PodSearch

ATP

683: I Didn’t Want to Melt My Rug

 

00:00:00   It's been a, it's been a busy, uh, busy few, few weeks.

00:00:03   All right. Well, we'll talk about that later. And speaking of, we have a lot to cover,

00:00:07   so we're just going to plow right into follow-up and I will start. Um, I don't know,

00:00:11   Marco, if you choose to dub in, we are the champions. Can we use that outside of the

00:00:14   context of the Mac pro? I think we can. Um, but you're welcome world, uh, because we have

00:00:20   apparently caused Apple, just us, nobody else has caused Apple to ship a more sane workouts app

00:00:27   in watchOS 26.4. Uh, I haven't actually tried this myself, but allegedly, uh, in the new version of

00:00:34   watchOS, you can actually press on a workout type to start that workout. You don't have to wait for

00:00:39   the stupid animation for the stupid play button to show up. You can actually just hit the giant green

00:00:44   thing that says, you know, like traditional strength training or whatever, or outdoor run or what have

00:00:50   you. Uh, allegedly that is now fixed. You're welcome world. Yeah, that's, uh, that's very

00:00:55   promising. I'll also, um, a couple of quick notes. Um, when, when I was last ranting about the workout

00:01:00   app, I did, I mentioned that I wish they had some smarter, smarter settings around like reminding you

00:01:07   to unpause workouts when you like, you know, when you start moving again, this actually is a built-in

00:01:13   feature. They do have reminders automatically to remind you to pause a workout. If you've like

00:01:18   stopped moving for a while and to unpause it. And in fact, there's even a feature where we'll try to

00:01:23   automatically pause and unpause it when you start, you know, if you like stop in a, in a run, um, and

00:01:27   then you, you resume again, it will actually try to remind you. What I found though, is that those are

00:01:32   first of all, very annoying. Like when you, if you're actually on a run, I have used those and I have found

00:01:38   them to be just so aggressive of just like, if you stop for a second, like if you stop like at, at an

00:01:44   intersection, tap, tap, you want to pause your workout? Like, no, I'm just, I'm going to be here for

00:01:48   eight seconds. Like no. Um, so I, I don't like those for that. But then also I have found that

00:01:53   in walking workouts, they tend to be a lot less, uh, accurate. And I don't know if they're just

00:01:58   looking for bigger changes in speed. Um, so I have, I have not found the current implementation of those

00:02:03   to be very good. It also doesn't address things like ability to undo a bad GPS point or undo a missed

00:02:09   save. Um, there is however, at least a setting that I turned on the, the end workout button.

00:02:15   there's an option to have it prompt you before it officially ends with a confirmation step that

00:02:19   helps accidental endings become less likely. Um, but it's still, it still needs a lot of just kind of

00:02:26   polish. Like there's what we like about Apple, what got us all to Apple in the first place

00:02:32   was lots of little smart design and lots of little delights throughout much of their user interface and

00:02:38   many of their features of their, of their software and hardware. And the Apple watch seems to have like

00:02:42   none of that. Like it just seems like the Apple watch has was designed in a very rigid way to have

00:02:49   exactly these features, nothing else. And again, part of that was, you know, for hardware constraint

00:02:55   reasons that were very good reasons, but so many of those kind of like little nice delights of like,

00:03:00   Oh, that was smart. And I needed that. Thank you. Those things seem to be mostly missing on,

00:03:04   on watchOS. Um, and so the workout app is, is one example of many where like there's a lot of room

00:03:09   for that. And I wish they would add more of it. What they have is a nice start, but considering

00:03:14   we're like, what, 11 years into the Apple watch platform, something like that. Um, it could be a

00:03:19   lot better. So I hope it continues to get better, but this is a good start.

00:03:22   All right. Let's enter a formula one corner for hopefully just a couple of minutes. Um,

00:03:27   do formula one tracks have corners or are they all like rounded?

00:03:30   They sure do have corners. Although I think you can come up with a better name for this instead of

00:03:33   calling it formula one apex. I don't know. I don't know. I don't know. I don't know enough formula

00:03:37   one terms to come up with something good.

00:03:38   We'll workshop it. But anyways, uh, first of all, I wanted to very briefly, uh, spoil last,

00:03:44   this past Sunday's, uh, China grand Prix and say that this really so far is a very interesting

00:03:49   season to start watching formula one to briefly recap. There's all new rules, all new regulations.

00:03:53   The cars are totally different. And, uh, my favorite team McLaren, they have two drivers as

00:03:58   every team does. And neither of their drivers made it to the starting line. This, this last

00:04:03   race, which stank, but I'm going to claim that that left the space for a very interesting podium.

00:04:09   So the winner of the race, uh, is a 19 year old who races for Mercedes. His teammate was, uh, I believe

00:04:18   second. And then former Mercedes superstar now at Ferrari, who has not done anything good in the last

00:04:24   season or two, Lewis Hamilton, arguably the greatest of all time, potentially, I think by wins, he is the

00:04:30   greatest of all time was also on the podium and it made for a really adorable moment where you have these

00:04:34   two actual current Mercedes drivers and a former Mercedes driver and they're the race engineer for

00:04:40   the winner, which is basically the dude on the other end of the, um, radio was formerly Lewis Hamilton's

00:04:45   race engineer. And so, you know, it was a really, really lovely podium, even though my particular team

00:04:51   didn't even make it. Um, and so again, really good time to pick up formula one, if you're at all

00:04:56   interested, but that I will stop trying to sell it. I'll stop being a shill at this point.

00:05:00   With regard to follow-up, uh, we have a little bit of information about cameras. John and I got

00:05:04   ourselves wrapped around the axle with regard to cameras. When we talked about it last week, Matt

00:05:08   Rigby writes in the TV broadcast, I'd say that onboard cameras are perhaps 15% of the coverage at

00:05:13   most. And they mostly use the track mounted cameras like any other sport, which is exactly true. I should

00:05:17   have made that more plain when we were talking about it. Uh, but again, you know, you can choose if

00:05:22   you'd like to view other cameras, uh, a account that goes by the name, the racing line. Uh, if an onboard

00:05:27   camera fails during a race and no action is taken, the only time an onboard camera failure

00:05:31   has had a direct impact on a race was the 1995 Italian Grand Prix. Ferrari was leading first

00:05:36   and second when Jean Alessi's, I hope I had that right, camera detached from the lead car and

00:05:41   hit Gerard Berger in the second car, breaking the suspension. Whoopsie-dipsies. Alessi's leading

00:05:46   car broke down a lap later with an unrelated wheel bearing failure. Ferrari throwing away race

00:05:50   wins is a constant in F1 history, can confirm. Uh, also, uh, the racing line pointed us to a really

00:05:55   great about 10 minute YouTube overview of all the cameras in the car of which I think there's seven or

00:05:59   eight or something like that. The video said it's very, very interesting, very cool stuff.

00:06:03   Additionally, uh, something I should have talked about when we were discussing cameras and whatnot,

00:06:08   uh, Eric Fox reminded me that one thing you didn't mention was that the other, that the other two might

00:06:13   find interesting as in Marco and John, uh, for follow-up is how the driver feeds also give you

00:06:18   access to their radio communications with their respective teams. Of course, the best radio messages,

00:06:22   red, the whiniest, uh, get put on the main broadcast, which is true. And I don't think I mentioned that

00:06:27   at all. And that's a really great point from Eric. And that can be very fascinating. Normally it's not

00:06:31   that interesting, but it can be fascinating. And then finally, with regard to F1, uh, Brendan Webb

00:06:36   writes, I'm not sure why Apple hides the Sky Sports replays on the main navigation, but you can search

00:06:42   for Sky Sports China or whatever race you're looking for and they magically appear. So I should have said a

00:06:46   little context. I had said in the last episode that, uh, as far as I could tell, at least in replays,

00:06:51   there's no way to get the Sky Sports feed, which is mostly about the commentators. Um, I, I like the,

00:06:58   the Sky Sports commentators. I find them to be egregiously, uh, UK biased, you know, Lewis Hamilton

00:07:04   and George Russell, you know, and, and Lando Norris can never put any foot wrong anywhere for any time.

00:07:10   And that's, I mean, hi, I'm American. I know how that goes, but, uh, it gets to be a little

00:07:14   much from time to time. Well, anyways, a lot of people prefer it because that's what F1 used to,

00:07:18   or excuse me, that's what ESPN used to broadcast. Well, apparently instead of just using it as like

00:07:22   a different audio track, which I think is all it really is, Apple treats it as an entirely different

00:07:27   broadcast. But I think the only way I've been able to find it is exactly what Brendan says. You have

00:07:32   to actually search for Sky Sports, whatever, in order to find it, which again, is just bananas to me.

00:07:39   I really think that Apple does have a lot of potential here for F1 coverage and being,

00:07:45   you know, the American partner for F1 coverage, but golly, it's been slow out of the gate,

00:07:49   but what are you going to do? All right. Let's talk Rosetta. Colin McKellar writes with regard

00:07:54   to processor emulation and how long it lasts. I did some Wikipedia diving and put together this post

00:07:59   about how long Apple's processor emulation lasts on the Mac. This looks like a three mile long post,

00:08:04   but it's actually a bunch of charts and stuff. It's pretty good stuff. Um, and John, I think you've

00:08:09   extracted some stuff, which we'll talk about in a second. Yeah, that's Colin continuing here.

00:08:13   Yeah. So Colin continues and writes, in short, 68K emulation was supported on PowerPC for as long

00:08:18   as classic macOS existed, including when it lived on as the classic environment in macOS 10. PowerPC

00:08:24   emulation, the original Rosetta, was around for five and a half years after Mac's transition to Intel.

00:08:28   According to Apple's present plans, Rosetta 2 will last about 18 months longer than Rosetta 1.

00:08:34   Yeah. And then I pulled out the, some more interesting info from, uh, this post, which is how long, uh,

00:08:40   each architecture was the current architecture for the Mac. So 68K was the, the architecture for the Mac

00:08:47   for 10 years and seven months. PowerPC for 11 years and eight months and Intel for 14 years and four months.

00:08:53   And so far, Apple Silicon has just been five years and four months. So Intel, uh, the Mac has been Intel

00:08:58   longer than any other processor. Um, and maybe that's why I feel like, uh, even though they're,

00:09:03   according to these, uh, stats, it's going to be 18 months longer than Rosetta 1 lasted. It just feels

00:09:08   shorter because Intel has been around so long. Maybe it's just because of the prevalence of Intel

00:09:11   software, like going to Intel opened up this whole new world of stuff that became easier to compile for

00:09:18   Darwin essentially. Um, but yeah, uh, it's, uh, you know, I, I, my recollection of 68K to PowerPC is

00:09:26   basically what they said. Like it basically never went away. It didn't go away until classic Mac OS

00:09:31   went away. Even then it was classic, classic environment inside Mac OS 10. Uh, but PowerPC to

00:09:36   Intel, PowerPC didn't stick around that long at all. The other thing was Intel was so much faster than

00:09:40   PowerPC. And I guess that's also true of, uh, Apple Silicon with, uh, Intel. So anyway, um,

00:09:45   it's really is just a feel thing. Like it just feels like it's a little bit too soon because

00:09:49   circumstances are different, but it will in fact be 18 months longer than one of their transitions,

00:09:53   but shorter than the other. Then related to Tom Obarsky writes on Reddit. I saw that a user received

00:09:59   a pop-up warning of upcoming deprecation from a fresh install of final cut pro non-subscription on a brand

00:10:06   new Mac book Neo. That's something else. And so, uh, we'll put a link to the Reddit post, but the little

00:10:10   notification reads support ending for Intel based apps. This version of final cut pro will not open in a

00:10:15   future release of Mac OS. Learn how to update on to an Apple Silicon version.

00:10:19   That's a terrible message because I'm assuming it's not final cut pro is not the problem.

00:10:23   Uh, anyway, you can keep reading.

00:10:24   Yeah. So, uh, Tom continues. There's some speculation as to whether this is because the

00:10:29   A18 pro is not seen as an M chip and therefore is assumed it must be Intel or whether a component

00:10:34   within an FX library could still be flagged as Intel and thus would invalidate the whole install

00:10:39   weird edge case to say the least. If I were to guess the, hi, this is Casey. Uh, if I were to guess,

00:10:44   I would say that it's probably a library or something like that, but I mean, I don't think

00:10:48   that the, the A18 not being seen as an M chip is definitely wrong. It's plausible for sure.

00:10:52   I don't think that's very plausible. I mean, there's a bunch of ways to look up the architecture

00:10:57   and they, you know, the answer is, is it arm or is it not arm? And the A18 pro is very clearly

00:11:01   arm. Uh, if you were to look it up on a phone that ran it, I think it would, the answer would

00:11:04   come back as arm. So I think it's got to be a library or a plugin or something. And speaking of

00:11:09   that, I don't have the notes, but Apple just bought whatever that, uh, that Final Cut Pro

00:11:12   plugin company. Oh yeah. Yeah. I do wonder if they're, uh, you know, the best way to make sure

00:11:17   that any Intel only plugins are no longer Intel only, just buy the company and make sure they're

00:11:22   all Apple Silicon before the deadline.

00:11:25   We are sponsored this week by Lisa. So a few months ago, I was looking to get a new mattress

00:11:31   for Michaela because we bought something unremarkable off of Amazon. And despite being well-reviewed,

00:11:35   it's a pile of trash. I felt terrible about it. And I, at this time that I was looking about for a good

00:11:41   mattress for Michaela, Lisa reached out and said, Hey, would you like a mattress so you can talk about

00:11:45   it on the ad read? And I said, yes, please. And I got Michaela a natural hybrid mattress and then

00:11:53   immediately regretted my choices because I wanted a natural hybrid mattress after sleeping on it for

00:11:58   a night or two. Uh, it is genuinely very, very good. And so what makes it good? Well,

00:12:02   Lisa has a whole lineup of beautifully crafted mattresses tailored to how you sleep. Each mattress

00:12:07   is designed with specific sleep positions and feel preferences in mind. Their mattresses are

00:12:12   meticulously designed and assembled here in the United States for exceptional quality. Plus they back it

00:12:16   all up with free shipping, easy returns and a 120 night sleep trial. Lisa isn't just about sleep

00:12:23   either. It's also about impact. They donate thousands of mattresses each year to those in need while also

00:12:27   partnering with organizations like clean hub to help remove harmful plastic waste from our oceans.

00:12:31   And who doesn't like that? So like I said, we got Michaela a natural hybrid mattress. This features

00:12:36   some moisture wicking organic cotton that apparently is from New Zealand, uh, on the top of it,

00:12:42   which keeps you cool, but also warm all at the same time. And it gets that sweat away from you.

00:12:45   It is genuinely very, very good. And I'm pretty darn sure a Lisa mattress is in my future. So

00:12:51   if you like me are thinking about buying a Lisa mattress, go to Lisa.com for 20% off mattresses. Plus

00:12:58   get an extra $50 off with the promo code ATP exclusive for our listeners. That's L E E S A.com promo code

00:13:04   ATP for 20% off mattresses. Plus an extra $50 off support our show and let them know that we sent you

00:13:11   after checkout Lisa.com promo code ATP. Thank you so much to Lisa for sponsoring the show.

00:13:17   Some teardowns have been released recently. Uh, the iPhone 17 E has an iFixit teardown. We'll put a

00:13:26   link in the show notes and it appears that a lot of the parts are the same as we saw in the 16.

00:13:31   Yeah. And interestingly, one of the things they pointed out is that, um, say you've got a 16 E

00:13:35   and you wish it had MagSafe, you can just put the 17 E back on there with MagSafe.

00:13:40   And it mostly works like, so there's a few points, like it doesn't, so MagSafe will attach to it and

00:13:47   charge it, but you don't get the little like animation, you know, the little circle filling

00:13:51   thing or whatever it is because like, there's a little, it's, it's a little bit of a Franken phone,

00:13:54   right? It's not, not a Franco phone. So to be clear, so, so the 16 E did support Qi charging,

00:14:01   just not MagSafe. Right. But that means, but the, the animation that you get for MagSafe is the,

00:14:06   like the green ring that appears or whatever, but it will charge it. So anyway, many parts are,

00:14:11   I think basically every part is the same except for the, like the, the logic board or whatever. And

00:14:17   they, they, they made a phone out of a mixture of a whole bunch of parts. They were very excited

00:14:21   about the part interchangeability and the fact that they didn't get hassled by the software for like

00:14:25   parts pairing and stuff like that. But, uh, yeah, 17 E is just a 16 E with a different logic board and a

00:14:31   different SOC and a magnet on the back. All right. Let's talk about MacBook Neo. Uh,

00:14:35   first of all, Gruber had a post with regard to the on camera and also Mike, uh, uh, on screen

00:14:41   indicators and, uh, Gruber linked to Apple's platform security guide, which states the MacBook

00:14:47   Neo combined system software and dedicated Silicon elements within the 18 pro to provide additional

00:14:53   security for the camera feed. The architecture is designed to prevent any untrusted software,

00:14:57   even with root or kernel privileges in Mac OS from engaging the camera without also

00:15:01   visibly lighting the onscreen camera indicator light.

