Meta's Big Llama 3 Release, Google's New Culture, MKBHD vs. Humane

0:00Meta has introduced its new big Llama 3 AI model, Google fired 28 protesting employees for taking over its offices, and do reviewers like MKBHD have a duty to not kill companies whose products suck? All that and more coming up right after this. Welcome to Big Technology Podcast Friday edition, where we break down the news in our traditional cool-headed and nuanced format. We have a great guest for you today. Alex Heath is fresh off a conversation with Mark Zuckerberg, talking about the Lama News. He's the deputy editor of The Verge, and he's the author of the Command Line newsletter. And he's here with us today to break down everything that's happened this week.

0:35Alex, welcome to the show. hey thanks for having me alex always good to have a double alex pod two alexes uh this should be an epic week first of all like we're continuing our cycle of every friday or every week having a big celestial event happen we had three weeks ago an earthquake last week an eclipse this week taylor swift's new album is out the tortured poets department so we're keeping up this streak have you have you gotten a chance to listen to it today you know i actually started playing it um like an hour ago. And I don't know, man, I really feel like I shouldn't weigh in here unless you have something like overwhelmingly positive to say about Taylor Swift.

1:16I've realized it's best to just not put opinions on the internet. Well, I'm not going to be afraid with my thought here. And I'll tell you this. I hit play. I was eager to see what she had. It's obviously been a very hyped album. And I realized that I am Taylor Swifted out. Between the exposure at the NFL and the fact that she's become this mega even bigger than mega celebrity i am i think i'm good for the time being on new taylor swift music and i'm not ashamed to admit it it doesn't mean she's not great at what she does but i've had my share of taylor swift i'm ready for something new r.i.p your mentions alex all right well we'll take your your silence here as uh as telling but we won't put you on the spot uh something that you will be able to comment on is this new metal llama 3 model obviously it's really when we talk about like Llama 1, Llama 2, Llama 3, it's a large language model, but it's really more than one model.

2:09They've released two models this week, two smaller models, and they're working on a massive one that's set to come out this summer. And of course, you know, we have GPT-4 and Claude. These are not open source models, but Meta's are. And, you know, it's a very big strategic moment for them in terms of putting these out. So we'd love to hear your reaction to what Meta is doing here and where it puts them in the conversation in the broader AI world. Yeah, I mean, I'd also be curious to hear what you think. I think they have been considered a leader in open source AI, especially with Llama 2 last year.

2:47But now with Llama 3 and the assistance and putting that everywhere in WhatsApp, Instagram, if you start using the search box in Facebook and Instagram, you're going to see Meta AI, at least in the US and a bunch of other countries. So that's a big bet. That's a lot of real estate. And I think, you know, I talked to Zuckerberg this week about it, and I think he wants to use kind of their massive distribution to kill ChatGPT in the cradle if he can, so to speak. I mean, I think he sees this area of, you know, AI chatbots, agents as something that Meta has to be in. And I don't know, man, it's not, it's not super obvious to me that that is a thing that makes sense inside a social media app.

3:35I think the jury is still out on if people want to be, uh, using like a chat GPT, like experience next to their friend chats and WhatsApp. right so I think they've got to prove that this is now they're in the like I don't know this week was the like firing gun of the race and now they have to prove that they can execute and that people actually want this and so I think that the pressure is on well let's talk about that right from the start and I definitely want to talk a little bit about these models and their capabilities but I think the productization is really the key here so your point about whether this should live within a social media company well it's a messaging company, I think, first and foremost, right?

4:15Meta had this pivot to privacy, basically, the company figured out that people don't really want to share as much on news feeds, they want to share privately within messaging groups. And so they've, you know, you think about what Facebook is today. And the big blue app is much less important, I would argue, than the messaging apps, Messenger and WhatsApp. And to some extent, Instagram is a messaging app with the media sort of front end. And so if you're thinking about where one of these bots is going to live, naturally, wouldn't it be in a messaging app where you would effectively, you're messaging with your friends, you can call it in, you know, to conversations with your friends when you're like trying to find something to do or trying to answer a question.

