We unfolded the iPhone Duo

31:12Vee Song:So this is enabling a new readiness feature, which, you know, again, the wearable nerds out there will be like, everyone else has had a readiness feature for like five, six, seven, forever ago. There's no new sensor, right? It's not like this unlocks some new thing that they can track. But because they have more data, they actually are able to interpret it in new ways and get stuff like heart rate variability and this readiness scores. They went hard on heart rate variability. Is that a big deal? The thing about wearables that I want our readers to understand is that a lot of those times it's not measuring anything super directly.

31:53Vee Song:Like even your heart rate with the optical sensors, it's not measuring it directly. It's shining light into your skin, finding out the rhythm of your pulse based on how the light reflects back. And from that, an algorithm is estimating your heart rate. And we've gotten to a very high degree of accuracy with optical heart rate sensors compared to when wearable technology first came out. But heart rate variability as a metric, it's a proxy for your autonomic nervous system, which is kind of like a stress indicator how covered you are from certain things. And as werewolves shifted towards sleep and recovery over the last several years, it's become a more prominent metric, especially among athletes.

32:35Okay, while we're talking about mysterious health metrics and uncertain science, some of this stuff sounded scary. They were like, we'll tell you your longevity. We'll tell you the age of your body. And I'm like, oh, God, like, can it actually tell that stuff? Does that mean anything?

32:59Vee Song:Basically, these companies are all measuring different metrics, and they're weighting them in different ways in their proprietary algorithm to basically kind of give you an estimate as to whether your physiological age matches your chronological age. And the idea is you want to be younger or on par or younger. And if you are older, then maybe you want to change certain levers in your health so that you can be physiologically younger than your chronological... I think it's a load of crock. If you're using it in the way that I think it was originally intended, to see areas of your life that statistically, with the literature, we can say it gives you a better chance of living longer.

33:45Vee Song:Okay, sure. You can be like, maybe I'm not doing enough in terms of sleep. I should really change my sleeping habits so that it's more regular. or, oh, I'm really like kind of eating pizza and fried chicken all day. And maybe I should eat a vegetable every now and then. In that sense, it's fine. But, you know, you kind of get into it. Like I've talked to several readers, several people out in the wild who are just like my whoop age. I got obsessed with lowering my whoop age. And then, you know, I had one commenter who was just like, and then I got a blood test and found out I was really not doing so hot on that blood test.

34:19Vee Song:So it's like your health is holistic, but it's just very much like, guys, let's use common sense. eat nutritionally most of the time exercise regularly and try and sleep consistently and your health age should be fine like you could get hit by a truck tomorrow and it won't matter that your health age was like 10 years younger than your physiological it's this is the thing every single year i watch these apple watch presentations and i'm like oh man this is a major improvement over last year and i've gotta wear one of these things or else i'm gonna know nothing and I won't be healthy. And then like three days later, I'm like, this is probably fine.

34:59Like I probably don't need it. Does this seem like a big level up to you? Does this seem like it's going to change what the Apple Watch can do? Is it going to displace, you know, the whoop for whoop diehards or is this actually just like another iterative? OK, they're they're adding more expected health features kind of year.

35:18Vee Song:I think it's more towards the latter with an Apple twist. Like Apple doesn't really care about being the first to these health features. They care about executing it in a way that they think is scientifically valid and actionable. So like last year, they introduced a sleep score, right? Every other wearable has had a sleep score for like ages now. Like it's not new. And so like this health app redesign that's coming later in the year is very much kind of bringing it more in line with what we're seeing from other health tech companies. So in a sense, it's catch up, but with an apple twist. So one thing I do give them credit for is that they do try and take a much more conservative approach so that, you know, you can just engage with it to your degree of comfort.

36:03Vee Song:You know, I'm looking at their longevity tab that was in the keynote, and it's less objectionable than some of the other ones that I've seen in this space, which are really just kind of pushing you to the frontier edge. it's a lot more like, hey, we're going to take the generalized approach. It's going to be a lot more holistic and top level, which I think for the average person is the right choice. But, you know. Yes. You can see there's so much stuff in there now. So those are the health updates. To me, the more unexpected and more differentiated and wild update was what they did with AI. So Apple announced this thing called audio intelligence, which is just another AI acronym.

36:54And there are a bunch of sub features in this. But the two that I think are most striking are called Live Rewind and Siri Recap. V, can you TLDR what these two are for me? Yeah.

