Meta's New Llama 3.1 is great, OpenAI's GPT-4o Mini & Adobe's AI Update

AI4H 68
===

Gavin: [00:00:00] Meta's new Llama 3. 1 model is here. It's big, fast, and it's open sourced. And it smokes the competition, like fine meats. Smoking these meats here. Smoking meats. Smoking these meats. Why do I suddenly feel like wakeboarding? Plus, OpenAI and UBI, Adobe upgrades its AI, XAI makes Elon say, oh my my. This! This is not a normal show.

Gavin: Because I'm here and it's AI for humans! Ha ha ha ha ha! You're not gonna like this episode.

Gavin (2): Welcome, welcome, welcome everybody. This is AI for Humans, your guide to the weekly wonderful world of generative AI.

Gavin (2): Kevin and I are in person this week.

Kevin: in hot, hot, HOT Lana, friends! That's right! I, I can't, I like to punch and kick during the intro, but I'm not wearing shoes right now. And we do have a Patreon, not an OnlyFans. Also,

Gavin (2): no one's going to see that because this is for the audio only. So

Kevin: just to be clear. I'm a master of all mediums.

Kevin: They can, they can hear these feet, they [00:01:00] can sense these toesies. The point is, we are in Atlanta. For the GBTA, we talked about it a couple of weeks back on our pod. We are so glad that they brought us out. We got to talk AI with a bunch of business and travel types. I had a really good time. Yeah,

Gavin (2): it was fun. It was really interesting meeting a lot of people out in the wild who are still learning about AI

Kevin: who did not like us?

Gavin (2): No, they did like us, Kevin. They did like us, which was really nice. So, you know, if you ever, if anybody out there is listening to this podcast and you think, Hey, I think it would be cool to have two people come to my business and talk about AI stuff. We do that. We definitely

Kevin: mitzvah or quinceanera. We are cheap.

Speaker 10: what is your 

Gavin (2): plan?

Gavin (2): What did you say?

Kevin: You flew spirit.

Gavin (2): I thought you said something else. I'm glad you didn't say that. I think, but it's all

Speaker 6: good. Okay,

Gavin (2): Yes. All right. We should get to everything here. So again, really quickly, thank you for listening here. If you are a fan of our podcast, you are listening, please leave us a five star review on Apple podcast.

Gavin (2): That makes a big difference for us also, we do this podcast on YouTube. And if you go to YouTube, you'll [00:02:00] see our news video that comes out every Thursday, the same time the podcast does. 

Gavin (2): And we have a Patreon that everybody can go join right now. We've talked about this before. We're not going to go too far into it. We are looking for help to up produce the podcast and do other stuff. And if you feel like you're a fan of the show and you want to help us, great. If you don't leave us a five star review, do all this other stuff. Kevin, I think we do have to get into it today though, because there is some super, super big news

Kevin: this week.

Kevin: Let's get into the

Gavin (2): news. Oh, in person. That feels good.

Kevin: Right? You don't want to wait for the three second lag.

Gavin: The big release today is llama 3. 1 and we're releasing three models. Um, the first time we're releasing a 405 billion parameter model. So it's by far the most sophisticated open source model that I think anyone has put out. It really is competitive with some of the leading closed models and in some areas is [00:03:00] even ahead.

Gavin: Why is this a parameter measuring contest? What is happening? Okay, so we have to, let's step back here and say, Meta released a new large language model, Lama 3. 1, the 450 billion parameter model, meaning that it is the best. Biggest model. 405 billion. 405 billion. Yeah, yeah. This is the problem. This is the problem with this entire scene.

Gavin: 405 billion. Parameters are like neurons in the brain. They are what manage connections. I am distilling this so poorly that someone is going to leap down my throat on YouTube, but that's a lot of parameters. That's, okay, so that is a lot, that's a lot of parameters, but let's talk about what it means.

Gavin: First of all, the benchmarks have come out and they're pretty good. Like compared to some of the other stuff, the benchmarks are looking like the new llama model. Is really one of the best models, if not the best. Now it's not the best in every benchmark and the benchmarks are a very fine group of things that they put these AI models through to try to see how good they are at certain elements, reasoning, math, reasoning, communications, probably a wizard role play.

