{
  "video_id": "vCP7C8CVw28",
  "channel_slug": "ibmtechnology",
  "channel_handle": "IBM Technology",
  "title": "Apple’s new CEO & how AI understands intent",
  "duration_seconds": 2423,
  "url": "https://www.youtube.com/watch?v=vCP7C8CVw28",
  "upload_date": "20260424",
  "transcript": "Apple really has this AI moat working for them. So I think Apple\nhas a competitive advantage here, but I'm really curious to see if they're only going to, you know,\nstay in that hardware lane. All that and more on today's Mixture of Experts.\nI'm Tim Hwang and welcome to another episode of Mixture of Experts. Each week, MoE brings together a\ngroup of the nicest, smartest people working at the frontiers of artificial intelligence to lead\nyou through the week's news. On this week's episode, we've got Sandhya Iyer, Associate Partner,\nAI and analytics, Bri Kopecki, AI Engineer, and Kush Varshney, IBM Fellow. Welcome to you all. Thanks\nfor coming on the show. We've got four big stories today that we're going to cover another gigantic\ninvestment deal coming out of Amazon. We're going to talk a little bit of some interesting results\ncoming out of the Adobe Summit this week. And we'll talk about Mythos, the Anthropic model\nleaking. But I actually want to start with perhaps one of the big\nleadership changes of the week. News has come out that Tim Cook, after leading Apple for many years\nnow, will be stepping down and retiring, with John Ternus coming up and leading the company.\nAnd his background is really in the hardware part of the business. He was involved in the launch of\niPad and AirPods and many generations of iPhone, Mac and the Apple Watch. And so\nI want to kind of take this discussion because obviously there's been a bunch of speculation\nabout, like, Tim Cook's legacy and with, you know, Steve Jobs be a supporter of this. But obviously\nat MoE to spend a lot of time thinking about the AI aspect of all this. And so maybe Kush, I'll\nkick it over to you first. How do you read, uh, Ternus' kind of ascension to the CEO\nof Apple? You know, is this is this ultimately a kind of bull signal for Apple's continued push\ninto AI, or is it almost kind of like a step away from it? Maybe. Yeah. I mean, I think at the end of\nthe day, Apple is a hardware company, and it makes a lot of sense that they chose their their\nhardware guy to, uh, to take the lead. But, um, as I was going through his, um, background, his,\num, sort of accomplishments. One thing really stuck out to me, which is, um, his focus on kind of, uh,\nreliability, durability, repairability, accessibility, these sort of things. And, um, I think,\nuh, look, if he can bring that to AI, I mean, like, I think that would be really great because, um, I\nmean, Steve Jobs obviously said, uh, computing being a bicycle for the mind. And, um, I think,\nuh, one interesting thing about bicycles is that, like, they're very accessible. They're very. Great. I\nwant the wheels popping off. Yeah, they're very repairable to, like, if the wheels do pop off, you\ncan go in anywhere. You can get it fixed. Right. So, um, uh, hopefully that that ethos comes through and,\nhopefully in the AI side of it as well. And that's actually something I did want to touch on a\nlittle bit is Sandhya, I think we've talked a little bit about this in the context of sort of\nAI being a discipline, which is often like guess and check, and these aren't like perfect systems\nthat you can get to run in exactly the way you want hardware to run. Um, and, you know, there was an\nargument that I think someone was making on MoE a little while back that, you know, this is really\nlimited Apple, because they're so sort of conservative in some ways. Um, and I guess kind of\nquestion for you is like, is that the wrong way of thinking about AI? Like, are we now kind of\napproaching the point where it is going to be sort of reliable and like a hardware CEO like\nTernus is going to be able to, like, really fully get the benefit of this technology, or is it just\nultimately still kind of like a conflict between, you know, almost ways of doing engineering work? I\nthink in this context, I mean, especially in Apple's context, I getting or\nmaybe I'm saying this wrong, but letting AI be rogue doesn't make sense and I\nwould expect someone like, as what Kush just mentioned, someone like John Ternus to\nactually come with that specific idea of, okay, reliability, stability and all of that.\nNow the the interesting part here is we don't have I mean, I know Apple has\nan AI. Um, AI is a discipline under software, but it doesn't have AI as the front runner or\nit's not AI first yet, at least in terms of the leadership. I personally am like looking forward\nto see how they approach that, because then having like a two in a box, um, would be interesting.\nAnd that's, that's the that's probably the right move for Apple. Um, yeah, that's my take on it, Bri.