00:15:03   And the context here is that the MacBook Neo doesn't have a little green led next to its

00:15:08   front facing camera. Like all the other MacBooks do instead, they have, um, an icon in the menu

00:15:14   bar, like the green little, like FaceTime, the icon, and also like a green dot in the corner

00:15:20   of the screen. And that dot is basically a stand in for the light. And last time we mentioned it,

00:15:25   I'm like, I'm sure Apple has done something to try to make that more secure, but it's obviously

00:15:29   more difficult when you're trying to secure software versus trying to secure it in hardware,

00:15:33   where in theory, the led, um, if the camera's on the led is on business, like electrically hardwired

00:15:40   connected to it. And you'd have to stop that from happening. You'd have to, uh, get physical access

00:15:45   to the, uh, laptop and cut a, you know, trace somewhere to prevent, uh, you know, that from

00:15:50   working. And even that might be difficult. Although there was a bug with that. Um, I couldn't find this

00:15:55   link, but the first time Apple did this, they tried to say, and the light will always be on as

00:16:00   long as the camera's on. And they messed up the implementation, but that was a while ago. The

00:16:03   current implementation is pretty solid, but that's all in hardware. So how, what do they do to try to

00:16:06   make the software more secure? So this first bit from Gruber is there from Apple security document

00:16:11   saying, even if someone like, you know, puts a root kit on your machine, they have super user

00:16:16   privileges. They have root kernel level access to your Mac. Uh, still they can't stop that green

00:16:23   dot from appearing in the corner. Indeed. So Gruber got talking with friend of the show,

00:16:27   Guy Rambeau and Guy wrote to Gruber, the software-based camera indicator light in the MacBook Neo

00:16:31   runs in the secure X-Clave part of the chip. So what is almost as secure as the hardware indicator

00:16:36   light. It runs in a privileged environment separate from the kernel and blitz the light directly into the

00:16:42   screen hardware. All of that applies to the mic indicator as well, which is a bonus compared to the

00:16:46   camera only hardware indicator. And then Guy also provided a little footnote. X-Clave's run on a

00:16:52   completely isolated real-time operating system that communicates with the kernel and user space

00:16:55   using a very limited API surface, not to be confused with the secure enclave, which is a

00:17:01   totally different thing. Yeah. So to actually stop the light from working, they, they need more than

00:17:06   just access the kernel level access. They need to hack the X-Clave, which is a lot harder to do

00:17:09   because it is very limited. It is a whole separate thing running a whole separate operating system.

00:17:13   And it talks to the main operating system, but it doesn't go through the main operating system.

00:17:18   It seems like to get those dots onto the screen, it just does rise them directly into the video

00:17:22   hardware, old school style. So that's pretty cool.

00:17:24   Yeah. Yep. And then also Gruber linked to random Augustine's post about the Apple X-Clave's,

00:17:30   which you can read over on medium. If you can hold your nose to long enough to read something on

00:17:35   medium. All right.

00:17:37   Hey, do you want to log into medium?

00:17:38   Oh God, it's the worst. CPU benchmarks. Apparently we got something a little bit wrong. John,

00:17:43   can you tell us about that please? Yeah. The iPhone 16 Pro Max with the A18 Pro, which is what I was

00:17:47   using for CPU benchmarks because the Neo had not been released to anybody yet, has 46% higher single

00:17:53   core in Geekbench, not 30%. I just plain did the math wrong. So it was 30, 34, 28 versus 23, 47. And

00:18:00   I think those are the numbers from the earlier episode, but I just didn't do the math right. So 46%

00:18:05   faster in single core, not 30. Thanks to Philip Sommer for the correction.

00:18:09   Then with regard to SSD benchmarks, Vito Treino writes, your MacBook Neo SSD benchmark figures in

00:18:15   the last episode came from the Verge's benchmarks, which only measured sequential speeds. While those

00:18:19   are important for large file transfers, random speeds are a better reflection of using the computer,

00:18:23   you know, OS boot, logging in, launching apps, et cetera. Sequential speeds may be more visible to

00:18:28   the user in finder progress bars, but random speeds are arguably more important to the user experience,

00:18:33   contributing to the feeling of snappiness and ultimately the longevity of the machine.

00:18:38   This seems to me to be a far more important metric for the target market of the Neo. I would agree with

00:18:42   that. Well, so before we go on, so obviously, um, then the Neo has an SSD and not a spinning disc back

00:18:49   when spinning discs existed, uh, random access versus sequential access was a key way to measure

00:18:56   performance because, uh, as Vito noted, um, you're doing random access when you're doing most stuff.

00:19:03   You're mostly not just copying one giant file from location A to location B, although even that

00:19:07   could be random depending on fragmentation on a spinning disc or whatever, but spinning discs had

00:19:11   to move actual disc heads on an arm to get to the track where the data was, wait for the arm to settle,

00:19:18   wait for the, this, the sector that you want to spin underneath the heads of the track and then do

00:19:23   that all over again, back and forth and back and forth, making all those lovely noises that we remember

00:19:27   from the spinning disc days.

00:19:28   And so random seeks on a spinning disc were murder because you spend most of your time waiting for

00:19:34   physical items to travel, to be aligned and settled. And then a brief reading of some magnetic

00:19:40   information. And then you start the whole process over again. Most of your time was burned on overhead

00:19:44   SSDs. On the other hand, do not have any moving parts in that way. They don't have arms that move.

00:19:50   They don't have discs that spin. They never have to wait for an arm to move to a different section of

00:19:55   a disc and they never have to wait for a disc to rotate several more degrees and they never have

00:19:59   to wait for the things to settle. And so you would think, well, isn't random access on an SSD

00:20:04   basically a constant time operation? Like it doesn't vary. It doesn't care if you're reading

00:20:08   address zero or address 4 million. It's, it's all just chips. Well, there are things that can affect

00:20:15   random access because the chips have to be read in certain size chunks from certain regions at a time

00:20:19   and so on and so forth. But so that's why you still do random tests on SSDs. But the difference

00:20:24   between random and sequential is not nearly as pronounced as it used to be. And what most people

00:20:30   feel like they're waiting around, when do you feel like you're waiting around for a disc?

00:20:33   It's when you're copying a big file and you see a big progress bar. Like if you're copying some huge

00:20:37   movie from one disc to another or whatever, if you're not limited by your internet download speed,

00:20:41   or you're not limited by the other disc, um, especially now that with APFS that you get the,

00:20:47   uh, you know, the clones, like when you duplicate a gigantic file on APS, it's like instant, right?

00:20:51   Because it doesn't actually copy the data. So that is entirely eliminated, but copying to and from

00:20:56   multiple volumes or to and from multiple, you know, physical disc mechanisms. That's when people say,

00:21:02   oh, shouldn't this be going faster? It's two SSDs. And that's where you'll see the difference. So I agree

00:21:08   that random and sequential are not are two separate things, but I really think that sequence, the reason

00:21:14   people show sequential is, well, first of all, it's the number that's going up. So it's fun to show

00:21:18   benchmarks. Look how much faster it is. Uh, and then also I think that these days with SSDs, that's the main

00:21:24   time people feel like they're waiting. Like I wish my SSD was faster because I'm copying this giant video

00:21:30   file from this external SSD to my internal one. And I'm looking at a progress bar and I'm looking at my watch.

00:21:35   All right. Uh, so V Vito continues. I've only found random seek benchmarks in one review,

00:21:41   which is Andrew Mark David's review at 11 minutes, 17 seconds. We'll put a timestamp link in the show

00:21:46   notes, but the random speed seemed to be on par with what we've seen from the M1 MacBook air.

00:21:50   They also, from what I can tell, don't seem to be too differ much across the entire Apple Silicon

00:21:54   lineup. So I looked at that and I think they do differ across the Apple Silicon lineup. So we'll put a

00:22:00   link in the show notes to the, the benchmark app, which I had never heard of, but you can download and try

00:22:04   it if you want. And then I graphed the numbers for the R and D 4k QD 64 read and write and R and D 4k

00:22:13   QD one read. I don't know what that means. I'm assuming random 4k, but I don't know what the QD

00:22:17   means, but anyway, two different read and write benchmarks for, uh, three machines, the MacBook

00:22:23   Nioh 256, the M1 MacBook air 256 and the M5 MacBook pro 512. And especially on the R and D 4k QD 64,

00:22:32   uh, read and write tests, the MacBook pro is way faster. It's like twice as fast. And these were in

00:22:38   random seats as the Nioh. So I wouldn't say that at random, they're all about the same. I mean, granted,

00:22:43   when you get down to the, the, whatever, the R and D 4k QD one, right one, maybe they're a little bit

00:22:47   closer, but look at, look at those graphs. I'm going to say there's still a, uh, appreciable

00:22:52   measured different difference between, uh, the fastest and the slowest. So for example, I don't

00:22:57   know what these numbers are, but they come from the disc benchmarks, but the, the Nioh is 582 and the

00:23:02   M5 MacBook pro is 1180. That's close to double. Um, and similarly on the, on the other, like the

00:23:10   closest benchmark is 31 to 45, still somewhat substantial difference. So yeah. Um, I, I agree

00:23:16   that people should be still testing random, not just testing sequential, but I think just people,

00:23:20   especially in this YouTube, I've complained about this all the time because my entire childhood was

00:23:25   spent arguing about benchmarks and nobody argues about benchmarks anymore. We're just YouTubers just

00:23:29   used Geekbench and like, well, the number says what the number says and nobody. So I, I applaud

00:23:34   Vito to be out there saying these benchmarks are BS. You're just measuring sequential. There's much

00:23:38   more to the story. I agree. And there's much more to the story than random. It's like, well, random,

00:23:41   I don't care about your benchmark. Your benchmark has this problem. We need to test the launching of

00:23:45   Photoshop, but not that version of Photoshop. And yeah, uh, that's the culture I come from. So I applaud

00:23:50   Vito for calling it out here, but I will say that I think random, random access still differs

00:23:54   between the less expensive and more expensive, uh, SSD MacBook Pros or MacBooks.

00:23:59   All right, John, you went on a field trip.

00:24:01   Yeah, I was, happened to be in the mall. Um, well, I actually, I took the trip to the mall. I was

00:24:06   driving my son there, um, get his haircut. And I, the reason I drove him is so I go to the Apple store.

00:24:11   I went to the Apple store to look at all the new stuff. Um, and I ran right to the MacBook Neos

00:24:18   and I don't, I don't know what I was expecting. Like I had this image in my head of the product.

00:24:22   Obviously I'd seen it on Apple's website and we talked about it on the show and I knew all these

00:24:25   things about it. Um, I just, I, I didn't, I didn't go in with any explanation of like, I'm just going to

00:24:31   go in and look at them and they're going to be exactly like how I thought they would be. And they

00:24:34   weren't, I picked it up and played with it and touch it. And I'm like, they have pulled off something

00:24:41   with this product that is tricky to do, which is, I mean, think, think about the laptops. What are

00:24:47   they? They're, they're like flat rectangles made of metal, especially when they're closed and you're

00:24:52   not using them. Like what is there to the flat rectangle made? How many flat rectangles made

00:24:57   of metal has Apple made since the unibody laptops started all those years ago? Just so many of them.

00:25:03   Surely they're all the same at this point. Is there any room within the flat plank of aluminum

00:25:09   to do anything literally? Uh, and I think there is because they've turned, I mean, I don't know if

00:25:17   they did this on purpose or whatever, but like one of the things that separates the Neo from its more

00:25:23   expensive brethren is, and we'll get to this in a little bit. Um, this, the screen, the screen part

00:25:29   of it is not as skinny as it is on the other models because it's more expensive to make them thin,

00:25:35   I guess, right? It's, it's a little bit fatter in the screen part. It's a little bit fatter overall,

00:25:40   uh, like a millimeter. And one of the things you can do when the screen part is a little bit fatter

00:25:46   than it might be on like the old MacBook air that came to like a really sharp, like the M1 MacBook air

00:25:50   comes to a really sharp point at the, at the part that you pick up, uh, like the lid, you know,

00:25:54   the, the screen lid. One of the things you can do when it's a little bit thicker is you can put a bigger

00:25:59   corner radius on all of the corners. And I think, and also by the way, this, the Neo is,

00:26:06   I forget the exact measures, maybe like a quarter inch, a half an inch, uh, less wide and deep. You

00:26:12   know what I mean? Like if you, if you put this on top of, uh, a M5 MacBook pro, you'd see the M5

00:26:19   MacBook pro sticking out around the edges. Cause this is smaller. So this is a little bit smaller

00:26:23   and it has a bigger corner radius on all the corners. And the result of that is picking this

00:26:30   thing up and just handling it as a thing that you tuck it under your arm and carry it to your,

00:26:35   you know, I don't know, to the library or to your class or whatever. It feels so good and so solid and

00:26:43   so friendly in a way that the more professional laptops do not, because the more professional laptops

00:26:49   are sharper, they're thinner a little bit, but they're also sharper, like literally sharper on

00:26:55   the edges, uh, because that's just the way they're designed. Rounding this thing over makes it feel

00:27:02   pleasant and solid and friendly and approachable. And I know this sounds so dumb. It's like, it's a

00:27:07   rectangle of aluminum. How friendly and approachable is an extra, you know, millimeter. I'm telling you,

00:27:12   this was my impression upon picking it up. I'm like, this thing feels great. And it feels great in the

00:27:18   same way. The only analogy I can have is like the, um, the iMac G4 with the little, uh, like the,

00:27:22   the metal arm with the floating LCD display on it. Remember that one, that arm looked and felt so good

00:27:30   on that machine that you were like, I can't believe that if I pay this amount of money, I get this.

00:27:35   And that's the impression I got from the Neo that like this thing feels like you're getting more than

00:27:41   you paid for. You're getting more than your money's worth. It feels solid and expensive and

00:27:47   nice in a way that I did not expect it to. I don't know. I don't know if I expected it to feel janky or

00:27:53   something, but it like, I know going on about this and you're like, so it had rounded corners and you

00:27:57   were wowed by that. Who cares? This, this is my impression that they really knocked it out of the

00:28:02   park that it feels solid. And the second part of this is the trackpad, which I, I did know going in,

00:28:08   I was thinking this trackpad is going to be a little bit janky. Nope. It's, it's great. Like

00:28:13   maybe it will get bad over time. I don't know. This is a floor model. And like the first day it came out

00:28:17   or the first week it came out feels so good. I pressed everywhere on that thing. It is pleasant

00:28:22   and easy to click everywhere on the thing. And it doesn't feel loosey goosey and wiggly and tilty.

00:28:28   It feels great. So I think if you got somebody, one of the, a Mac book Neo, they are going to be so

00:28:34   happy, especially if they've never had a Mac before and they've only had PC laptops. Oh my God, this is

00:28:38   so much better than I thought it was going to be in terms of fitting to finish. And I don't know why I

00:28:43   had these, these doubts in my mind, but I was blown away by it. And obviously it's still got eight gigs

00:28:47   of Ram and a bad screen, blah, blah, blah, but it's like, it's $600. So Mac book Neo, they think they

00:28:52   knocked it out of the park with this machine. Coincidentally, I happened to also go into an Apple

00:28:58   store yesterday and also for, with the goal of like, Oh, like I was literally, I was walking past it.

00:29:04   I'm like, Oh, I should go in and check out all this new stuff. The first thing I did was walk over

00:29:07   to the Mac book Neo. I must've looked very strange to the staff because I was basically like,

00:29:12   fondling it like a different edges, like trying to figure out like, at what point is this going to

00:29:19   feel cheap? And I did not find that point. Yeah. Yeah. Like it. So I was going over all the corners,

00:29:27   all of the, like the, the rounded edges, the sharp edges, the finished, I, I, you know, I, as I often

00:29:33   do in Apple stores with, you know, picking up new products, I closed it and I moved it around and I

00:29:37   like, what is the bottom of like, what are the feet? Are the feet worse? I love even the, even the feet,

00:29:41   the feet on the Mac book air are sharp little cylinders. Like they're not domes. Like they were

00:29:46   on the M one Mac book air, the current M, you know, M two M three M four Mac book air. The feet are kind

00:29:50   of like sharp edge cylinders. The feet on the Neo are rounded and friendly with flat bottoms. Amazing.

00:29:56   I move my hands over every inch of that thing. And I, like I would, I opened it up. I felt even like

00:30:03   the, the screen, like the interior screen hinge, like the, like above the keyboard. I even felt that,

00:30:08   like, where's the sharp edge. Where is the sign of cheap manufacturing? I could not find one.

00:30:16   It just feels like any other Apple laptop. Like I even, I forget, I'm sorry. I forget who it was.

00:30:22   It was somebody. I think I massed on somewhere. I said a few days ago, like, Oh, you can tell there's

00:30:26   like a different, like a, not as good of a grain pattern. I looked, I couldn't tell. I ran my hands

00:30:32   over the flat parts. Does it feel different? No, it doesn't. I picked it up, moved it around. Is the

00:30:37   weight balance off? Does it, does it feel less solid? No, it just feels and looks great. I was

00:30:47   shocked. I really thought there would be some kind of like noticeable lower quality level about it

00:30:53   compared to the Mac book pro and Mac book air. There isn't, I was not able to find one at least in the,

00:30:58   in the Apple store with a good few minutes of it. Yeah. And I actually, I hope they take some of

00:31:02   these decisions and bring them to the Mac book pro. I'm not saying they need to make it as rounded as

00:31:05   this. Like I do like the thin lid on the Mac book air, micro pro stuff like that. But like some of these

00:31:11   decisions are just, they're not, obviously they're not more costly. In fact, they save money, but

00:31:15   they're just better decisions. Like at various times we have complained on this show about exactly how

00:31:19   sharp certain edges are on Apple's laptops. And I think the Neo shows that if you're willing to round

00:31:25   things over a little bit more, it makes for a more pleasant to handle machine. It also helps that

00:31:29   this is smaller and it does look like a little baby because it's not like 11 inch Mac book air size.

