EP 230: AI News That Matters - March 18, 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 Google DeepMind building agents. And Apple finally has a large language model, kind of. And there's a new humanoid robot that uses ChatGPT and interacts in real time flawlessly. There's so much going on in the world of AI news and you can spend hours every single day trying to keep up, or you can just tune in with us. We do this almost every Monday. So welcome to Everyday AI. My name is Jordan Wilson and I am the host. And if you're new here, well, Everyday AI, it's for you. It's to help everyday people learn what's going on in the world of generative AI and how we can all leverage that to grow our companies and to grow our career. So if you're new here, joining us on the podcast, thank you. We normally do this every Monday, as long as, you know, we're not on the road somewhere else. Speaking of being on the road, we are on the road, but more on that here in a bit. So let's get going. And just as a reminder, if you haven't already, go to your everydayai.com. you know, if you can't catch us every day, don't worry. We always recap each and every show on our newsletter as well as we now have more than 220 back episodes on our website. So it is a free generative AI university. Go check that out at your everydayai.com. all right, but let's get into the AI news that matters. And if you're joining us live, thank you. Like Peter Scuda, thanks for joining us. If you have any questions, get them in there. It should be a fun show. There's a lot going on, so let's just dive straight into the news that matters for the week of March 18th. All right, so first and foremost, Google DeepMind has introduced SEMA, an AI agent that plays video games. So Google DeepMind has announced this new AI model called SEMA, and it can play video games based on natural language instructions. So this model has been trained with the help of eight video game studios and can perform over 600 basic skills, making it more efficient than models trained on one specific game. So SEMA is a generalist AI agent that can play video games based on natural language commands. So the ultimate goal from DeepMind here is for the AI to be able to play games by itself. So you might be thinking, all right, why does this matter. Well, this is kind of, I think, the first official foray from Google into what could be considered agents. Right. So this new initiative from DeepMind is widely being reported as Google's first semi official foray into agents or smart AI that can make human, like, decisions in real time without any human input. Right. So that is kind of the big trend of 2024. So so far is autonomous agents. And, you know, this AI being being able to work without much human input really at all. So this thing from Google, DeepMind, a lot of people are just looking at this as, okay, Google's, you know, creating an AI that can play video games just to showcase its ability. But it's really much more than that. This is kind of a fun and engaging testing ground for what we all expect to be the next phase of generative AI, which is agents. All of the big companies are working on this. And when you combine it with humanoids. Yeah, it's going to get wild. Some more on that here in a second. All right, our next piece of AI news. And if you caught this a couple days ago, you might have been scratching your head kind of like I was. All right, so OpenAI's CTO gave some mixed answers on training data for its AI image model, Sora. All right, so if you don't know AI, OpenAI's new text to video model, Sora, has sparked controversy for the sources used for its training data. And CTO Mira Marathai avoided questions, questions about the subject in a recent interview. All right, so if you don't know Sora, don't worry, I'll give you the super high level overview. So they just started, OpenAI just started releasing kind of new outputs just about a month or two ago. So right now it is not open to the public. But OpenAI researchers, as well as a select few kind of designers and motion graphic artists have use or have the ability to use SORA right now. So essentially you can, with text prompts, create some very impressive AI videos. And kind of the big thing that has separated SORA from other, you know, text to video kind of offerings so far is the quality. Right? So we've seen from Runway, we've seen from Pika Labs and others the ability to create some still, you know, very usable, very inspiring, I might even say text to video with their different generative AI models. However, Sora, it blows it out of the water. So we've, we've, you know, shown comparisons before. That's not what this is about. But, you know, in general, a lot of people Were wondering because of the high quality of the output when OpenAI just released this is, you know, how did they train this? And in this interview, it's, it's a little interesting. So we're going to play a clip here. But it was in a recent interview with the Wall Street Journal. So OpenAI CTO Mira Mike Muratai avoided questions about the sources