\nOne of the things that. Also kind of comes to mind is we've talked a lot about kind of Apple, like\nreally wanting to control every aspect of the experience. Um, but, you know, Apple has also been,\nyou know, one of the most prolific companies that, for example, outsources a lot of their hardware\noperations and their manufacturing. They've also and in the AI space, you know, kind of very flatly\ndecided to partner with other big AI companies to do this sort of stuff. Um, and. So I guess. Kind of\nlike, um, maybe like the the main question I have for you is, I think ultimately, do you think, like a\nlot of AI is just going to be, you know, part of it's almost like an app store paradigm for, for\nApple that they're going to own the hardware gate. But and at the end of the day, they sort of see\nall these models as kind of like ultimately sort of fungible. Or do you think kind of like Apple's\neventually going to say, like, hey, we're going to do really just like specifically Apple AI? Yeah,\nthat's a that's a great question. And to be honest, I've been really underwhelmed with Apple's\napproach to AI over the last years. You know, we have companies like, I mean, every other major\ncompany that's been a game player, you know, investing heavily in AI. And then we have all the\nnewer companies like OpenAI or Anthropic really, really dominating. So I've been really\nunderwhelmed. But I think one thing we need to see very clearly is that Apple understands that there\nis short, there's a software part, and we have lots of things running in the cloud when it comes to\nAI, but Apple really has this AI moat working for them. So I\nthink Ternus really understands John Ternus really understands that this is an advantage to\nthem. And they're going to use that moving forward. Because you mentioned the Apple Store, right. And\nrunning different applications that run LLMs locally. What that requires is massive\ncompute. And this massive compute creates heat. So you're going to need some technology\nthat handles that heat creation or that's running in the palm of your hand essentially. So using 3D\nprinted titanium, which we all know is running in the Apple Watch, that allows you to act\nas a heatsink for those computers. So I think Apple has\na competitive advantage here, but I'm really curious to see if they're only going to, you know,\nstay in that hardware lane or also branch out and do bigger things, which I think they should do and,\nand will do. Yeah, I think the ambition here, I think, is one of the interesting things that I did\nwant to touch a little bit on. I mean, of course you're actually a perfect person to ask about\nthis. You know, one of the things I got really excited about when Apple started to really kind\nof turn the gears on, sort of AI, was the idea that they were going to achieve the dream of kind of\nlike distributed private AI, right? They had made privacy such a big part of, like their sales pitch.\nAnd so the idea that all the AI could happen, kind of like on device with very little going to the\ncloud was like genuinely pretty exciting architecture. You know, that dream seems to have\nfallen a little bit by the wayside. I mean, what's your assessment? You know, especially with Tim Cook\nstepping down now is, you know, is that still a technical feasibility? Like, is that a dream that\nwe should believe in, or is an apple kind of going to go the way of everybody else? And a lot of it's\njust going to be cloud? Yeah, I think I mean, this trend has been the case for for quite a while,\nwhich is that if you just wait a few months, um, the capabilities of larger models, um, are\npossible with smaller models and, um, I don't see that kind of trend slowing down. So, um, uh, yeah.\nI mean, just an 8 billion parameter model can do, I mean, stuff in like, in six months, we'll be able to\ndo things that, uh, let's say 120 billion parameter model is doing now. So I think, um, the edge\nstory is going to be there. And, um, I think coming back to this app store idea, I think it's also\ngoing to be models with adapters, um, to get them to be like more capable on specific things. And\nyou can have app stores of adapters, app stores of skills, um, these sort of things. Um, and uh,\nhopefully, I mean, Apple goes in that direction. I think it would make a lot of sense. So, Sandhya. What\ndo you think? Do you think? Kind of like, in some ways, it feels like a lot of this code gen stuff\nis going to lead to a world where it's, you know, the App Store. Like, does it even make sense to\nhave apps anymore in a world where you just tell your phone what you want and it can do it on the\nfly? Um, you know, I guess because we think about kind of the future of Apple. Do you think that's\nalso under threat in the current sort of AI era? Oh, yeah. I'm gonna look, I'm already creating my\nown apps, right? That's right. Why do I need to go to the App Store? Right. With all the open cloud\ncode and all of that being there. Um, I think we spoke about this weeks back when this was a topic.\nSo. Yeah, I think that's also true. The App Store is no longer. It's not going to be like the\ndifferentiating factor. Well, we'll keep an eye on it. Um, obviously John Ternus is going to have to\nprove to us here at MoE that he can do the job, but we'll be sure to talk to him going forward.