00:31:34   It's not, you know, the Mac book, uh, adorable Mac book one size. Oh, that was very sharp as well.

00:31:39   But the fact that it is a little bit smaller, I think is going to make it even more attractive

00:31:43   to somebody who doesn't want to spend two grand on a laptop. They just want a nice laptop

00:31:48   and not spend too much money. And this boy is just a nice laptop without spending too much money.

00:31:53   Yeah. I, I was blown away and I, and of course, you know, I, I opened it up and I, I did similar

00:31:58   tests as John. Like I was put, I, you know, knowing how the trackpad works, I was clicking,

00:32:02   not only clicking the trackpad, but also I was like, I tried to click it like on the corners,

00:32:06   on the edges. Like, does it feel bad if you click like diagonally from a corner? No.

00:32:12   Cause it used to like, I have, I have some Mac book pros with mechanical clicking trackpads that

00:32:18   did not feel good. And this does not have any of those problems. Yeah. It, I was shocked. And of

00:32:24   course, you know, opening up apps and I even tried like the keyboard and I was like, Oh, the keyboard

00:32:27   feels a little bit, I thought it felt like a little bit squishier, but then I went over to a Mac book pro

00:32:33   to compare and that felt exactly the same thing. No, that's, that must be just in my head.

00:32:37   Yeah. I think it's like the same keyboard mechanism. So yeah, it's, it felt like it. Um,

00:32:41   on the colors, I thought they all looked pretty good. The, the yellow and the pink were more pale

00:32:48   than I expected. My favorite color was actually the blue. I thought the blue would be a lot closer

00:32:54   to the Mac book air midnight color, which is so dark. It's almost black. Yeah. It's lighter than that

00:32:57   though. It's way lighter than the Mac book air. Um, and it's, it's not a light blue. It's still,

00:33:02   it's still like a dark blue, but it's, it's more like a, like a denim jeans color, like jeans color.

00:33:06   Yeah. And I, I really, I thought the blue Neo was a really nice looking computer. Um,

00:33:12   and if I were to need one of these, which I don't, but if I were to need one of these,

00:33:16   I would almost certainly go with the blue. Um, but I also respect people going for the yellow

00:33:20   cause that's, it's a really fun color. Um, but the blue is a little more of my style, but

00:33:24   I was blown away both from just sheer pride in this platform of the Mac that I love so much

00:33:32   that so many more people are going to be joining us.

00:33:36   it's going to be like, finally, like all the people who, who weren't able to afford

00:33:41   the other Macs, like we're bringing more people into the fold. That's a great thing for the platform.

00:33:47   As I said, it also made me feel so impressed by both Apple's hardware prowess and also just

00:33:55   how great computers are right now that everything the Mac book Neo can do all of the like high end

00:34:03   pro apps that like it can run. It might not be as fast as everything else, but like it can do it.

00:34:08   Your iPhone in your pocket can also do that. What a time to be in computers. Like, yeah,

00:34:16   it's not all roses. There's some problems that we need to work out. That's always going to be the

00:34:19   case. It always has been the case, but wow, what an amazing time we are in for hardware that not only

00:34:26   is it possible to make a laptop like that, that is that good for that price. And it's not like Apple's

00:34:31   like not taking any profit on this. You know, Apple doesn't do that. You know, they're making 20 or 30%

00:34:36   at least margins on this thing. And to also have all that power in our pockets all the time is an

00:34:43   incredible resource. Like we have amazing supercomputers available to us all the time and for not even that

00:34:51   much money. Like that's wonderful. And I think this product is a hit. And I think this is like when you,

00:34:57   when you look at what PCs do in this price range, Apple's going to kick their butts and it, they honestly

00:35:05   could use the, the butt kicking. So this is, this is just fantastic. I'm blown away by how good the,

00:35:10   the, uh, MacBook Neo was. Um, I also briefly got to handle the iPhone 17 E. It feels great in the hand.

00:35:16   It's just like, you know, just like the 16 E. Like you feel it and you're like, wow, that's really light.

00:35:21   That's awesome. Do I want it? No, but it's, but it's great for people who want it. Um, and I walk right

00:35:28   out the store and I didn't even realize until today, I totally forgot to even look for the studio displays.

00:35:34   Oh, come on. I was so blown away by the Neo. I was like, all right, you know, mission accomplished.

00:35:40   I'm out of here. Walk right out. I didn't, I didn't forget. I looked at the studio display XDR,

00:35:45   although unlike a Casey store, they had these back to back. So I couldn't see them at the same time.

00:35:50   Not that I needed to, but I played with all the settings on it. Um, I was looking for the thing

00:35:55   that lets you pick between like the reference mode and whatever. And I couldn't find it in my brief

00:35:59   looking there, but I honestly, I didn't know where to look. It wasn't, I didn't see it in the displays thing.

00:36:03   Um, they don't have a lot of, well, there's two problems. The Apple store is really bright.

00:36:07   So it's really tough for even 2000 nits to show, you know, cause they, the light sensor,

00:36:11   the ambient light sensor is going to put the monitor already at pretty high brightness.

00:36:15   So the best you're going to get is like two X because if the, if the monitor is at a thousand

00:36:19   nits and HDR is 2000, it looks less impressive than if the monitor is at like 300 nits because

00:36:25   you're in a dim room and then the HDR comes on at 2000. But anyway, their sample photos and the

00:36:30   photos app didn't really show off HDR that well. So you'd be forgiven Casey for not being able to see

00:36:34   what's going on. Um, I, they also didn't have it in the mode. They had it in P3, like the,

00:36:39   the display profile. They didn't have it in the combined P3 and Adobe RGB mode. I switched it to

00:36:45   that mode and then the screen blinked and then it blinked again and it had switched itself back.

00:36:50   I'm like, well, there's some bugs. And then I switched it back into the mode again. Eventually

00:36:53   I got it to stick. I couldn't see any difference, but I just, it was interesting that that wasn't

00:36:56   the default mode for the stuff, but otherwise it looks like a studio display, but it has better

00:37:00   black levels and it's brighter. Cool. It costs twice as much. Yeah. I mean, I, did you, did

00:37:06   you notice the refresh rate at all? Like when you were scrolling a webpage or something? I

00:37:10   mean, yeah, I could tell it was 120 Hertz, uh, for scrolling stuff, but I didn't like, that's

00:37:14   why I put it in, put it in adaptive, but I didn't know how to really test that. But yeah,

00:37:18   it's, it's, it's high refresh. It's a good monitor. It's a good monitor. The only, the only

00:37:21   thing it has going against it is it's not 32 inches and 6k. Everything else about it is

00:37:25   very good. No, the only thing it has going against it is that it's $3,000, whatever the

00:37:30   heck the price is. Well, I mean, proportionally given the difference in size and pixels,

00:37:34   that's not actually that bad compared to the pro display XDR.

00:37:37   We are sponsored this episode by Zapier. Zapier is how you actually deliver on your AI strategy,

00:37:45   not just talk about it. You know, these days there's AI everywhere. And honestly, the biggest

00:37:50   problem that most people have with AI is just figuring out how to use it and figuring out where

00:37:55   to integrate it into their workflow. So Zapier has this amazing AI orchestration

00:38:00   platforms. You can bring the power of AI to any workflow. So you can do more of what matters

00:38:05   and have AI help you out in all the other ways. So you can connect top AI models like chat GPT

00:38:09   and Claude to the tools that your team already uses. So you can add AI exactly where you need

00:38:15   it, whether that's AI powered workflows, autonomous agents, customer chatbots, or whatever else you

00:38:20   can imagine. You can orchestrate it all with Zapier. And it's easy to use. Zapier is for everyone,

00:38:26   whether you're a tech expert or not. So many teams have already used AI tasks with Zapier. Over 300

00:38:32   million tasks have been automated. And Zapier is just such an amazing platform. You know,

00:38:36   they've been around for a very long time in the automation space, way before AI. They're a very

00:38:41   mature platform. And they have always specialized in automation. Obviously, AI has huge advances in

00:38:48   automation possibilities. So Zapier is right there helping people actually deploy AI to their real

00:38:54   work across their real company with their real tools, all happening right now. This is not something

00:38:59   you just do in the future. This is something you can do right now with Zapier. Join the millions of

00:39:03   businesses transforming how they work with Zapier and AI. Get started for free by visiting zapier.com

00:39:10   slash ATP. That's Zapier spelled Z-A-P-I-E-R dot com slash ATP. Zapier.com slash ATP. Thank you so much

00:39:19   to Zapier for sponsoring our show.

00:39:21   All right, let's talk about MacBook Neo teardowns. First of all, Tech Renew did one where they did fast

00:39:30   forward a bit, but effectively they tore the whole thing down in 10 minutes. I think we talked about

00:39:33   this briefly last week.

00:39:35   I think it was six minutes they had an apartment.

00:39:36   Oh, okay. There you go. iFixit did a teardown. There's a video as well as a blog post about it.

00:39:42   They did some interesting things. They were talking about, hey, how is this, how does this weigh what

00:39:47   it weighs? In the good way, in the bad way, how does it weigh what it does? And so the Neo's bottom

00:39:52   case and keyboard is only each grams lighter than the MacBook Air's, despite being six and a half

00:39:58   percent smaller in a two-dimensional area, 101.3 square inches versus 95.1 square inches.

00:40:05   The Neo's screen and lid is 48 grams heavier than the Air's. And at 86 grams, the metal H that's

00:40:13   under the trackpad that assists in the way the trackpad works, that's 7% of the Neo's total weight,

00:40:20   which is bananas.

00:40:21   I mean, weight well spent because that trackpad feels good.

00:40:24   Yeah.

00:40:24   Yeah. And the Neo's full trackpad assembly is almost exactly twice as heavy as the M3 MacBook Air's.

00:40:31   And I would not have guessed that because you would think like, oh, the Taptic engine,

00:40:34   isn't that going to add weight and everything? But I guess having moving parts and like those

00:40:37   leaves, those steel leaf springs in there and the big heavy metal H, it's fascinating that

00:40:43   where is the weight? It's in the screen lid and the trackpad.

00:40:47   And metal's heavy.

00:40:49   I can't tell. So it's eight grams lighter. The bottom casing keyboard is eight grams lighter,

00:40:54   6.5% smaller. But in their video, they had like, you couldn't see, they put the stuff on a scale

00:41:00   with like the ones kitchen scales with a little like non-backlit LED readout or LCD readout for the

00:41:06   weight. I couldn't read the weight. So I don't know, is eight grams lighter at 6.5% smaller? Is

00:41:11   that proportional? I need to know the title, the total to do that math. But anyway, that's,

00:41:16   they didn't dive too far into it. But I would say that most of the weight difference,

00:41:20   the reason the Neo weighs exactly the same as the larger MacBook Air, despite having a smaller

00:41:26   battery, is the trackpad weighs twice as much and the lid of the, where the screen is, weighs a little

00:41:32   more as well.

00:41:33   Righto, with the trackpad, there's a screw in the center of the trackpad mechanism that lets you

00:41:39   adjust the force required to activate the membrane switch that triggers a click, which is pretty cool.

00:41:44   So you can, I presume.

00:41:46   Yeah, I mean, it's not for the user to do. It's, you know, if anyone needs to repair it or if the

00:41:49   trackpad ever started feeling wonky or got too loose or got too heavy or whatever, it's nice to

00:41:53   know that there's an adjustment screw there. They also did the usual thing of like doing whatever,

00:41:57   whatever scanning thing there. Is it a, is it a pet scan, cat scan? I don't know. They're doing some

00:42:01   kind of scan that shows the insides of the thing. It's like an x-ray, but like they colorize it and

00:42:06   everything. And so you can see the speakers I put in our show. You just look at the video. You'll see,

00:42:11   see all the images, but in our show notes here, I have, I'm showing the same speaker from two sides.

00:42:16   That's the same speaker from the top and from the bottom. I know in our notes, it looks like it's left

00:42:19   and right, but that's just the same speaker. So you can see how big the speaker actually is.

00:42:23   Everything else in those big black squares that are to the left and the right of the trackpad

00:42:28   is not decidedly not speaker. And good old I fix it opens it up. Although they did it so fast

00:42:33   in a sped up thing that I had to frame by frame to get these screenshots, but

00:42:36   they used like a, essentially a hot knife to cut open those black things. And yes, the speaker is

00:42:42   in, I don't know, let's say one quarter of it. And the rest of it is completely empty air. Um, it's

00:42:49   just hollow inside there. And I think I know why. Do you see these pictures here in the bottom one where

00:42:53   that wire is dangling? That's just like laying on top of it. That's not, there's the wire is not inside

00:42:57   there. That's just the wire. Like, you know, from the, you see from the top, right? Anyway, there's

00:43:01   nothing in there. It is empty air and it is not air that is channeled to like, like a base port to try

00:43:06   to like have a tuned length of tube to, nope, that's not it at all. It is just plastic, but it is not

00:43:13   flat plastic. It is plastic with those little, uh, diagonal ribs for cross bracing and X patterns.

00:43:21   Like there's, it's, it's separated into little boxes and within each box there is an X all in the box and

00:43:26   the X are made of ridges of plastic. And then you see the four holes, um, where screws go in.

00:43:31   This is what I think of these black plastic things are doing. And probably also doing the same thing

00:43:36   in the iPad. These black plastic things are essentially, uh, stiffness braces because that

00:43:44   square of plastic has nothing to do with the speaker, but is braced in this way and then screwed at four

00:43:50   points into the chassis to make it. So the corners of the chassis don't twist and bend. It's like for

00:43:55   torsional rigidity. It is a mechanical stiffener. That is my theory. And I'm sticking to it. Unless

00:44:00   someone from Apple's design team tells me that's not what this is for, because they're definitely

00:44:04   not air pockets to increase base because they wouldn't put those cross bracing in there. That's

00:44:09   not the, you know, what the inside of any, you know, base port ever looks like, but you do put things

00:44:13   like that in to stiffen things. So, uh, MacBook Neo and possibly also the iPads as well. I haven't seen

00:44:20   inside their black things. It has plastic stiffeners bolted to it to make it so the corners aren't floppy,

00:44:27   which is maybe why it feels so expensive.

00:44:28   Yeah. Cause otherwise, like if you had like battery in there, battery is rigid. And so, you know, you,

00:44:34   you probably wouldn't need to add more stiffeners to that. But in this case, when you don't have

00:44:38   enough battery to fill the cavity inside the case, yeah, that makes a lot of sense. So that that's,

00:44:44   I think that's a very good theory, especially looking at these, you're right. Like this,

00:44:46   you wouldn't design a ported speaker enclosure with that pattern in it.

00:44:50   I can't even tell if that is open to the air, like, cause they didn't do a good job of peeling

00:44:54   off all the plastic. Like this went by really fast in a few frames. So, but I, it's just totally

00:44:59   separate. Another thing. So, uh, we just talked about how they, you take this thing apart and the

00:45:03   battery isn't glued in, it's screwed in or whatever. And there's been some talk about, um,

00:45:07   obviously for repairability, you don't want to deal with glue and stuff like that, but also like

00:45:11   that perhaps, uh, this makes it less expensive to assemble. I don't know a lot about assembly,

00:45:18   but I'm thinking that 18 screws versus sticky stuff, the sticky stuff might be cheaper to assemble

00:45:24   because you just put the sticky stuff on and you slap it in there. Like a machine can do it that

00:45:28   much more easily than a machine can do 18 of these tiny precise screws. Um, but I don't know,

00:45:33   maybe machine does all of them. Maybe machine does none of it, but at any rate, 18 tiny screws.

00:45:38   And also this is another thing that may contribute to the weight. The battery in the Neo looks like

00:45:44   it's in like a metal frame with flanges and those flanges are screwed down with 18 screws.

00:45:49   And so the, when you have the sticky stuff, you don't need a metal frame to hold the battery.

00:45:55   And then because you don't need, you're not screwing anything down. The battery itself is literally

00:45:58   stuck to the chassis with glue or in the case of the iPhone, that electrically releasing glue thing.

00:46:04   So I think that metal frame adds weight, those 18 screws add weight, uh, and it is much more

00:46:11   repairable, but it also might be a little bit heavier and might actually be more expensive to

00:46:16   assemble because you got to put an 18 individual screws versus one sticky thing and slap and you're done.

00:46:24   all right. Tell me about these main boards. Yeah. Just to compare like what, what is the

00:46:28   MacBook Neo logic board compared to the MacBook air M three logic board. And I mean, the Neo logic board

00:46:33   looks more like an iPad logic board. It's a skinny little thing. It looks like a, an extended phone

00:46:39   logic board or you can, you can compare it to the, uh, the iPhone 16 pro, which has the same SOC. And yeah,

00:46:44   you can see it's the same SOC, but the iPhone is just massively miniaturized. I can't,

00:46:49   I think they're just showing two board. Is that two boards or two sides of the same board? I can't

00:46:53   even tell on this, uh, I fix it thing, but you can watch the video and see the size comparisons.

00:46:57   Suffice it to say that the logic board portion of Apple laptops has just been shrinking and shrinking

00:47:03   and shrinking over, over the decades. And it is now disappearing to the point where now it looks like

00:47:07   an iPad logic board. And I wonder if in five or 10 years, it's going to look like a phone logic board.

00:47:11   the unthinkable happened, uh, particularly to, to Marco, actually, uh, AirPods max two are out.

00:47:19   That is not the AirPods max with USB-C that came out a few months ago, whenever it was,

00:47:24   this is an honest to goodness, AirPods max two. And John, you get to do a victory lap.

