IoT Coffee Talk
Welcome to IoT Coffee Talk where we talk tech and all things related to the Internet of Things. This program was started by a group of friends who just happen to be thought leaders, experienced practitioners and business leaders in the technology and communications services industries and more. This pod will feature weekly recordings of our casual conversations on the tech topics of the day with the hopes of delivering engaging and thought-provoking insights to our audience. We hope you enjoy program.
IoT Coffee Talk
#311 The Next Generation Turing Test
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Nice little run there.
SPEAKER_00I thought you were heading down the Jimi Hendrix path fully there for a moment.
SPEAKER_06That was He was getting there. A little Pacific Northwest uh speaking of Tip of the Cap.
SPEAKER_01Making coffee. I was inspired by you guys.
SPEAKER_06That's right. It was a subliminal, subconscious thing.
SPEAKER_02Exactly.
SPEAKER_01Yeah. Jimmy. Do people know that Jimi Hendrix grew up in Seattle? Seattle. Everyone in Seattle goes.
SPEAKER_03Yeah, everyone in Seattle based on Favorite Son.
SPEAKER_06Actually, it used to be uh at the Mopop uh museum in downtown Seattle. It used to be a really good, I don't know if it's still there, but really good Hendrix um kind of whole exhibit.
SPEAKER_00They turned it into something else, didn't they?
SPEAKER_06Yeah, I don't know. I haven't been there in a while, but the Hendrix exhibit used to be. That was classic.
SPEAKER_01By the way, um we got a jam sometime, Pete. I know, yeah. I was playing yesterday.
SPEAKER_06I got some new gear yesterday.
SPEAKER_01I was like, Yeah, we should just be the friggin' entertainment at the next day.
SPEAKER_02Sure, yeah, that'd be good. Why not, right? Yeah, yeah. You know, I can join in.
SPEAKER_05There you go.
SPEAKER_02I love that. That's like a flying big ukulele.
SPEAKER_05There you go. Wow. Classic.
SPEAKER_02Can still be broken rolling big ukulele.
SPEAKER_05Sure, why not?
SPEAKER_01So hey everyone, welcome to IoT Cobby Talk. Remember, don't take us seriously. If you do, it's at your own risk, at your own peril. Uh, we highly recommend the insurance policy that you should take out if you do so uh because it's not a good idea. And uh yeah, it's another week, and this is gonna be episode 311. Can you believe that? That's pretty friggin' ridiculous, right? It keeps showing up. It's like yeah, and this is how we detox at the end of the week after being basically injected overdose on AI, AI, AI, AI, AI. Yeah, I know.
SPEAKER_00If you sound like the Mongolian throat singer, if you do that, if you do that fast fast enough, you go on LinkedIn these days, it's just like uh it's like a tsunami of stuff.
SPEAKER_01I mean AI slop.
SPEAKER_06Well, I don't know if it's slop, but some of it's slop, but yeah, it's just impossible to to keep up with this stuff.
SPEAKER_00I've had a whole I've had a whole quite amusing conversation this week with a with a clearly AI generated um or AI um fake recruiter who has been I mean I've been getting it was very clear it was fake to begin with. Like um it was like $2.2 billion IoT platform company looking for a new CEO, blah blah blah. I'm like in the Bay Area. I'm like, I know that there's not a $2.2 billion revenue IoT company in the Bay, you know, you failed at the first gate, but this thing's got it's it's probably I'm on the 15th email back and forth, I think. Because I'm like, let's see how far this thing can go. Wild. It is it's wild, actually. I mean it's it's oh yeah, yeah. It's purporting to be a member of a recruitment firm. And I contacted someone I know of the recruitment firm and said, just say you know, this is kind of like someone's someone's kind of doing this with your with your name. But um as an offshoot of what's happening on LinkedIn, Pete.
SPEAKER_06I mean, on LinkedIn, so much of it's well generally well, it's kind of ironic with all of the cape AI capabilities out there. You you're there's more spam and obvious, you know, spam more than ever. And you'd think so. They're using AI for the spam, but they're not using AI to counter the spam. Even though I'm paying LinkedIn like uh whatever, the premium thing. Yeah, to filter. There's a super premium non-spam.
SPEAKER_01Yeah, that would be great, actually. Uh you know, I would I would pay to filter out all the yeah, no, I would.
SPEAKER_06I would little boost privacy booster, whatever.
SPEAKER_01Yeah, if that's the feature of the subscription, uh I would I would pay. I mean, you know, think about like for instance, um, on the iPhone, they introduced uh the the call for screen code screener, yeah, a call screener. Love it. And actually screened out uh uh yesterday a call, which was obviously uh a spam uh phishing call where somebody was a per you know the bot was impersonating uh an agent from Amazon and was wanting to talk to me and uh obviously I think this was a voice, it left a voice message. That was the creepy thing, right? And um was asking me to uh call them back uh to validate that I had and verified that I had made a purchase of an iPhone. And I went to my Amazon account, there's nothing there, right? Right. And chances are though, whoever they called did buy an iPhone filled out. Uh well I'm glad these things are are filters that take away all this nonsense, the distractions out of our lives. And um uh definitely not divers value in that.
SPEAKER_06My favorite spam call I used to get uh noise. You know, you ever get you get the call from like Microsoft Tech Support, they call you to help you. So they called me once. I was at Microsoft, I was in the building, and they're calling me, and they're like, I'm calling from Microsoft Tech Support, whatever. I'm like, Oh, cool, what building are you in? You know, obviously I'm playing along with it, and we're just going on for like a half hour. I'm like, okay, I'm in 119, so where are you? Let me connect to you. Are you are you you know, I'm like, I'm trying to talk him through like, can you go to the the MS Web thing and we can connect? And where are you in the directory? And you know, uh yeah. So we we kept going for a while, then he hung up. So but yeah, I would I would get calls inside of Microsoft from the Microsoft tech support spam thing.