\nGreat. I'm gonna move us on to our next topic. Really interesting A story. Another huge monster\ninvestment announced by Amazon this week. They stated on Monday that they'd be investing an\nadditional $5 billion in Anthropic, with kind of sort of milestones that Anthropic can hit that\nwould raise that investment to $25 billion. And a big part of this, which we've been watching for\nsome time, is this interesting relationship where Anthropic, like every other lab in the space, is\nextremely compute constrained. And so part of this deal is that Anthropic literally turns around\nand buys lots and lots of cloud services from Amazon, and is part of a dynamic that we've seen.\nYou know, NVIDIA will invest in a company. That company brilliantly turns around and buys NVIDIA\nchips. Um, I guess, Bri, what's your take on this kind of increasingly tight relationship, I guess,\nbetween Amazon and Anthropic and I guess specifically like in between the cloud business\nand what Anthropic is, is doing. Right, yeah. I don't think it comes as a surprise. Um. Anthropic\nis essentially choking on their own success, which is a nice problem to have, right? But, uh,\nit it, um, forces you to be to look for resources and partnerships in the market that can allow you\nto, to continue to soar as they have been. So this thing here, I mean, it's to me, it's kind of like a\nquid pro quo, uh, you know, unfolding in front of our eyes, you know, uh, Anthropic investing,\nover $100 billion in cloud services, uh, from from AWS. Um, it's a massive\ndeal. Um, but it also we have to look at the underlying technology that, um, uh,\nthat Amazon is providing here, which, uh, we already know, NVIDIA has been, you know, providing\nGPUs, which are great for general, uh, computations. Uh, the Trainium 2 chip\nuses ASIC, which stands for, uh, application specific integrated integrated circuit, and it\njust allows for more specialized computations. And along the way you can save a lot of\nresources, which if we're looking at the workloads and the the compute demand that Anthropic is\nfacing, um, you're going to be making a lot of savings, uh, along the way. So it's not\nsurprising, um, interested to see how that partnership will continue to unfold, though. Yeah.\nFor sure. Yeah. We should get into the compute aspects of this, which I think are genuinely\nreally interesting. Um, the ASICs point that Bri, you just made, I think is, is fascinating because, you\nknow, GPUs in some ways are great, but they're very general. ASICs, you get a lot of performance, but\nthey're very specific. Um, and I kind of look at that and I say, okay, well, Anthropic is obviously\ngoing to immediately become one of Amazon's biggest customers in this front. And so is this a\nworld where Kush, like, almost like these clouds are going to become a lot more almost model\nprovider specific over time. Like you can imagine Amazon having like an Anthropic-specific ASICs\nwhich like make it work at just like a way higher level of efficiency otherwise. The downside, I\nguess, is that there's a little bit of lock in, right? That like it's hard for Anthropic to move.\nAnd it's also hard for Amazon to move from that relationship. Yeah. Um, so it's a great\nquestion, right. I think, um, when you had, uh, I mean, people selling shovels to gold miners, the shovels\ncould be used by any of them, right? But the like, the basic, as you're saying, I mean, it does. Then\nstart getting that, uh, that lock in the match to, um, uh, to to. Yeah. I mean, the workloads, the,\nthe Cuda, I mean, like all of that sort of stuff. Right. Um, so, uh, I think maybe\nthere will be some lock in. But then what has also happened in the history of computing is there's\nbeen these layers that translate trade. Um, so whether it's compilers or like any sort of other\nthing that kind of like makes the that job easier of going between the hardware and the software. Um,\nuh, so yeah, I don't know. Um, we'll see. But, uh, clearly it's part of the strategy for, for\nAmazon. They do want to get this sort of lock in going. Yeah. For sure. Um, so I guess\nmaybe the question I have for you is where who has the power in this relationship? Um, because I\nhit a little bit of the the news story here, right. Which is, on one hand, Amazon says we're going to\ngive you $5 billion. We might even give you as much as $25 billion. And I look at that and, you\nknow, your eyes pop out and you say, well, what a huge number. And you read a little further down, it\nsays, uh. Anthropic also agreed to purchase more than $100 billion worth of Amazon cloud services.