37:10Vee Song:So Live Rewind is if you're in a situation and you're like, oh, what did someone just say? you press the crown twice and it like goes back 15 seconds in time and produces a transcript of what was said. So you can just check it on your wrist. And if you choose to save it, it will go into the Siri AI app. There's a new hardware chip that it's recording nonstop to enable these features. Kind of. There's like the way that it's described in like the privacy documents that it's come out because I feel like Apple was very like, oh, people are going to have questions. So here is a white paper that was ready to go by the time the keynote had ended.

37:54Vee Song:There is something that they call a secure exclave on this chip that the audio is processed through. Supposedly, Apple can't access it. You can't access it. There are no audio recordings saved. So the raw audio file gets sent to this exclave beep-boop-bop AI processing happens. That's a scientific term, beep-boop-bop AI processing. And then that audio file is immediately deleted. There is a visual cue that is recording, that you are recording something on the wrist when it pops up, as well as an audio chime that also happens at the same time so that people around you are supposedly aware that this feature has been enabled and you have to manually enable it at all times.

38:37So at no point are you going to be able to be like, what did someone say 30 seconds

38:42Vee Song:ago? Like that's, you're not going to be able to pull up a transcript of that. It's just the boop. You got 15 seconds. That's it. You got 15 seconds. So it's like the second someone says something and you're like, oh, didn't catch that. You have to activate it for this thing to come up. And who knows if it's going to be any good, right? Because it's a mic on an Apple Watch. It has to be kind of, this mic doesn't have like long range frequency. It's not like you're going to be able to point this thing at something someone said very far away. It's kind of more what's in your immediate vicinity. The social use case for this is very mysterious to me.

39:21Like I both intuitively understand it where I'm like, oh yeah. There are definitely times when I'm like, I could not get that word. But what is the use case for, and so I will press a button on my wrist, which will then announce that I missed what you said, and then read what you said, instead of just asking you to say it again. I'm sure there are some accessibility use cases there. That feels very clear, and Apple did highlight that. But for the, you know, day to day for the millions and millions of people who have these, that feels very strange. And that's only one of them. So then there's Siri recap, which I think takes us to a much grander level.

40:04Vee Song:Basically, like when when it's on, you'll get a recap of the conversations that you have. So like it'll detect that they're covered. Or it's not fully accurate to say it's always on recording because you can customize when this feature works. So you can customize it to be like only during work hours or never on during later hours. So like you have some customization there. It is opt-in, so it's not turned on by default. So it wouldn't be accurate to say that every Apple Watch that's out there is suddenly listening to everyone. That is not necessarily the case. But it has the ability, correct? You could set it to always listen.

40:43Vee Song:You could set it to always listen. And when it detects that a conversation is happening, it'll produce high level meeting summaries that then appear in your Siri app and then auto deletes after seven days. So it's kind of like, you know, how you're on a Zoom or a Google Meet call and Gemini just like produces a summary of that meeting. Which I gotta say, I love that. I like, those are really helpful. Like in certain contexts, I think there's going to be a lot of people who find value in that particular thing. It's just, and this is a kind of an existential question that's been on my beat, is there's, I know that I'm using it for a non-creepy method.

41:34Vee Song:People around me don't know the parameters that I've defined that this feature will work. They don't know how I've configured it. They don't even know whether I have a model that's compatible with this feature because it does require the secure exclave on the chip. So it's not coming to previous models. And so in the whole world of Apple Watch users, the vast majority of people who will not upgrade this particular year because they have perfectly fine and functional Apple Watches, they don't have this capability on their device, right? So it's only people who get the new ones that have the potential to do that.

42:13Vee Song:And even then it has to be opt-in. And so there's just like a lot of like calculation that's going on in my brain right now as to how this feature will work for most people. But as I understand it, it's not a transcript. They strip attribution out. They don't differentiate speakers. So like us as journalists, you can't record a meeting and be like, oh, yeah, I'm going to go find quotes in there now. Like you can't get to this weirder place where it's like, OK, so you get this generic summary of all the conversations you had throughout your day. And so I'm going to get a summary being like, you told somebody at length about GTA 6.

42:51They didn't care. Like, what am I going to do with that? Like, I just on one hand, I find this so wildly impressive from a technological standpoint. The fact that we're at a point where we can just leave a microphone running nonstop, we can use AI to make sense of it and summarize like that's very cool. Whether that actually produces useful insights, like we're, I don't know that we're there yet. This feels like we're doing it because it's cool.