used for sora's training data, claiming that it was just publicly available and licensed. All right, so I'm gonna go ahead and play a clip here and let you all hear kind of in real time. So this is reported from the Wall Street Journal asking OpenAI CTO Mira Muratai about, about Sora's training data. All right, let's, let's go ahead and take a listen. What data was used to train Sora? We used publicly available data and license data. So videos on YouTube. I'm actually not sure about that. Videos from Facebook, Instagram, you know, if they were publicly available available yet publicly available to use. There might be the data, but I'm not sure. All right, so we'll, we'll go ahead and link this out from the Wall Street Journal reporter in today's newsletter. But yeah, a lot of people were kind of taken aback by, by her response, as was I. You know, it almost seemed like she wasn't fully prepared for that question. Probably knowing that that would be one of the main things that someone would be asking you about when sitting down, just because, you know, all of the conversations around OpenAI's Sora has been how are they so far ahead of everyone else in terms of quality? So the, the training data has been a topic of debate before this interview with the Wall Street Journal. So, so definitely, it's definitely worth talking about. And yeah, if you're, if you're listening on the podcast. So Juan here joining live. So thanks for joining. Juan said the facial reaction says it all. So yes, when asked about the training, the training data, Mira had a very uncomfortable look on her face. And yeah, it's already been making the, the meme rounds on the Internet. But yeah, this to me is a pretty big miss from OpenAI. I think if you're gonna go out and, and have a sit down interview about a model that has really taken, you know, the creative and the advertising, the marketing, the AI worlds by storm. You have to be prepared to answer some basic questions about training data. So yeah, the facial reaction, it kind of like what, what Mike Forgy here is saying live as well, that mouth drop is a guilty signal. Yeah, it was A little, it was a little cringy if, if you were watching on the live stream. But yeah, don't worry if you didn't see that yet. We're going to be linking that out in the newsletter. All right, let's keep going with more AI news that matters. And a big one. A big one. I actually intentionally plopped this in the middle of the show just in case a couple of y' all joined in late. All right, so Apple has released a paper previewing its large language model, mm1. Yes, that's right. So it's an interesting approach here on Apple's release of this, but let's just go ahead and go over the details first. So Apple researchers have released a paper previewing their new large language model, or actually it's their new MLLM or multimodal large language model. So, yeah, we're going to be hearing and shifting that conversation probably over next year from LLMs to mllm. So the difference being, you know, with multimodal inputs. So Apple's research team has developed a new highly capable multimodel large language model called mm1. All right, so interesting naming so far. So we're not sure if this is going to be the name when Apple, when and if Apple finally releases this, however, a little confusing, you know, in, in, in my opinion, just because Apple's, uh, new chips, you know, that they debuted about, you know, two or three years ago, were called the M1. So there was a lot of talk around M1 chips. So now you have the M1 model. So again, we're not sure if that's what it will ultimately be called, but is what is being referred to now. So right now this model comes in three size three sizes and outperforms most competitors on tasks such as image recognition and natural language reasoning. But it still lags behind Google's Gemini and OpenAI's GPT4. So right now, MM1 is a multimodal large language model developed by Apple, like I said, in three different sizes. So they have a 3 billion parameter model, so much smaller, 7 billion parameter and 30 billion parameter. So even the, the larger one, at least right now at 30 billion parameters, is still a fraction of the size of, you know, Google Gemini Ultra 1.5, as well as GPT4 Turbo, which is reportedly 1.8 trillion parameters. All right, so again, right now this is as far as I know, at least as of, you know, a couple hours ago when I checked last. You know, you can't go out and use this model, at least right now it is not publicly available. So all we have right now, which is important to talk about and which I think it hasn't been grabbing so many headlines just yet, is because this is just a research paper kind of showing, showing some different results, inputs that they generated and outputs that they were able to get from those inputs and kind of looking at the multimodal aspect of it. And I think one of the things that Apple is seemingly stressing about its models kind of focus in the paper at least, and as you'll see on screen here, is the ability to better kind of both work