4:55And then also it is a conversational interface. So why not have that live within a messaging app? Yeah, I guess just the mental model that I have for these chatbots is that there are things that you go to separately, right like the chat the chat gpt app um and i think of them more as like utilities right now than i do like another friend almost um that would live next to like where my mental model is when i go into whatsapp is like to talk to people or groups of people and so adding in like a chat gpt type experience that feels a little strange i know like in the virgin newsroom there was a lot of heated opinion about whether this was a good idea.

5:37And people had a lot of strong opinions that they didn't want this in their WhatsApp. So they have to prove that people want it. And what Zuckerberg told me was they see the feedback loop of this. So basically getting this out to more people, having more people engage with it, and then learning from that and how they can improve from that as being one of their key differentiators over time versus like an Anthropic or a chat GPT, which just has less surface area and less potential users to reach by default of just not having a network of 3 billion plus daily users that Meta has. So it makes sense why they're doing it from like a, you know, if you really believe that these AI agents are the future and that's how we're going to be interacting with computing, you want to get it out there.

6:25Meta has a mixed track record on this stuff, right? And, you know, I think, but I think Zuckerberg sees this as like you've got you had stories you know uh which was like a format that snap invented and then they grafted on to to metas apps and then the same thing with reels which was you know tick tock basically which has works now um and has has done really well for them and i think he sees the assistant as the same thing for these ai chat bots um and i mean he's very clear like he wants this to be the most used AI assistant in the world. And they have a shot just by default of the distribution.

7:03It shows how distribution is still king. And even if you have cool tech and you start the race like OpenAI did, you might make it just come down to distribution at the end of the day. It is really important, like the type of personalities that these things take on in different settings, right? So in meta, that might be more social. Within ChatGPT, it might be more informational. Claude, you might want to like talk with your documents there. Like there's different versions of this. And it may be that, you know, because of that distribution, they can sort of steer what this becomes. And this is sort of what you talked about, how distribution is so important, right?

7:41We have a comment here coming in. People don't come back to ChatGPT. And it's true. I mean, ChatGPT, the usage from the data that I see has really leveled off, like hit that hundred million user benchmark pretty quickly and then hasn't really blown past that. In fact, it seems to have shrunk. And so even if meta isn't the perfect place for this, it still seems like it has a chance to live out that vision that Zuckerberg wants to see because again, because of that distribution, right? Like let's say chat GPT is at 200 million users right now, right? It's nothing compared to the billions of people that now have access to meta AI.

8:18Yeah. I mean, that's the bet. We'll see. Right. I think it's been a running joke inside Meta for the last six months or so because they first debuted this assistant in September of last year. Right. And it was only in the US, but it's kind of a joke inside Meta that no one uses it and not even like Meta employees use it. Um, and it was also kind of hard to find though. Like you had to kind of search it out. Right. Um, I was actually shocked in the Virgin newsroom. How many people didn't even know that this assistant already existed in their WhatsApp and messenger. Um, and that's why meta is now putting it, I mean, literally inside the Facebook feed.

8:58So if you're scrolling and there's a video that may like recommend a prompt based on what's in the video, like, do you want to learn guitar like this? Um, and same thing with the search. I mean, the IG search box is probably one of the most, you know, traffic surfaces of that app. And the fact that the assistant is going to be right there in a lot of Meta's biggest markets is a big bet. And it's, they would normally test this for a really long time, A, B tested in different countries with different cohorts. And they're just turning it on for everyone, which means it's a top down huge bet. And yeah, I don't know.

