ecosystem positions the company as a strong competitor in the chatbot market, challenging leaders like OpenAI's Chat GPT and Google Gemini. So essentially here's, here's what you need to know for this. So previously with previous versions of Meta's Llama. So with Llama 2, its previous most powerful model, it wasn't available for the general public. I mean it was, right? So you could go in and you could download the model. But that did require a little bit of tech, know how it required a pretty fast computer with a good gpu. So most people, you know, you and me, average people, you couldn't really run it, right. Even on most of, you know, I use a couple computers, I think Meta's Llama could only run on my one fastest computer. Right. So this is different now. So Meta I think is really shifting its strategy here with llama3 and now Meta AI. So yes, anyone can go to Meta AI and try the model out. What I like is, you know, it's, you don't even have to log in. So you can log in if you want to save all of your chats. So this is, you know, now very similar to Chat GPT where you can just go on and start using it for free. So that's the other thing to know right now is Meta's models are free to use. We'll see because there's another one coming which I'm going to talk about here in a second. But right now if you go to Meta AI, you can Even log in with your Facebook account. If you want to save your chats to go, go back and use them later. However, right now you don't even need to, so you can go and give llama3 a try. So let's look at some, some things I think that are important here. So like we talked about, you can go in and for our live stream audience, I kind of have an example here. Like Chat GPT and Gemini, when you log on to Meta AI, they give you kind of some example prompts. And then also a couple other things worth noting is the knowledge cutoff is December 2023, which is pretty nice, right? When we talk about using large language models, one of the biggest things you have to worry about is hallucinations and is it giving you up to date information? Is it just making things up? So the December 2023 cut update is pretty good, right? So that's only about four months ago now and right now it's only chat GPT's newest update that has it to December 2023 as well. So Mata, I think making some, some noise here with, with the open source model and, and we're going to get into this a little bit more on the website because I don't want to go into the details on like, is it truly open source? You know, yes and no. So we'll get into that in the newsletter. But here's what else this means is you can download the entire model right now and you can essentially fork it, right? So what that means is you can kind of create versions, you know, or variations of this model. You can connect your company's data to it as well. You know, you'll have to have a little bit of, of tech know how, you know, to take advantage of, of that and you know, using RAG to bring your own company's data in. But you know, already meta is Meta's. Llama 3 is benchmarking. I won't say off the charts because it's still on the charts, but for a, a pretty much open model, this is pretty like, I'd say it's unexpected, right? Like you, you normally don't expect open source models to compete with closed proprietary models from, you know, OpenAI from, from Google, from Claude, but let's actually look at that. So for our, for our podcast audience, I'm sharing a chart now, but essentially right now we have the two different variations of llama 3. So you have their 8 billion parameter model and their 70 billion parameter model. So let's just call these small and medium. All right, there's a large. I'll tell you about that here in a second. All right, so right now, the 8 billion parameter version of llama 3. So kind of the comparison that Meta shares in its benchmarking, this is its own internal benchmarking. Right, but we'll share a little bit more on that here in a minute. So right now it is hitting well above kind of Google's Gemma 7 billion parameter model and also Mistral's 7 billion parameter model. So, you know, when we talk about MMLU, which is one of the most popular, you know, benchmarks that we talk about a lot here on the show, which is the multitask language understanding. Sorry, the massive multitask language understanding, it's essentially like, hey, is this model as smart or smarter than a human? Right. More or less. So right now the meta 8 billion parameter model is, you know, out punching its weight class by far for those smaller models. And then we look at the kind of quote unquote medium model. So that is the 70 billion parameter. And what Meta is comparing it to here is Gemini Pros 1.5. So that's from Google. So that's kind of its middle, middle model as well as anthropics middle model. So anthropics Claude Sonnet. So as well, that has 70 billion parameter model there, just barely getting above Gemini Pro in the mmlu. And then, you know, a couple points ahead of Claude's Sonnet. So what, what most people were talking about when this came out is, oh, well, what about Chat GPT, you know, GPT4? What about, you know, Claude's most powerful model, Opus? What about Gemini's Gemini Ultra? Right, so that's the thing. These are technically Meta's medium models. Okay. So they do have another model as part of this llama 3 release, which is a 400 billion parameter. Right? So we have the 8 billion, 70 billion and 400 billion parameter models. So without getting too technical, that's just the amount of data that these models carry with them. And obviously when you think about downloading and using these models locally, 70 billion is already a pretty big model to try to run locally, right? I mean, the smallest model