Gavin: [00:04:00] There's a bunch of different benchmarks out there, but by and large, this thing outperforms GPT 4. 0, the best model by open AI. That's okay. Keep going. Keep going. I don't, what wizard roleplay are you thinking about playing? What should we play? We're in person. We might as well play some wizard roleplay.

Gavin: What's your name of your wizard? What? What? Is that a real thing you play? It was the first thing that popped into my head. You said Yeah, it's roleplay. Let's keep moving. Okay, the other big thing, Kevin. Is that it is being fully released open source, which a lot of people weren't sure was going to happen with this model because it is essentially Meta's biggest model and most cutting edge.

Gavin: They are going to release it open source and let developers fine tune it and see everything, see the guts of it, understand how it works. Yeah. And this is very different. Then both Claude and OpenAI, this is Mark Zuckerberg's hill to die on. He believes very strongly in open source AI. And we believe in it as [00:05:00] well.

Gavin: The very exciting thing about all of these LLAMA releases historically is that they come out. They wow you with benchmarks, but you just got to wait, let it simmer for like a week or two, and then a community will go and make it uncensored, add in a bunch of capabilities, fine tune it for math or coding, in three weeks time, OpenAI's got to release something.

Gavin: Yes, yes. To answer. So this is the other thing to think about with this is before we get into all the stuff Zuckerberg has said about it, OpenAI's been sitting on something, right? They've been sitting on something they've been working on for a while. We know they've been sitting on the GPT 4. 0 voice model.

Gavin: But I think they're sitting on something else. I think we're gonna see something by November or December of this year. You have said post I think that's a really smart thing. But we might see a GPT 4. 5 or something. But what this means for all of you is that a free large language model that competes with the best in class paid model.

Gavin: Zuckerberg just mic dropped it on the community and this will, [00:06:00] over time, make its way to all of the meta apps. So this is what you'll have when you go to Meta AI. Over time, it's there. Well, it's already there in WhatsApp. They're rolling it out to Instagram and to Facebook. Booyah! Did you really? I did, I did, I used it.

Gavin: Are you sure you used that one? It says 3. 1. Wow, because they said that it was coming later. It may not be the 3. 1 405 billion, and that's the thing. So, it's sometimes a little difficult to key into exactly which model you have. But this thing We'll be in all of these apps, but not just the official meta apps, Gavin.

Gavin: These will be in every AI app that wants to build for something that's free and open source. And that's what Mark Zuckerberg talked about is the idea. He's going to put it on Azure. He's going to put it on Amazon, all these places that that's going to allow developers who are the people that are going to make the apps that will build with this are going to have access to it.

Gavin: We built it for ourselves and before Llama 3. 1. You know, we, we kind of had this instinct that if we made it open source, there'd be a community that would grow around it. And that would actually extend the capabilities and make it more valuable for everyone, including us. This is a real [00:07:00] shot across the bow to the companies like Claude and companies like OpenAI.

Gavin: But we used Meta today and you were surprised at how capable it was. The one really interesting implementation that's out right now that you can go play with is on Grok. Not G R O K, not Elon's Grok, but G R O Q, which we've talked about in the show before is an AI hardware company that's doing some really interesting ways to process AI on Their own proprietary hardware.

Gavin: And it's always been known for being very fast, but I actually used grok to, to rank sausages. Of course you did. I asked it to rank sausages and it gave me a really good ranking. We'll put up the video here. Hey grok, tell me about the best sausages in the world. Okay. Now tell me, uh, make a list of the ingredients in these sausages.

Gavin: And then I actually wanted it as a table. So it formatted that data as a table. And by the way, it is taking pages of output and. Nearly instantly because of the grok custom silicon. Handling these requests in a way that made me go, Ooh, Oh, that's interesting. We've seen the [00:08:00] speed many times with grok, but the capability matched with the speed was magical.

Gavin: You know what made me think of when I did that? Cause what I was doing was you, you just saw it here in the video. I was doing audio call ins and it was giving me text back, but it did make me think about GPT 4. 0 voice mode. Because I bet that in some ways what they're doing, right? They're like processing it so fast that it can feel like a real time conversation.

Gavin: That's right. So you asked for sausages. Yes. You had it grid those sausages. You had it rank them by, I think, deliciousness. Yes. And by usability as a weapon. Yes, which is, I, you know, sausage fight. You never know what's going to happen there. We don't want to do that. Let's not do that. Do you know which wizard you want to call if you have a sausage fight?