9:36They want to play to win here. And they have a really great research group. It's definitely, I wouldn't say it's considered to be as elite as open AIs, but they're definitely in the top, you know, three or four research groups in the world. We've got Google, OpenAI, VEM, Anthropic. So they've got a lot of the right ingredients and they've got a lot of GPUs. So I'm really interested to see with Llama 3, this 400 billion parameter model that they're training. I know you talked to Ahmad, their head of Gen AI on the pod this week. And that's going to be a big deal if they open source that. There's not been a model that large and that complex that's been open sourced.

10:24And I'm curious to see from a bigger than meta perspective picture, what pressure that puts on OpenAI and others to either open source or not. So Alex, you've brought up a lot of the skepticism here in terms of whether this will work. What is your sense on whether it will work or not? I don't know. I know you're a reporter and you're going to beg off this question, but just handicap it a little bit. Like the assistant or Lama 3? Well, let's go both. I mean, Lama 3 is a big deal. I mean, it's already, I think it's been, they said it'd been downloaded like over 100 million times. It's been used in a lot of apps already.

11:10It's a huge part of the developer ecosystem for AI already. So the third, you know, Lama 3, especially the 400 billion one when that comes out, it's a big deal for the industry. In terms of the consumer applications of this as an assistant in Meta's properties, I don't know. They have a good shop, but it has to be a good product. And I'm looking forward to trying the assistant more now that it's been upgraded with Llama 3. I think there was a sense that Llama 2 was, it was barely like chat GPT 3.5 level performance. So these bots are really only as good as the models that power them. And so now that there's a model that is more approaching GPT 4, and when the 400 billion one comes out, maybe even exceeds it on some areas, that will make the assistant more compelling.

12:06You know, they've also got Google in there now, which Google's providing real-time search results, which I think they're the only chatbot besides Google's own Gemini that has that. So they've got Bing and Google. I'm sure they'll build some other hook-ins. They've got to keep building it out and making it – I think it's pretty bare bones right now. But they've got to make it more personalized as well as something I really want. Like I've heard that they will probably let you eventually be able to, to generate images based on your likeness on like Instagram, for example. Oh, that's definitely coming.

12:39Yeah. Yeah. And like that, that's cool. And that's unique to what they do. So they need to have more kind of unique wedges that compliment the fact that your internet presence is already on their apps. And they have a lot, that's a lot of valuable data that they can use to personalize the assistant to you. um and like if i'm you know searching you know on instagram for uh you know ideas for like a japan trip that i'm taking next month i would like the assistant to know that and to know what i'm already looking for and like if i ask it a question be like well you looked at this spot in kyoto this is another one that looks you know very similar um stuff like that so i think it's like we're in the very very beginning of this stuff uh especially for meta and they they have to move really quickly because open AI is moving really quickly.

13:26And if I would talk to Zuckerberg about this, if open AI puts out GPT-5 later this year, does that leapfrog them again and reset everything? So it's a fast moving space. That's why it's so fun to cover. And what did Zuckerberg say to that question? The thing, I mean, he can't know, right? I mean, no one really knows what GPT-5 is going to hold, But he made the point, which is fair, that they put out Llama 2 after GPT-4 had already come out. And now 3 is coming out before GPT-5. And they're about to train version Llama 4. And they're already roadmapping Llama 5. They're moving very fast. With this model, they said that they're using 10 times the amount of data and 100 times the amount of compute that they use to do 2.

14:15And that's what Ahmed Eldela told me on Big Tech War Stories, which is the show I do through bigtechnology.com, like the newsletter separate to this one. It was a very interesting conversation where we talked about the making of these models. And of course, like, again, like we've talked about it a couple of times in this conversation. There's like the two smaller models that are out now and this bigger 400 billion parameter model that's supposed to come out this summer, which is like four or five times the size of what they have now. and I want to talk about that because there's questions about whether they're going to open source it but I think this conversation also gets to the value of where gets to where value is going to be created with these models right because you have of course the actual development of the model themselves and that's what meta and open AI and Anthropic and Google are competing on and then you have the way that it gets built into products and there's been so much focus on the actual models.