seamlessly within text and images and also images within text, which I think, you know, if you want to have a highly capable multimodal large language model, it has to be able to both read and understand text in photos as well. So it seemed like that was one kind of key differentiator that Apple was really pushing with its new M1. But yeah, I'd love to hear, yeah, a big agree with Carolyn here saying it's a big reveal. So I did kind of mention that because it was interesting, right, because we've been hearing now for months. Right. And I've said it all along, Apple is never first to the party, right? So, you know, we didn't expect, you know, Apple to release a large language model, you know, months after ChatGPT or anything like that. Apple has been historically known to not be the first person at the party, but to be the coolest kid and to usually have the most polished interface, the best user experience, et cetera, etc. So there have been many reports that Apple has been spending millions, yes, millions with an S. Millions of dollars a day on development of their generative AI, on development of their large language model. So again, we're not sure if this is all they have, if this is just an early iteration and what may eventually make its way to our devices. Right. That's what is most important. And I think, you know, the obvious thing on why this matters for all of you out there listening is because of Siri and because of the future of Edge of edge computing, Edge AI or on device AI. Right. And I think that's, you know, even by looking at the parameters of the model, you have to think that that's where this is heading. Right? So the reason why, you know, something like a GPT4 Turbo is so incredibly powerful because it is a large, it is a huge language model with, you know, reportedly 1.8 trillion parameters. So when you look at these kind of three reported sizes or these three kind of published sizes that were in the paper of 3 billion, 7 billion and 30 billion parameters, presumably these are models that could fit on devices. So similar to how Google's Gemini Ultra now lives locally on their Samsung S24 phone. And that really changes what you can do with a large language model. It changes the capabilities of generative AI by to be able to run something locally. Right. Like as an example, I was, you know, over the last two weeks I've been on probably 25 hours worth of, worth of flights or maybe 20 hours worth of flights. I can't do the math right now. But you know, I couldn't use a large language model. I probably would have liked to, but you know, you have to have a very fast, you know, Internet connection which as an example, you know, airplanes don't. But you know, I think when you talk about Apple, where this really is important is the next iPhone. Right? So we've been talking about the wwdc, the worldwide Developer Conference from Apple in June. Presumably they will be announcing what they're going to be doing with this, whether it is this MM1 large language model, whether they're going to be releasing a couple. But presumably what's going to happen is this. You're going to have a Siri that is actually smart. Right? So you know, I, I know I'm hard sometimes on, you know, smart assistants like Siri and Alexa, but if you use large language models a lot, like I know a lot of our audience does, and then when you use something like Siri or Alexa, there's a lot to be desired. However, I do think even just by looking at the parameters and the sizes of these three flavors of MM1 from Apple, you have to think that this is coming to Edge, this is coming to on device AI and to be able to run a model locally as an example on your iPhone, on your smartphone is, is, is really a game changer in terms of product and also how we all interact with generative AI. And that's something that I talk about a lot here on the show because I don't think people fully realize or understand even the importance of something like prompting. Right. So what happens now when we have these, you know, maybe mm1 coming to our phone? You have to be able to learn how a model works and how to be able to work with it to get the most out of it. So this is really going to be a common theme that we're going to continue to see. But we did expect, you know, kind of some more announcements from Apple, but it was an interesting announcement that they did this via a research paper release. It's not the normal Apple playbook, right? So normally you get complete silence and you get a big reveal at a conference or maybe you get some leaks ahead of time, something like that. You know, a spicy promo video that makes people ooh and ah. So this is an interesting approach here from Apple to go the scientific paper route. Personally I like it. Right. We even saw with, with, with Google with their original Gemini I, I think they, they really botched its release. You know a lot of people, myself included really criticized Google heavily for a heavily, you know, heavily edited and marketing and not to tell you the truth, not super truthful representation