SPEAKER_00I thought well anyone in Microsoft knows that there is no tech support in Microsoft.
SPEAKER_06Well, yeah, it's it's uh it's a self-evident thing, right? What are you talking about? Microsoft support. It must be an impersonator.
SPEAKER_00I think LinkedIn, I mean, I I think that they're making a trade-off between uh content that generates clicks and therefore you know meets what I think is an increasingly fake measure of engagement and actually applying the filters that would remove this stuff because Chat GPT for one is extraordinarily effective at identifying when something's been written with AI. I mean, very I'll take emails I get from people and stick in ChatGPT to say what tell is was this AI generated? And it comes back with a broken down analysis of exactly why and what.
SPEAKER_01Um but I do uh but yeah, I mean I I wouldn't give Microsoft too much credit. I know that they sort of try, but you know that in Outlook, there when you get a fishing an obvious phishing email, especially the ones from DocuSign that are completely hijacked identities and just look really authentic, you know it's fake. You you report it as phishing, it comes back and it'll tell you, oh no, this is legit, no problem all the time, every single time. It it doesn't recognize uh phishing emails. Uh and uh I don't know. The I I just feel sorry for the world that doesn't know what's coming or is already here. It's really really bad. And I think people are just they're either ignorant or in just massive denial about right.
SPEAKER_06Well, I mean back in the day, remember when there were there weren't computer viruses and then they were, then there was the whole McAfee Norton industry of people that you you paid money to counter the so we'll have people paying money for AI screening of AI spam, you know. AI spam.
SPEAKER_00I think there's a whole space where AI will continue to try and get better at fooling human beings or understanding human beings. Where uh whereas I think more in the industrial spaces, someone said to me last week, they said, Well, I I don't know why people are spending so much try time trying to figure out how to get robots to robots to get along with human beings in a factory environment, because the gay the aim is to remove the human beings from that equation. So, you know, and it was it was a conversation about tactile technology to you know help teach robots to touch, etc. And and it was like his his view was like, well, the whole goal here is to you know for the humans to not be in that environment to be a lights off factory. So why are we spending all this time trying to make humanoid robots that are somehow more accessible to humans?
SPEAKER_01Who gives a best of you, but you know yeah, but but it's because the machines can't do what humans do, right? Um I think that's the big problem. And when you talk to folks who have been doing quote unquote physical AI for a long time or even robotics, humanoid robotics, the biggest problem is the tactile stuff.
SPEAKER_03Yeah.
SPEAKER_01Um, you know, and I think we mentioned this uh maybe a couple of episodes ago. But without the tactile, the vision is limited, right? It's essential, or actually maybe it's not even essential if you think about it, but the tactile part is the toughest part to get. And then um you know the these physical models that we're we're that people are building, they're based on largely vision. Right? And so they I mean, you know, what about smell, right? I mean, i if somebody farts in the room, there's your AI. No. Right? Mine does. But no, I mean that that's like the essential question.
SPEAKER_00Uh I think in the near future, is like if if you're if you fart, does your AI near But do you think if you think about video signal is so much easier to ingest and process than tactile signal or sense signal or true.
SPEAKER_06I think but I think Leonard's invented the new Turing test, which is uh can your AI detect farts? I think that's the yeah. Then we've reached AGI once we I mean just move down into that.
SPEAKER_04They kind of have already had that, right? I mean, all you're doing is introducing methane sensing into the AI realm.
SPEAKER_06So that's true. I think there was someone who built like an artificial someone built an artificial nose.
SPEAKER_00I think it's out there. Yeah. Yeah, I remember that. That was one someone that was a guy who um Josh who works on the Microsoft guy. Yeah, yeah, yeah, yeah. French guy. Can't remember his name actually. Um Bertrand, something other stuff.
SPEAKER_06Put that up to uh Chat GPT and Ural set.
SPEAKER_01Yeah, I mean, but Bill, I would argue it depends on what you ate, but yeah, you're right. But it all starts with sensors. We're gonna be at sensors converge, right? Yeah, good segue. Great segue.
SPEAKER_06Speaking of farts, speaking of sensors converge. Speaking of sensors converging, yeah, exactly.
SPEAKER_04That is that is the ultimate convergence right there.
SPEAKER_06There you go. Um, yeah, sensors converge is next week, right? Is that like the fifth through the seventh? Uh I think that's where it's in Santa Clara.
SPEAKER_01Santa Clara, yeah.
SPEAKER_06Um everyone going on for decades there, that show. That one is gonna be it's gonna be a good one.
SPEAKER_01Yeah, and um there it it's great. I'm on the board, and it's great to have um Edge AI Foundation as a partner. I know that everyone loves it.
SPEAKER_06Yeah, and uh we have a big pavilion there, and uh I'm doing a keynote entitled How Edge AI Will Save the World. So what? That's the keynote, yeah. Are you serious? So you gotta tune in for that because you want it, everyone wants to know.
SPEAKER_01Oh yeah, now I really want to know. Why don't you just tell everyone now?
SPEAKER_06No, you gotta you gotta go to the show.
SPEAKER_01Oh, no way the ticket.
SPEAKER_06So yeah, doing that. We're doing uh doing a fireside chat with ST. We have a panel too on uh kind of embodied AI. Ah so yeah, it should be a lot of fun next week. So a lot of sensor stuff. Sensory sensors.
SPEAKER_01Yeah, hey, you know, oh geez, what's his name again? The con the the um the actor on your shirt. Oh, John Wu. That's not John Wu.