Gavin: I don't want to do that. I don't want to do that. Okay. I did then. I even asked it goofily. What sausage would be best to use as a weapon if you were in a fight. And we're gonna do a deeper dive. This just came out the day of our recording and we've only had minutes to play with it, but we even reused some older character prompts that we had written.

Gavin: And it was so much better. So much [00:09:00] better, so capable. It was, it made him laugh. Yeah. When writing a mini comedy script for something that we're going to have coming out pretty soon. So again, the takeaway for all of you is that this thing is out. Yes. This thing is going to be free and everywhere. Yes, this thing is good.

Gavin: Yes. Okay. So the other thing that happened is that Mark Zuckerberg wrote a very long letter titled open source is the path forward for AI. And I just want to say, he also did a 35 minute interview with Rowan Chung. So shout out to Rowan Chung, the AI rundown, a great newsletter. Well, he really thinks that open source development and just so everybody in the audience is familiar.

Gavin: I know we have a lot of tech experts in our audience, but a lot of people that aren't. Open source development means basically everybody is allowed to take software, work on it. They might put it back into the open or they might make their own versions of this that they become closed. But it allows for a community of people to work on a piece of software and Zuckerberg has really been championing this for AI because he believes it is safer than closed development in his mind.

Gavin: One of the biggest issues with AI becomes unintentional harm. And the [00:10:00] idea behind unintentional harm means that the A. I. goes rogue on you, right? So, the famous paperclip problem where it's like you tell an A. I. to make paperclips and then that's all it wants to do and it ends up killing people because it has to make more paperclips.

Gavin: With open source A. I., what Zuckerberg is saying in this letter is that you don't run into that problem because people will see that before it comes. Now there is the other side, which is intentional harm, where people have the ability to do bad things. And a lot of the bigger companies are using that reason for not releasing their models.

Gavin: That creates an environment where there's that clearly the haves and the have nots. I have my model. It's closed down. It is all seeing all powerful. Sorry, everybody else in the world where that is the open source communities rallying cry. Our software needs to get as capable as the big leagues that are closed and Usually there is a significant lag time between the best models and the open source models, and Meta is closing the gap.

Gavin: Do you ever wonder with Meta, like, when Mark does one of these things, do you like, like, do you think he like just gets the group together and he's like, Let's go [00:11:00] boys. It's time. You know, it's like, it's like, cause you know, that's a big launch. Yeah. Like they're launching across every single platform. And the other thing we've skipped, which we used to talk briefly about the avatar?

Gavin: That's what I was going to say. That's what he's doing with his bros. Yes. Yes. Get together. We're going to smoke some meats and we're going to see me as a gladiator. How do I look, boys? So basically, Mark Zuckerberg teased, which I tried to do and I couldn't access, I don't know if you tried this, but I don't want to scan my face for meta.

Gavin: I've scanned so many things, I don't care anymore. Stop. Stop. Stop right now. We're talking about this thing. So it, what Mark Zuckerberg teased, we'll show the video in the YouTube. You basically go into the app, and he does it within Instagram, and you scan your face very quickly, and you say, imagine me as this, and he says, imagine me as a gladiator, imagine me as a streetwear guy or something.

Gavin: That's right, yeah. I'm pretty excited about this latest update to Meta AI. It basically lets you generate images of yourself, in any style, and doing anything that you want. Let's check this out, because I need a new profile pic. All right, so you take a few photos of yourself, then you wait for it to [00:12:00] upload.

Gavin: And then you can enter whatever prompt you want. Imagine me as a gladiator.

Gavin: All right. Boy band. And it's stuff we've seen in stable diffusion. We've seen in all sorts of different things, but it is one tap easy. It's within Instagram, which is billions of users and it's in people's hands faster than Apple intelligence got that feature into people's hands, which is again, the big thing I think, even in his letter, he goes off on Apple.

Gavin: I think Mark is trying to. Push past Apple. I think that's the goal he wants is like Apple's in his sites. Not really open AI, not really any of these other AI companies. Cause they're kind of small compared to what Facebook is. He wants Apple. All right. So there are a bunch of more big AI stories. GPT 4. 0 mini just launched.