of their Gemini model. You know, in their original marketing video that came out I believe like December 5th it was kind of showing all these, you know, capabilities or reported capabilities for Google Gemini that it could not actually do. So I don't hate the approach here from Apple but an interesting approach nonetheless to go the scientific paper route. All right, and hey, let me know from the live stream audience, what do you guys think? What do you guys think of Apple's approach here? Do you like it? Is it confusing? Have you read the paper again? We'll be linking to the paper. If you can't find it, don't worry, we already dug it, dug it up off the Internet. So we'll be linking to that in our newsletter today as well. So make sure you sign up@youreverydayai.com. all right, our next piece of news and this one, as crazy as it sounds, this one might be the biggest piece of news that we're talking about today. So figure has demoed its ChatGPT powered humanoid bot feature, figure01. So we've talked about figure01 here on the everyday AI show a couple of times but up until, you know, up until a couple days ago when this demo was first released, all we really had or all we really saw were were specs. You know, all we, we saw and heard were promise. So anyways let's, let's go into a little bit on, on figure 01 and then we'll, we're actually going to play a little demo as well so you guys can, can hear and see and watch this as well. But figure 01 is equipped with an advanced AI and visual recognition capabilities and it showcases its ability to perform tasks and engage in conversations with a human. That's the key differentiator and potentially revolutionizing many industries. All right, we have to pay attention to this. It is going to I think be a, I hate using the word game changer So I will just say this. I think it is going to greatly impact the future of how we all worked. So the figure 01 robot is equipped with what figure is calling a new visual language model. So VLM, we have so many acronyms today, right? We have LLM, MLLM, and now we have Visual language model. So the visual, the visual language model allows Figure 01 to smoothly interact with humans and perform tasks simultaneously. So its developers over there at Figure have included those that have worked at companies like OpenAI, Boston Dynamics and Tesla. And they have very ambitious goals for their humanoid robots to control and yeah, billions with a B. Billions of robots. Yeah. Crazy, right? But this is also why they have some of the largest investors in the world, including I believe we talked about this on the show two weeks ago. But Microsoft, Open AI are both huge investors. Nvidia Salesforce, I mean it's a who's who of who is investing in Figure and their new figure 01. So let's just go ahead and let's just show a quick little demo here. And if you're listening all the podcasts, I'm going to play about a minute of this. Like I said, we're going to go ahead and link this full link this full video in today's newsletter. But hopefully even if you're listening on the podcast, you can kind of understand what we're talking about just from this short little clip. Let's go ahead and take a watch and take a quick listen. I see a red apple on a plate in the center of the table, a drying rack with cups and a plate, and you standing nearby with your hand on the table. Great. Can I have something to eat? Sure thing. Great. Can you explain why you did what you just did while you pick up this trash? Are you still running in circles trying to figure out how to actually grow your business with AI? Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Genai. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to to use gen AI. So whether you're looking for chat, GPT training for thousands, or just need help building your front end AI strategy, you can partner with us too. Just like some of the biggest companies in the world. Do go to your everyday AI.com partner to get in contact with our team, or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on gen. So I gave you the apple because it's the only edible item I could provide you with from the table. Great. So based on the scene right now, where do you think the dishes in front of you go next? The dishes on the table, like that plate and cup are likely to go into the drying rack next, y'. All. Is anyone else. Was anyone else just extremely impressed with that demo or is it just me? So let's, let's talk about if. If you weren't joining us live. If you're listening on the podcast, you can probably assume what was happening there. But again, this is. Is said to be real time, which is always important because, you know, again, like we saw with the, the Google Gemini and with just all new generative AI systems in general, you always have to know exactly what is going on behind the scenes. So presumably this is in real time. Yeah. And like Juan is saying here and, and everyone. Yeah, everyone's saying, wow. Monica's saying, unreal cecilia with the PG13 language edited though. So I love it. Yeah, it's sign. Yeah. Tara