is pretty big model to run locally. So this should be, we should be hearing more about this 400 billion parameter model because, and I did share about this, you know, about an hour after it came out, you know, if you follow along on LinkedIn, I shared a pretty long, long breakdown actually of this new unreleased model for meta, what it could mean and with the potential benchmarks, because Mark Zuckerberg Meta's CEO did already share some kind of in progress benchmarks for the 400 billion parameter model and he was saying that it was already hitting near GPT4 and some of these other models. So pretty interesting news here from Meta and yes. So Jackie, yes, you can go download it. So yeah, we'll have those links in the newsletter. So, you know, for our live stream audience, if you are new to Llama, yeah, you can download it. But also unlike previous versions of Llama, you can just go on to Meta AI right now and use the, I believe that's the 70 billion parameter flavor that they have available on Meta AI. So yeah, you can download it and you can play around with it and you know, there's obviously different apps also that you can download for your desktop that makes it much easier to run these, these models locally. And we'll be sharing about those in the newsletter as well. All right, let's keep this thing going. Actually, no, I, I do want to talk about this as well. So if a lot of people don't see this, but Meta did say that primarily that Llama 3 was trained for English. So if you go on to the chat arena leaderboard, which I'm probably going to have a dedicated episode on that once, you know. So essentially it takes these ELO scores, which is kind of an old scoring. I, I wouldn't say old, but it's a scoring system kind of borrowed from chess in other competitive games. But you know, essentially an ELO score is when you kind of pit two models blindly side by side and you know, there's been tens of Thousands or actually now 500, 000 votes on, you know, which model is stronger. I do this all the time. Anyone can go on. We'll share the link in today's newsletter and people essentially, you know, it's, it's like the blind Pepsi taste test. I don't know if anyone remembers that, you know, marketing from like 20 years ago. You know, you get outputs, you can put a prompt in. It'll give you two outputs from two different models. They don't tell you which model, you vote which one's better. But already Llama, if you just look at English, right? So if you don't look across all languages, LLAMA is already beating out the most powerful available model from Google Gemini, which is their Google Gemini Pro. And also anthropics Claude 3 opus. So already llama is hitting heavy. All right, our next piece of AI news, this one also kind of scary, but Microsoft has unveiled its VESA1 image to talking video model. I don't know if that's what we're calling it, image to video, but it is for essentially real life human avatars. All right, so they did this via research paper. So Microsoft's Vasa1AI research paper has introduced a cutting edge framework for creating hyper realistic talking faces by converting a single portrait photo and an audio file into a live animated talking head video. So VESA1 showcases impressive lip sync, realistic facial features and head movement, setting a new standard in AI driven animation technology. So unlike other existing tools, VESA1 can work with photos, photos facing various directions and incorporating factors like eye gaze direction, head distance and emotion for enhanced control and realism. So potential applications of VESA1 include advanced lip syncing for games, creating virtual avatars for social media, and enhancing AI based movie making for more realistic content. And I'm curious for our livestream audience or you know, hey, you can always check your show notes if you're listening on the podcast, hit me up. Have you all seen this? It's actually wild. I'm going to show it here in a second, so if you haven't seen it, don't worry. But despite this being just a research preview, BASA1 has demonstrated exceptional performance like it's really scary good included perfect lip syncing to songs and handling different image styles like the Mona Lisa, generating a 512 square pixel image at 45 frames per second in just about two minutes using an Nvidia RTX 4090 GPU. So this part's important. Right now there are no immediate plans for public release, thankfully. Right. Like I don't think you want to see something like this, something this powerful released right now, especially right before the election here in the us. But the technology's potential impact on industries like gaming, social media and entertainment is significant. And this is hitting at a future where AI driven animation could become more accessible and widespread. All right, so let's go ahead for our live stream audience, let's go ahead and take a look here. So I'm going to go ahead, let's just do this. I'm trying not to block. We'll just do the middle one here. So essentially this is. Let me recap and tell you what this is in a nutshell. So based off one image, just a single flat image, this new VESA1 technology is able to create these videos. So it does need one base image as well as a voice and then from there it can create this. So I'm going to go ahead and play a couple examples here for our live Stream audience. Again, this is based off one still photo and a training voice. So you do need a voice to start with. I. I don't know if the VESA one has kind of default voices that they use, but I'm pretty sure it's. A person will upload their own voice. So let's go ahead, take a watch and a listen. We'll just do two examples. We'll just do about half of them because they're each about a minute.