Gavin: Here's the really exciting thing. If you're building in the AI space, if you just assume that the frontier, the foundational models are going to get cheaper and get [00:13:00] better. You're gonna win, and that's what's happening now. This was massive for OpenAI, they cut costs by 60%. So on OpenAI's end, they passed a lot of those savings on to the end user, and suddenly if you had a project that was getting thousands of users and costing tens of thousands of dollars a month, it just got cut by 60%.

Gavin: Yeah, so suddenly your margins go way up, you have a much better chance. Sam Altman said specifically, he tweeted this, he said way back in 2022, the best model in the world was Tex DaVinci Oh, oh, remember this one? The DaVinci model, yeah. We used it, we used it a while. It was much, much worse than this new model.

Gavin: It cost 100x more. That is two years ago. Two years ago, there was a model that was way worse and it cost a hundred times more. So what you see is a curve of exponential capability and you see a drop in price. And going back to my grok experience, When you have always on super smart AI that is very fast, the world changes.

Gavin: [00:14:00] Okay. Really quickly, we have to talk about UBI. So there was a really interesting study that Sam Altman partly funded about UBI, universal basic income. And the idea with UBI has always been, if you give people money to live, a monthly stipend, a thousand bucks a person, yes, that it might make their lives better in some way.

Gavin: And that also, the reason it connects to AI is because Transcribed With AI, there's a lot of people worried that it's going to put people out of work and that UBI is the only way that that people will actually be able to live in the future. One thing was it got a lot of negative press early on because it actually people worked a little bit less.

Gavin: You're right. That was the headline. People want to say, Oh, clearly UBI doesn't work. This was a longer term study than most. And. They basically found that people worked, it was like 1. 6 or 1. 8 hours less per week. Yep. It basically totaled over the course of the year, you got six or eight vacation days. Yes.

Gavin: Which I know is reason to take to the streets. Yes, but, but I will say the interesting thing was two, two things they figured out. One is they [00:15:00] spent more time with friends and family, which That's amazing. Secondly, these are all people who are not making a lot of money and it allowed them conceivably to go do other stuff and to build wealth.

Gavin: In fact, when they came back and looked at their savings after year three, this was a long term study. The people who got the UBI actually had significantly more savings than the people who didn't. Now, all of this is early stages, but I do think this is something we're going to have to seriously look at as a human being.

Gavin: Many of the other studies show that it is incredibly effective. They are smaller studies. I think this ultimately shows that. I know people wanted to cherry pick certain facts and shift the study in one way. But even Sam Altman himself was really excited by these findings. And also, levels of altruism went up.

Gavin: People used their extra money. to help friends and family to buy goods and services for them. I think it's just a net positive. The UBI of a thousand dollars is not that much when you think about it. But of course, for people who don't have homes or people that need to help in a lot of ways, it's a big deal, but imagine a UBI of like 50, 000 a year, right?[00:16:00] 

Gavin: There's a lot of people who may be out of work that could pursue all sorts of interesting things. Now, all of this is kind of a pipe dream right now because nobody knows they're only tests and. I still doubt that I can see it working in America per se, but we'll see. I will say, if you're worried about job displacement with AI, you should probably get into something hyper specific like being a vector artist, because we see a lot of AI tools, if you don't know what vector art is, it's Kevin, I hate to tell you something, Adobe just introduced some new AI tools and one of them is you creating vector art in Illustrator.

Gavin: Uh, don't be replacing artists, Adobe. So Adobe just rolled out a bunch of new Firefly AI tools in its things, and one of them is a very cool way that you can create art within Illustrator. And Illustrator Is there a professional art program so they rolled this out? It allows you basically to generate a shape and then it you can fill it in generative fill like what they would normally do But vector art, so it's unlimited resolution.

Gavin: You can go to any scale with [00:17:00] it and it's like very fine detail vector art It looks It's really good. And you can see all the little vector points. And it's layers. Right? It creates layers. So the layers is a big thing. If you're an artist or an illustrator, you understand AI art never really creates layers, but this creates layers for you.

Gavin: So it's a very, very big update. Yeah. But if you're going to have really incredible tools like vector art, bird generators, or super fast chat bots, I guess you need like a super villain sized layer with a super cluster of compute. Kevin, could this be in Texas? Could I put the supervillains there in Texas?