SPEAKER_04Well, no, I mean you you so you this is so this is from the movie The Killer?
SPEAKER_01Yeah, yeah, yeah. Um god, what's his name? The guy on the left. You're right. Yeah, yeah. You know what? It's like weird, those old Hong Kong movies are really violent, man.
SPEAKER_04Dude, you this is a classic t-shirt.
SPEAKER_01Yeah, yeah, yeah. It is. It is.
SPEAKER_04It it's uh it's the John Woo collaboration with uh The Killer, and uh Supreme put it out. You can probably you can probably pick it up for a few hundred bucks. Yeah, that's uh he is he was in uh a few hundred bucks.
SPEAKER_00You could come get my my son doesn't have that one, but he has like a wall he has a closet full of Supreme stuff, which is like uh yeah, the the online kind of gray market in Supreme is is is off the hook. It really is. Yes, one hundred percent.
SPEAKER_01He was in uh Crouching Tiger or Hidden Dragon. Jeez, how can I not remember his name? I'm getting old. Um anyway.
SPEAKER_05Yeah, that's right. Rely on AI for your memory.
SPEAKER_01I'm relying on uh uh Chao Yun Fat. I have to re you know, I'm gonna okay, credit to AMDV. It was a database. I just did like a simple lookup. You know how I mean that guy was huge. Huge, yeah. I mean, he's freaking superstar in Asia. Yeah, yeah. It's cool, man. So uh oh my god. What's up? What's up? So um embodied AI, what the hell is that?
SPEAKER_06I know. Well, it's another another term, you know, about AI, you know, in things. Um, you know, that uh it's like, you know.
SPEAKER_02No, I don't know.
SPEAKER_06This is actually this is actually the first question of the panel. The first question of the panel is what is embodied AI?
SPEAKER_00Did you did you do an adult warning, adult content warning at the beginning of this? Because I feel that we're headed into dangerous territory.
SPEAKER_06There, um, you know, it's you know, people would think about it as like uh AI that's like infusing capabilities inside of an object, um, like a robot or things like that. But it's it, you know, people use it for like smart glasses and stuff. Personally, I'm not I don't think it's really like that great of a term because it's really squishy and weird. Yeah.
SPEAKER_00Um but uh but it's AI and things, but it's uh you know that's kind of does it mean we all have to go back and try and start thinking about embedding RFIG chips in our wrists because the GPU is a lot bigger, and I can imagine that being about the same protocol.
SPEAKER_06Well, that's now you're talking about implantables. So you have you have wearables, hearables, and implantables. So you're talking about implantables, which would be using like neuromorphic or spike in neural network, you know, which is kind of infinite battery life that you'd be inside your in, yeah. So implantables. Actually, on a more serious note, implantables are being used for like deep brain stimulation and things or Parkinson's and other stuff. So you so you'll see actually some interesting uh applications of implantables uh over time that can sort of self-tune their uh things. Right now, when you do things like that, like pacemakers and stuff, you kind of set it up and you kind of put it in there and see what happens. But imagine if you could read the feedback from the body and adjust your electrical output and things like that. So it's kind of uh implantable is actually a pretty cool, pretty cool area uh if you look at folks that are doing that. Tough space, obviously. It's fairly challenging to do that.
SPEAKER_01But you know, you can you can start to solve big problems with it, then maybe that's uh that's one of the well you already see.
SPEAKER_06I mean, it's not implantable, but a lot of people weren't wearing the stuff on their skin to monitor their glucose and whatever. So you know, there'll be some implantable in there that I mean I think they have that thing with the CPAP thing. Instead of a CPAP, you can get a thing in there to uh stop your snoring or something.
SPEAKER_04Well, maybe that's where the lot of the it it's it's so weird. We're using tech to solve problems that we've created by bad diets. That's true.
SPEAKER_02I know that's right.
SPEAKER_06I want to keep eating my cotton candy, so how do I do that? Well, you know, use this tech to uh monitor your sugar level.
SPEAKER_01Well, I mean, that's like literally the entire pharma industry, right? I mean, you know, diabetes, it's a self-inflicted uh disease for the most part, right? I mean type two, type two, type one is not type two, not yeah, okay. Thank you.
SPEAKER_04I mean, but it's type two is yeah, if you think about it, like for those of us that are Gen X, uh you know, we were we were advertised a bunch of different things. I mean, right? I mean, breakfast in the morning, depending on where you were, it was like fruit loops, sugary cereal, and everything like that. Count chocolate, count chocolate sugar, and then you crash in class. Yeah.
SPEAKER_06Lucky charms to get your lucky charms in there with your marshmallows.
SPEAKER_04You're right. It's like those of us, I mean, I was one of those. It's like, give me the lucky charms. I'm tossing those little grain things out, and I'm keeping the marshmallows, right?
SPEAKER_06You go with marshmallows, yeah. Yeah, no, then you hit uh 10 in the morning and you crash. Oh my god.
SPEAKER_01So that explains it. I always wondered why I was so tired by that time one o'clock rolled around. I was like, I needed to have a siesta. There you go.
SPEAKER_04We're being we're being marketed to about all of these different things, and then then you you know these medical things that are just I'm like, we never had commercials that were like healthcare commercials that hey, go ask your doctor to give you this. Oh, yeah. By the way, you might have anal itching and uh deterioration and uh no, the best is the warning, all the warning labels is like don't take this if you're allergic to it.
SPEAKER_06It may cause may cause death, it may cause whatever. And uh, you know, so it's like uh makes sense, it's crazy. The the yeah, these days it's you're inundated with Jardians and Ardeans and all kinds of weird names.