\nAnd, you know, I'm no financial expert, but you see 25 billion going one way and 100 billion going\nthe other way. Who's who's in control? Is this a cloud story or is this like an AI story in some\nsense. I think from Amazon's side it is about the fact that I have, um, I have these\nchips. It's it's them trying to like, say that I have these chips that, you know, this frontier\nmodel is being powered by. So you don't it's my statement against NVIDIA or trying to\nplay the chip game. Um, so I think it's that and from Anthropic and it is about\nsaying that, okay, I am committed to the Enterprise Award because it's it is all about reliability\nand stability for them. Capacity for them. Um, so it's a two way thing and saying that. Yeah.\nSo I believe it's a two way victory path. It's not cyclical. But yeah. I\nthink you got a good point there. I think uh, the numbers though speak for themselves. I think uh,\nAmazon definitely has the bigger hand right now for the time being because if Anthropic doesn't\nfind the infrastructure and the technological, um, capability of of its\nmassive models. And I know we're going to talk about Mythos here in a second as well. Um, so\nAmazon has the upper hand here, and Anthropic is, uh, uh, lucky to, to\npartner with them for the time being. Um, they are yeah. They depend on that infrastructure and\nAmazon is providing it. So, um, I think it's it's clear that they have to dance the\ndance with them. Yeah. For sure. It could change in the future. It could change. But for yeah, those\nnumbers speak volumes to me. Yeah. And that's actually you're touching on briefly. Exactly the\nkind of last item I wanted to talk about here was, uh, you know, kind of feels like, uh, it's like Game\nof Thrones, right? Where, like, it kind of feels like there's this alliance now between Anthropic and\nAmazon. And of course, on the open AI side, it's like open AI and Microsoft like, it does kind of\nfeel like there's these kind of like stacks forming where, you know, these very tight\nrelationships between cloud and model provider and then even like hardware. And then I think\nincreasingly maybe even energy like it kind of feels like you'll almost have these kind of two\nindependently operating ecosystems in some ways with all these alliances forming. Is that the\nright way to think about where this market is going, or do you think it's going to be a little\nbit more promiscuous, like you said? Like maybe things change and and Databricks one day going to be\nlike, actually, we want to go do more stuff with Azure. And it's actually a lot. It's not as locked\nin as I'm making it sound. Yeah. Um, yeah. I think you're on to something here. I think we're\nwatching the the founding fathers of AI. Right. So just kind of like how America was formed\nand you had the the original founding fathers that stepped forward and formed alliances and, you\nknow, conquered the land. And this is what we're seeing unfold in front of our eyes. Right? So I\nthink for the time being, we're going to see some massive collaborations like this, which I think we\nneed to be careful with the word collaboration because both need each other's is exactly what a\nSandhya said, but they also heavily depend on each other. I think once this, once the dust settles\nand the space of AI will be more democratized. And I'm really hoping we will have lots of\nthought leaders who are making sure that we democratize the space. Then we'll see more of that\ncollaboration that you mentioned. That's my my guess. Um, based on history and in,\num, in technology and all other spaces. That's great. Well, it brings up maybe like, a final item.\nUh, Sandhya, I think, um, you know, we're talking a little bit about how some of this hardware might\nbecome more specialized with time. Do you think that's going to be a threat to the open ecosystem?\nI think that's maybe like a final thing that kind of comes to mind is that like, well, you imagine a\nworld where the chips are really designed for whatever specific models OpenAI is doing, or the\nreally specific models Anthropic is doing. Does it mean that, like, in some ways, open will become more\ncompute constrained because it's like not going to be necessarily specialized for these, like\nchips, which are kind of like really kind of going into proprietary models. What I actually read\nabout, especially the Anthropic and Amazon one, is that Amazon will tightly couple this or design\ntheir chips specifically for Anthropic or for cloud and that. So is that\nthat actually goes away from the open ecosystem. Comment. And but in a way, I'm I\nguess I'm not answering this question, but I'm tying it up to the previous question because I\nfeel because I feel that, um, that's good for Anthropic. I mean, it's it's a\ntight partnership. Now, now it's just that they coexist and, uh, and\nAnthropic gains all the optimizations in the cost effectiveness that this, you know, partnership\nleads to, or this tight coupling of the design chip design specific to them is good for\nthem in terms of everything. Yeah. Um, so I guess in in what I just said, we're going\naway from open ecosystem. Um, and of course, you could think maybe. Yeah.\nI mean, you know, my general perspective about. Oh, I know, yeah. Bad thing. Right. Um, bad thing. Right. Yeah.