Have you ever had, maybe you're in that place right now where you want to turn your life around and you know, somewhere deep in your soul there could be some decisions that you have to make. Like, you know, like, it's like things were. Something was decided for you and instead of trying to make something that is done work, it's like the invitation is to make the decision, commit to that.

Y', all, have you seen this? This is wild. This is wild. Like, okay, if you're listening on the podcast, obviously you can tell the quality of the audio. Right. And even I wouldn't say stutter, right? Like sometimes I'm stuttering around and mumbling around, but it has this inflection in the voice that you don't get right now from any, you know, text to speech, AI, text to speech generators, any obviously video generators. The quality is mind boggling. Yes, this is wild. Yeah. And, and I do know, like, you know, Matthew here shared like, you know, an example of, of hey, Gen. So, yeah, there's a lot of programs right now that do this, but I will say right now, at least compared to the VESA, this VESA1 model, which is not publicly available, this, this model is so far ahead of. Of others, I would say pretty. Yeah, scary real. Yeah. TC here says that's terrifying. Matthew says, but my video editor friends are scared. Y. Scary good. Let's take a look at one more, one more example here before we go on with, with more AI news. Here we go.

The first thing we need to look at is the letter H. So the H sound at the beginning. It depends what country you're from. But many native languages have a problem with putting too much tightness in the throat and it can become more of a H sound. So it's very important not to over exaggerate this sound. It's a very soft, very relaxed sound. It's in English, so just softly release the sound. Help. Okay, so that's the first thing.

All right.

It's.

It's almost like. So it's. It is. Yeah, it's. It's so good. It's scary almost. Right? So in, in that example, and again, we will have the links that you can actually go and download these, these, these videos if you want. Right. So I do like that Microsoft is making this available as well as they have just a lot more information. Right. So kind of scrolling through here, this paper talks a little bit about how this new technology works, how it takes a single image, an audio clip, and then essentially control signals that you can put in there as well in terms of if or how you might want the avatar. It's so weird calling an avatar because it looks way, way human, but you can kind of have certain level of controls. All right, let's keep this thing going. All right, our next piece of AI news for today. Boston Dynamics has unveiled a fully electric Atlas robot for real world applications. All right, so Boston Dynamics has introduced a fully electric version of its humanoid robots, Atlas designed for real world and commercial applications. So Boston Dynamics, a leading robotics company, aims to showcase the enhanced capabilities of the Electric Atlas in a lab setting, factories, and everyday life. So previously, Boston Dynamics, I'm sure most of you have seen and heard they had an Atlas robot. Before that they kind of retired, making way for its new, much more humanoid but electric iteration of Atlas. So the transition from hydraulic to electric power was made possible through a partnership with Hyundai, focusing on advanced automotive manufacturing capabilities. The development of the new Atlas robot showcases the continuous evolution of AI technology in the field of robotics, paving the way for enhanced capabilities and applications in various industries. The Electric Atlas boasts increased strength and a wider range of motion compared to its hydraulic predecessor, enabling it to handle diverse manipulation tasks in various environments. And despite the technological advancements, the release of the Electric Atlas has sparked skepticism on social media. Obviously, people drawing parallels to science fiction scenarios like the Terminator franchise. So let's go ahead, we're gonna do this one as well. Give everyone a quick little 20, 30 second video here that Boston Dynamics shared on their, on their Twitter. So let's go ahead and give that a play. So I don't think we really need the audio here. There's not really audio. So essentially what we have here is, yeah, a robot, humanoid robot doing a backbend in a very scary fashion and then walking, turning its whole body around, displacing its hips. Right. It looks like it can walk in all direction. There's no front or back, presumably. I mean, there's a lot of AI in this latest iteration of Boston Dynamics. So we'll be sharing about that, obviously, more in today's newsletter. All Right. There's more. Yes, Big Apple news. Hey, any, any other day, any other day this could have been our lead story, if I'm being honest. Hey, thanks for what Matthew just said here from the live stream said Jordan is helping us manage information overload effectively. Well done and fun to follow along. All right, well, hey, thanks. Thanks Matthew. Make sure to check out the the daily newsletter as as well today. That'll make it