00:47:29   Yeah. I was just a happenstance that I saw this because some listener said, um,

00:47:34   was asking us to do an ask ATP in the future, which I don't know if we'll do, but it's saying like,

00:47:38   you should list all of your periodic reminders as an after show or a member special or something.

00:47:43   Um, and I actually have a reminders category called far future. So I don't, so those things

00:47:48   don't get jumbled up. And what do I put in far future stuff like this? I went to the far future

00:47:52   saying, I wonder what's in there these days. And the very top item was, uh, a check if there's an

00:47:58   AirPods max two for the Marco bet, blah, blah, blah, like whatever. I'm like, you know what? AirPods max

00:48:03   two did come out today. So I gathered all the info in ATP episode 604 at

00:48:08   one hour, three minutes and eight seconds. Marco said, after I had said something about,

00:48:12   uh, the, uh, AirPods max and a revision said, John, I think you're being optimistic.

00:48:18   Yeah. You can make a bet, add it to your calendar. I don't think there's a chance in hell the AirPods

00:48:22   max will get another update at least for two years. I did add it to my calendar and there it was.

00:48:27   And I added it to my reminders. There it was in my far future reminders. Well, guess what?

00:48:31   That was September 12th, 2024. Apple got it just under the line. Although I didn't actually agree

00:48:36   to the two year bet. I actually, uh, uh, agreed to take the bet at three years, but either way I won

00:48:41   and the world gets AirPods max two, uh, which are so very different from their predecessor.

00:48:49   Yes. They're very, very different or something. I mean, they're different in the way everybody

00:48:54   wanted because when the one with USB-C came out, you're like, that's it. They just added USB-C.

00:48:58   Where's the revision? I want one with an H2. Well, guess what?

00:49:01   I mean, so it's, it's very similar in the approach to the vision pro update, which is like, all right,

00:49:10   here's this product that immediately upon launch, everybody was like, well, this part's good,

00:49:18   but these parts really suck. Here we are like, you know, many years later and Apple has released

00:49:23   the next version of it. And literally all of the problem people have with it are unaddressed.

00:49:30   Totally just left exactly the same. And it's like, all right, if you liked what we shipped before,

00:49:35   here's an update with essentially a spec bump, nothing else has been touched. It's like, okay.

00:49:42   Now the difference between the AirPods max and the vision pro is that people buy the AirPods max.

00:49:48   Honestly, the AirPods max, this is not a product for me, just comfort wise. They are extremely

00:49:54   uncomfortable for me because they are so heavy. Um, and they did nothing to change that.

00:49:58   They did nothing to change the clamping force. They did nothing to change the annoying,

00:50:03   weird case thing that they come with. None of that is, is different. They did. They didn't change

00:50:09   the kind of awkward control scheme on them because they insist on digital crowns and everything they

00:50:13   make. They probably do still sound great because the first one sounded great. That's, that part's

00:50:17   probably fine. It is nice that they are matching the, many of the features of the newer AirPods that,

00:50:22   that have come out. So that's that, that's all great. Um, but if they didn't work for you before,

00:50:28   they won't work for you. Now, if you love the AirPods max, and maybe you have like an older pair

00:50:34   where the batteries are starting to wear out and maybe you're annoyed at the lightning cable or

00:50:37   whatever, then this is, this is a great time to upgrade. If you still want to use them. Um,

00:50:41   they are also still very expensive. Uh, but I see a lot of these out in the wild now, maybe just

00:50:49   cause you know, people in New York have a lot of money, but like I see a, a lot of AirPods maxes out

00:50:55   there, especially on young people, like teenagers. I'm like, your parents bought you a $550 pair of

00:51:01   headphones. Damn. But, uh, but I do see them a lot. I do think people are buying them in reasonable

00:51:11   numbers. Not enough. You know, I'm sure it's nothing compared to like the AirPods pro or the

00:51:15   regular AirPods, but it does seem like it is a successful product that has a market, the price

00:51:21   and the physical downsides of it keep that market smaller than it could be. But Apple seems to not

00:51:29   care. So for the market that Apple has chosen to stick with for this product and seems to have no

00:51:35   interest in trying to expand that market, this is an update. It's not even a great, it's just,

00:51:40   this is an update. Okay. It took, it takes them from being like embarrassingly far behind the feature

00:51:46   lineup to roughly on par with the feature lineup and it solves nothing else, but okay. They did an

00:51:55   update. Yeah. So what's changed? Well, now we have the H2 in it, which means you get adaptive audio,

00:52:01   you get conversation awareness, you get personalized volume, you get live translation and you get voice

00:52:06   isolation. Additionally, you get one and a half times more effective, uh, active noise cancellation.

00:52:11   You get a new high, high dynamic range amplifier for even cleaner audio. Uh, you get reduced wireless

00:52:17   audio latency for, uh, game mode and iOS, macOS and iPadOS, but you get the same color, same price,

00:52:22   same case, same, not folding headband thing, basically all the same stuff.

00:52:28   Yeah. I mean, it's a good thing. They just, that's what everyone wanted. There's like, why doesn't it have

00:52:31   the H2? It's kind of like when the, uh, vision pro came out with the M2 and everything else was on the

00:52:35   M3. Um, so it's just a catch up thing. Uh, it should have the H2. It should have the H2 a long time ago.

00:52:41   Now it does. And it's got all the features.

00:52:44   We are sponsored this episode by one password. Now, if you're a small business, it's easy to

00:52:48   assume that you might fly under the radar security wise, but small businesses are being targeted more

00:52:53   and more by bad actors because cyber criminals know that lean teams often lack the resources to prevent

00:52:58   a breach. So teams of any size can be a target, but even the smallest of teams can foil cyber crime with

00:53:03   one password. One password provides simple security to help small teams manage the number one risk

00:53:08   exploited by bad actors, weak passwords. One password provides centralized management to make sure your

00:53:13   company's logins are secure. It's a simple turnkey solution that can be rolled out in hours, whether or not

00:53:19   you have dedicated IT staff. And it might, no matter how complex your security needs may get as your company

00:53:24   grows, one password will stay with you every step of the way. But of course, the earlier you start, the easier

00:53:29   it is to build your business on a secure foundation. Compromised passwords are the number one way bad actors

00:53:33   attack companies. So a password manager should be the first security purchase you make for your team.

00:53:38   I personally use one password, both for my own personal stuff, but also for my business logins. Now I have

00:53:44   multiple businesses. One of the, you know, many of them are just for me and I'm the only person. I still

00:53:47   use one password for that because it helps me, helps me generate all these like amazing passwords and keep

00:53:52   all the accounts separate. But where it really shines is with my family and with my business that I do with

00:53:57   other people. So we have a family share plan where my wife and I, and even my son can share passwords and

00:54:02   we can assign what's shared and what's not. And then for the business, for the restaurant, we share passwords

00:54:07   with the managers and my wife and me, all different access levels, all super easy to manage. And it's

00:54:13   wonderful to have everybody on the same page, to not have to like have post-it notes with passwords

00:54:17   written on them and for all those passwords to be really secure and centrally managed. I love using

00:54:22   one password for this. It is fantastic for this. So take the first step to better security by securing

00:54:29   your team's credentials. Find out more at onepassword.com slash ATP and start securing every login.

00:54:36   That's onepassword.com slash ATP. Thanks to One Password for sponsoring our show.

00:54:41   Overcast is shipping in beta, in beta. Overcast is shipping transcripts. Tell me what the heck is

00:54:53   going on here because this sounds pretty freaking great. I've been working on this for some time.

00:54:57   Apple Podcasts launched transcripts probably what a year and a half ago now, maybe two years ago. It's

00:55:03   been a while. I knew back then, oh crap, that's a really good feature and I'm going to have to do

00:55:11   that someday. But how am I going to do that? Apple Podcasts operates at a much larger scale than I do

00:55:17   with much more resources. And I just had no idea how I could possibly ever match that.

00:55:24   Last summer, with the release of, with the beta of iOS 26, Apple launched a new speech transcription

00:55:34   API on iOS. They basically opened up the iOS speech model that is used for iOS's built-in transcription

00:55:42   of things like, when you speak to Siri, it's that model. When you transcribe notes,

00:55:47   you know, it's, it's that model. What this meant is that it ran on device. It was very optimized and

00:55:55   very fast and very lightweight. Now, I've been looking at different ways to do transcription

00:56:01   for a little bit before that, especially once OpenAI's whisper models had come out. That was,

00:56:07   that was a game changer in transcription models because the accuracy was so much higher than what

00:56:12   had come before. It was shocking how good whisper was. Um, but the problem with OpenAI whisper is that

00:56:19   it just, it's a really big model. It's really slow. And so it's great for, you know, one person using on

00:56:27   their computer, um, or some very specialized app uses, but like for the most part, like it's great if you

00:56:32   are transcribing like your own podcast that you make once a week whisper is great. Um, but if I was going

00:56:39   to try to offer transcripts and overcast for like all podcasts or many podcasts that wasn't going to

00:56:45   scale. And there are also, you know, if you're operating at a smaller scale, again, like if, if you, if you have

00:56:51   like a, maybe a podcast hosting company and you wanted to transcribe, you know, audio for your customers that

00:56:56   they upload, if you're dealing with, you know, maybe that's hundreds of uploads a week or something,

00:57:02   that's a very different scale that might be operating at. So there are things like transcription

00:57:07   provider or transcription APIs from the AI providers. So, you know, OpenAI has a transcription API.

00:57:13   The problem with those is that they're not really designed for podcasts. They're, they usually have

00:57:19   audio length limits or size limits that many podcasts would exceed and just cost wise, you would be

00:57:28   talking about hundreds or thousands of dollars per day, uh, at the scale that overcast would need those

00:57:35   to be. And overcast is not going to take on a thousands of dollars a day API cost. That would,

00:57:42   let's just say that would require me to raise the price of premium higher than most people would be

00:57:46   willing to pay. Um, what happened last summer is when Apple released their iOS 26 and all the OS 26 is

00:57:55   when they all had this new API in them for Apple's on device transcription model, I ran some tests.

00:58:00   It just blew me away how incredibly fast it was. Um, so one thing I noticed like on a regular M4,

00:58:09   like on a Mac on an M4, uh, I was running this on my, on my laptop and it was able, cause I wasn't

00:58:15   going to put, you know, Tahoe on my main Mac, but I, so I was doing all this development on my laptop

00:58:20   over the summer. And I noticed that, you know, it was able to transcribe. If I ran a few jobs in

00:58:24   parallel, I could have one M4 Mac transcribe audio at about 200 X the audio's playtime. So in other words,

00:58:35   about, about 200 minutes transcribed for every real time minute that has passed.

00:58:40   goodness. And that was so much faster than anything else I'd ever tried. I noticed I'm okay. Well,

00:58:45   if one Mac can do like, you know, 200 X real time, well, how many podcast minutes are there? Let me see,

00:58:52   like, what if I just start transcribe? If I let, let me get like a couple of Mac minis and just have

00:59:00   them start transcribing the most popular podcasts. How many can they get through? And I, you know,

00:59:05   I have some information of like, I can see what the most popular podcasts are. I can see how many

00:59:10   episodes they release, you know, per, per month or whatever. And I can do some division and figure

00:59:14   out like, you know, usually I can, I can download a copy of what they serve and I can see how long it

00:59:18   is and start analyzing things. I'm like, okay, let's, let me try. What can a couple of Mac minis do?

00:59:24   I was blown away by how effective that was. You know, I'm like, okay, these, I can't just keep

00:59:32   burning these in my house and let me see what I could do. So put them next to the water in the

00:59:36   closet. It'll be fine. Right. Exactly. Yeah. Um, so I went to, uh, Mac mini, uh, vault Mac mini,

00:59:42   which is, um, I think the other actual name is cyber link. That's the actual name, but it's Mac mini

00:59:47   vault. Um, and cause what they are, I think they're the only provider where you can, you can like rent a

00:59:53   Mac mini from them for a hundred bucks a month. But if you buy your own and mail it to them,

00:59:59   then they'll host it in their data center for only 50 bucks a month. And I was like, well,

01:00:04   it doesn't take that many months to come out ahead if I'm just buying the base model. Um, and, and it

01:00:08   turns out, and I analyze, by the way, I don't know if listeners might recall, I recently have had strong

01:00:13   opinions about the, uh, the bang for the buck for, uh, M four family chips and what the, what the

01:00:20   maximum bang for the buck for processing power is. This is why I did all this

01:00:23   analysis last summer. And it turned out the base model Mac mini was like way more bang for the buck

01:00:29   for processor performance than any other configuration or any other computer. Okay,

01:00:33   great. So I got two Mac minis and I sent them off to Mac mini vault in Wisconsin. And I even,

01:00:38   I leased one. Um, there's a company called green mini host in the Netherlands. I'm like,

01:00:43   if there's certain podcasts that are like EU region locked that I found a Mac mini host in the EU.

01:00:49   I think the Netherlands is in the EU. Uh, anyway, so at least in Europe and, uh, and, uh, so I'm like,

01:00:55   great. All right. So I had those three and this, this required me to learn a bunch of stuff. Like

01:01:01   I had to, first of all, you know, first of all, I'm installing the Tahoe betas and every single

01:01:06   time there's a new beta, the API changes a little bit or some limit is imposed or some limit is released

01:01:10   or some bug is fixed. So all summer I'm updating these, these Mac minis with the betas. And that's,

01:01:16   that was another question I had to like first ask Mac mini hosts. I'm like, can I,

01:01:19   or Mac mini vault rather, like, can I run betas on your Mac minis? Like, is that,

01:01:24   is there any reason why I couldn't do that? And no, there's no reason. So it's fine.

01:01:26   So I'm running the betas. I'm doing everything via remote desktop, like, you know, like VNC,

01:01:32   like to, to the Mac minis, like from my home computer. And meanwhile, also doing it all like

01:01:36   on my laptop locally and updating that every time. And then I did like this whole thing. I had to,

01:01:41   had to learn how to basically run Mac servers, which is something I really had not done beyond just

01:01:46   occasionally having a Mac mini in my house to be like a NAS or a streaming thing or whatever.

01:01:50   There was just a link that I was going to put in a future episode, but it seems relevant now that I,

01:01:54   let me see if I can find it. There was like, I think in 26.3 or something that file vaults now lets you,

01:02:04   unlock the file vault encrypted drive over SSH. Yeah. Like before your thing boots. That would

01:02:11   have been nice. I mean, presumably these hosting companies will do this for you. That's what you're

01:02:17   paying them for or whatever. But for people who don't have that, but yeah, apparently we'll put a

01:02:21   link in the show notes to the Apple support article that even before the OS boots, when it's just at the

01:02:25   thing where it's like, Hey, enter your password to unlock this file vault disc. So I can boot the OS.

01:02:30   You can apparently SSH in remotely and enter the file vault password. So it will continue with the

01:02:35   boot. Yeah. And I've, I've learned so many things about running Macs remotely. Like, you know, all

01:02:38   the different, like I have a big text document of all the things I have to do, you know, for a fresh

01:02:43   install, like to prepare it and everything. Um, because you know, every time there's a software

01:02:47   update, I got to do the whole, I got to do all this again and go, go to all of them. And, and of

01:02:51   course in the summertime, like during the beta period, it's every two weeks, every two weeks,

01:02:55   I'm going and updating all of these running the tests, like making sure I can still do

01:02:58   what I want to do. Um, and meanwhile, I have to also integrate this, you know, this, these

01:03:04   three Mac minis into overcast infrastructure. So I have to build in a job queue performance

01:03:11   and health monitoring, like from, from like the overcast transcription app that I'm running

01:03:15   on these Mac minis that integrates with my other performance and health monitoring on the

01:03:19   server side. And I have to, you know, build on things like, all right, well, if, if the Mac

01:03:23   mini takes a job and then crashes, then I have to have some kind of like repeat job

01:03:27   queue entry to re queue that after a while and have some other Mac mini handle it. Like

01:03:31   it was, it was such a journey, but it was working. I want more coverage. I don't want to just do

01:03:37   like the top hundred podcasts. I want to do, you know, anything with more than X followers

01:03:42   to it or X subscribers to it. Great. So I got two more Mac minis in August. So, you know,

01:03:47   started out like really, really simple, you know, just the three, like the two that I own

01:03:52   and the one I leased in the Netherlands. And then now I'm up, now I'm up to two more,

01:03:56   now I'm up to five. And what five could do was great. You know, it was like, okay, all right now,

01:04:02   but you know, the cost is starting, you know, now I'm, you know, $250 a month plus the hundred,

01:04:07   like it's, you know, it's like, it's starting to get like, okay, this is, this is starting to get to

01:04:11   a decent amount of money. And I thought, okay, let's see how the five do the five. You know,

01:04:17   it ramps up pretty much linearly. Okay. How many more podcasts can I cover? And it was, you know,

01:04:25   I'm starting to cover like a somewhat reasonable percentage of popular podcasts, but then I'm like,

01:04:31   all right, well, I want, I also want to do like some of the back catalog. Like, can I go back like a few

01:04:35   months? Can I go back a few years? Like how, how much can I do since I don't really need a lot of

01:04:42   like help from the host, I'm just running these anywhere. I wonder if there's like a local data

01:04:50   center that I can co-locate. Like if I can, if I can get like, you know, people, people make these,

01:04:54   these brackets that hold six Mac minis into like a three U space on a rack. And I'm like, if I can get,

01:05:01   let me, let me see, like, what would it cost to host a three U rack of Mac minis in a data center?