43:21Vee Song:It's very much a thing that I think is useful in a work context or in a planning context. Like if you're a parent and you have a lot of kids, you have to attend a lot of meetings for kids that have a lot of action items that come out of it. In these contexts, I think it can be useful. And we already have the kind of technology in like this particular context where you're video conferencing and you're like, ah, what was the action item from this meeting that I need to do? Like in those contexts, I think that's what it's meant for. Like you attend a lecture and you want to know what the notes from that lecture are and the points that were salient.

43:58Vee Song:But from our like day-to-day conversations where you're just talking with friends, do you really need like a bullet point of what, who said what, what and where, why? Like, not really. And also, again, like no attribution is going to be given in these in these notes. So the way I understand it is like. I say something that the meeting notes from the Siri recap is never going to say the said this. There's no said in there. If there is a name that pops up, it'll be like. Amanda did her homework. It's just so interesting that they're like, we're not going to make it too useful because because then it would be really creepy.

44:40And so it's like, we don't want to go too far. But that is the risk here, right? Like this is the summer that the Meta Ray-Bans switched from being this, you know, kind of hot up and coming product to being called perv glasses, like quite widely. And now Apple is like, oh, by the way, you know the smartwatch we make that millions of people buy every year, all of them now have the potential to record audio 24-7. I don't know how people are going to feel about that.

45:17Vee Song:Based online, not particularly happy about it, which is sort of like interesting because in my comments, everyone has been like, Metaglasses, not well meta makes them. I'll wait for Apple's version, which, you know, like is sort of like a soft launch here with a device that is like already used by a lot of people, trusted by a lot of people with a lot of sensitive information, right? Like smartwatches, they learn a lot about you in terms of your health and like a lot of data is collected from them. So in one sense, Apple is sort of testing that philosophy, I think, about AI wearables. And like if we come out and have a very strong privacy first thing here, we have a secure exclave.

46:00Vee Song:We're not recording audio. Everything auto deletes. We're trying to keep in mind it's not only your privacy, but the privacy of bystanders around you. Like it is one of the most thoughtful implementations of a privacy policy that I've seen in this class of device. And at the same time, I'm like, how? I mean, obviously, they kind of want the press to communicate how it works and educate people. But I'm like, it's very difficult to have that stick in the public's mind. It's very hard to communicate this feature. Yes. Yeah. And I think there's this important distinction between can Apple keep your audio secure?

46:38I fully believe they can, right? I'm sure that this is quite well thought through. That's one side. The other side is, do people want to be recorded? And I think for Apple, this is like, this is an area that they're stepping into in a big way. We're expecting them to release a pair of AirPods that have cameras on them next year that are going to probably do a very similar thing. And I think that's going to make people uncomfortable in a very new and unusual way. They're allegedly, you know, tinkering with an AI pin in their labs too, which is a thing that it's different when Humane does it and they sell a couple hundred of them versus when Apple does it and this thing becomes a true mainstream product.

47:20I think this is actually going to make the Apple Watches one of the most interesting products of the year. So V, I'm very excited for you to test these and see whether this stuff even works at a basic level. I think there's going to be so much more to talk about here.

47:37Vee Song:I think it's also very telling that these features don't launch until later this year. So like, that is interesting. Yep. Still more to work on. Some of the audio intelligence features are launching, but it's kind of like the lower stakes ones, like Shazam is going to work better. Yay. Or like audio, like it'll be able to tell on a doorbell, right? Yay. Like, okay. The pixel has had like a 24 seven Shazam feature for, for like a decade now. And it is like one of the single best features in the pixel. If they had just said the Apple watch now is 24 seven Shazam, baby, I think they would have sold so many new Apple watches off that they should have, they should have kept it simpler.

48:20Vee Song:It's still the underlying concept, right? It's the same underlying concept that it's always listening so it can always shazam. It's just telling me whose song, I'm not, that one I'm not stressing about. But you're right. They've all been listening the whole time. All of our, I'm not gonna say it, but you know that wake word? Like, your phone is listening to you right now. It's very weird and there's a gradient and like Apple just took a large step forward. It just did a push. Yeah. I have to say that I really feel like they are so aware of the conversation that's happening right now because of the documents.

48:58Vee Song:Those documents were ready to go. Right. They put out like a white paper. Yeah. Before the keynote ended, they were like, and here's how it works and here's how it's private and here's how it's secure. And it's sort of just like, oh, they're aware. And I think they're trying to do the thing that I've been asking Meta to do for two years, which is to lead and show how privacy can be done. Whether or not it's going to be effective in the court of public opinion is something that I'm going to be watching just because it's hot privacy summer, hot surveillance summer with the flat cameras and all of this stuff.