even easier. But let's talk about this new report. So Apple is going to be int on device generative AI features for iPhones. So Apple is developing its own large language model to power on device generative AI features for upcoming iPhone series, according to a recent report from Bloomberg. So this new AI model is expected to run entirely on the device. So we're talking edge AI on device AI, not cloud. So enabling offline functionality without Internet connectivity. So Apple plans to leverage the neural processing unit or NPU of the Apple silicon chip to offer new AI capabilities. Apple. So again, this is where it's like, wait, which direction is Apple going? Right? Because we talked about here on the show like eight months ago a report that said Apple was spending millions with an ass millions of dollars a day on its own generative AI technology, presumably for this exact reason. And then a couple of weeks ago we got a report that Apple was actually no, they're partnering. You know, they're going to be partnering with maybe Google and you know, either using Google's Gemini Nano, you know, model or maybe one of its Gemma models. So there has been obviously Apple reportedly still exploring partnerships with Google, with Microsoft back open AI and with China's Baidu for AI tools integration into its next generation operating system. But this Bloomberg report is saying no, they're actually going in a different way. So you know, they're not saying if this is going to be for the next iteration. So we could see as an example, the next iteration of the iPhone may be partnering with a Google or an OpenAI or China's Beidou for the model and then maybe the next iPhone or the next device after that might have Apple's own internal model. So we'll have to see. But regardless, Apple's approach to marketing its AI features will focus on how they can assist users and managing daily routines rather than emphasizing speed and power. So we should see a lot more news or kind of which direction Apple is officially going on this during the worldwide Developer conference. So that's WWDC this June, starting June 10, and where Apple is expected to unveil not just the next generation of Its operating systems presumably powered by AI, but also now hardware. Right on Device AI. Pretty big news. But Apple is also planning reportedly to integrate these generative AI capabilities into series, into Siri messages, Apple music pages, keynotes essentially across all of its ecosystem there. But Apple's reported aim is to enhance the user experience with features like text summarization suggestions and more thorough AI integration. All right, let's, I mean, what, what do you all think though? I was taken aback by the report about a month ago that said that Apple was looking into exploring third party partnerships. I don't know, I've been extremely underwhelmed by Google's model and that's what the last report said, is that Apple was going to be kind of, you know, partnering or working with Google's model. And I was like, I don't think that's part of a good, good call. So we'll see if, if that's what's happened or if, you know, who knows, maybe Apple is just going to have a, a lighter version of generative AI features, you know, announced in June and maybe there will be a more in depth one, you know, the year after or you know, at a later announcement. All right, so our next piece of AI news. Oh yeah, here we go. We're talking Internet conspiracy theories. Yeah, I don't know if you guys saw this, but when I was sharing, sharing my, my screen here on the Boston Dynamic story, you'll see on the right hand, so this is on Twitter or X or whatever people call it now, but you'll see it says, oh, trending, what's happening GPT5. And you might have seen this and you're like, wait, why is that trending? Well, let's, let's talk about it. So there's been a lot of speculation over the last week because Open AI posted a few photo of a throne with the number 22 last weekend. I'll get to what that means. But essentially all these, you know, geeky people like me, but Internet conspiracy theorists are convinced that that means chat GPT's GPT5 will be released today. April 22nd. Right. Throne with a 22, 22nd. Why, why, why today? Well, today is Apple or sorry OpenAI CEO Sam Altman's birthday, April 22. So that's what everyone's been, you know, saying is like, oh, it's coming today. So yeah, if you're on social media, seeing this on, on Twitter or LinkedIn or anywhere else, I'd say this, don't believe it. It's, it's not happening because guess what else you notice there, there's a basketball. All right, so also if, if you look at the, the caption or the prompt that was used to generate this, it's a queen's throne. So what does that mean? A basketball plus a queen's throne plus the number 22. Does that ring a bell for anyone? Maybe Caitlin Clark. Right. Because also along with this timing and this was obvious, this was posted to open AI's Instagram story last week. I, I don't, I don't follow you know, anyone on Instagram. I just saw all the, you know, rumors swirling around, but it just so happened to be the exact same timing of Iowa had kind of a celebration