\nBut, um, I mean, we definitely need to support the open ecosystem, but I think there isn't\nthat much difference in the model architectures, right, when you're looking across the board. So, um,\nlike, how much can you really specialize? I really don't know. Um, so that's another sort of point on\nthis. Well, great. I'm going to move this on to our next story of the\nday. Um, Adobe Summit was just the past just happened recently. This is kind of like the big\nsummit that Adobe holds where they do a lot of their product announcements. Um, and it was also\nkind of a cool opportunity. IBM Institute for Business Value. IBV released the report in\ncollaboration with Adobe, specifically looking at dynamics on customer intent and customer loyalty.\nI want to dive into those numbers in just a second. But Sandhya, you were actually there and so\ncurious about your impressions about the Adobe Summit and anything you were excited about kind\nof trends that you think our listeners should pay attention to coming out of that event? Sure. I\nthink like every other AI summit that we've had, this story was all about orchestration. It was all\nabout the AI infrastructure, was not about Adobe's next Firefly model or something\nlike that. It was all about the ecosystem that supports agentic AI. And their\nannouncer announcements were all screaming about that. And I'll call out some specific things that\nI was specific. I was personally excited about the Adobe Enterprise, um,\nPlatform that they released. It's pretty interesting in a way that it is designed to, uh,\nsupport the infrastructure orchestration that's needed for end to end marketing AI\nworkflows, which is very cool. And it is something that IBM is also currently working\non. And it was very nice to see that. That's the and we've been working on that within\nourselves and also with our enterprise clients. And so it was nice to see it was obvious that\nAdobe would come up with that design, which supports all of it. And it was also nice to\nsee, uh, Jensen Huang, who shared a very friendly keynote. Uh, I'm not sure that anybody\nsaw it. If if you've not seen it, it was worth watching. Um, and the other interesting\nthing for me was, uh, again, GenStudio, um, this is because I am very close to content supply chain\nworkflows. And this is also and I had the opportunity to work on Adobe\nGenStudio, but I think the new one is, um, it has a lot more features from ensuring that your end-to-\nend creative workloads are like managed orchestrated well across. Which is great. I mean,\nI'm really excited about what all of us do with it. Uh, the I mean, I actually wrote a blog\npost about this saying the technology is ready to serve. It all depends on how we put put together\nall of this. And there's a lot of organization like governance and organizational change\nmanagement and bringing in all the right data into it. Again, I'm passionate about this. I'll keep\ngoing on and on. But that's it. Yeah, that's my five minutes spiel on it. That's\nreally good. And yeah I think I don't know. We don't talk about Adobe very much on the show, but\nI actually think in some ways they kind of are like a sleeping giant in this space in the same\nway that something like Microsoft Office is like a sleeping giant in the space, right? It is the\nsoftware that's like, so integrated into the workflow of so many businesses that the ability\nto like flip a switch here and there, in terms of how people work has like these huge ripple\neffects through the economy. Um, and it sounds like they're investing a lot in like, how do you\nactually get like, you know, the really basic stuff of like, how do you get AI to make like, the\ncontent pipelines much more efficient? Um, what do you think the, the next bit of work is? I mean, you\nsuggested a little bit like, it feels like the next bit is change management. Like it feels like\nthat's going to be where a lot of the next leg of work is to, like, even get the benefit of these\ntools. Um, do you think that's kind of where we're going in this space? No, I think the first thing is\nlike making sure you have the right data. The data work is the main thing. And then\nonce you have the right data coming in, you've figured out all your workflows. It's like your\nprocess transformation. I think it's it's a clear process transformation effort. And that's where\nyou begin with and then going into the change management governance aspect of it, which is which\nis a huge orchestration that's needed by itself. So yeah. And uh,\nthe I just wanted to like call out, uh, things around the fact that, um, visibility was like the\nkey theme of that of Adobe Summit. And I'm saying this from\nvisibility, not just from a brand perspective, but visibility from what is happening within your\nworkflows. So I think that's what it is. Yeah. Well, it feels like it can become a real shared problem,\nparticularly as all this agent stuff picks up. It'll just be like, do I know exactly what my\ncommunity is doing? Yeah. Well, let's dive a little bit into. The numbers on this, uh, IBM\nIBV study. Um, Bri, I guess I'll start with kind of like the top line number, right? So I think