SPEAKER_00And uh but that's that's the the interesting Sam, you talk about embedded tech and neuralink and all that sort of stuff, but the other thing I think I was really this week was the kind of the alternate approach is using gene therapy, and there's this thing called the Amanaka factors. Um you can use a specific set of genes to um reprogram cells um within the body to make them revert to or reverse back in time toward being a stem cell. And they've been doing various kind of questionable experiments, I think, on on rats and what have you, severing rats' optic nerves, exposing them to these Yamanaka factor uh factors, and then the optic nerve regrowing um completely and and and the rat then being able to see, and similarly reverting going from being grey haired to you know young and bouncy. So I think there's there's a there's a kind of race between is this a medically driven, genetically driven kind of uh uh strand of innovation, or is it something that's facilitated by hardware and embedded technology? Yeah. My money's on the gene therapy side, I would I would say, but yeah, the cancer risk is uh self-evident if you're tell telling yourselves to do things you know outside the normal form, then so why is it that as you were describing that, I kept thinking of aliens in the movie.
SPEAKER_01I don't know why. Superhuman. No idea why.
SPEAKER_06Xenomorphs you're talking about.
SPEAKER_01Well, you know, it ends up you know, either we're a product of that already, an alien experiment, or we create we create the guy that works in the research was also the first person to make a human and uh human ape chimera that they um they killed the embryos at like 20 days.
SPEAKER_00So the guy's kind of out there and is not you know not well respected, but he's he's one of the folks who's getting vast amounts of funding from Bezos and others to go for it. Yeah.
SPEAKER_01I mean, there's like things that yeah, there's things that you simply shouldn't probably you shouldn't do, but you're gonna you know there's you're gonna do it anyways. Right. There's some it's pronounced Frankenstein, by the way. Frankenstein, not Frankenstein. Yeah. Yeah, okay. That's thoroughly depressing. Okay. Yeah.
SPEAKER_00I mean, that's I I see that as my role to kind of be depressing. Yeah, thanks.
SPEAKER_01Yeah, I didn't think we're all screwed. We're all screwed.
SPEAKER_06Started with farts and we ended with xenomorphs.
SPEAKER_01Yeah. Always the way. Always the way. Yeah, maybe that's what we'll call this episode the deterring test. Uh but uh that's that you know, okay. So going back to neuromorphic though, the yeah, you know, medical devices are probably where a lot of this is gonna start, you know, solving problems that we've caused for ourselves with bad diet as as um, there's no shortage of demand for medical solutions these days.
SPEAKER_03Well, there's yeah.
SPEAKER_06I mean, there's interesting, there's the there's the obvious of these, you know, we can talk about Ozempic and all that other stuff, but actually taking the tech and making it more applicable worldwide so that you know uh a lot of uh you know societies that don't have uh the the richness that we have, like in the US, that can now get more access to better tech because they're using AI on it is going to be interesting. So um so that's that's pretty cool. So as the cost uh of these things comes down, uh I think we'll see you know more proliferation of of healthcare initiatives.
SPEAKER_00And that's for like you know, infant mortality and all that stuff is uh much more interesting for me than Ozempic face or whatever that what is that uh I mean it's your same thing that it's basically giving people tools to carry on doing the stupid things that they shouldn't do uh without any consequence, it takes away the consequence of terrible dietary decisions and and not exercising and all the present. But I don't know, it's it that I I agree. I AI, I think that the that potential could to kind of democratize access to science, I think, is incredible. Though you know, we've talked before about the impact on on folks having you know no junior developers anymore, etc. Though I did hear someone say this week that they were confront they were shocked by the fact that a junior developer is actually cheaper than the amount of tokens that they're consuming to do this.
SPEAKER_06Well, that's it, yeah. I saw that too. And so it was like, oh, we're hiring junior developers to save on our token budget. I'm like, isn't like wasn't it the opposite last year? It's like we're we're firing people to use AI. Now you realize the opex for AI. Um, I was talking to someone the other day. It's like, oh, we want to like you know, uh you know, analyze all of our YouTube videos and come up with all this stuff. Turns out like doing this stuff is like really expensive. Yeah, you know, oh yeah, yeah. Uh it's it's just not feasible. Going back to the LinkedIn discussion, probably one of the reasons LinkedIn doesn't use AI to filter spam is that it would be a lot of operational costs for them.
SPEAKER_01Even though they do, I know like Google does with YouTube because they they go and look for any kind of mention of election-related stuff. Yes, they will ban ban that content.
SPEAKER_06And they do like copyright checking and all that stuff.
SPEAKER_01Yeah, yeah. But I think a lot of that is algorithmic, and so this is the problem. A lot of folks think that everyone's using genai and using NBL72s of whatever variety to do a lot of stuff. Actually, a lot of this stuff is uh as cheap as possible to run um ML, right? Uh I mean, when I was at um NABCO, the ML for the win, you know, hardly anyone is using generative AI except for uh just experimental stuff like uh meta tagging, you know, do uh using scene detection for meta tagging. But even then, um nobody wants to have this uh always on AI monitoring, right? And I don't know where I mentioned this. Maybe it was last week, um, but it costs too much. So what a lot of customers are asking for as they see these massive bills is a cheaper way. And a lot of that ends up that cheaper way ends up being more of a reactive modality rather than a predictive. Sound familiar. And it's about how do we just improve our our ability to detect and respond, well, more of how do we respond quickly and remediate uh when an incident does happen. And what they ended up doing is reducing their um token consumption by 95% and you know, dropping the cost of the solution by 95%, which of course the hyperscaler wasn't too fond of, but guess what? That's where practical lies. And so, you know, I just really scratched my head wondering what's going on. You know, we we see like the hyperscalers claim that their AI businesses are going up, or the cloud guys, neo cloud guys, but then you you have to really wonder, okay, how much of that boosted, and I posted this, how much of that boost boost is for AI generated stupid cat videos, right? Um, because you know, Nana Banana kind of went bonkers um in the last three months, right? Uh uh so how much of that is because of you know that kind of use, right? And how much of it is because of doing stuff, right? I mean, yeah, a lot of that is just creation of AI slob for the most part, right? Right.