for Caitlin Clark and the Iowa women's basketball team as well as she just went number one with the over overall number one pick in the WNBA draft. So most people I think are looking at this as kind of like a, you know, an ode to her and her success. But yeah, so be watch out everyone. Everyone out there today is saying, oh, the new version of GPT5 is coming out. You know, look at what Llama just did. And you know, a couple weeks ago we got updates from Google's Gemini and you know, Claude 3 now has been out for about two months. So people are like, oh, today's the day. I'll tell you this. No, absolutely not. I don't know, I could be wrong, but I give it about a 0.0% likelihood that we get a new GPT5 today. Sam Altman's talked about on, you know, some recent interviews that there's going to be multiple kind of refreshes or updates to their platform. He has signaled that yes, there will be a new model this calendar year. I don't think we're getting it today. I don't even know if we're going to be getting any updates from OpenAI today. I, I wouldn't think so. But presumably we will first be seeing updates probably with OpenAI's whisper voice translation, maybe with their dolly three. Maybe we'll get a dummy down version of sora available. So OpenAI's, you know, AI video project. But yeah, I don't see anything happening there with today in GPT5. All right, in our last story, so TED Talks went all in on a SORA looking future. So TED Talks just posted this over the weekend. A little AI video made with Sora. I'm going to show you guys just, just, just the front half of this here because it's, it's like a three minute video very Very impressive video. But let's just take a look at this one. So give me a second here. Live stream audience, you can take a look. I will put the audio on for this one. We're just going to let it play for 20 or 30 seconds. So you can see another great example. So this is TED Talks and we'll link this tweet in our newsletter. So it says, what will TED look like in 40 years? For TED 2024, we worked with a couple artists, it looks like Paul Trillo and Open AI to create this exclusive video using Sora, their unreleased text to video model. All right, so let's just take a look at what this looks like. All right. And then hey, we'll, we'll end there on this strange, you know, looks like some lab grown meat or something that we have, that we have going on there. So yeah, it didn't get the best reception. Right. So I, I, I, I think the, the video was actually surprisingly well done. Right. And if you're listening on the podcast, you can always go back and you know, watch, you know, today's episode, the, the video of it. So you can see or check it out in the newsletter. But essentially it's flying through, you know, a set of corn fields and then through these, you know, futuristic buildings. Presumably, you know, people are giving TED talks in these various settings, but it's just kind of a continuous one shot kind of flying through, you know, a series of inside and outside but very futuristic looking venues. But didn't get the best reaction, didn't get the best reaction here on Twitter when someone said this is what TED would look like in 40 years. A random generated video that doesn't make much sense. Plus meat slabs. So yeah, I, hey, I, I thought it looked pretty, pretty cool, but didn't get the best response there in the Twitterverse. So what, what do you guys think? So much happening today in the world of AI news. So hey, we'll just give you the quick 30 second recap here. So Meta has introduced Llama 3, its new pretty much open source model that is benchmarking off the charts as well as its Meta AI kind of chatbot and rolling it out across its different platforms. Microsoft unveiled its scary good Vesa one image to talking video model. We showed you some examples of that. It is not publicly available luckily because it's too good and it would not be a good time right before the election here in the US So you can go read the paper and download some examples that we'll have in our loot in our newsletter Boston Dynamics unveiled a fully electric Atlas robot powered by a lot of AI after retiring its hydraulic version of Atlas, which has been, you know, been out in the wild there for a couple of years. Apple Report, according to a Bloomberg report, is introducing on device generative AI features for its upcoming iPhones, which kind of goes against some previous reports that they would be, you know, partnering with maybe a Google or a chat GPT instead. Rumors are swirling that chat GPT is going to release or OpenAI is going to release GPT5 today after a not very cryptic, but you could say it's cryptic, you know, story on social media. I'd say no. I'd say that's just more Caitlin Clark. And last but not least, TED Talks kind of getting roasted a little bit for its take at a look of the future with Sora. So that is it. A lot more in our newsletter. So make sure you go to your everyday AI.com and please join me tomorrow. We're going to come super hot for Hot Take Tuesday, so join me tomorrow and every day for more Everyday AI. Thanks y'.

All.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.