the\nstudy took a look and said, okay, how are people using sort of AI assistance in making up their\nmind about what to purchase and the basic finding. In some ways I was like, yep, that that squares\nwith my understanding, right? Which is they find that the use of this for purchasing decisions\ngoes up by like 62%, you know, in the last two years. Totally makes sense. I think the next little\nbit I found really interesting, because the way we normally tell this story is like, oh, young kids\nuse AI, they're doing crazy stuff with like TikTok. And that's where all the interesting purchasing\ndecisions are happening. But what they find is what, like the boost among Gen X and boomers, is\nreally where a lot of the action is, it seems like. And that actually there is this kind of like\ngenerational slant where older users are using this technology more for purchasing decisions. So\nhelp me think through that is very counterintuitive in my mind, because I'm kind of\njust like, I don't know, my parents would never use this stuff for purchasing, so. Well, my mind you. Aha.\nYeah. And actually actually your mileage may vary. No. Every everyone's everyone's going to have a\nlittle bit of a different personal experience here. but I thought that was really interesting as\nwell. So we have a an adoption of this technology. So interacting with website, making\npurchasing decisions, um, in with Gen X I believe was an 82% adoption\nand boomers even 92%. And you mentioned her parents. Um, actually, my dad is a\nhe's incredible when it comes to AI because he has a very low tolerance. And I think a lot of\npeople of that generation, um, he has a very low tolerance of dealing with, um,\nuseless, uh, hostile environments online. Um, so there's all this AI that you can use\nto, to interact with. And another thing that they have is they have a low tolerance of dealing with\nthat, and they have disposable income, which Gen Z does not have. Okay. So if you have money and\nyou're sick and tired of navigating hostile environments, yeah, you're going to make use. You're\ngoing to tell an AI agent. Hey, this is the thing I want. Get it for me. And they're going to choose\nthe path of least, at least resistance. And that's the thing companies need to understand. Um. And I\ndo this even less than my dad does. So I am reminded every day that this is very\nmuch the trend. And I think Gen Z will eventually, um, catch on. But that's something that, uh,\nyou know, companies nowadays need to be aware of and need to understand how the intent, where the\nintent is coming from, from the consumer base and how to give them what they want. So it's no longer,\nyou know, people going on the website and searching for the best product. And, you know, you\ncan catch them with a nice ad campaign. Um, it's very much agents talking to agents, and that\nchanges the game. Yeah. Kush and I think this is one thing I was curious to get your opinion on is,\nyou know, when you used to hire an ad agency, the ad agency would be like, how do I advertise to\npeople who are going to buy the product? But now it's like, you know, particularly with where MCP is\ngoing. And like the design of websites that are specifically for AIs to access, I'm kind of like.\nIs the future of advertising just convincing agents to, you know, buy your product and like, does\nthat totally change the game in terms of like messaging and what you say because you're not\nadvertising to the AI that will advise the human versus the human itself. It seems like. Yeah. Um,\nthere was a paper by Microsoft Research last year about the agent economy, and it was very well done,\nactually. Like the first time I read it, it was like quite eye opening. And it's exactly talking\nabout all those sort of things. It's like agents advertising to other agents, ad agencies,\nadvertising to agents, optimizing instead of search engine optimization, optimizing for\ndiscovery. And um, kind of yeah. What's the metadata that you need to have up there? And I think what\nBri said is exactly right. I mean, you don't like these agents aren't going to tolerate, like, all\nthe fluff, right? I mean, just give me the numbers. Like, um, uh, I was trying to figure out our next\ncar purchase. Right. And, um. Uh, look, I don't need to see all of, like, the, the, the ram\nad with, like, all the trucks and the mud and stuff, like, just, um, tell me the numbers and, um, I mean,\nthat's like the the thing, right? So, um, I think maybe that that is where we're going. And, um, uh,\nthe other thing is that LLMs are good for understanding intent, right? I mean, that's the kind\nof starting foundation for the study and, um, uh, the computing, uh, again, like the history of\ncomputing has this new element which, um, uh, the LLM, which is, like, perfectly suited for, for\nunderstanding this thing, um, tapping into, um, kind of like the, the mind perception and so forth. So,\nyeah, I mean, look, this is exactly the evolution that I would expect. If I may add, um, just based on\nthe research and things of that that are happening in this space. It. Is actually a combine.