SPEAKER_06And um, you know, uh well it's yeah, this is it's hitting the fan if you look at the uh open AI IPO analysis, you know, the the one of the interesting metrics I've seen is the revenue per gigawatt, right, for data centers, right? So it's something like 10 billion per gigawatt today, um and something like that. But the the data center costs about you know 40 billion, so you really need to run the data center for about you know um four or five years to recoup your investment on the data center. So the the revenue per gigawatt is not not there, it's just not there.
SPEAKER_01Um well and I I think that's exaggerated too, because they're just doing simple math. The thing is, is um you know, semi-analysis came out with I don't know if they even coined a term, but it's a concept we've already talked about on IoT Coffee Talk. It's about um uh they call it token efficiency, but for what you know what we've been referring to it as you know mapping actual outcome value uh to the number of tokens actually required to deliver that that uh value, which varies, right? And so um token efficiency uh actually is the wrong way of looking at it because you have to look at it in terms of the actual consistency of pricing, being able to price an outcome and manage it make sure that on the back end you're profitable, right? Because the cost can be variable, and I I don't think people see that yet. In fact, I talked to one, you know, that company that you referred me to um about that topic. And it it's a big blind spot.
SPEAKER_05Yeah.
SPEAKER_01Uh and and so yeah, it's the the all a to you can't just take prevailing token price, okay, multiply it by how much your gigawatt of AI compute can generate, and then think that that's revenue. That's that's total at that one given point in time. But what happens as the price of tokens continues to drop when companies like Deep Seek and others uh they continue to um you know force the pricing down, right? Yeah, yeah. Which is what they've done.
SPEAKER_06Right. But there's the cost per token, but there's also the definitely is going down, which is great, but the inefficiency is the frontier. But the other question is where's the revenue? Is there enough revenue being generated for a $40 billion data center? Like basically, you need $40 billion of revenue to pay for a $40 billion data center, right? I mean, at the end of the day. So is that being generated? Is there enough? And this is where the open AI IPO, I think, is going to be is running into some speed bumps, right? Is like they they sort of missed some of their revenue targets recently. And people are questioning, you know, can you really IPO uh in this environment with OpenAI? And I think I'm afraid that if they end up scrapping their IPO, it could have some other follow-on kind of market consequences and get people skittish on all of their own.
SPEAKER_00Well, I can things down at this, but I I've not been following the um the the the Musk versus Altman trial um too much this week. But I mean that's that's obviously clearly a potential spanner in the IPI plan. But I wonder if the IPO were to I kind of think it's unlikely that it will, but were it to falter, you know, what what impact would that have? It would definitely shake confidence, but maybe it would also cause people to be a little bit more focused on what's real versus what's not.
SPEAKER_01Well, I mean, think about how much everyone is uh you know, how much everyone has loaded up their RPO uh RPOs on supposed um open AI related contracts. Yeah, yeah.
SPEAKER_04Yeah, yeah, but aren't we aren't we still negating the the real issue, right? I mean data centers. Yeah, I'm I'm sorry. Welcome to welcome to FA and FO Friday.
SPEAKER_01I mean I was about to say that they're all FA and O uh FO-ing.
SPEAKER_04Yeah, well, yes, because I mean we're still skirting the real issue. Damn grid resilience. How are you I mean at least it takes energy to do all of this?
SPEAKER_06Right. So now you have to build your own power supply for each data center and not rely on the public grid, which now boosts your costs to like the moon, basically. Um, so that's a whole other this is why the cost of the data centers, even though the cost per token may be going down, the cost per producing the token cost per data center is going up. Um because even if you can get through the planning process to yeah, assuming you can get through the you know not buy backyard protests and uh everything else going on, right? It's uh it's expensive and it's difficult.
SPEAKER_04Okay, so so you you gotta build you gotta build your own microgrid with some level of sustainable energy, right? So you got some wind, you got some solar, you need your lithium crystals, you know. You need to cool this shit. So wait a minute, let's run some water through there. Where's that water coming from?
SPEAKER_06Yeah, the infinite resources of the ground.
SPEAKER_04That's that's that's I mean, it it's comical. I mean, you know, from my perspective, it's comical for those that understand energy delivery to solution being provided and what the dependencies are each step of the way.
SPEAKER_06I saw some some analysis that I think someone was trying to justify their the data center and the water usage, and they said, Well, you can't use you know a million million gallons a day, and they said, Oh, don't worry, we're gonna use the water at night. So it's like that's still a million gallons a day. It's just at night. It's yeah.
SPEAKER_00Well, the sad I mean the sad thing is at the moment when you say renewable energy, that would be a great thing, but you know, what you've seen is is companies that I shall not name putting you know uh diesel generators in containers out of the back of the data center and then saying that the local the local community that's that's suffering a particular pollution, oh no, that's not that's not happening at all.
SPEAKER_06No, they're uh they'll use diesel, they use whatever fuel they need, but uh yeah, that's that's a big uh knot to untie. And going back to the original point, if you're not generating forty billion, fifty billion per data center over four or five years to pay it back, then you know, is it even worth it?