Gavin: Gavin's dear friend and spiritual guide in this world, Elon Musk, has just announced that he has the world's biggest supercomputer for a Grok 3 and some people are saying he's lying, that it's actually not fully online right now, like not even close to it, so to say. You've got the most superpower thing is a little bit of a stretch.

Gavin: Elon has been known in the past to stretch the truth a little bit, but let's hear him talk about his new XAI [00:18:00] supercomputer in Memphis, Tennessee. XAI is a fairly new company. It's just a little over a year old. Um, so we, we really need, we have a lot of catching up to do to companies relative to companies that have been around for five or 10 or 20 years.

Gavin: Um, we're catching up fast. I think with the, the. Velocity of improvement of XAI is faster than any other company out there. Um, we just completed the, um, we were just able to install and bring online, um, a massive new training center that we, as mentioned, we're building in Memphis. Um, and it's, uh, from getting hardware installation to it beginning training was only 19 days.

Gavin: And that's the fastest by far that anyone's, uh, been able to do that. Pretty impressive. He's pretty good at getting engineers to do seemingly impossible feats. Yes. But he's going to use this to train Grok 3, which is the AI assistant on X. Grok with a K, not with a Q. Yeah, Grok And that should be out, he's saying sometime in December.

Gavin: So we'll see how capable it is. [00:19:00] Honestly, again, I think the end of this year is gonna be bonkers. You're gonna see OpenAI's new thing, you're gonna see Claude's gonna have a new thing, Llama will be fine tuned by a bunch of people by that point, and then you'll have Grock. I don't know, maybe we're gonna get an Apple Intelligence, maybe we'll have some stuff leaked out by then, although they say later on.

Gavin: Either way, it feels like a pretty big rest of the year. It feels like things are certainly not slowing down, which everybody said, and it feels like there's still plenty of performance to eek out of all of these things. 

Gavin (2): Okay, everyone. , one of our favorite segments on the show. Every week we call out some of the coolest things we've seen. Yeah. It is time for AI. See what you did there. 

Sometimes you're scrollin without a care, Then suddenly you stop and shout. Hey, I see what you did there. Hey, I see what you did there.

Kevin: Some augmented real time AI artwork stopped us in our tracks and made us [00:20:00] say, Hey, I see what you did there. Smallfly over on X did a massive post about a deep dive, around an image to image diffusion pipeline using stream diffusion and SDXL turbo.

Kevin: Using this pipeline, Hughes was able to take live video, , , and have Stable Diffusion re imagine what it was seeing as AI art.

Kevin: And he did this really beautiful treatment with it, where he had this overhead camera, and a bit of a, like a magnifying lens,

Speaker 10: really cool. Yeah,

Kevin: and as he puts the lens over the world, , you can see the images going in via stream diffusion and stable diffusion XL changing what is within that magnifying glass so that a bunch of abstract shapes or play doh suddenly look like vivid imagery and as you move it around in near real time the images are updating and it's just a stunning

Gavin (2): it's very cool because it reminds me of like how specific and unique AI tools can be, right? This is an AI tool. This is a tool that he created to use This sort of thing, but it's doing something that's so artistic and so interesting. , one of the coolest ones was [00:21:00] when you see a little dropper going across three different things and each time it shows up somewhere else and , he's actually literally putting drops of a dark liquid. Yeah.

Kevin: like water, and as the ink expands, like a digital Rorschach, it is drawing what it

Speaker 10: So I 

Gavin (2): think this is a really interesting thing to look at for the future of AI art when it comes to how do you and I use the tool to do things that aren't just creating an image out of the blue, ? This guy is actually creating a new way to look at reality through an AI art lens, 

Kevin: And if we talk about this near ish world of augmented reality Gavin, that's what it's gonna be Like you're gonna put on the glasses. You're gonna pick your Checkpoint what your model is and as you look around your world's gonna turn into play doh or pixar 

Gavin (2): all right, everybody. Other next up. A really quick thing from a guy named Nico Christie, , who invented a new AI agent called Roshi.

Gavin (2): And Roshi is an in game AI app that he created in Minecraft that kind of is like having a friend [00:22:00] or a sibling to show you how to play the game within the game. So imagine like an AI agent that's like a tutorial wizard that can go around and do stuff with you. 