49:34Vee Song:It's all related. And so this space is how I feel for the last 24 hours. I think it's going to be so interesting to test these things and see what actually happens. but particularly to see how people respond once they're in the wild. V, thank you so much for joining us to break this stuff down. I'm sure we're going to have a lot more to discuss later this year. All right. We've got more Apple to talk about, but not about this week's event. The other week, Apple announced some new Macs, and they went all out marketing them as AI powerhouses, which we thought was probably just some marketing bluster.

50:11And some listeners told us maybe it's not. So Nilay Patel and David Amell called up an AI researcher to break down what's real and what's hype.

50:20Nilay Patel:So like a week ago, we did the VertiCast and we were talking about the New Mac Studios. And in the way that we are often want to do, we dismissed Apple's marketing as out of hand. They were like, these are the best computers for AI. And I was like, that's silly. NVIDIA, because I'm Team Jensen, as everyone knows. And all these people told me I was wrong. Look at that. So I was like, OK, we got to get this right. I'm clearly missing something. The Vergecast audience doesn't often just tell us we're wrong for no reason. Well, usually there's something there. So we called up CJ Trowbridge, who's an AI researcher.

50:53Nilay Patel:I've been following his TikTok for ages. He's really smart. And he is going to explain to us what's going on in local AI. It is more surprising than you think. And whether or not Apple's claims about these Macs actually hold up. CJ Trowbridge, you're an AI researcher. You have one of my favorite TikTok accounts about AI, specifically local AI. Welcome to the Vergecast. It's nice to be here. Nice to meet you. I am a huge fan. I'm very happy to have you here. So here's the main couple. Last week, we were talking about the new Mac Mini and Mac Studio and Apple's very hypey press release about how they're the best computers for AI.

51:28Nilay Patel:And I said something along the lines of, I bet NVIDIA has something to say about that because probably NVIDIA chips in a desktop PC are better. And then we got 10 million comments saying I was wrong, which I appreciate. A lot of them about unified memory. Some of them about daisy chaining all these Macs together on Thunderbolt. There's some reports that OpenAI and Anthropic are actually buying these devices and the tens of thousands. So I figured I should bring on someone who knows a lot about local AI and the state of the art there to explain exactly how I got this wrong. And if Apple's claims are true, tell me, is Apple correct here?

52:02Nilay Patel:Are they the best devices for local AI? What is the state of the art there? I think they're one of the only options for a specific use case that most people don't have alternatives for, especially for people who are trying to run large models at home. And you can see it in the price pressure around those particular SKUs versus the high-end consumer NVIDIA SKUs. It's interesting because there are all these class action lawsuits against Apple for misrepresenting their AI capabilities. And I think that has fed into a lot of the speculation that their hardware is all bad at AI. And it's true for the phones.

52:42Their phones are not very good at running AI. But for these computers that specifically have hundreds of gigs of RAM, they can't. Apple is now finally starting to market these Macs as the best Macs for AI that they have made. So I think my question here is like, they've had the Mac Mini for a while now, right? The M4s are the ones that went crazy for OpenClaw. Now they have M6. Why is the M6 tangibly better? Or is it? Because it seems like, okay, because Apple is starting to actually market it as the best AI inference computer that they have. Yeah. So is it, it's not better? So like the inference happens using typically it's integer eight math or smaller, which there are a lot of structural issues that pop up around that.

53:29And so like for one of them with Apple, they actually don't have the capability of doing integer eight math. So like they copy an integer eight number into the register and have to like convert it to 16, which is true for many GPUs as well. And they did that because it cuts the memory bandwidth in half. So you can advertise twice as much, right? Nvidia does the same thing. So a lot of these, I think, theoretically, you can say, this is a single-digit percentage improvement over the previous generation. And it's like, well, yeah, that's true. It is the best. It's better than the last one, but like 3 % or whatever.

54:04Nilay Patel:So head-to-head with an Nvidia GPU, what would you take? A new Mac Studio or a PC with an Nvidia GPU? It depends what you want. Yeah. If you're trying to run a single model that's like less than 24 gigs, then something like an RTX 6000 Pro is going to be a lot faster like a lot faster but if you want to run something that's bigger than 24 gigs it's not going to work at all right it'll run in the system RAM at the speed of system RAM so I think that's the big advantage that they have is like if they want to put a lot of RAM you can put bigger models in the RAM but it's not going to be as fast let me ask you a really dumb question And this is like the root of this whole thing.