SPEAKER_01Well, yeah, and then you know, consider that um, yeah, it was funny. Uh if you listen to the Microsoft um earnings call, the first question that was asked by the analyst was uh how is all this gonna be paid for? And Amy uh Hood and I don't think Amy or Satya did a really good job of answering, you know. Basically, we have no freaking idea. And this is like three and a half years after these all these guys are making these huge this FOMO bet, right? And yeah, the math doesn't work out at all, right? Um, you know, even if you assume that advertising is going to subsidize a lot of this stuff, right? The advertising industry in totality is only like I think it's around 900 billion, like a uh no billion. So it's not a trillion, but you can't assume that uh AI quote unquote AI the TAM for AI is that entire market. It it's not because the AI is just a tool, right? Um how are they gonna make their money, right? Uh it unless enterprises are really getting a lot of value out of this stuff, but you know, there's conflicting signals. How much of the quote unquote demand, and there should be a study on this term demand, what the hell does it mean? Because you hear a lot of the Wall Street guys talk about well, there's so much demand. It's like, yeah, demand for like the lower level stuff, but what about demand up here? What is it? Well, it's infinite demand. Well, yeah, of course, if you give stuff out for free, if you give cocaine out for free, guess what? You might yeah. Um, of course, it's gonna look like there's infinite demand. But as soon as you start charging a crap ton, you know, um if you can price it like cocaine, then yeah, you might you might you might make it. But if you can't, because what you're selling is kind of worthless.
SPEAKER_00Well, and the numbers we're talking seeing, you know, when we talk about kind of the investment in data centers and the investment in electricity, I read the other day, the whole global annual investment in generating electricity, fuel operations, building new facilities tops out around 1.5 trillion. And if you look at that against the numbers that Microsoft are quoting, that that you know that um OpenAI are quoting in terms of infrastructure spent on data centers, I mean kind of like it's a shocking amount of money that's being being budgeted. And like you said, if the if there isn't a clear if there isn't a clear path to a return on that investment, then how long does the patience have uh how long does people's patients last?
SPEAKER_01Well that's the thing, is I I think um, you know, people's patients are starting to run short. And so now, I mean, what was a there's a fortune article that came out about Google's um earnings recently, and um there's like I think about 27 or 37,000 in 37 billion in other income, and that happened to be uh unrealized um securities gains from their position in anthropic, right? Which you know the article talks about how they can throttle that value based on how much they invest, so that's cash going in that inflates other income. So um I mean, what's what's going on here? Do you know what I'm saying? It's like how you how how profitable is the venture? Because one of the things that you do notice the more these guys claim that they have an AI neo-cloud-ish kind of business in their cloud portfolio, the faster their operating margins go down, and the faster I assume their gross margin goes down. So, you know, yeah, have said this also on IoT Coffee Talk for actually a couple years now. AI is a bad business for hyperscalers, and they know it. They should if they don't. Um, but I have to hats off to Satya though, because last year, this is after he got on the um what is it? Um I think it was uh Darkesh Patel show. Yeah, he and um uh uh Dylan Patel, they went and did a tour of uh the the um Microsoft data centers and they interviewed Satya. One of the things he was very honest about is the depreciation, accelerated depreciation, and how they have to be very measured and how the investments are scheduled over time. And you know, obviously there's an anxiety for them to get as much capacity uh because of their FOMO agenda uh to secure as much capacity as possible and deliver that capacity to monetize. But you know, they none of these guys ever ask themselves, it seems, the question of how profitable is that monetization? So you make a trillion dollars at a negative profit. What is what does that equate to?
SPEAKER_00I think I think it's those similar business. It's a similar argument to what you saw in the cellular industry when the 3G license, when the first kind of really expensive license set that was released, at least in Europe, was the 3G license set. Yeah. And you know, I was in in telco, I was in in 02 at the time. It came down to it was a cost of doing business that if you didn't have a 3G license as a telco, clearly you were out of the game. Um, and and and the math of what they cost versus you know what they would generate, we looked at it in 02 from two perspectives. We looked at all the lovely little revenue we're going to get from content and blah, blah, blah. That was one model. And then the other model was value of the stock price with a 3G license, value of the stock price without a 3G license. And the difference between those two numbers was more than the cost of the licenses. And we spent 20 billion, 16 billion, something great on licenses across the UK, Europe, etc. I think the same is true now. If you don't have a honking great big investment number in your financials that says this is what you're going to spend on AI, and you don't have a you know a chat GPT, or you don't have a dog in the fight, then are you in the game? So it's almost, I don't think it's even a question about profitability. It's about buying a chance to be you know at the table in the next in the in the next round, at least. Yeah, yeah.
SPEAKER_01But you know, that's that that is the but then that's the gamble right after.
SPEAKER_00And sometime at the m sometime the money will come in in the future. Like, okay, when it will be worth it. We'll look back and it'll be worth it.
SPEAKER_06Um just have to find the right use cases, and that's what's still, you know, when you build the general tech, it's like, well, we'll see what we're gonna use for. I mean, obviously, taking meeting minutes, you know, document development, there's like some use cases that are really sticking. Um, what I like about edge AI, just to pivot to that for a second, is that it typically starts with a business problem that needs to be solved, and then you apply the tech to solve the problem. So you kind of know like what you can spend to solve that problem to begin with, as opposed to let's build a general purpose chatbot thing and see what happens. So that that's why I think on the cloud side they're struggling with finding monetizable use cases. Like, is it worth the tokens? Is it worth the operational cost to solve the problem in this way? And in some cases it's not yet.
SPEAKER_00So yeah, I think well, sorry, that one of the challenges I see is that you know when you look at it from a cloud provider's perspective, you want that. That edge AI component to be commoditized. You want the value to accrue to your, you know, to the back end analytics, the back end magic. And you know, the guys who are sitting in that back end have very, very deep pockets. And the guys that are sitting at the edge, with a few notable exceptions, generally don't. And so when it comes down to the battle of who wins, yes, I think edge AI has a much clearer business case, but if you've got a bunch of hostile folks over here who are trying to gut the price of what you're providing, can you, as an edge AI provider, be successful or will they kill you off? Um because free free is a big persuader. And I think a lot of a lot of a lot of the the platform providers are looking at the edge and thinking that stuff needs to be free. You know, we need to make that free so that the the generation of data is no longer, you know, it doesn't cost anything. And we need to do that.