01:05:06   If I can get that for a few hundred bucks a month, then I might, you know, I might be able to come out

01:05:11   ahead of what it's going to cost me. If I'm going to, if I'm going to, I was thinking, you know, maybe

01:05:16   I'll have like 10 or 12 of these, like that could like, that could do a lot. If I can, what I was doing

01:05:22   with three and then with fives, like if I can get like 12, like, Hmm, this, this could really be

01:05:28   something. And so I bought a few more and I'm like, as I'm, as I'm doing this, I'm like, let me see,

01:05:35   like, here, I'll paste you in the Slack. This is, this is where this went.

01:05:38   Before you send whatever you're going to send, I would just like to say that I am looking forward

01:05:43   to the day that I get a shipment from you padded in Mac minis.

01:05:47   Mac minis is packing material when the, when the M5 Mac minis come out.

01:05:52   All right. So you've sent two photographs that I will describe. One of them is two Mac minis with

01:05:56   little stickies on them that say, what does that scribe one into? And then a second photograph

01:06:02   that has one, two, three, four, five, six, seven Mac minis stood kind of upright, if you will,

01:06:07   with a post-it per Mac mini, uh, connected to, I guess that's a UPS of some sort.

01:06:12   No, that's a switch on the left and that's an old iPad on the right. Just as just basically

01:06:15   because you know, when you run them, they get hot and I didn't want to melt my rug. Uh, and

01:06:20   this was my, my thermal management solution.

01:06:23   The one on the right is on the rug though.

01:06:25   Yeah. Well, when you just have one, it's not that bad, but when you have seven, you know,

01:06:28   it starts to add up. Um, so this was like, all right, now I have a few more and it's going well.

01:06:35   And I'm like, okay, I ordered, I'm like, as I'm like looking at different data centers,

01:06:39   I was like, I want to know like, you know, what are my situations for, um, you know,

01:06:43   for racking me somewhere. So I got a rack enclosure to see like what it is. And so here's this next

01:06:48   photo here. This is six of them in a rack enclosure made for Mac minis. It's a very simple,

01:06:55   you know, just metal bracket.

01:06:56   Yeah. It seems like it's a shelf with brackets to mount it on the rack, you know, and very little

01:07:01   else. There's not too much here.

01:07:03   Yeah, exactly. Um, and so then, so I started talking to data centers and

01:07:07   wait, I'm sorry. There are so many Mac mini boxes on the right hand side of this.

01:07:11   You can build a little fort out of them.

01:07:13   I mean, it was really getting ridiculous. Like, cause there's, I've now at this point bought

01:07:17   seven Mac minis and like, it's, you know, it's starting to get, you know, I'm like, do I keep

01:07:21   the boxes? Like, what do I do?

01:07:23   Everyone thought it was a, it was open claw that was causing the run on the Mac minis. It's

01:07:27   really just Marco.

01:07:27   Yeah.

01:07:28   For the record, I'm looking at this rack mount and I don't know if there's, if, if there's a

01:07:33   reason behind this or if this is really the, the, the right iteration numbers, but I see

01:07:37   scribes, seven, eight, nine, 10, 11, and 12.

01:07:39   Well, yeah. Cause there were already some already in the data centers.

01:07:42   Oh my word.

01:07:43   So at that point, one through six were in the data center or no one through five were in the

01:07:48   data centers. Six was that extra one off to the side of the previous photo. That was just

01:07:53   one that I just had, that was like my beach Mac mini. Like, so that was just, I, I, I conscripted

01:07:59   it into, into service for, for this purpose.

01:08:01   You should start looking on eBay for a M4 Mac minis.

01:08:03   Maybe I'll be, I'll be supplying them soon.

01:08:06   But anyway, what, what is your power bill for the month or months that these were all in

01:08:11   the house? Holy crap.

01:08:12   I think all those together are less than my Mac pro.

01:08:14   Oh, by far.

01:08:15   Probably true.

01:08:15   It's probably less than your monitor. So the, the combined power usage of six Mac minis

01:08:21   doing transcription is something like 250 Watts.

01:08:25   Oh, that's really not bad at all.

01:08:26   They're not doing much GPU work and they're all just the base model M4. So the, the top

01:08:31   of the power envelope is not that high. Um, so it's, yeah, it's like, I think it's like

01:08:35   about 50 ish Watts or 40 ish Watts each. It's yeah. It's like 250 for the whole, for that

01:08:41   whole thing. Um, cause I was curious too. I measured it. So that's, that's how I know

01:08:45   that anyway. So I started talking to data centers and it was very hard to find any data center

01:08:50   that would even take like, you know, just one guy, because it turns out what I've learned

01:08:55   is that most data centers, their businesses, like big companies or the government, you know,

01:08:59   so it's like, they don't have like not a lot of data centers willing to talk to one guy.

01:09:02   And I found one locally that was, it's on Long Island and it was, it was willing, you know,

01:09:08   they were willing to talk to me and I was like, all right, so, you know, I'm looking to,

01:09:11   I'm looking to get like, I'm like, I, I, what I need would be at minimum, like, you

01:09:14   know, three U plus whatever I would need for like a, you know, network year or router or

01:09:18   whatever. So like three, maybe three or four U of rack space, you know, 300 Watts of continuous

01:09:24   power usage and whatever you can give me on the internet side. What, what would that cost?

01:09:27   And I learned basically that was too small for this data center to, for any data center to

01:09:33   care about that. Like I was, I was, I've never done anything with data center before. So I

01:09:37   had no idea. What is it like? I was picturing like, you just rent, you know, space by the rack

01:09:43   unit, like by the, by the U basically. And I thought I could just ask for like, for you and

01:09:48   like, Oh, how much will that be? And like, it turns out it's not worth them dealing with that.

01:09:52   The amount of space that we have on long Island is ample. Um, but power and bandwidth are the

01:09:59   expensive parts. And so I, they're like, okay, well we can give you like, maybe we can give

01:10:04   you half a cabinet, but we mostly just do full cabinets. And I was like, like, like what? Like

01:10:09   the full, the whole thing, the whole tall rack, like the, you know, 48 or whatever rack use that's

01:10:14   and I'm like, that's, Oh God. Okay. And I'm like, well, how much is that? You know,

01:10:19   you need another 15 Mac minis then. Well, I'm like, I'm like, this is like, I don't know how much

01:10:25   this is going to, this is going to be a lot to afford. So I'm like, all right, well that's,

01:10:29   that's way above what I, what I can do right now. So I'm like, all right, let me, let me figure

01:10:33   something else out. Um, and so for most of the fall, and I just ran that rack there, that,

01:10:41   that enclosure of the six Mac minis. I had like the, the six in the data center and then I ran these

01:10:46   seven just in my house. In the water closet. It's calling to you. You know, it is the ultimate,

01:10:51   the ultimate, forget about the, the, you know, entire server cabinet. You've already got one.

01:10:56   It's got water bottles on the bottom and you just shove this thing in. You close the door.

01:11:00   Yeah. Well, the good thing, since they're quiet and since they work as a nice little heater also,

01:11:04   um, although not much, I mean, a 250 watt space heater is a very small space heater, but I just

01:11:09   had them on top of that file cabinet you see on the right there in that picture. I just had them

01:11:12   sitting on top of that for all fall, like all the whole, it was just sitting there, um, just running.

01:11:16   And especially cause like, you know, as I was coming back on the mainland for the school year,

01:11:20   like, well, I have this perfectly good internet connection over here. It's not going to be used

01:11:23   very much all winter. Um, so might as well use it. And so I'm like, all right, if I put

01:11:27   six of them here, maybe I can put another six of them in the restaurant. Cause the restaurant

01:11:31   also has an internet connection that is idle all winter long. Like, Hmm, like this, I start,

01:11:37   I started like scaling it up in my mind. I'm like, all right, if I can, if I can make this

01:11:42   work, like I can put them all over the place. Like in, I have, I have two houses in a restaurant

01:11:47   and maybe I can like send some to Casey or something.

01:11:50   two turntables and a microphone, right? I'm just going to pick up your couch cushions

01:11:54   one day. Casey, there'll be some Mac minis under there and it will scribe 27, scribe 28.

01:11:58   Like what the hell? Oh my word. Um, so anyway, so in, I, I had, you know, there have been in,

01:12:06   in the fall around black Friday and there've been a couple of times Mac minis went on significant

01:12:11   sales and I was like, well, if I can get them for like, I did the math and I'm like, if I can

01:12:17   get them for like $500 or less, like their retail price is 600, but if it was like 500

01:12:22   or less, I would look at it. And sometimes you get them for like 450. That's getting pretty

01:12:29   like, that's getting pretty good. So there were a couple of times, uh, where I'm like, I splurged

01:12:35   and I'm like, you know what? If I get six more, I can do a lot more transcripts.

01:12:40   Oh my word.

01:12:41   And so I would get six more. And that happened a couple of times. Well, it, it soon became,

01:12:47   apparent that I, I was probably ready for the next level. Um, and so I called the data center

01:12:54   back. I'll take a cabinet. Yeah. And it turned out a cabinet was not that much once you had

01:12:59   like 18 Mac minis. And so I got myself a data center contract. Oh my word. Um, and I had to

01:13:09   learn a lot about how do you host things in a data center? I had never even been inside

01:13:16   one. Yeah. I've never, I had never seen one. I was picturing a very different thing than

01:13:21   what it was. I was picturing everything like black, dark, like, you know, loud, awful. Like

01:13:27   it wasn't loud. It was moderate volume, actually. It used to be louder. Let's say I got this contract

01:13:33   with this local data center. They were very nice, very helpful. Cause of course, everyone

01:13:36   who works in a data center is a nerd like us. So, you know, everyone's, everyone's very helpful.

01:13:40   And I think they were very happy to see a customer who was not just like a boring, you know, bank

01:13:44   or something. Um, so anyway, things escalated a little bit. Did they provide networking equipment

01:13:50   or did you provide that? I provided that. So I'll get to that in a moment. Um, but I will just

01:13:55   show you the final state that I have reached. Oh my God. This is what? 36 Mac minis. Um,

01:14:03   and I still have the six that are in the, that are in the beach house and the few that are in

01:14:07   the coat, the, uh, Mac mini vault. I now have 48 Mac minis. Oh my word. You have 48 Mac minis.

01:14:14   Did you even look at like AWS Graviton arm, you know, uh, like how much did you price out a cloud,

01:14:21   uh, base thing for this or no? Not really. Like I was looking at like APIs from the AI providers and

01:14:28   stuff, but like actually do like, so what I like about this setup is the recurring cost is almost

01:14:36   nothing. When I build overcast, I build it for the long haul because I've been around the block. I know

01:14:41   like anything I sign up for, that's going to be like $3,000 a month that really adds up. But this entire

01:14:49   thing is like $1,000 a month and I have a lot of headroom. I had to pay up front for the Mac minis

01:14:56   obviously, but they're not that expensive. And I was able to get most of them on sale whenever they

01:15:02   were like, so like I, I didn't pay that much for this amount of capacity. And each one of these has

01:15:08   16 gigs of Ram, an amazingly capable processor. Now it has professional, you know, full-time networking

01:15:14   in a data center, power redundancy and everything. So actually when I compare the processing power I

01:15:19   have at my, at my disposal here versus what I'm paying Akamai for since they bought Linode, like

01:15:25   this is, this destroys anything else I've seen in terms of capability per dollar. Um, this is way better.

01:15:32   So I had to learn everything about this. So as you see there, as you mentioned, I have my ubiquity

01:15:40   light up cables because of course I got a ubiquity light up switch and ubiquity router, um, because I

01:15:46   know how to run those. I, and like, I already have in my ubiquity site manager, I already have multiple

01:15:50   sites. I have been at the house of the restaurant. Like, so I'm like, I, I already manage multiple

01:15:54   ubiquity sites. I know how to do it. Their equipment is rock solid. I had to learn a few other things that

01:16:02   So number one, I had no idea that power in data centers is 208 volts. Oh, I, right. You didn't

01:16:12   know either. Right. So, you know, us, you know, we have the 110 or the 220 or whatever, whatever,

01:16:17   208, whatever it is, like it's the 220. It's the, it's the, the, the high voltage. Fortunately,

01:16:21   like all modern computer equipment has universal power supplies that it can take in anything from

01:16:27   like a hundred to 240 or whatever. Like the, like it has a wide range of the power supply levels,

01:16:31   but everything else you have to deal with. Like, so another thing I learned is that power at a data

01:16:37   center is usually provided with two different inputs. And I, it's, they, they were great. They took me on

01:16:42   a whole tour of the data center and I saw like the, the entire infrastructure of the place. There's two

01:16:48   of everything, two generators. There's two power regulation sides. There's two HVAC sides. Like there's

01:16:55   two of everything for redundancy and it's, it's a very nice data center. Um, and so everything has

01:17:01   to be able to fail over. And when you are connecting to their power, the back of the rack has a whole

01:17:09   bunch of plugs. Oh, and by the way, they're not just like the U S you know, NEMA 15, whatever plug,

01:17:17   like what you see on the back of a computer, like the, I think it's the C 13 or the C 14. They're that,

01:17:23   that is what data center plugs are. Here, I'll send this picture. You can see, um, this is the

01:17:28   back of the rack. Yeah. That's what I wanted to see to see how well you did on cable routing.

01:17:31   Yeah. So the, the colorful plugs you see on the back of the rack, there's a left bank and a right

01:17:35   bank, and that's two different power inputs. And so at first I was like, okay, well, I'll just plug

01:17:41   into one of them. Cause you know, if these, the, like these are not serving traffic, they're work

01:17:45   consumers. So if some go offline for a while, it's not a big deal. Like the queue will just build up.

01:17:51   And then when they come back online, they'll just start working through the queue.

01:17:54   I told the guy who was giving me this program, I go, that's because all my stuff only has one

01:17:57   input. I don't have redundant power supplies, any of this stuff. And he's like, well, once a year,

01:18:02   we like take down an entire side for, you know, a few hours to test things. And so you don't really

01:18:08   want to be plugged into that side during that time. And then, you know, we do the other side. So I'm

01:18:11   like, okay, well, what, I'm like, what do people, how do people solve this problem?

01:18:15   They buy server hardware where they can plug both cables into redundant power supplies.

01:18:19   Well, that too. Or it turns out there's a device called an ATS, an automatic transfer switch. And so

01:18:24   you can plug in the ATS into both sides and then it, it offers you a single set of outlets that is,

01:18:30   that is redundantly powered basically. So you can plug in all your single ended stuff into the ATS.

01:18:34   And since all of my stuff is single ended, that's what I did. So I had, I had to learn that.

01:18:39   Like there's so much about this that I just had no idea. Like this is how servers were like physically

01:18:45   run. Yeah. I didn't know. You're really not taking advantage of this rack. Those Mac and minis are

01:18:49   just barrel. They're like the size of the face plate on the old server. And the whole rest of the rack has

01:18:54   just empty space. Yeah. Well, it's a lot of space. Like, yeah. Cause servers are long. Remember even how

01:18:59   long the X serve was, which is probably the only rack map server you may have been familiar with. They're

01:19:03   really long. Yeah. I mean, there's nothing stopping me if I wanted to like, I could use both sides of the

01:19:07   rack. Like I could, I could just put a second set of them on the back rack mount holes.

01:19:13   You want that, you want the hot side going into the hot aisle. You don't want to be pushing hot

01:19:16   air into the middle of your rack. So yes, technically you are correct. Um, in practice,

01:19:21   like as I walk through the, the aisles of the, of the data center, there's no heat coming from your

01:19:25   rack minis. No, there, there's some, but like in practice, like I'm like, I walk, I walk past,

01:19:29   you know, as I'm walking down the, the ostensibly cold aisle, some people have installed their

01:19:34   servers backwards and are blowing hot air into the cold aisle. So like it doesn't seem like it's that,

01:19:37   you know, strictly necessary, um, at this scale. Um, and yes, you're right. Like the

01:19:41   Mac minis are just not producing that much heat. Um, you'll see like the big black things I have

01:19:47   kind of like at top and midway. Those are fans. Um, there's a, there are rack mount fans that suck

01:19:51   air in from the bottom and blow it out the back because the Mac minis are just, you know, having

01:19:55   the, having the hot air just go straight up. Yeah. The Mac minis are not blowing air out the back

01:19:58   of them, unfortunately. Right. They're just kind of, you know, it's wafting upwards from them.

01:20:02   And so these fans suck it in and blow it out the back, but even the, like, I don't think it's that

01:20:06   necessary at this scale, obviously, or honestly, but anyway, so I had to learn about 220 volt. I had

01:20:13   to learn about ATSs. You know, they, they provide the internet via what they called a DIA. And I had

01:20:21   to, of course, I'm instantly going to like chat GPT and Gemini. What the heck is a DIA? How do I,

01:20:25   how do I connect whatever that is to whatever I have? What should I have? What equipment do I need?

01:20:30   What settings do I use? They give you like a yellow fiber wire and they're like, here you go.

01:20:35   I ran it to your cage. Okay. Now what do I do? And I had to like buy like a little fiber transceiver,

01:20:43   but it had to be exactly the right kind to go into the SFP port on the ubiquity, whatever router I got.