\nIt's a combination of having a lot more creative content across platforms. Um, making sure that you\nhave information about your brand in Reddit and all the aggregator sites because like the AI\nsearch engine is actually going against. How many mentions do you have credible mentions of your\ncontent across all of these platforms? Do you have YouTube content, etc.? And also, do you have\nstructured metadata that I can, you know, read and understand? So it's a combination. That word brand\nis so interesting. I mean, I only briefly worked in the ad industry and it feels like the usual thing\nwas like, oh, well, on one hand you have got brand advertising, which is like the big glossy piece,\nlike, you know, you know, page that you buy in Vogue magazine or something like that, and then you've\ngot programmatic advertising, which is like very data driven and banner ads and that kind of stuff.\nI mean, does the idea of brand advertising to agents even make sense anymore? Because normally\nthe idea of brand advertising is like, oh, like, uh, Nike just do it right. It's like it's trying to\nkind of tap into something like cultural and very subjective and very normative. You know, I don't\nknow our agents susceptible to that kind of persuasion. I think it's like one of the\ninteresting questions here. Um, I don't know, Kush. I see you nodding a little bit. Curious about what\nyou think about that. Yeah, I think, um, it could happen, but I think it's in terms of, like, how the\nmetrics are set up for the agents. Right? So the agent designers are setting up, um, kind of what\ncounts as success. And, uh, I think, uh, I mean, click through rate used to be a metric. Now it's maybe\nthe completion of the actions. Um, because if the autonomous agent isn't, um, actually, like, finishing\nthe job, then it's not succeeding. So maybe, um, whatever side effects or whatever, like latent\nvariables, it picks up on that eventually. Let it complete the job. That's what's going to going to.\nSo a final element. Of the study I wanted to touch on and maybe I'll kick this question over to you,\nis the study also looks at um basically like how quickly various industries can capture customer\nintent, which I think is a fun question, right? Is like, once you get a signal that someone wants to\nbuy something, how quickly do you like take advantage of that? And they do this kind of fun\nstudy which benchmarks across a bunch of industries. And I guess, like maybe building on\nwhat Kush just said, you know, the the word intent is really interesting here because it almost\nsuggests a world where, like, you know, I'm browsing online, looking to, you know, buy a sandwich. Um, I,\nyou know, click on, you know, a website that's like a sandwich review website. And the question is, who\nserves me the ad quickest so they can take advantage of that. But in a world of agents. Is the\nidea of. Intent kind of disappearing or dissolving, like normally they're it's like, oh man,\nTim wants a sandwich. I got to sell him the sandwich right now. But in a world with agents,\ndoesn't the agent just kind of have persistent intent? Like, it's just trying to get me the\nsandwich. And so, I guess kind of question for you is just thinking a little bit through how, like,\neven intent changes in a world where AIs are kind of like just trying to do something on my\nbehalf versus me literally being the one that goes around the internet kind of clicking and\ntrying to find the product that I want. Yeah. Um. That's a good question. So I think, uh, companies\nand brands, uh, the brand relationship between the final consumer is going to change because of\nagents. Um, you're no longer catering or trying to get to the eyeballs, the human eyeballs that are\nlooking at the product. Um, you're trying to plug your, your, your brand, your offering,\nusing APIs into the agents. Um, I'm probably not\nreally answering your question here, but I think, uh, that relationship around, uh, intent\ndelegation, um, is is is going to change. Um, I think it's going to\nit's going to be simplified a lot more in terms of, okay, so this person is interacting with this\nagent, uh, that's, you know, plugging into the brand and in the offering of the brand. Um, so it's\nsimplified in terms of what does the user want? What can I get it less of? If a, if a human just\nscans through a website and is trying to establish an understanding of the brand. Kush\nmentioned shopping for a car. It's more of a feel-based. So I think eventually it might become\na little bit more touchy feely and there's going to be emotions worked into because brand is\nemotion, branding is how do we feel about something? Do agents feel anything at this point? I\nthink they feel like they they look for intent. Uh, they are, uh, uh, designed to deal with\nintent. But, um, yeah, I, I guess I don't have a definite answer for you. Yeah, totally. The\nfuture you're describing is, is very interesting in some ways, is like the agent itself needs to\nhave a sense of how long my intent lasts for. So it'd be like, Tim always gets really hungry, but\nthen, you know, he gets distracted really easily. So, you know, if I'm an agent trying to buy him lunch,\nyou know, if it's lasting more than half an hour, I should just not not, you know, deal with it. And,\nlike, you know, the agent is going to have to model that in some sense, which I think will be really\ninteresting. Right? And sometimes it might just be, uh, the hunger for information and not necessarily\na purchase. So understanding those nuances, um, and, you know, we have some smart brains, uh,\nworking on those agents. I think they'll, they'll understand those nuances and they'll\ninteract successfully with the end user.