SPEAKER_04But it's it's it's not that difficult to to prove that out, right? I mean, if if we just take um again, yeah, I and I I I agree 100% with Pete because he's he's talking very sensical in terms of edge AI and the business problem that wants to be solved. So we clearly understand what problem it is that we're trying to solve, and and we're applying the tech to it. And you don't have to say, oh, it's it's AI right there. You you are solving a problem. It's not about the tech. If we're having the conversation about the tech, we're having the wrong conversation. Yeah, we're literally saying there's a problem here, and we're going to solve it by getting as close to that as possible, and we're gonna do everything there. Now, if you're doing it right, you're reducing the amount of data that you're sending back to the center cloud, yeah. Yeah, and if you're doing that, you're pissing off the CSPs because they they're they're they're not getting those workloads, and you're pushing those workloads out to the edge.
SPEAKER_01Yeah, but that um I mean, yeah, and so the point I was trying to make earlier before Alistair like friggin' it interrupted me. Just kidding, dude. Um, not really, but uh was you know neuromorphic. Um a lot of folks that think that's a a cloud thing that oh hey, it's gonna replace GPs, blah blah blah. But it's uh it's actually solving or it has potential to solve constrained problems at the edge. So when you look at medical devices, right? Novel medical devices to help solve or monitorable lashes. Right, right, exactly. Um, this is where a lot of that next generation innovation and invention is gonna come from, right? Um and um and actually uh this is kind of what we saw with um AI. AI really blew up on the smartphone, not in the data center, right? Data center was just doing some you know ML training, blah blah blah. But where value was being exhibited was like in friggin' compositional photography and shit like that, you know. I mean, awesome value, huge differentiation for a lot of the players in the space. Uh and and so I, you know, the place to look is at the edge in many instances because some of the you know everyone thinks that, oh yeah, you know, you're doing all this in innovation and it's coming down to the edge. Actually, there's a lot of stuff that goes up, right? SoCam, for instance, that memory collaboration uh with um micron between Micron and NVIDIA, a lot of that is based off of like you know, mobile memory tech, right? Low power, um, you know, high throughput, that kind of stuff. So yeah, um anyway, I just wanted to drop that one out there. Um because and just because I think there's this misconception that advanced technologies are incubated in these large data centers. That's not true. Actually, a lot of this shit happens to Bill's point, solving really hard problems at the edge, and then they scale up, right? Or they trace their way up. And then, you know, like so now um because data centers you're sort of past this whole point of brute force, and now you have to figure out how to optimize. A lot of the principles of the edge and IoT are becoming very important, especially for operational data centers, because hoses leak, you know. Um uh, you know, freaking sockets can overheat, right? Or so you have to have all this physical monitoring of the entire infrastructure and environment. And guess what? A lot of this technology is coming from the edge.
SPEAKER_06Yeah. Yeah. The uh, you know, I know Qualcomm had been talking about um some kind of interesting new memory architectures to solve this issue. Like in the phone space, you know, um uh a lot of you know, language models are memory bound more than tops bound. And so coming up with better memory architectures is really critical. And so if you solve that in the phone, then hey, guess what? Maybe we can solve that on the server and then in the data center, and all of a sudden you can start to get a lot more efficiencies. And get that cost per token down even faster and blah, blah, blah. So yeah, the edge is kind of where you know, when things get commercialized, as you know, you first get them working, and then if you actually want to commercialize it, you have to cost reduce it so that it actually makes business sense.
SPEAKER_01Well, you know, and and uh this whole idea of like unified memory on SOC. Uh I mean, you basically have memory as close as possible to almost all the different kinds of cores that you would require to do AI. Like this heterogeneous compute concept, it's on your smartphone, right? And that's why, like, think about it. What Apple's done with um uh Apple Silicon, they've taken it basically the you know, the A series chip that powered your smartphone and scaled it up to uh a desktop, a laptop, and uh, you know, sort of a workstation scale. And guess what? It's great for all kinds of different AI, right? And that's why all these people now we're talking about agentic are scrambling to get one of those friggin' uh Mac minis. Those things are like you know, according to the Apple folks on their call, those things are gonna be sold out, they're tapped out for crazy, as far as they can tell. They finally found their product market fit. Ironically, it's AI, and guess what? They're gonna make money off of it. I know, yeah, they're making money. You know, think of dude, their their capex went down like 25%.
SPEAKER_03It's crazy.
SPEAKER_01But yet they are writing an AI, um, they're writing the AI hype trend by selling hardware.
unknownYeah.
SPEAKER_04But I mean, but think about it. That that's what that's what the that's what it is, right? The whole thing. Look look at the conversation that we're having. CPU versus GPU. Yeah. That's everything builds from there.
SPEAKER_01Yeah, yeah, yeah, yeah. Well, yeah, and but now there's the new frontier of like um uh yeah, uh I I think the MPU has a lot of potential. LPU, whatever you want to do. Yeah, LPU, yeah, yeah. Um, but especially the MPU, uh, because it can do it can do um uh matrix multiply, scalar, and vector, right? So these are all different operations. If you look at the diversity of uh inference that's coming our way, whether it's the vision stuff, the language, whatever, uh you do kind of want to have a general purpose optimized, highly optimized um you know uh IP uh uh that uh you can run that stuff off of. And I think the MPU or neural, if you want to call it neural processor or whatever, uh it's gonna find its time. And inference might actually be that opportunity because you know, like AIPC has really struggled. Nobody can figure out what the hell do we do with this thing. Well, inference hasn't really, you know, from the data center hasn't come down quite yet. But you know, especially uh Qualcomm has been ever since like ChatGPT became a big thing, they've been first to market with a lot of stuff related to to um you know um optimizing small or tiny ML stuff, right?