Kevin: I am not a Minecrafter. , I have tried to dip in as an adult and I felt Woefully underprepared for what that world was and how it works 

Kevin: it would be great to be able to spawn an NPC that can hang out and help guide me through the experience.

Kevin: 

Kevin: How many times are you playing any game and you have to pull out your phone and go to a second resource and figure it out? We might find in the near future that there's, ,

Speaker 10: game

Kevin: in game NPC or whatever, yeah, that's gonna guide you through it. But that's really cool.

Gavin (2): it's very cool. 

Gavin (2): Okay. . And then finally. Interdimensional, , TV created. What is one of my favorite , new AI videos? It is called runway of power. Which basically takes what we think of, we should kind of walk through what we think the guy did with this.

Gavin (2): When you watch the video, what you're seeing is world leaders. Walk a fashion runway with very specific outfits on them as they go further and further down.

Gavin (2): And you see the starts with the Pope and then you see Kim Jong un, you see Vladimir Putin, you [00:23:00] see Biden in a

Speaker 6: He's rolling. 

Kevin: No one's hating. You see Elon do a weird transform into a future space pirate of some sorts. , Donnie Trump , and the Louis V.

Gavin (2): what's cool about this is , we think that this guy used runway gen three video to make the runway footage probably of people walking down, but he's also doing a couple other things.

Gavin (2): He's probably doing face fusion to swap faces

Kevin: a face swap in there, but it's the clothing where I don't think runway is generating clothing that detailed, but there are other posing software and clothing swap software where you can, , upload an image and it will apply it.

Kevin: to, , a model and, interpret the motion of that model and make the clothing move and flex around it. So I think this is like probably three, maybe four tools all bundled into one, which is a really stunning

Gavin (2): That's right. And my favorite one is the Tim Cook has an outfit on where he's essentially just wearing an old MacBook. It's not even like a new MacBook or maybe it's an iPad. It's an iPad.

Gavin (2): You're right. It's an old iPad. Uh, it's not even like it's got the rounded corners in a weird way. 

Kevin: It's really [00:24:00] well done. It kind of popped, , on the , internets.

Kevin: And we showed this artist before because they did celebrities in mortal combat back in the

Speaker 10: Do 

Gavin (2): you know he did four Celebrity Mortal Kombats? We may not have seen all of them, but like, oh no, number four is Britney versus Cosby. Come on, Interdimensional TV. What are we doing here? All right. I see what you did there. 

Gavin (2): So Kevin, this week, since we're in Atlanta, we haven't done a crazy amount with AI because we've been working and on the road, but we have done a few things because being in person, Kevin and I end up doing some weird stuff.

Gavin (2): Next week, we are taking a mini break, a one week break from the regular show. So we won't have a regular show. But we are going to have a special treat. A special video that we have made and I will just say Kevin and I spent a lot of time trying to generate Babies with so no i'm not giving anything away.

Gavin (2): It's just a little tease babies with tattoos with runway gen

Speaker 6: Yeah, 

Kevin: we've been content flagged a few times by Runway, but I think we got enough [00:25:00] generations through that we're going to have a nice fun video for everybody as we take a down

Kevin: Wow, what a world of content spoiler I've slept a total

Speaker 10: yeah. Sorry. We didn't say that at the top.

Speaker 10: Well, I

Kevin: I mean, figure it out I've slept a total I had an oopsie on this trip and I slept about three hours in the last two days I don't think it's healthy. It's not healthy,

Speaker 10: it's not

Kevin: can just say straight up, it's not healthy. So, don't take red eyes, friends.

Gavin (2): Don't take red eyes. Kevin didn't sleep. . Thank you so much for listening. So again, we will not be in this feed next week. We are taking one week vacation. It's the middle of summer. We're taking one week off.

Speaker 6: something

Gavin (2): Unless something crazy happens, which is very possible. So thank you, Podcasts. We love our audio listeners so much.

Gavin (2): We hear so many awesome things from y'all. So please keep listening and please keep telling people about the

Speaker 6: podcast. 

Kevin: And hey, sweet patrons out there. We see you. We appreciate you. It's going to be more than a tip jar soon. 

Kevin: But in the meantime, thank you to everybody who is giving us dollars to support this operation. It's something we [00:26:00] didn't want to ask, but we're delighted that people are showing up for that. So so much. And we'll have more there soon.

Speaker 9: Bye everybody.