says, sign me up. Love his voice. So, I mean, what is literally happening here is presumably this is in real time and the figure 01 is powered by ChatGPT. So aside from. Yes, this seems like a very realistic conversation. But what's wild to me is, you know, when we talk about AI in the capabilities of AI, I think this is something that we overlook is having someone that can both help with everyday tasks. So on that side, hey, who wouldn't love, you know, a humanoid robot? I think all it takes right now is like a quarter million, quarter million dollars and probably a minimum order. But who wouldn't love, you know, a humanoid robot to be able to go around and put your dishes away, right? Like, that would be fantastic. Or to be able to fold laundry. So speaking of that, I do have to now draw a comparison because with figure 01, at least with the demo so far, it can operate independently and is not necessarily pre programmed. Speaking of folding laundry, it was Tesla's Optimus bots that went viral a couple months ago for simply folding a T shirt, but that was something that it was programmed to do. So I, I can't, I can't state loudly enough how impressive this demo is with Figure 01, powered by Chat GPT. Here's why we've been talking about, you know, robotics on this show for a while. We actually have a great robotics show for you later this week, which I'll get to here in a second. But I think that for the most part, you know, kind of these, these humanoids or, or robots or whatever we're supposed to be calling them, they've been very limited in scope, so they've maybe been trained to perform a series of tasks, but for the most part, that is not a relationship per se, or that is not something that is really applicable across different fields, whether it's in home work, automation, manufacturing, etc. Because guess what happens in real life? Things don't go according to plan. So as an example with Tesla's Optimus bottom, if it was folding laundry and, you know, there was a huge gust of wind, could it continue to fold the laundry? I don't know. Presumably it could, Right? But that's why I am super impressed so far with figure 01 and what it's been able to do. Because it is, as you just saw there in the, in the demo, or as you just listen to it, is having a real time conversation with a human. Again, presumably it's real time, a real time human conversation and an interaction. And not just that, but being able to explain its rationale. The thing I love is in that demo, when the gentleman asked figure 01, kind of why it did what it did, he asked it to do it while putting trash away. Guess what? I'd hate to admit this. I wouldn't be able to do that. I would make a mistake. I am terrible at multitasking, right? Like, if my wife is asking me to, you know, a question while I'm putting dishes away, I'm definitely either going to put the dish away in the wrong spot or I'm not going to be able to fully process her question. So that is why I think it is extremely impressive because in this demo, figure 01 was able to not only complete a task, but also to hold a conversation about something that was unrelated to the task. So when we talk about the future of generative AI, when we talk about even, you know, the real, like the real world application of large language models, right? This is powered by Chat GPT. This does not work without, you know, a large language model like Chat GPT. So it is using, you know, its ability to process information and then to speak the information as well. And if you didn't check out our kind of AI and 5 video last week. OpenAI and ChatGPT did just unveil a whole lot of, of new features for its kind of, its ability to speak back with you with some much more realistic human voices. So if you heard that, that, that voice there, super realistic. It doesn't sound very AI generated. Kind of a smooth voice, right? Yeah, it's kind of like, like what Juan's saying, having a Rosie the Robot back from the Jetsons. Yeah, I can't. Tanya also can't wait till I can afford one of these to clean my garage. Yeah, I would love for it to clean my garage. All right, let's, let's wrap this up and talk about one more piece of AI news that matters, and that is the Nvidia gtc. All right, if you are joining live, you might know my, you might notice my setup's a little different here or maybe even if you're listening on the podcast, maybe my audio quality isn't as crisp as normal. But that is because I'm on the road right now, just a couple blocks away from Nvidia's gtc. But let's talk about what's going on. So Nvidia is having its first in person GTC conference in five years to be held right here where I am today in San Jose, California. And this event will bring together AI leaders and top companies across various industries to showcase breakthroughs and advancements in AI. You know, we have the Jensen Huang keynote today, which should be extremely exciting. I'll be there. So if you have questions, I'll probably either going to be on on LinkedIn or maybe on on Twitter. So make sure that you're following us