\nAll right. Final story of the day. I'll keep this a little last a little bit quick, but I think I\nwould be remiss not to mention it. In some ways it is the the funniest AI story of the week. Um, so I\nthink just an episode or two ago we were talking about Mythos, you know, the rumored next model for\nAnthropic, um, Anthropic later made an announcement saying, yes, indeed, They're working on a model\ncalled Mythos, and it is indeed so powerful and so scary with regards to what it can do in\ncybersecurity that we're actually not going to release it at all. We're only going to allow a few\ncompanies to take a look at it, and otherwise it's going to be very under control. Um, there's a story\nthat came out of The Verge and kind of leaked around that it turns out that a relatively small\ngroup of people allegedly hanging out in what must be a discord somewhere, sort of just reverse\nengineered access to Mythos and had just access to this model for a period of time. Um, so of\ncourse, help me resolve this split screen. On one hand, you have a model that seems so dangerous\nthat you can't release it. On the other hand, the security around it seems so soft that, um, you know,\na bunch of nerds hanging out a discord can seem to get access to it. So what are we seeing here? Is\nthis is this just normal kind of cybersecurity failure? Is it, you know, a proof that Mythos is all\nrhetoric? Like, what should I make of this? Yeah, I think, uh, as humanity, I mean, we're all about\nstories, right? Oral narratives and so forth. And even the name Mythos implies it needs, needs some\nmythology associated with it. Right? And, um, I mean, there's always this, um, like role in, uh,\nmythology from all over the world, right? Um, there's some trickster who goes and steals fire\nfrom the gods and gives it to humanity and stuff, whether it's Prometheus or, um, uh, in the\nGreek world. Um, Maui, in the Polynesian world, whatever it is. Right. And, um, this is that story,\nright? And I think it's just building up that mythology and, um, uh, I mean, in those stories,\neventually what happens is, like, humanity loses its innocence. And like, I mean, these sort of\nthings. Right? So maybe that's what this is indicating. Um, so, like, I don't want to go into,\nlike, the security failures and stuff. I think the bigger story is the story part of it, the\nstorytelling. And we'll see how that plays out. Bri, any predictions on whether or not this makes\nit more likely or less likely that they will let all of us get the benefit of Mythos soon. Not too\nsoon, I think. We now know. I think we know a project Glasswing well the very same day this\nleak happened. Uh, you know. Anthropic also released the model to the, uh, the many\nmass, you know, giant companies that, uh, were collaborating under Project Glasswing, uh, with them.\nSo I think there's there's a strategic collaboration happening. Um,\nI don't think this is meant for the public just yet, which is ironic because the public very much\ngained access, um, through just a group of, you know, nerds on a discord group, uh, who were essentially\nthey were just they went they went around, there was a massive, uh, data breach, and then they were\njust rattling digital doorknobs until one opened and boom, uh, it opened. Um, so it's\nironic, and I think it's also just, uh, It reminds me, and it should remind all of us how\nincredibly messy technology actually is. You know, as much as, you know, code is very structured. And\nat the end, if there's a small syntax error, your program will not work. Um, so we think of\ntechnology being extremely, um, clean and and well structured and\nextremely organized, but it's actually a very messy space. I like the philosophical approach\nthough. You're taking Kush. Um. Uh, one has to do that sometimes because there's just so much\nhappening. That's great. Well, words to live by. Uh, Kush. Uh, Sandhya Bri, thanks for joining us on the\nshow. And thanks to all you listeners. If you enjoyed what you heard, you can get us on Apple\nPodcasts, Spotify and podcast platforms everywhere, and we'll see you all next week on Mixture of\nExperts.",
  "transcript_chars": 39749,
  "ingested_at": "2026-05-21T19:10:56.703765+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 3702,
    "like_count": 85,
    "channel_id": "UCKWaEZ-_VweaEx1j62do_vQ"
  }
}