54:47Nilay Patel:Maybe I've just been poisoned by the hype. But, you know, a big part of the modern AI moment is NVIDIA saying these GPUs are not just for graphics. They're for accelerated computing. This is like Jensen's speech. And it turns out AI really needs accelerated computing for the training and inference. And we need to build, we need all the electrons in the world to achieve AGI because only NVIDIA GPUs can do it. And we're going to buy Hugging Face and all this other stuff. like, right, NVIDIA is on this terror because of the power of their GPUs, right? That is the message they are sending. And, you know, again, maybe I just bought the hype, but I don't associate Apple with super powerful graphics.

55:26Nilay Patel:Is it just the memory architecture? I mean, Apple famously went to unified memory ages ago. They did it with the M1s. Is it just that they got lucky in picking that memory architecture? I don't think it's even that. I think, like, for example, when you look at the secondary market for used enterprise inference cards, even the ones from 10 years ago are hard to get. Because chips really haven't changed that much in the last decade in terms of the capability at the tasks that they were already doing 10 years ago. The performance is like a single digit improvement since then. There's new extensions, there's new shortcuts, there's new duct tape for really specific applications where you have like new kinds of tensor cores, stuff like that.

56:08but the actual like raw compute capacity is not super different with the 10 years ago cards versus today. And actually the inference cards, like a, like a P40, for example, um, has more compute power than any of the, uh, M series max chips. Right. And these are 150 bucks. Um, but there are cycles and sometimes it's like, you can't get them. I think the big difference there is that it gets you from not being able to do it at home to being able to do it at home. So it's true that there are better ways, there are more expensive ways you could get like an RTX 6000 Pro versus a Mac Studio, and it's going to be a lot faster.

56:48But it's much more limited in the RAM, right? So you can potentially get much, much larger models that will run on a Mac Studio with 500 gigs of RAM versus something like even the highest end RTX 6000 Pro that only has 96 gigs of RAM. So I don't think it matters as much the memory architecture, although they do have higher memory bandwidth than many of the other chips that are available. But the reality is that any CPU can do anything that any CPU can do, right? Or any GPU can do. GPUs can do more things at once. They're faster with certain kinds of tasks like matrix multiplication. but if you look at like a like a pixel 2 right a pixel 2 gets like five tokens per second on many of the new quin models right that doesn't sound like a lot but that's a quarter of a million tokens a day so like how many old phones do you have in a drawer somewhere that you could be getting a quarter million tokens a day out of you can run the newest ai models on like a like those never obsolete e-machines like it's still it's a cpu it's got storage it's going to be really slow but like really any computer would work.

57:58So there was this article recently that came out that said a 10-year-old Xeon is all you need, right? Like a play on the original Transformers paper because those old trash can Mac Pros are like a couple hundred bucks on eBay and they have 128 gigs of RAM on some of them and they have a Xeon and they have two GPUs and they're like small GPUs but it's big enough that you can put the parts of the model that need to go fast on there. So something like OLAMA or llama.cpp, they'll do that automatically. They'll take the encoder, the decoder layers, the KVCache and stick that out in the GPUs, run everything else out of the system RAM.

58:36And suddenly you can run these huge models or even from the hard drive, right? You don't even need the RAM to hold the model. Would you say that this frenzy to buy all these Mac minis and Mac studios is overblown? Because I see so many AI influencers, if that's what we're going to call them. um daisy chaining all of these mac studios together they're buying ten thousand dollar mac studios five of them putting them together you're saying the speed if the speed is not the most important thing in the world to you you can ostensibly run these just you know while you're sleeping and they'll eventually on your pixel too do you think that we don't really need these mac studios in these mac minis it's it's all about how long you're willing to wait yeah it's 100%.

59:17So especially the idea of like daisy chaining, like doing like multi-system clusters to do inference is insane because it's going to run at the speed of your network, which is like slower than just running it off of an SSD. Like it doesn't...

59:30Nilay Patel:Apple wants you to use Thunderbolt, which is reasonably fast, right? That's a PCI level connection. Well, still, I mean, PCI is so much slower than RAM. Fair. Right? Like if you've got it, if you can put it in RAM, that's all you need, right? But at the same time, like you can get superior results with these much smaller and more efficient models that are distilled and focused on the particular task that you're trying to tackle. So I always say like, don't try to cluster for inference. So if better means, can I solve the problem? Can I do the task? You don't need the big models. You don't need to daisy chain multiple systems.