15:11But I guess like one of my thoughts here is that the model race is going to matter less and less as these things converge because the real money is going to be made in terms of how people turn this into products. And it's almost like Meta has this week even pushed that home even more because it's Llama 3 model is good enough and it's going to be, you know, it's going to be free to everyone to use it through open source. And it's good enough that like you can get it and And it's not like there's going to be riches if you build something incrementally better than it because people will just go with it.

15:44So that seems to me like we've always asked, like, where's the economic benefit going to come from this AI revolution? And more and more, it seems like it's going to be with semiconductors like NVIDIA and then the way that you build products on top of it. So like whether LAMA 3 is as good as GPT-4 or not, it seems almost like beside the question. Like the real point is the one we started earlier, which was, you know, is this going to actually deliver value in a product for meta? And I think everybody who's building with AI, including Microsoft, is asking, is AI going to deliver value in our products?

16:16And that's the big question right now. What do you think about that? I think that is exactly right. I mean, I cover this stuff and I don't get a ton of value out of these chatbots. You know, the hallucination problem is a big one for me. and being able to trust what it tells me. Like if I'm using it for research, for a story, and I want to like compare a bunch of financials or something, it would be actually kind of hard to find all this and SEC documents, etc. And I ask the model to do it in a very specific way. And it gives me an answer, but it's like, I can't trust it. And therefore, why am I using this to begin with?

16:57Because like, then I'll have to go find all of it anyway. which I was trying to avoid by using the model. So the more it can be grounded, the more things like search, um, are integrated and you can, it can learn from you. I think the more valuable this stuff will become. I mean, clearly people want to use this stuff. I mean, I don't think meta would be putting this across all their products like this. If this was some flash in the pan interface. Um, it's just, it's just early, man. I know I said that already, but I just feel like the industry is moving so fast, even though consumer, um, interest in actually using these tools as valuable parts of everyday life, uh, is relatively early.

17:45Um, and I think, I think Zuckerberg knows that. And he hopes at least that the meta AI will be the first time that millions of people are introduced to conversational generative AI like this. Because even with ChatGPT hitting that 100 million user mark as quick as they did, I think that was monthly or something. It wasn't daily. There's a lot of people, a lot of millions of meta users that have never used a chatbot, right? And it reminds me of stories, right? Like we both covered that era with Snap and the competition there. And they really kneecapped Snap and hurt its growth by taking this kind of magic thing that Snap had and introducing it to a lot of people where it's like, at that point, it's like, why do I need to go download another app?

18:35If I have a really good chatbot in my WhatsApp, why would I download the ChatGPT app? Exactly. And last year, I wrote this story that said, Mark Zuckerberg is coming for Sam Altman and OpenAI. And it's like Zuckerberg, first of all, he's very good at seeing a product that's taking off and has mass user appeal and baking that into Facebook. And I also think there's an element of like, he wants to be the alpha dog in the tech world. And it almost like, it's so interesting because there's been all these different attempts for Zuckerberg to shape his image, whether that was like the tour that he did around America or, or the different speeches that he gave defending free speech, all that stuff.

19:15But it seems to finally be working for him. And I think one of the signs that I've seen is that there's been this image floating around social media this past week that sort of Photoshop of his actual announcement video where he gave the Llama 3 update cleanly shaven, but with a chain and a t-shirt. And someone Photoshopped his beard onto him. And it's been going wild. And there's all these great memes of it. Like somebody wrote, I think they posted the picture and they did quote, you're the only girl I'm talking to Zuckerberg and the beard. And it is interesting. It sort of goes to the image of Zuckerberg, which is kind of an important part of this whole thing.