there and ask questions too. What do you want to know? I, I believe I'm going to have an opportunity tomorrow to do a kind of informal Q and A in a small group with the Nvidia CEO. So I'd love to hear from you. What do you want to know from Nvidia? What questions do you have? Where do you see everything going? So pretty, pretty exciting. And this is, I would say, one of the most anticipated kind of tech conferences. I'd say it's got to be top five in the last 10 years, right? Nvidia hasn't had an in person conference since 2019 because of the pandemic. But also, I mean, here's here, here's the reality as well, right. Five or six years ago, a lot of people maybe thought of Nvidia just as a chip maker, right? Maybe you thought of Nvidia if you're a gamer and you need a certain chip, or if you're a video editor and you need a better graphics card, right? That is not what Nvidia is anymore. I had an episode last year where I literally told people that Nvidia was the most important company in the United States and it was the most important company to the American economy. Guess what happened after that show? Nvidia has almost tripled in market cap in less than a year, which is historic. And the reason why this is such a highly anticipated conference is because right now, Nvidia has an unfair advantage over everyone else. All of the large language models, right? So even presumably the figure 01, because it's powered by OpenAI. OpenAI, one of its biggest partners is OpenAI. So everyone is. Or, sorry, Nvidia. Everyone is using Nvidia's GPUs to power the future of AI, to power large language models, to power generative AI, image, video models, everything. So Nvidia is actually the epicenter of the future of tech, marketing and business and how business is getting done. So it is a highly anticipated conference. I'm extremely excited for today's keynote as well as the conference. All right, so speaking of the Nvidia GCC conference, if you haven't already, go ahead and check out our show notes. We have a link on there where you can register for free. Unless you're from San Jose or San Francisco, it's probably too late to come in person, but maybe you're listening and you are. You can still buy tickets, but you can also attend the free virtual events. All right, so again, check out the link we have here in the description of the live stream and our newsletter and our podcast. You can go ahead and sign up. And at the same time, you know, we have instructions there to enter into a giveaway for a free gpu. So, yeah, maybe. Maybe your computer's a little slow, or maybe you want to run Nvidia's new chat with rtx, but you need a certain. A certain chip to do that. So go ahead and, you know, just by signing up for the free conference, you can enter into the giveaway. All right, so here's what we have this week. Speaking of Nvidia, I told you, I'm in a different location. Yeah, so here's what. We have a lot going on this week. Some special shows, some exclusive interviews that I am extremely excited about. And also, if you are an avid, you know, viewer of our live stream, can't thank you enough. But today or this week, we're going to be doing double duty Even today. So make sure to, you know, check out our newsletter that we sent out last night with our complete schedule. But tomorrow we will be having Malcolm Deo, the vice president of Global Financial Services at Nvidia, talking about making money moves and how Nvidia is using AI to change financial services. Wednesday, we're going to be talking with an Nvidia partner, Evan Sparks, about the how to create and capture value throughout your biz with generative AI. That is one I am going to be extremely excited about. Everyone's always asking how can I create value or how can I actually use generative AI? And then how can you tell of its impact? We're going to be bringing that to you on Wednesday and then on Thursday. Speaking of robotics, we are literally talking with the director of Robotics at Nvidia. This is going to be a great conversation with Amit Goel talking about robots among us, how Nvidia is building the future of robotics. And then last but not least, on Friday, driving the future forward. Nvidia's vision for an AI powered transportation. All right, so many great conversations planned for this week. We're going double time. So maybe we, we had a little break from, from live streams. Maybe you missed, you know, the live streams, so we're just doubling down. We have a lot of great shows planned for you this week. I hope this show was helpful. If so, please consider sharing this. Repost this if you're listening here on LinkedIn. We spend hours sometime for each show to put it together. It takes you about 10 seconds to go ahead and click repost or maybe on Twitter, etc. Or if this is helpful, please leave us a review. So that is it. So I hope to see you back even today and every single day this week. For more info and more action and announcements and breaking news from the Nvidia gtc. And we hope to see you back 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.