1:00:09You just need to figure out how many tokens a second do I need to reach my deadline or whatever? Like, how fast do I actually need this to go? And I think what I've found is that for most people, it's really a question of their interaction with the model is the speed that they're concerned with.

1:00:26Nilay Patel:Is there benefit to buying cutting edge hardware to run these models? Because that would be why you would buy a new Mac Studio. And we are seeing, I mean, there's a headline in the information. It's like how Apple accidentally won AI. Because to your point from the very beginning, the demand is actually very, very high for these cutting edge systems. Is that being driven by anything or is it just buys? Well, so like the MBA answer is like VCs don't want to pay for CapEx, right? So like all these companies that are, all these startups that are doing AI stuff are supposed to be renting services instead of buying equipment.

1:01:01And so if you want to rent, you have to rent from, you know, any of these inference providers in the cloud versus, I think it's a different, a different perspective is, people, especially at home, people who are technical, maybe they're laid off, typically all kinds of different use cases where there is this feeling that they want to be able to run it themselves for whatever reason. Because you don't see, I think, a lot of software startups doing that. I think it's more consumers doing that. And to some degree, people are buying it. But the difference with CapEx versus OpEx there is that if you're renting in the cloud, you're renting something that's subsidized.

1:01:48So if you have transaction economics where it's like, I want to buy a million tokens as cheaply as possible and get them used in the cloud, you're going to get a lower unit price from renting. but if you're someone who wants to be able to do it at home and you can afford a couple hundred watts of electricity every hour the difference is huge if you can pay 150 bucks and get a p40 card and have competitive performance with the cloud models just with the cost of electricity that doesn't make sense for a company but it makes sense for an individual because the unit economics are totally backwards for capex versus opex considerations uh and the same kind of consideration for a person at home where it's like i don't really care about a little bit of electricity i want 50 tokens a second at home without having to pay 20 bucks a month for it and i think there's there's there's this like particularly in the american um cultural context there's this like sovereignty thing.

1:02:55People want to have a garden. People want to do inference at home. Like it's the same kind of mindset. Like, why am I renting that from some sketchy person? And it goes down when I need it. I could just do it at home.

1:03:07Nilay Patel:Americans want to do inference at home is like a political slogan that will completely scramble the data center debate in this election cycle. We got to get T-shirts printed. 100%. I've heard you talk about your thesis before. I know your other big prediction is that these frontier labs are just not sustainable right that it all of this will come back and be run locally in different ways put a timeline of that for people because i know you've talked about that a lot well maybe like if they were serving the best models we have the chinese models they could have favorable unit economics and i think the question when you look at like for example there was this recent uh report from dirac about open router and you see this like 60-40 swap of like people were using 60 % of their tokens going to the US labs, 40 % to open source models.

1:03:54And it swapped in the last like six months. And so now the vast majority of the market is for these open source models being served on open router by various inference providers at favorable unit economics, which is something the US labs are not capable of doing. so I think it's that same that same divergence between like we want to serve you know trillion parameter models and lose 40 times what we make versus like people who are actually getting serious about inference as a commodity business the question becomes like well okay is it easier for me to to run quen models on open router and pay whatever it costs to do that Or should I buy a trash can Mac and do it in my garage?

1:04:36I think that's sort of the tension that will continue to emerge because these frontier models are the open tiny models at this point, right? And it's only going to keep being true and getting more true.

1:04:47Nilay Patel:All right, we're going to have to have you back after I've built a trash can AI rig in the garage. This has been amazing. Thank you so much for being on the show with us. Yeah, thank you. Nice to meet you guys. Thanks, Nilay and David. We're going to take a break. When we're back, Allison is joining us again. and it's time for the return of the Thunder Round. Propel Fitness Water with Gatorade electrolytes, zero sugar, and vitamins. Propel hydrates better than water to help you get the most out of your workout and get back to your best self. What propels you? Propel with Gatorade electrolytes.

1:05:20Right now, when you buy more carpet at the Home Depot, you save more. During our Buy More, Save More event going on now, upgrade your carpeting with the top brands and save up to 15 % off installed carpet projects. Installation starting at just 49 cents per square foot. Select from a wide assortment of carpet to fit your room, lifestyle, and budget during the Buy More, Save More event at the Home Depot. Offer valid August 27th through September 13th, 2026. Base price for standard installations only. Excludes Traffic Master and Stock Carpet. For licenses, see homedepot.com slash license numbers.