20:02He's almost become, and I'm curious what your take is on this, this like kind of tech icon in a strange way where he's being worshipped in Silicon Valley as someone I think who's taken a beating, kept on shipping and has this kind of don't give a, you know, fuck attitude in terms of like what he's going to do. And that's sort of like the new Zuckerberg. And it seems to be working. What's your perspective on that? Yeah, I mean, I think it's certainly working. I think his comps team is thrilled. um yeah it's i don't know really what to attribute to i mean you said he's like an icon i don't know if icon has positive or negative connotations i think it can have both but like i think he's always been an icon i think he's always been like but there's a different level to it now well now he people like him of him going through the timeline right yeah people in tech like him now which is um different and he's seen as like a innovator which is what he desperately wants and needs for recruiting and for all these reasons and also just personal, you know, gratification and ego, right?

21:05We all want to be loved. Um, yeah, man, I, I don't know. Like he, uh, he's in his, like, what is it like chat era? I don't know. He's like, he's, he's really just letting his hair down literally. Um, and it's working from a perspective of like, he seems to be more out there and willing to engage in a way that he felt really robotic and really closed off for a very long time. And we're both members of the press. I wonder what you think about this idea that this might be part of his ability to have, he took a real beating from the press over the past few years. Like there was a moment where you could say anything negative you wanted about Zuckerberg and no one would come after you.

21:52And sort of some of it became a little excessive. So he took this beating from the press. You know, he's continued to remain relevant, relevant through Facebook. He made these difficult year of efficiency decisions within Facebook. And then like, you know, almost to solidify his reputation as a fighter, right. Started doing UFC stuff or whatever his, his MMA. Do you think that's part of it? that just that that's all connected the MMA yeah yeah for sure I mean I think the MMA literally was like it was the pandemic and he found a hobby a little bit but um yeah I mean I think he you know there's something he said I interviewed him around their connect event last year in the fall and there was something he said at the end about I just want to be like building awesome stuff again like i think he'd felt really um pulled away and distracted by the everything from cambridge analytica on through like 20 really into the pandemic that time period was all about politics uh scandals government um he was spending a lot of time in policy world and not like he is a engineer and a product guy at the end of the day and i think he feels like he's finally able to like focus on that because the reputation of the company is it constantly under fire like it used to be they're always like one scandal away from that happening right like they're i know they're they're terrified of a gemini diversity type scandal with the assistant for example like if that were to happen that's a whole nother you know the metaverse stuff was rough in terms of kind of getting out ahead of their skis on messaging on that and then everybody kind of realizing there's not a lot there yet.

23:35But now that he's moved on to AI and that's what everyone cares about, you know, they're able to position themselves as like a key, and it's true, they are a key leader in the space. Yeah, I don't know. I just think he's like becoming more comfortable. Oh, it definitely appears that way. Let me tell you where I think the potential misstep here is. their commitment to open sourcing these models. Now, the argument that they're giving is that if this technology is going to be super powerful, better to give it in everybody's hands versus allow one sort of unchecked actor to have it. And, you know, that sounds good when you say it as a line.

24:21But then when you think about it a little bit deeper like this, the fact that this technology, which is so powerful, can really be used by anyone. if you open source this, open source it, and you have really little recourse if someone goes against your rules, like that is, that seems like a potential vulnerability. And when I was speaking with Ahmed Aldela about the, I'm sorry, the metagenerative AI head about it, I said, well, are you going to open source this 400 billion parameter, right? This massive model that's going on, that they're going to build and release in summer, they're training it right now it wasn't a definitive yes and that i think that even shows this sort of discomfort that meta might have with what open sourcing all these models might have now i'm pro open source i think it's obviously good it'd be silly to be anti-open source but it also goes to the point where like where you're when you're spending these hundreds of millions of dollars maybe billions of dollars to train these models and they can do crazy things uh how open do you want to make that and And I think there are potential downsides, like potential things that can explode in a negative way if this stuff is so freely released.