01:20:48   Um, the, I think it's the, one of the, the dream machine pro max ultra. I don't know. Uh,

01:20:54   one of, one of the high end rack mount ones. Um, like I had to learn all this stuff. Like,

01:21:00   and this is, you know, for people who work in data centers, I'm sure this is trivial to you because

01:21:03   you've been doing it forever, but like when you've never seen it before, that's all brand new. So I had

01:21:09   to do a lot of just learning. And again, thank God for like the AI tools and YouTube tutorials here and

01:21:16   there for certain things. I was able to figure it all out eventually. And again, the data center people

01:21:19   were very helpful with the parts that they could help with. Um, so it was a really great learning

01:21:26   experience. It was very fun doing it. I don't get a lot of physical projects and like to be able to like

01:21:34   wire up a whole bunch of computers and, you know, run all these network cables bundled together and,

01:21:39   you know, all like have all, have everything all organized. I even like, I, I recently set it up

01:21:44   so that like, if one of the Mac minis crashes in a way that I can't reboot it remotely, I can now log

01:21:50   into the power thing and remotely just kick off an entire bank of them and like, have it, like have

01:21:56   it power cycle an entire, you know, set of six. Um, so I built a manual kicking machine. Um, and it's

01:22:02   reference signals. It's just, it's a lot of fun. I'm so happy I did this. And what this has enabled

01:22:08   is overcast is now transcribing every podcast that has more than one listener, every single one,

01:22:19   every podcast. Well, there are language restrictions. Um, because I'm using Apple's API or Apple's, um,

01:22:26   you know, models for this, their models only support English, French, German, Japanese,

01:22:33   Italian, Portuguese, and there's one more. Oh no, that's it. That's six. Yeah. If I actually look at

01:22:40   what overcast top languages are, um, the one I really want is Dutch. They don't support Dutch yet.

01:22:46   And that, that is one of overcast top languages in terms of like podcasts being produced that people

01:22:50   are listening to. Um, but otherwise like that covers most of what most people are listening to

01:22:56   among those podcasts that, that I can support with this API based on the language. I am able to

01:23:02   transcribe every episode that comes out, you know, easily catching up every, every podcast that has

01:23:07   more than one listener, um, that is public. And then I was like, you know, what I want to do the

01:23:13   back catalog. And so I started kind of raising the numbers of like, all right, let me go back again,

01:23:19   start out, go back a month, go back to, and then, you know, the, the, you know, jobs in the queue would

01:23:24   spike up and then they would slowly churn through them and slowly work through. And I'm like, okay,

01:23:28   that's, that's good. That's getting there. Um, and then I was like, why the number, uh, I'm just not

01:23:37   keeping up. Like what's, there was like a couple of weeks where like, I'm just, I was not able to keep

01:23:41   up. I'm like, what is going on here? I, I bought 48 Mac minis. Why can I not keep up? By the way,

01:23:48   when I said a few episodes ago that I've seen the Tahoe welcome tour a few times.

01:23:53   Oh, gracious. This is why I have seen it so many times. Welcome liquid glass. I would love to never

01:24:02   see that screen again. Anyway, so I was like, what is going on? And I, I looked at, I built a whole

01:24:08   logging feature to see it for every, so every server could log every action that was taken on, on a

01:24:12   transcript. It's like, what is going on here? And I found after a little while, there was a bug that

01:24:19   was causing a lot of the transcript jobs to not be deleted afterwards. So they would recur after a few

01:24:28   hours. Oh no. So I fixed the bug and within a couple hours, the queue went to zero. And I was like,

01:24:35   Oh, I have more capacity than I thought I had. Oh no.

01:24:42   So then I thought, I wonder what it would take to transcribe private podcasts too, because that's

01:24:51   something that Apple podcast doesn't do. Wouldn't that be cool if I could do them too? And so I built

01:24:57   that. It turns out I'm able to support almost all private podcasts too. So now Overcast is transcribing

01:25:04   every public podcast with more than one listener. Almost every private podcast, like above a certain,

01:25:10   I forget where I, I think I set that threshold at something like 10 or something like, so if you

01:25:13   have a membership that has at least 10 people, something like that, um, then, you know, in

01:25:16   Overcast, then that'll be transcribed as well. And all of this is running on these 48 Mac minis.

01:25:23   Do you really need 48 at this point? I'm not trying to snark. I'm genuinely asking you,

01:25:29   like, what do you think is your steady state need if in order to handle the load that you have

01:25:34   before I fixed the bug that was causing them to repeat themselves? Um, I, I thought I could never

01:25:40   get enough. Now when I've reached a steady state, I do drop below a hundred percent use, but what I've

01:25:48   been doing is just going back more and doing more back catalog work. And whenever my queue would hit

01:25:53   zero, I would just like lower the threshold. It's like, all right, now, now go back five years. Now go

01:25:59   back seven years. Now go back 10 years. And every time I would do that, the number of jobs would spike

01:26:03   up and then they would slowly churn through them and, and, and they, they prioritize new releases and

01:26:08   they prioritize popular podcasts. So there's not much cost in having the queue be full of a bunch of

01:26:12   work to do. Eventually I would, I would catch up with at the new level because all those back catalogs

01:26:17   will have been transcribed. And then I just, you know, lower, lower the limits again. The next,

01:26:22   the next thing to do is lower it to podcasts that have one listener. Cause right now I'm at like,

01:26:27   if you have two or more listeners, again, more than one listener, that's one thing. But if,

01:26:30   if I go to one listener, then that like, I think that like doubles the number of podcasts or something

01:26:35   that I have to do. So that that's its own decent size jump. But once I'm done with a lot of the

01:26:40   back catalog work, I think I will be probably ready for that at this level. Now I also, I realized like

01:26:48   I will have a surplus of computational power here before too long. Um, so what do I do with that?

01:26:54   And well, I can start moving work to them. First of all, like things like imagery sizing that I have

01:27:01   right now, my main Linux servers at, at, uh, Akamai there's, you know, they're doing a lot of that kind

01:27:06   of work. I can move that to these. Why not? They're super fast. They have all this great stuff. I even

01:27:11   actually tried earlier in the winter, you know, how a lot of podcast apps will use the dominant color

01:27:18   in artwork to tint the interface in some way. I tried running that on these, like where like one

01:27:25   of the jobs they would do like, so one thing they do is language detection. As I look into more like

01:27:32   smart transcript features, I can do things in the future that I haven't done this yet, but I can do

01:27:36   things like, you know, automatic chapter detection or topic detection or summarization, things like

01:27:41   that. Well, now I have all this computational power I can do that with. And it's not going to be,

01:27:46   it's not going to be the same as if I was running like, you know, uh, chat GPT's flagship model or

01:27:51   Claude's flagship model. It's not that kind of hardware, but it, it is basically a whole bunch

01:27:56   of computational power that will be almost free to operate over time. And so there's a huge incentive

01:28:02   for me to move work to them or to find work they can do that would help the product. So I'm not worried

01:28:06   about that. I think I will have plenty of work for them to do over time. And also I had to solve

01:28:12   the issue with dynamic ad insertion. I need transcripts to reach a certain minimum level

01:28:19   of functionality for me. One of those is like, all right, I definitely want them to be time synced.

01:28:24   You know, I don't, I don't just want a static transcript that just dumps all the text of one

01:28:30   version of this, but then the version that you have that has different ads in it doesn't line up.

01:28:34   Everything has to line up. And I thought, well, how do I do that? I spent a lot of the summer just

01:28:40   trying, developing and testing different algorithms to basically make signatures of the audio to analyze

01:28:48   the audio to say, all right, where, you know, what kind of waveforms are there or what kind of

01:28:54   frequency pairings or frequency patterns or like, like I tried all these different algorithms, like

01:28:58   somehow characterize what this audio is. So that way, when the server transcribes the version that

01:29:05   it gets, and then you on your phone download your download of, of that podcast, but yours has

01:29:11   different ads spliced into it. So the, you know, it might not line up or it might have things inserted

01:29:16   or removed. I want my transcript generated from the server to be able to be lined up with what your

01:29:21   phone got from, from its different download of the same podcast. So how do I do that? What actually

01:29:27   is like a unique way to identify different parts of the audio? Honestly, I got to say AI was very

01:29:32   helpful in this. Um, AI did not write the algorithms for me, but it did tell me what kind, like it,

01:29:39   you know, I, I was asking AI models last summer, like how do different companies solve this problem?

01:29:43   Like what, what do different apps use for this? What kind of approaches are there? Um, and it was like,

01:29:48   oh, here's what, you know, uh, Shazam uses for its fingerprinting. Uh, here, here's how Spotify does

01:29:53   there. Here's like, they tried to take some guesses on how YouTube content ID works. Like there's all

01:29:57   these different things like, okay, this actually very helpful. Um, I was able to over the course

01:30:03   of a few months, get a signature algorithm that worked, that was able to read the same file.

01:30:08   And I could like, you know, I was doing tests. I would like, you know, download something from

01:30:12   here, then download something through the Netherlands Mac mini that I was renting. And like, all right,

01:30:16   we have different ads now. Um, compare these, like, can the algorithm find the parts that are

01:30:21   the same and line them up correctly? Um, can it detect which parts were inserted or removed so I can

01:30:26   then adjust the transcript in those areas? Um, and it took months, but I finally got there and

01:30:33   between the transcript algorithm and the signature algorithm, all of those things were just barely

01:30:42   fast enough to plausibly run on the iPhone. And they all use APIs that were built into the iPhone.

01:30:50   So I'm like, wouldn't it be great if I was able to offer on device transcription for any,

01:30:57   for any podcast episodes that I didn't do myself server side. And so from that point, I was developing

01:31:04   the entire pipeline as something that also compiled and ran on iOS. And I'm happy to say I was also able

01:31:13   to ship that so that in, in the beta today, like any podcast that I'd haven't transcribed,

01:31:19   you can just hit transcribe on your iPhone and it'll, it'll do it. It might take like, you know,

01:31:23   a good, you know, three to five minutes depending on how long it is. Um, and it uses the new background,

01:31:27   the background, um, activity API that introduced an iOS 26 for final cut exports. It uses that one

01:31:33   where like you can do like the long running processing task. If you start it from the phone

01:31:36   or from the app in the foreground, and then it puts the progress in the live activity.

01:31:40   So if you have iOS 26, you can transcribe on device too. And critically, you're making the same

01:31:47   transcript. Like it isn't like the ones you make on your phone are worse than my server ones. They're

01:31:51   the same. It's the same models between the Mac and iPhone. It's the same models with the same quality

01:31:55   and running, you know, all the same process. Um, so that I think is another great feature of this.

01:32:01   Um, but the, I gotta say the signatures took the longest by far. Um, and I had to learn all sorts

01:32:07   of stuff too, about like long running Mac processes, you know, obviously memory leaks, weren't going to

01:32:13   be tolerable, but that was, that was easy to get around or to avoid rather. But things like, how do I

01:32:19   make, like, what do I do if my process crashes? And the answer is launch D like you, you basically have

01:32:24   like a launch D, you know, thing that you, you create that says like, just automatically always make

01:32:27   sure this app is running. And if it crashes, just open it up again. Um, when I was

01:32:31   doing the, the image color detection that I was saying earlier, like detect the dominant

01:32:36   color in this image, there was a part of that code that kept like seg faulting and a really

01:32:42   deep part of the accelerate framework. And I could not figure out why it seemed like it's

01:32:46   actually part of, part of Apple sample code. And I couldn't, it was so deep in the stack.

01:32:50   It was like, I think it's just a bug in like the, in the OS about this call that I'm calling.

01:32:56   There was no way to avoid this crash. So I learned XPC and I made it its own little XPC

01:33:02   child process so that when the image detection would crash, it wouldn't crash the whole transcription

01:33:07   client. Uh, like I had to learn, I had to learn things like, you know, Apple remote desktop,

01:33:13   which is how I meant right now. This is how I'm managing all the servers is I literally have,

01:33:17   I have like a setup script, but then, and I can SSH into all of them. So like when I update

01:33:21   the transcription app, I use a script to just SSH, copy it all over and relaunch it.

01:33:27   But if I have to like change Mac settings or things like that, I have to like log into the

01:33:33   screen of the Mac with remote desktop and do it that way. I'm sure that Mac, like it admins

01:33:40   know of way better tools to do this. I would love to hear about those because this is ridiculous.

01:33:47   Um, and I don't want to log into 48 max to do like OS updates down the road. Um, so I would love to

01:33:52   hear about those, but just lots of learning about like how to run fleets of Macs and Apple does not

01:33:58   make that easy. They seem like they don't really spend a lot of time considering that, but it has

01:34:05   been running just fine. It is working. One trick I learned, um, for a long time there's been in,

01:34:11   in like what used to be called Mac OS server and is now just built in. Um, one of the things in the

01:34:16   sharing panel is called content caching, content caching, basically like max detect if there's a

01:34:21   content caching server on the network. And if there's like a software update that comes through

01:34:25   the content cache Mac will download it. And then when every other Mac in your house does that software

01:34:33   update, it'll just pull that copy over your local network. Instead of, you know, if you have five

01:34:37   Macs, instead of each one downloading Mac OS, you know, 26.2.1, five times, you'll download it once

01:34:44   to the content cache server and then all of your other Macs will download it directly from that one.

01:34:48   Well, that matters a lot when you have 48 Macs on a network. So, you know, I designated a few of them

01:34:54   to be content caches and, you know, just little stuff like that. Like it, it's been a really great

01:34:59   overall experience. It's been a massive undertaking. Just like the amount of work this has been on so many

01:35:06   levels, the signature algorithm, the transcript algorithm, like building out the whole queue processing

01:35:13   system for external servers that aren't in my, my main Linux server network, the health monitoring, the remote

01:35:19   desktop, like all there's so much here. And, you know, then serving the transcripts, using them in the app,

01:35:25   lining them up. Like when you do get a different signature, like how do you line up that signature with the one that

01:35:29   the transcript has? There is so much work behind this. But where I've gotten is I got to learn a bunch of

01:35:37   cool server stuff. I got to do some really tricky programming problems, which that's always fun for a

01:35:43   programmer. If we're honest, we love, we love tricky problems, even when they're frustrating and take months

01:35:48   to figure out. But we do love when we do, when we finally do figure them out. And now I have a very

01:35:54   capable computational cluster that I can throw a bunch of work at and it gets it done very cheaply.

01:36:02   And I believe I am the only podcast app to do private podcasts, at least server side. I also believe

01:36:12   I'm the only one where you can like do it on device if they don't have it. I'm certainly the only one

01:36:17   besides Apple that's going to do it for free. But but that's I I'm so I think I think this is all

01:36:23   worth it. And the feature is great and will only get greater over time. Right now, it's again, it's

01:36:30   very basic right now. It's you know, you can you can view the transcript and you can tap or it automatically

01:36:34   scrolls with the audio. It highlights what's being said. You can tap anywhere in it. You can scroll

01:36:39   around, tap anywhere in it to seek to that point. There's a lot of kind of obvious features that

01:36:45   like, well, once you have transcripts, you will obviously also want X things like transcript

01:36:50   search. That's an obvious like number one, like, yeah, you should be able to search transcripts.

01:36:54   I'll get there. Chapterization summaries of sections. There's a lot of that kind of thing. Like, yeah,

01:37:00   that I should do that clip sharing where you can share a clip like, yeah, that that should be

01:37:06   transcript based and the output to include the transcript. Yes, I agree. Like there is so much

01:37:11   that now follows from this. But this has taken me almost a year and I wanted to get this like I got

01:37:17   to ship something. I got to get this out there. And even this like there were a couple of days last week where

01:37:24   I broke the build on my own phone. I broke its ability to use transcripts. So for a couple of days after I had

01:37:30   been using them for a while for a couple of days, I couldn't use them. And I hated it. Like I noticed

01:37:36   immediately because what you know what I do is like I swipe over to the transcript view when like whenever

01:37:41   I'm listening to a big show that has stupid DAI ads, like I'll hear of great another ad for the Apple

01:37:46   card. Awesome. Oh, another ad for pure leaf iced tea. Okay. And I can I can go over the transcript and I can

01:37:53   skim skim skim and just tap right after because it's very obvious where the ad ends. That's great.

01:37:57   Like features like that, you get used to that or even just like a what did they say? If you listen

01:38:03   to a podcast and a bus drives by and you miss a couple of words, you don't have to seek back and

01:38:08   repeat the whole 20 to 30 seconds. You can just swipe, swipe, swipe and see. Oh, that's what they

01:38:12   said. Okay. Like stuff like that. It's really great for that kind of thing. It's also really great for

01:38:16   like, you know, if there's like if there's a podcast that like, oh, I know somewhere in here,

01:38:21   they mentioned something like this, you can quickly skim around and find it.

01:38:25   Or if it's like, oh, I got to listen to this episode before tonight, but I don't have that much

01:38:31   time. But there's information in it that I would like to know generally what they talked about.