1:05:50Sierra has all the best active and outdoor brands for the super athletic stuff, like running gear for cruising up the trail. Woo! And the super athlete-ish stuff, like fishing gear for chilling by the creek. Nice cast. Fitness apparel to push for higher reps. You got this. And golf balls priced so you can afford to lose one. Or a few. Head to Sierra or Sierra.com for the brands you want at the prices that let you do it all. From athletic to athletic-ish, Sierra's got it. All right, we're back, and Allison is back with us. But you know who's not here? Nilai and David and that means that I have unlimited power once again which means Allison it is finally time for the return of the Thunder Round oh my gosh that is right it's back one night only the Thunder Round returns no one can stop us this is also good because we have a time crunch Allison needs to get to the airport that's right the Thunder Round differs from the Lightning Round in this way we go through stories quickly.

1:06:58There is a timer and our producer, Eric Gomez, will strike us down with the power of thunder if we go too long. Let's do this. Okay, Allison, I have to tell you about the OpenAI math drama that happened this week. Have you heard about this? I have not. This is bananas. Robert Hart has been covering it on TheVerge.com. I encourage everyone to go read that. He's much more thorough than I'm going to be here. But there's this mathematics competition competition known as the Millennium Prize. In the year 2000, they picked a bunch of very difficult and complicated math problems that they would award a million dollars to the person who solved them.

1:07:38And they're supposed to, you know, advance mathematics and encourage new discovery. So OpenAI, not thinking about Millennium Prize at all. And then on September 1st, so about a week ago, they hear some rumors that one of them might be solved or close to solved. And so they go, huh, let's see what we can do. And so they decide to throw like 10 ,000 AI agents at solving all of the Millennium Prize problems that are still open. And they get kind of close on one of them. And then they're like, let's throw the entire swarm at that problem. And then they solve this problem. but in doing so they scoop the mathematician who had made a major advancement here so the mathematician had not solved the entire thing he had solved a sub problem of it but now they're in a beef because the mathematician had used codecs for part of his work and so he's like yo did you guys maybe train on my work in your system and they're like, ah, we can't rule that out.

1:08:51Oh my gosh. It's really nasty. It's really weird. I don't like this at all. First of all, the thought of a math problem you can't solve makes me feel queasy and I don't even want to look at it. That's all of them for me. Directly, yeah. They're very simple math problems that I need AI to help me with. Okay, that's wild. I'm so mad for that mathematician. Yeah, it's really weird. There are lots of like very messy ethical questions here. And I think like the math world is, I think, like rightly kind of outraged right now. It's this weird thing where on one hand, very cool and good that we can advance mathematics in roughly a week from a group of people who don't care that much.

1:09:41And we're just like, let's see if we can do this. Like, that's that's cool and powerful. And on the other hand, the mathematics community is like, hey, normally we work together.

1:09:54Nilay Patel:Right. Normally we share research and support each other. And yeah, that that did not happen here. It's really what. Oh, we're going to. All right. Allison, what have you got? Okay. Well, we're going to talk about the Trump phone. I'm so sorry to our colleague, Dominic Preston, who probably thought he was done writing about the Trump phone. The Trump phone is$250 more expensive. It is a victim of Ramageddon. It was$4.99, which was the launch pricing. it felt kind of unexpectedly generous, I think is how Dom describes it. Because it comes with 512 gigs of storage and 12 gigs of RAM. That stuff doesn't grow on trees.

1:10:44And sure enough, Trump Mobile had to, well, they raised the price. So it's$750 now. $750 for what is effectively a, is it a two-year-old phone? because it came out this summer, but it was supposed to come out a year before. So it's actually like at least a year old or more. And now it's$750 and it was already a bad phone. Yeah. Yeah. That's the gist of it. He does know. So it's still running Android 15, which is a lot of money to pay for a phone running software. It's two years old. But Dom was pleasantly surprised that the phone got a security patch. Yeah, or it was moved forward from the February 2026 security patch is now on the May 2026 security patch.

1:11:43Okay. Which is one more than we thought it would get. The horrifying thing about this is that raising the price implies that people might still be buying this phone. That's not good. We must emphasize, even as a joke, it should not be done. It's not good. Please don't. Please don't. There are so many phones you can, even in this economy, for$750, you can get a good phone. does that mean like we our apologies we bought a few of these we had to find out does that mean we can flip these can we flip these for more money or at least just pull out the RAM I don't know okay Allison are you an over ear noise cancelling headphone fan I feel neutral towards over ear headphones oh my god okay well I think the legends in that space are the Sony WH-1000XMs.