25:32I talked to Mark about that, and I do think they're going to open source the big model. They have to go through safety evaluation once it's done training. So they just can't say that until in case they're I mean, he basically said, barring any like really unforeseen anomaly in the output of the model, it'll be open source. He wants to open source. They just can't definitively say it until they've evaluated it, which I think is the responsible way to approach this. I think it would actually be even scarier if they were saying we're going to open source any tech we make, no matter even if we don't know what the model and state is going to be.

26:10But he did say image generation, for example. They made some big leaps in image generation with the new assistant in Llama 3. and he was saying, you know, different modalities of these models when you get out of text, so video output or text output, we may not open source those. And he specifically called out, you know, that it's an election year and they were concerned about the image generation being kind of just out there and freely available for developers to use in potentially, you know, more nefarious ways. So they may open source part of these models, but not the multimodality, which I think is interesting.

26:49But yeah, he's not like he wants to be out there as the open source leader, but he doesn't want to be dogmatic about it, which I think is actually a nuanced position that I don't know. I'd rather him be thinking about it that way than being like dogmatic about it and kind of open source no matter what. Yep. And it's been interesting, like the financial impact on the company. The stock jumped a bunch yesterday when they released the model and today down three and a half percent. So we'll see. I think just to recap, looks like this model definitely puts them in conversation as one of the providers of the best models out there.

27:27I mean, we'll see what happens when GPT-5 comes out. You're right, there might be this race. They get the distribution in the product. It's good for recruiting. And again, it just kind of shows that it's got to be productized. If it's productized well, it's valuable. If it's not productized well you're not you're not left with much yeah i mean i think you nailed it i remains to be seen if they will if they'll make it a good product that's kind of what they have to do now and if they don't this is a massive overreach in terms of what they thought they were there is there like reality lab spending going from vr to ai i don't think it's very expensive i don't think it's that simple um um we actually i talked about when he bought all the gpus he did so they're i think maybe the first or second largest customer for nvidia i think there may be tied with microsoft 650 000 gpu equivalents by the end of this year so he bought all of those like at the very end of i think 2022 so before gpt4 even came out and this was when meta stock was at like 90 it had bottomed out and he was placing this massive multi-billion dollar gpu order and he said it was actually for reels um for doing recommendation videos for reels they were crunched on that and they needed more compute.

28:41And he was like, but I just doubled the order because I didn't want to be in this position again, where there was a big new thing that we needed to build for him. We didn't have the compute we needed for it. That turned out to be pretty smart. Turned out to be pretty smart. Something that really only like a company at that scale with like founder control, founder control could do. I can imagine at Google that getting shot down by like the CFO. Um, but he made the call, they did it and yeah, I, now they have all this compute, um, that they're using for llama three and four and five. Um, and it was prescient to, to, to, to kind of get that compute because now it's very hard to get H100s.

29:22Um, and so, yeah, I don't know. They, they're definitely well resourced. Um, they just, they need to be seen as the leading place for the best talent in the industry to come. And I think that's part of why you see him so out there doing all these interviews with me and others is, I mean, he was open about this with me, you know, that the best people want to work on the biggest problems with the most impact, right? And so he wants Meta to be that place. Yep. And so you mentioned the Gemini thing, and obviously Google has been working behind the scenes to try to clean that up, clean up the structure, right?

29:55They promised structural changes and those structural changes came this week. and there was a number of things that happened. First of all, Sundar Pajai had this like pretty remarkable memo that he sent out to Google and they published it, which means they want the world to see it. It's called Building for Our AI Future. And one of the sort of least noted things that I think is really important is they're moving this responsible AI team and research to Google DeepMind. And Sundar says it's to be closer to where the models are built and scaled. Now, I don't know this for sure, but if I was to take a guess, I think this group had a lot to do with some of like the safety things that were placed into Gemini after the, after it was built within Google DeepMind, at least that's what I've heard and sort of led to this, you know, embarrassing moment for Google where this thing was just like getting everything historically wrong in terms of its image generation.