01:38:36   That's all you can. You can skim the text of what they are talking about faster than you could hear

01:38:41   it. So you if there's like if they're talking about baseball, you can skip it. You know,

01:38:45   so there's like there's all sorts of little benefits like that, you know, much of the benefits

01:38:51   of chapters in podcasts that do it, but just applied to all podcasts. It's it's a great experience. And

01:38:59   it is very clear to me, having developed this and now using it, it is extremely clear to me

01:39:06   that this feature is table stakes, that all podcast apps need transcripts. And once you are used

01:39:15   navigating a podcast with transcripts, using seek forward and seek back buttons feels like you're

01:39:22   a dinosaur. So it is very obvious to me that podcast listening and podcast navigation require this for

01:39:33   any app that wants to be like a serious podcast experience, you know, to be a good listening

01:39:38   experience. You need robust transcript features now, and that will only get more so over time as

01:39:45   we kind of evolve our UIs and our playback experiences to use them more and to offer more,

01:39:51   you know, offer more utility that's powered by transcripts. So I think this is, you know,

01:39:56   I don't I don't always arrive to the party on time with this kind of thing. But I'm I'm here now and

01:40:04   I'm very glad to be here. And it was a long road to get here. It was a ton of work and a ridiculous

01:40:10   amount of, you know, I don't know. It was an astonishing amount of work and expense to to a

01:40:22   not small degree. Um, one of the things I keep wondering as you talk about this to the degree you

01:40:28   are willing and capable of sharing, how are you storing the transcripts? Because, you know, the obvious

01:40:34   question as well, you know, the obvious answer is, oh, it's just a bunch of text. Well, no, it's not because

01:40:40   then you have timestamps and well, but I mean, where is he storing it? Well, that too. Yeah. The cloud,

01:40:46   obviously, of course, of course. Yeah. But like where, where are two cloudflare are two. Okay. And what is the data

01:40:53   type? What? I know some of this is special sauce. You don't want to get too specific. But to the degree you're willing

01:40:59   to share, like what, what is being stored for each of these transcripts? Well, you know, me, of course,

01:41:05   it's a custom format. Of course. I was going to say it would normally be JSON, but knowing Marco is not

01:41:10   going to be JSON. There is JSON under there somewhere. Um, but of course I'm like, well, if I customize it,

01:41:16   I can compress it to be a lot smaller than that. And so, yeah, there's, there's compression, there's

01:41:21   shorthand, there's like, you know, Delta encoding. The total storage space for all the transcripts of all

01:41:26   the podcasts in the world in text form, gzipped can't possibly be that big. No, well, I wasn't

01:41:32   always storing them on R3. I started out storing them in a database and like, that was stupid. R2 or

01:41:36   not R3. But yes, sorry. Yeah. I think you're, I think your problem is going to be that the minimum

01:41:40   block size of R2 is probably bigger than all your transcript files. So it might actually behoove you

01:41:45   to combine them into multiple files if you cared about storage, which you don't. Well, but keep in

01:41:49   mind also that like these transcripts. So right now in, in the UI and overcast, it is displaying,

01:41:55   you know, each kind of like line of text, so to speak. It's, you know, roughly like one to two

01:42:00   sentences, maybe it's, you know, it's displaying them like a series of short paragraphs when it is

01:42:06   highlighting whatever is being spoken. It highlights the entire paragraph at once, but I actually have

01:42:11   word level timing. Oh, gracious. That was my next question. The, the actual data I have is much more

01:42:18   precise than what's showing. I have word level timing and word level confidence. So I'm actually storing a

01:42:23   good amount of data for each transfer. Like I, I knew like if the API is going to give me this

01:42:27   information, I'm going to store it. And you know, I don't, I might not necessarily use it for anything.

01:42:33   Maybe I, maybe I won't use it now, but maybe down the road I will. Like for instance, if you're doing

01:42:38   clip sharing, you know, if you look at what, like there's, there's other apps out there that will

01:42:42   export videos that have like live captions of what's being said. That tends to look a lot better

01:42:48   if you highlight word by word. Um, or if, if you're doing something worth exporting a video or a

01:42:53   picture and say like the, the block, the current block of text is just too long to fit like in the,

01:42:59   in the frame, uh, you'd want to be able to shorten it. And so you would need, you know, sub frame

01:43:05   precision for that. So I'm like, if, if the API has given me word level timing, I'm going to store

01:43:11   it. I don't store everything. Like there's some of it, it will give you like alternate, uh,

01:43:15   transcription options. If it's like, if it's not too confident what was said, it'll see like, well,

01:43:20   I think it's 60% sure it's this, but 40% sure it might be this. I did experiment with those at first,

01:43:26   but I found that the, the alternatives that it was suggesting, they were such low confidence.

01:43:32   Most of the time it wasn't worth storing. So I mostly did not store those on that front,

01:43:36   by the way, that's another one of my questions. So you've, you've got all these text transcripts

01:43:40   now. And as you noted, you're using models optimized for the fact that they can run on the phone and

01:43:44   they're inexpensive to run on the server and so on and so forth. And they're not going to be

01:43:48   up to the standards of the gigantic models that, you know, that you could run in the cloud or

01:43:53   whatever. Right. Um, but once you have the text, like, cause I was looking, I was testing this

01:43:59   feature earlier today. And I looked at the, uh, part of the previous episode where someone sang a line

01:44:05   about, uh, someone walking the length of the city. Um, and the transcription was, they say he walks the

01:44:12   lake L-A-K-E of the city because length and lake sound similar. But what I was wondering is, could you do

01:44:21   a post-processing pass with some L-M thing where you feed it the text? Because I think any kind of

01:44:30   good model would say the lake, the lake of the city doesn't make as much sense as walking the length of

01:44:35   the city. Like it could figure out from context clues in the English language sentence structure

01:44:39   that lake is the wrong word choice there. And especially if you have stored alternatives,

01:44:43   it doesn't have to hear the audio. It's not correcting the transcription saying, oh, I'm going to use it more

01:44:47   props. Obviously then you're just doing transcription twice, but I do wonder if doing like essentially

01:44:53   text grammar correction or best guess correction with like a system prompt that says, um, you know,

01:45:00   like if you had like a summary description of the podcast from the description and the fee,

01:45:03   this is a technology podcast. I don't know if that would help with length of the city, but anyway,

01:45:06   just a world knowledge model would pick out those lyrics and correct them and put in length.

01:45:12   I do wonder again for maybe for the top 100 most popular podcasts in the most recent episodes

01:45:17   that doing a second pass and saying, this is my transcribed thing by my little models. And then

01:45:23   chucking the text to another model and saying, fix this up and make it. So it doesn't say lake of the

01:45:28   city. That is something I've thought about. Um, and, and, and look, the accuracy of the on device model

01:45:34   is fine, but not amazing. You've got, you had to do the thing that everybody does, which is you put a

01:45:38   big warning at the top and say, you know, this is auto generated. There might be errors.

01:45:42   I said, there are probably errors because they are, there are probably errors. Um, you know,

01:45:46   cause, and you know, in my testing, it is good enough to know what's going on. It's good enough

01:45:53   to, to have a, have an idea of what's being said to get most of it and to be, to figure out a lot of

01:45:59   that stuff. It is not perfect enough to be like a document that you're going to show somebody or

01:46:05   especially with like proper nouns and stuff like that, where it's going to struggle.

01:46:08   Right. Exactly. Like it again, like it's, it's the iPhone dictation model, I think, or, you know,

01:46:13   roughly it is not, it's not like a world knowledge model. That's going to be like, Oh, I know the

01:46:17   name of that because that's the name of a country and the capital of that country is, but no, it's

01:46:20   not doing that.

01:46:21   Right. Like whisper is better at that. Parakeet seems in most people's testing, um, parakeet seems

01:46:27   like it is a little bit better at that. So that's why I'm going to look at that soon. But, um,

01:46:31   it's not like massively better because that, that is mostly just down to like model size and

01:46:35   complexity.

01:46:36   Are you, are you resigned to redoing all of your transcripts on M sevens with the more powerful

01:46:40   model in a few years?

01:46:41   I can't. So there is a versioning system in the transcript data. So yeah, no, I'm just saying

01:46:46   like, are you, are you already setting up to say like, I know I'm doing all this work, but at some

01:46:50   point in the future, I'm going to have to redo it all, like all the whole back catalog again, because

01:46:55   the models will have gotten that much better. And because the M seven will be that much more

01:46:58   powerful. I mean, yeah. So like I expect over time, this will get better, but like, so going

01:47:03   back to like, you know, if I like, you know, using like a big model somewhere to clean up the

01:47:08   transcripts that is on my list to investigate, I think the costs would be prohibitive to do it for

01:47:15   everything. But as you said, like if you, if I just do it the same way, like do have like a,

01:47:18   a popularity minimum, I do it for everything over this.

01:47:22   Yeah. I do wonder like, cause cause it's not transcribing audio, it's just text. I wonder if

01:47:27   it would just be trivial for it, you know? Cause like how much text is there? Honestly, it's a

01:47:30   tight, such a tight, it's like, forget about it. Like the text would fit in one thousandth of the

01:47:35   context window of these giant things these days. It's just, it's so much easier than transcribing

01:47:39   audio, you know? Yeah. So like what I'm going to look at next, uh, well, maybe not exactly next,

01:47:44   but what I'm going to, what I'm going to look at soon now that I have the transcripts, what smart

01:47:49   things can I do with them? And that, again, that would be things like finding like topic changes and

01:47:53   trying to make like a table of contents, you know, make, make chapters that aren't that, you know, if,

01:47:57   if, if podcasts don't have them themselves, um, you know, summarization of chapters maybe. So I

01:48:03   could say, okay, at this part, they're talking about volleyball, you know, like that's the kind of

01:48:06   thing I want to be able to offer. And my, my plan in my mind, which I have not tested anything yet,

01:48:13   but my plan has been to evaluate, like, that's when I go to the big API company or the big AI

01:48:20   companies and use their APIs. So, so I, I would use the open AI or the Gemini or the clawed API to say

01:48:26   like, all right, here's this block of text, find me where the topics are, clean up any errors that you

01:48:32   see that, that are probably unlikely to be what the person actually said, you know, stuff like that.

01:48:35   That is the next big step in making transcripts better. I'm not ready to do that yet. Like I had

01:48:41   to get this part out first, but can you use private cloud compute or is that only allowed to do like

01:48:46   by users from shortcuts? Like I'm wondering if you could. Yeah. So I could theoretically write a

01:48:51   shortcut and shell it out to private cloud compute, but I think I would get throttled really quickly.

01:48:56   Yeah. Yeah. Yeah. There is a limit on that. Now, I mean, I'm, I'm thinking about like on device,

01:49:00   like, cause you mentioned obviously like transcript search and yeah, you know, you full text search of

01:49:04   transcripts, blah, blah, blah. You know, you can do that all on device. Again, it's not,

01:49:07   it's not that much text. You're searching through one podcast worth of stuff, probably even if you do

01:49:11   all podcasts, whatever. But I also think there's a place for you to be able to do essentially LLM powered

01:49:17   search where you just chuck the whole transcript into the context window and let the person ask a

01:49:21   question and it finds the part. You know what I mean? So for, for one episode, that's easy. Yeah.

01:49:27   Doing it on device, doing that on the iPhone, the context window is only 4,000 tokens. So it's,

01:49:33   it's pretty tough.

01:49:34   Oh yeah. All right. Well, that's why I was asking about private cloud compute. If you could just

01:49:39   chuck the whole thing up to a cloud model, it was, I feel like it would fit in the context window of a

01:49:42   big cloud model.

01:49:43   It would absolutely fit. Like if I wanted to offer that, like with a big cloud model through an API,

01:49:47   like I could offer that that way.

01:49:49   Yeah. Bring your own API key or whatever.

01:49:50   Ah, that's cool. That's gross. I would just, I would find a way to do it cheaply. Like that's,

01:49:54   I would see like, what can I do for nothing or for near nothing? A lot of this, I mean,

01:49:58   first of all, like AP, like API costs for high end models are only going down over time. And it's

01:50:05   only, only going to continue. And what I'm asking for is not that complex. Like this, this wouldn't

01:50:11   require like a really high end, like thinking and logic reasoning model. Like it wouldn't require

01:50:15   that level.

01:50:16   Yeah. You could run this, like you could run this on your Mac minis with some like open weights,

01:50:21   like llama model thing. Cause you're really just doing fuzzier full text search essentially,

01:50:27   you know, and I'm not saying you should do this first. You should just do plain full text first

01:50:30   search first. Cause no, no AI, no anything, just full text search when people type in words and

01:50:35   you match and blah, blah, blah. But because it is just text, any kind of remotely reasonable,

01:50:42   small LLM powered thing that you can fit the transcript into the context window and then just

01:50:47   ask questions about it and get a, get a timestamp out of it. That would be amazing because as

01:50:52   someone who spends a surprising amount of time searching for transcripts, like when I tried to

01:50:56   find the, the, the quote that I just read about the AirPods pro max thing, I foolishly didn't put that

01:51:01   information in the reminder. So I'm like, Oh, sometime in the past, we talked about AirPods max and Marco

01:51:07   said something and I said something and I had to go find that. Um, just a little bit, just a little

01:51:14   tiny bit of fuzziness helps go, it goes a long way because, and especially for like,

01:51:19   are the transcripts going to get AirPods max? Is it going to get AirPods? No, that's a, I guess,

01:51:24   air and pods are two words and you have to think about sound alikes and stuff. And so I think any,

01:51:28   any little bit you can do to help pass the, the sort of very simple, uh, full text, you know,

01:51:33   non AI full text search will really make this feature more powerful.

01:51:38   Yeah. And I think like, that's, that's something to look at like, you know, in, in time as all these

01:51:44   things get cheap and I can run more of them locally. Like, you know, one limitation is that all these

01:51:50   Mac minis are just the, again, they're the base model. They're the 16 gig of Ram base model. So

01:51:55   there's only so much I can do on device when you start talking about bigger models. That being said,

01:52:01   there's nothing stopping me from say, adding a couple of high Ram Mac studios to this. And then,

01:52:08   you know, they would take on a certain type of job. Like I could do that.

01:52:11   What do you think Apple's using for its transcripts? Are they using their M2 ultra base servers?

01:52:14   Probably. Yeah. Cause like, yeah, whatever Apple podcast is doing. Yeah. Or they, they could,

01:52:18   they might've also done the same calculation and figured it's actually cheaper. You just do a whole

01:52:22   bunch of base Mac minis. Cause it, you know, it is like when you look at like number of, you know,

01:52:27   transcription minutes per minute that you can get on the base model Mac mini versus if you get like

01:52:33   the highest end M4 max, or I guess it's an M4 pro in the Mac mini. So the highest M4 pro it's not

01:52:41   twice. It's, or I think it's like, it's about twice as much at the most, but it's like four times the

01:52:47   cost. So it's not worth it. I have to imagine that they're using AWS servers to do their transcriptions

01:52:54   with probably with Parakeet models, but I don't know. Someone from Apple write in and tell us,

01:52:57   are you actually using Macs to do podcast transcription or is this all farmed out to, uh,

01:53:01   AWS? I mean, with Apple, it could go either way. Cause they have, they have different deals than you

01:53:06   would get with AWS in terms of pricing. Yeah, of course. Going through any, like doing all of this,

01:53:11   like any answer that begins with, why don't you just like, why don't you just use AWS? Why don't

01:53:15   you just use opening eye, whatever it is? Well, I just asked if you had priced it out. I think

01:53:18   what you ended up doing. Yeah. And the answer is thousands of dollars a day is what that costs.

01:53:22   Well, and you, you would also, uh, have to learn an entirely different set of things

01:53:27   than you learned for this project. True. Yes. Um, and I, one thing I also really like about the way I've,

01:53:33   I've built this, I can run all the same code that I've already written. Like, so first of all,

01:53:39   there's an advantage of Apple's APIs are really good for a lot of this stuff. So for example,

01:53:45   the music detection I'm doing, that's using Apple's built-in audio classifier. It can do stuff

01:53:51   like detect, like when dogs are barking, like, you know, I'm not using that part of it, but I could,

01:53:55   that's all just really easy. It's using platforms and languages, tools, and all of these things that

01:54:02   I'm already familiar with. I already, I'm using, I'm already deploying, I'm already like in it.

01:54:06   In this case, running these transcripts is using overcast code. The same code is running on the servers

01:54:12   and in the app. I only write that code once. I'm able to use Swift instead of like PHP or whatever

01:54:17   else I would use on the server. There's all sorts of benefits to having this just be one code base.

01:54:22   And like, as I am, as I'm, as I'm, you know, getting older in my career and as, as the world

01:54:29   is getting more and more, like there's more and more in the tech business, no one can keep track of all

01:54:35   of it. There's high value in specializing. And so in my case, like if I can write really good

01:54:41   Swift code, or if I, I mean, to the extent that we're still writing code in a few years, who knows,

01:54:47   but like if, if I can write really good Swift code and I'm really familiar with Apple's platforms and

01:54:51   Apple's APIs and Apple's like infrastructure and how to, how to run all that stuff, I can be, it's easier

01:54:57   for me to become an expert in that and to keep that up to date rather than like, I'm already falling

01:55:03   way behind on my like Linux server side knowledge, just because I don't do that much of it anymore.

01:55:09   I obviously, I still run a lot of servers for Overcast, but like they're not changing that much.

01:55:13   I'm not getting a bunch of new ones all the time. I'm not keeping up with like all the, the most

01:55:18   modern like toolkits and things that everyone's using these days. Like, so specialization is,

01:55:24   is I think is high value to me. The fact that I can just run, like I can do all this cool stuff

01:55:31   using Apple's tools and using the languages I already know and everything I run on the

01:55:36   server can run on the phone. That's amazing. And then down the road, if I'm still running

01:55:40   Overcast in 10 more years, at that point, the phones will be so fast. I won't need these Mac

01:55:46   minis anymore. Like I, my guess is I will probably never replace these by the time there would be

01:55:52   like a major upgrade in performance that would be worth replacing M4s. I don't even know if I'll need

01:55:58   these anymore because the phones will be so fast. They might be able to do everything on device. And I'm

01:56:03   like, there might be no reason to have all this work farmed out to these Mac minis.

01:56:06   Then you'll need a coordinate server so that it knows, um, you know, this person A has already

01:56:12   started transcribing the latest, uh, episode of the daily. So everybody else just hold off and wait

01:56:17   for that. Don't bother burning your batteries. Maybe let two people do it just as a backup or three.