1:12:50It is that lineup. I think they really came in and I think, you know, kicked Bose aside and have defined that space. They're now up to the WH-1000XM6s, but people loved the WH-1000XM4s. And so Sony announced this week, they're going to bring them back they're just like people don't like the six that much uh we'll bring out the fours so they're bringing back the fours at a lower price point which honestly just kind of rules like i'm like super into this like fan favorite gadget re-release yeah i have a a laundry list of digital cameras i'd like uh companies to do that with um let's let's bring back the panasonic gm5 shout out to that one I know I love it it's such a like sorry guys we did screw up here's the one you liked please do that more tech companies yeah I mean this is like just a total win it's like people like this people were John Higgins who reviews headphones and TVs for the Verge he's still like yeah no I would tell people to get the XM4s over the XM6s he's such an XM4 fan now personally I think the design of the XM4 I don't like them as much as the XM6 but I think I'm the outlier here clearly by virtue of Sony bringing them back so I'm going to say this is a win and I'm also going to take it over this win that Eric has not struck me down yet I'm going to pass it back to you Allison close us out what's our last story?

1:14:28Okay, right on cue alright we saved the best for last we are going to talk about the Ocarina of Time so Ocarina of Time remake is coming for a Switch 2 every elder millennial who is like socially awkward and spent a lot of time playing the Ocarina of Time instead of like going outside you know not speaking from experience or anything we're delighted thank you Nintendo so we're going to have this feature of being able to like, rather than just tap the buttons, you know, hum and play the songs that are integral to the Ocarina of Time that I'm sure are ingrained somewhere in my brain matter. But so playing instruments is hard, it turns out.

1:15:23So if you are challenged by by a manual ocarina. This company called Hori is going to have an electronic one that maybe defeats the purpose, but it's a little... It is a physical object shaped like an ocarina. It takes AA batteries, I believe, and will cost$50. But you'll be able to... Almost as much as the game. Yeah, when you put it like that. They're going to sell so many of these. Yeah. I think this is utterly genius. It's so, there's something sweet about this. It's like, you remember playing the ocarina on the, you know, playing it on the controller and you kind of knew it by memory. And then now we're making it like an analog experience.

1:16:14And then, but it's like, no, let's go back to the digital experience, but in a different weird way. I love it. I get that anxiety that's also like talking on the phone in the presence of other people. Like, I don't really want to like interact vocally with a video game. I fully wanted an Ocarina back in the day. I think like probably for the best that I did not get one. Yeah, same. My impression is those things are not easy to play for the person who just wants it as a video game accessory. So I'm like so in on digital Ocarina as an accessory. Super fun. Convenience. Yes. I wonder where it ranks against a recorder.

1:16:59Did you do the recorder in school? No, naturally. Yeah. Yeah. Because that was. But you can't sell as many recorders as specific accessories to a game. They're too widely available. Awesome. All right. Allison, thank you so much for joining us. Talking all about the iPhone. Going through the Thunder Round revival. long-awaited thanks to the fans who have demanded its return. That's it for this week's Vergecast. Remember to subscribe to The Verge for ad-free episodes of all of our podcasts, exclusive newsletters, and so much more. We'd love to hear from you. You can email us, vergecast at theverge.com, or call the hotline, 866-VERGE-11.

1:17:39The Vergecast is a production of The Verge and Vox Media Podcast Network. Today's show is produced by Josh Cajas, Eric Gomez, Brandon Kiefer, and Travis Larchuk. We'll see you next week. Labor Day savings are happening now at the Home Depot with select appliances starting at$399. Plus, save up to an extra$1 ,000 and get free delivery on appliance purchases of$998 or more. Get a Whirlpool laundry tower featuring industry-first UV clean technology designed to reduce bacteria in the wash without fading fabrics. Plus, with great prices at the Home Depot, you can save on select appliances designed to make laundry day easier.

1:18:14Shop Labor Day savings at the Home Depot today. Offer valid August 27th through September 16th. USLAC store online for details. The Viore Core Short moves with you. With everyday versatility and classic athletic fit. It's the one short for everything your day brings. Invest in your happiness and get 20 % off your first purchase at viore.com slash core20. That's V-U-O-R-I dot com slash K-O-R-E-2-0. Exclusions apply. Visit the website for full terms and conditions.