SPEAKER_00With edge uh edge compute. Yeah, yeah. I mean they don't they don't have much of a data center play, so I think some of it's kind of a whole pubs and streets.
SPEAKER_01Not yet, but it yeah, yeah. We'll see how that that plays out.
SPEAKER_06Ambitions, yeah, ambitions. I would say neo neo data center as opposed to data center data center.
SPEAKER_01So yeah, neo. That would go down on the edge to edge stuff. But the fact that they they they aren't just necessarily going forward with uh a GPU, I I don't think it's uh no NPUs.
SPEAKER_06I mean look at what DeepX is doing. These folks then these NPUs used to be CNN RNN vision oriented, now they're more language accelerated, right? So the new generation of NPUs are more uh language oriented, and then we're gonna see even more specialized silicon, I think. And then and speaking of memory, you have in-memory compute, right? So you're gonna see memory that's actually has AI acceleration capabilities in that. So so I think we're just at the beginning. We're still in the steam engine era of AI. That's that's part of my talk for next week, is it's still the steam engine. We still have the iron horse with the coal and all that stuff. We can see it's working, but it's it's super duper inefficient, right? So we this is the frontier.
SPEAKER_01Yeah, yeah. Yeah, and then maybe one of these days neuromorphic will will uh make its way up.
SPEAKER_06It will. It'll it's it'll inspire some interesting things, but yeah.
unknownYeah.
SPEAKER_00A lot of folks in the kind of formation stage of what at least the I see from in the in kind of pre seed um guys, they're in the formation phase of of figuring out like how do you actually execute this. The the the the uh appeal of neuromorphic computing is clear. Um and I'd I'd say there are a number of different you know uh uh approaches that are being taken, none of which is really proven yet. But it the necessity of of of edge, yeah, super efficient, power-efficient, edge-based, neuromorphic computing. It's about as clear as it gets. Um I think it's greenfield, so you know it it remains to be seen who actually you know who actually um uh takes the lead in that space. And there aren't that many spaces in in you know in in our world that are like that actually, that are kind of you know, this is kind of a greenfield space. No one's really figured out how to nail this, but the demand, the it is latent demand for that for that type of technology is very, very clear.
SPEAKER_01Oh yeah, yeah. I mean, it can definitely substitute. And you know, here's the good news for the neuromorphic guys, at least you're gonna be ahead of the quantum folks.
SPEAKER_06Oh, that's for sure. That's a whole other discussion.
SPEAKER_01Yeah, and you don't have to dump like you know, billions or trillions of dollars into supercomputing that uh and and solving problems that are money losers. You know, it's cost center stuff. People just uh they still don't understand um quantum. It's all like traditional solving traditional you know, supercomputing problems. Weather weather forecasting, basically, yeah.
SPEAKER_00Do you think that AI in its current is I has uh kind of overtaken some of the need for quantum? So the original goal of quantum was you can just churn through you can do what you do today, but you can do it a thousand times faster.
SPEAKER_01No, it it it it's literally you know, the way that it was described to me by a Princeton professor who's been studying this stuff for I mean developing this uh stuff for the longest time is think of it as a calculator, not a computer. So it will solve a really complex problem. It will not you're not gonna be able to do that.
SPEAKER_06Cracking the uh crypto uh uh encryption. Yeah, crack encryption calculation, right?
SPEAKER_01And and it will not replace traditional compute. In fact, traditional computing, bits and bytes are really good at what they do.
SPEAKER_06So, Leonard, are you saying that Microsoft's not gonna introduce the quantum PC next year?
SPEAKER_01I heard that that's uh Yeah, then that'll that'll come in maybe uh yikes. But like if Bob was here, he would say that it would be here in like uh whatever, you know, 20 years ago.
SPEAKER_06Quantum PC professional edition.
SPEAKER_01Quantum pilot.
SPEAKER_06Quantum pilot.
SPEAKER_01Yeah, Q Pilot, Q Bert. Yeah, yeah. So, anyways, no, that's good stuff, guys. So uh while we call it an episode.
SPEAKER_06Sounds good.
SPEAKER_01Hey everyone, uh thank you for making it this far. If you did, holy crap. You have more attention than any LLM or That's right, yeah, or any other. Maybe they just watch one of our little clips on YouTube, though. So that means that you have a context window of at least 20 billion tokens, so that's pretty good. So congratulations.
SPEAKER_00That's right. Helping them.
SPEAKER_01Yeah, we're subliminally twisting their minds. Yeah, I uh but no, we really appreciate your viewership and your listenership. And um, you know, if you don't um if you don't tell everybody uh about IoT coffee talk, we'll be quite disappointed. Guys, I don't know where I'm going with this.
SPEAKER_06R AI will spam you. See you at Sensors Converge next week.
SPEAKER_01Yeah, definitely. Sensors converge next week because the IoT starts with sensors. Without sensors, there's no IoT. You cannot sniff farts, and your AI will never know if you farted in the room. So remember that. And uh also, yeah, um, I think we're almost about to go live with uh Elevate Communities and uh check it out. It's going to be an effort to uh bring uh you know technologies, essential technologies to communities so that they can be resilient in um light of growing uh environmental threats and um you know other challenges that can plague a community. And we want to help communities all around the world, if possible, to um quickly leverage uh emerging technologies as well as established technologies uh for good. So, anyways, we will see you next week and have a great weekend.