01:56:21   Yeah. I mean, right now I already have that infrastructure in place. Like right now the,

01:56:26   the code is already there for the iPhone app to be a transcript worker. I'm just not activating that

01:56:34   because there's not much that an iPhone can do in like the, like there, there, there's a type of

01:56:39   background process. Uh, there's a type of background refresh task. Um, not the one that was introduced

01:56:45   in iOS 26, where you can like do the final code export with a live activity, not that there's a

01:56:49   previous one called a background processing task. And you can, you can request the system, wake up your

01:56:56   app in the background and give you a block of like high CPU time when it is charging. So that way you can

01:57:03   burn some battery power and it's not like you're not like doing a disservice to the user cause it's

01:57:08   charging. So usually that'll, that'll run like if when your phone is charging overnight. So I could

01:57:12   use that time, uh, on like everyone's phones to do some work. I'm not because that kind of feels a

01:57:20   little wrong cause that, that kind of feels like I'm stealing resources from people. Um, and the amount

01:57:25   of work you can get done on an iPhone in that time slice, they give you like three minutes or something.

01:57:30   It's not enough to even transcribe one episode of most things. So I'm, I'm doing all this work

01:57:35   myself, but in five or 10 years, when the phones are a lot faster, some of that math changes. Maybe I can

01:57:42   just do all of the work during people's overnight charging. If I wake up everyone's phone for one

01:57:47   minute and they can, they can all do all the work. Like maybe, I don't know, maybe people can opt

01:57:52   into it. I have no idea, but keeping all the code in Swift in the iOS code base, or at least

01:57:59   compatible with the iOS code base, even if I never do that scheme, that it still allows me to do things

01:58:04   like you can on demand hit the transcribe button and transcribe a podcast I didn't get to yet, or that I

01:58:10   won't get to. Um, so there's all sorts of benefits to doing it this way. So anyway, that's it.

01:58:15   I'm looking forward to you reinventing BitTorrent when you go the route of having all these phones

01:58:20   do like a minute worth of transcription at a time. I can just see it now.

01:58:24   I wasn't thinking of using them as like a, as a, as a, uh, uh, transcription farm, but rather instead

01:58:29   like on demand, you know, because obviously the phone is already downloading an episode because you

01:58:34   subscribe to a particular podcast, the RSS feed updates, the new episode downloads, you know, all that

01:58:39   stuff happens on people's phones and it's happening because they, you know, it's not like, oh, someone's

01:58:43   using my phone to download podcasts. No, it's cause you subscribe to it. Like you control that. The

01:58:47   user subscribes to a podcast and they want it to download the latest episode when it comes out so

01:58:51   that when they open overcast and hit play, the episode is hopefully already there. That's already

01:58:55   happening. I feel like part of that process could be, oh, and in addition to me downloading the

01:59:00   episode, I'll also transcribe it for you. Um, or cue it up to transcribe that as soon as the

01:59:06   overcast is in the foreground or whatever. And I was saying in that type of scenario, you don't want

01:59:10   70,000 people transcribing the new episode of the daily, right? When it comes out, because there's

01:59:17   a waste for all those phones to be doing the same work, but rather you just pick, well, these 10,

01:59:20   everyone who wants to do it talks to a coordinate server and says, Hey, I'm about to transcribe this

01:59:25   episode. Are there any other phones out there already transcribing? Because if there is, I'll just

01:59:29   wait until they're done because they'll transcribe, they'll upload the text and I'll download the text.

01:59:33   And you know, that type of sort of energy saving type of thing. But yeah, that's definitely down the

01:59:35   road when you have the computing power to do that. Basically with the consent that it currently

01:59:42   downloads, you know what I mean? Like it's downloading because they subscribed. I think

01:59:45   people will be fine saying, and also when you download an episode for me, transcribe it too.

01:59:49   And they don't care how it happens. They just don't want it to hurt their battery. They just want it to

01:59:53   magically be there. And it's not really possible to do that now. But like you said, in a few years,

01:59:57   I feel like if you did that, that would, people would accept that perfectly. They would accept it the

02:00:01   same way they accept the fact that overcast downloaded the episode for them.

02:00:04   Right. Exactly. And like, there's already like, right now I'm already, I'm already doing what I can

02:00:09   to share work where appropriate. So like, for instance, the signature, the audio signature that

02:00:15   like analyzes the audio and tries to line up with DAI, that is the same. If the copy of the file you have

02:00:23   it's a little bit heavy. It runs at like 600 X or so real time. So, you know, it might take a couple

02:00:29   of minutes or it might take a minute for, for a podcast episode or something. Um, and that does

02:00:34   run on the phone and it has, cause it has to, cause like the phone knows what it downloaded. I don't

02:00:38   know what the phone download. Do you, do you do a, uh, a content hash before you even bother on the

02:00:43   signature just to see if it's literally the same file or do you not bother? Yes. So first I do like

02:00:48   basically an empty five. Do you get the deprecation warnings from the stupid empty five library yelling

02:00:52   at you that it's a broken algorithm and you don't care? It isn't. It, I think it's, it's one of the

02:00:55   Shaw families, but it's one of the fast Shaw ones. Okay. I'm still, I'm still using an empty five and

02:01:00   Xcode. Yeah. And I know there's a way to change it that you can use a different API to get at the

02:01:03   same thing, but the different API is slightly slower. And I'm like, screw you. I'm just going to use

02:01:07   this API forever. Anyway, go on. Yeah. Anyway. So yeah, I first do like a file hash just to see.

02:01:11   And then I, and I go to the server and I say, do you have the signature for this file hash? Right. So

02:01:16   everyone who downloads ATP, since we don't use dynamic ad insertion, only one person is ever going to

02:01:22   generate that signature. And it's probably going to be the transcript servers that do it. Um,

02:01:26   cause they're also embedding the signature that they get in the transcripts. So for podcasts that

02:01:31   don't use DAI, the phones don't have to do this work at all. You need to add a button to the

02:01:36   ATP CMS that I can click that, uh, that forces immediate transcription of the latest ATP episode.

02:01:40   We don't need to, it's a, it's a popular podcast on overcast. So it'll be, it'll be prioritized.

02:01:45   Uh, I know you have the, the ping feeds API that people use. I'm just asking for the equivalent of

02:01:50   that for transcription. Yeah. Well, no, it's not going to happen. Not a public one just for us.

02:01:56   No, I don't, I don't need to. All right. We'll see. Yeah. Or you can download the episode

02:02:00   onto your phone and transcribe it for everybody. But yeah, so the signature thing,

02:02:05   yeah, it's like it, it checks that it hashes the file first and in the fast way, ask the server,

02:02:09   Hey, do we know what the signature of this hash is yet? For most podcasts, most of the time they'll

02:02:14   have that. The only reason you would have to transcribe, you would have to use the signature

02:02:17   yourself or generate yourself is if the version of the file that you got has different ads than

02:02:25   anybody else has gotten, then your phone will generate the signature for that hash and submit

02:02:30   it to the servers. So we are, so I am doing work sharing as much as possible there. So like the,

02:02:34   this, the work for the signatures isn't being duplicated. The work for the transcripts isn't,

02:02:38   isn't being duplicated. Um, but you know, it's still, it has to happen sometime, but yeah,

02:02:43   usually you're not facing that on your phone.

02:02:44   So if I get a copy of 99 PI that has, that's got local Richmond ads and nobody else has seen these

02:02:53   ads yet, how are you, how are you trans, uh, how are you making the transcript from that? Like,

02:02:58   are you uploading my MP3 to your server to, to churn on it?

02:03:02   Oh God, no, no. I mean, that would, that would be impossible. So in that context, what would happen

02:03:08   is your phone would download the transcript that Overcast already has for that episode.

02:03:14   The transcript would be, as you mentioned, for like a different set of ads that was injected in the,

02:03:19   in the copy that my servers got. And your phone would generate its signature for the file it has.

02:03:24   And then it would look at its signature and the transcript signature, and it would like find the

02:03:28   common ranges between them. Like where is, so basically like whatever is not different between

02:03:33   the two copies, it finds those aligns the server side transcript to those on your local copy.

02:03:39   And then any ranges in your local copy that were not on the server copy, it just doesn't transcribe

02:03:45   those. So you'll see gaps in the transcript with like an ellipsis there, which I believe is the same

02:03:49   thing that Apple podcast does on their transcript. Uh, anyway, so your phone is, is doing the bare minimum

02:03:54   work it needs to do to figure out what audio was it served. And then how to, how does it line up the

02:04:00   transcript to that audio? Gotcha. You can do a local transcription of just the ads.

02:04:04   I could, I mean, why? Yeah.

02:04:08   We're sponsored by true leaf iced tea or whatever it is. The same ads over and over again.

02:04:14   I'm so, I'm just so glad I don't listen to shows that have the dynamic ad and search on that.

02:04:17   Cause it's just, I hear everyone talking about it and it's, I mean, I guess I've heard a few of them

02:04:20   and it's not great, but yeah. Yeah. I, I signed up for the Verge's paid thing, which has ad free

02:04:25   podcasts finally. Thank God. But they still have the ad bumper. So it's weird. I mean, I know we have

02:04:30   the ad bumpers too, but I feel like it's incorporated in a way that you wouldn't know. It just sounds like

02:04:34   a set, a segment break, but their thing totally sounds like places where ads should go. I'm also

02:04:39   a Verge subscriber. Yeah. Yeah. Like I, I have like the paid search engine podcast one and their ad

02:04:44   bumper is like 45 seconds long. It's huge musical science. I have to like skip forward past it every

02:04:51   time. Um, but like I've, I've been, even though I have the Verge membership, I've been still keeping

02:04:56   the Verge cast public feed so I could test how my transcripts handled DAI. And I've heard the same

02:05:04   ads so many times between that. And like, I like Trevor Noah is probably the biggest, like the most

02:05:08   popular podcast I've probably listened to, I think. Um, and that of course is full of DAI. Like all,

02:05:12   all the big shows are full of DAI. Um, and I'm kind of looking forward to having this finally be

02:05:18   out there and having the signature stuff be, I think done so I can stop listening to all these ads and

02:05:24   just go back to the membership versions. One, one final quick feature request for the olds in the

02:05:29   audience. Uh, you need to make that text bigger or at least an option to make a text bigger on the

02:05:33   transcripts and also make the margins fatter. It is following the body text size that you set.

02:05:38   Um, so it should, it scales up with dynamic text.

02:05:41   There needs to be a separate setting for that. I mean, I think just, I, I'm thinking of like

02:05:45   the lyrics and Apple music, how huge they are. I'd be happy if they were that big, but I understand

02:05:50   music lyrics tend to be shorter than podcasts. So maybe there's some limits there, but my old eyes

02:05:55   say, why is this text so small? And also maybe do bolds for the current section instead of regular

02:05:59   for the current section and faded gray for the other ones. Anyway, those are my two quick feature

02:06:03   requests.

02:06:03   Yeah. I, I, I, the, the, uh, UI for it is very much like a 1.0 version of this UI.

02:06:09   Yeah. And, and bold, bold the current word since you have that info, you know, the whole deal.

02:06:12   Well, the, I wasn't, the reason I didn't do bold is that bold causes the text to change width and

02:06:18   therefore reflow.

02:06:19   there's no San Francisco variant that doesn't do that.

02:06:22   Um, actually I don't, I don't know if that, I mean, obviously there's monospace numerals, but I,

02:06:27   I do wonder if there's a variant that, uh, does not change metrics when you bold it, but I don't

02:06:32   think.

02:06:33   Well, then I will check then, then just change the color of the current word. Cause you've got the

02:06:37   info. That was like, that'll, that'll really help like the read along type thing of seeing the word,

02:06:41   like the little blue word or whatever color you make up, up, up, up, up, up. That'll be great.

02:06:45   Yeah. Well, and like, and what Apple does with their transcripts, um, is they, like the whole

02:06:50   thing, it's almost like the, um, the genie effect or the, um, what's the zoom effect on the dock

02:06:54   when you hover over things in the scale? I thought it was genie. Genie is when the minimized window

02:06:59   comes out in a little. Oh, sorry. Yes, yes, yes, yes. Sorry. Magnification is what you're thinking

02:07:03   of on the dock when the icons get bigger. Like we, yeah, when you scroll your cursor over the dock

02:07:07   icons and the one you're on grows up in size and the other ones shrink around it. Yeah. They do that

02:07:12   with their transcript text. So the current line of the transcript is bigger and bolder and the rest

02:07:18   kind of like fade into smallness and into the sides. Yeah. That's not great. It does solve some

02:07:23   problems. I think creates others. Um, but you know, so I will play with the UI. Um, there's, there's a

02:07:29   lot, a lot of iteration to be had here. Um, but gotta start somewhere. No, this looks incredible. You

02:07:36   should be very proud. I know this was a lot of work. Um, I can't believe you're announcing it when it's

02:07:41   beta though, because now everyone's like, when is this going to be released? Yeah, that's true.

02:07:44   Well, the good thing is like, I, so I, I mean, honestly, it's a big enough deal. I considered

02:07:48   skipping the beta and just going directly to the public. Not a good idea. Yeah. I realized that was

02:07:52   not. I know you're anxious to get it out, but do the beta. But I, I don't, honestly, I don't expect

02:07:57   this to be in beta for very long. Um, cause my goal here is not to revamp the whole feature with beta

02:08:03   feedback. My goal here is make sure it works to the level, like to the 1.0 level and then

02:08:08   get it to everybody. Um, so we will, we will see. All right. You said it would be half an

02:08:14   hour. It was 70, 82 minutes. That's it. Pretty, pretty, pretty accurate. If you add the Marco

02:08:20   multiplication factor. Yeah, that's true. I mean, I, I thought I was going to have this done

02:08:24   by the fall. So, you know, there, thanks to our sponsors this week, one password, Zapier and Lisa,

02:08:31   and thanks to our members who support us directly. You can join us at atp.fm slash join. One of the

02:08:36   many perks of membership is ATP overtime, our weekly bonus topic this week in overtime. We're going to

02:08:41   be talking about goals and features and the future of liquid glass in Apple's 27 series OSes coming out

02:08:48   this summer and fall. Uh, that'll be, that'll be fun. I'm looking forward to that. There've been a

02:08:52   couple of rumors recently on that. So we will talk about that. And over time,

02:08:54   atp.fm slash join to join us, to hear that. Thank you so much, everybody. We'll talk to you next week.

02:09:01   Now the show is over. They didn't even mean to begin. Cause it was accidental.

02:09:10   Accidental. Oh, it was accidental. Accidental. John didn't do any research. Marco and Casey wouldn't

02:09:18   let him let him. Cause it was accidental. Accidental. It was accidental. Accidental. And you can find the

02:09:26   show notes at atp.fm. And if you're into Mastodon, you can follow them at C-A-S-E-Y-L-I-S-S. So that's

02:09:39   Casey Liss. M-A-R-C-O-A-R-M-A-R-M-T. Marco Arment. S-I-R-A-C-U-S-A-S-Y-R-A-Q-U-S. It's accidental.

02:09:52   Accidental. Accidental. They didn't mean to. Accidental. Accidental. Tech podcast. So long.

02:10:05   All right. So there is some new news that broke. It was spoiled, I think, yesterday as

02:10:11   we record and is officially news, I believe today. The BMW i3 reveal has happened. This

02:10:18   is not the little ne-ne-ne-ne-ne-ne-ne-ne-ne-ne-ne-ne-ne-ne-ne car.

02:10:22   Yeah, not TIFF's car. Is that the sound it makes?

02:10:24   Yep. That's it. That's right. Well, only when the range extenders are on.

02:10:27   No, actually, the old BMW i3 was before EVs were required to make all those noises. So

02:10:33   it actually makes no noise. In any case, the i3 has been announced. So this is the

02:10:39   Neue Klasse or something like that. I'm sorry, Germans. That is basically, let's make an EV

02:10:45   from the ground up rather than taking the petrol cars, the gasoline cars and retrofitting an EV

02:10:51   into them. Even though, I mean, again, I'm an apologist for the i4 because I thought it was

02:10:57   incredible. But this is from the ground up. Let's make an EV. And so reading from a couple of

02:11:03   sites starting with the Verge, Sebastian Kroes, BMW's head of interior design for the Neue Klasse

02:11:08   cars, told me that the iX3 was designed with an emphasis on verticality to make it look taller.

02:11:14   The i3, on the other hand, has an emphasis on a horizontality. I don't think I'm pronouncing

02:11:19   either of those words right. I don't think they're words either, but I'm doing the best I can.

02:11:22   If you're a German designer, it is. Yeah, this is design ease. Yeah. Most directly seen in the

02:11:28   series of lights that span virtually the sedan's entire nose. The company hasn't quoted a formal

02:11:34   capacity for the i3, but I'd expect it to fall somewhere around the 109 kilowatt hour usable

02:11:39   battery capacity of the iX3 SUV. Enough for what BMW says is 440 miles on a charge. And 400 kilowatt

02:11:46   charging should mean that adding over 200 miles of range happens in about 10 minutes.

02:11:51   Neither of the two motors in this car uses permanent magnets, which has a few advantages.

02:11:56   First, you'll find no rare earths here. Secondly, the i3 can disable its motors and coast without

02:12:02   needing a disconnect system, unlike the Mercedes, which apparently does. From Mars Technica, weight

02:12:07   distribution is close to 50-50 with a low center of gravity thanks to the battery pack. And torque

02:12:11   delivery is rear biased out of corners and under regenerative braking, the rear axle regens more

02:12:16   than the front at first to stabilize the car. So technology wise, this looks really, really good.

02:12:23   Like this looks incredibly impressive. Visually, I'm not as convinced. So to my eyes, the profile is

02:12:34   excellent. The front is a vast improvement from the pig nose of the i4. You know, when I, when I told you,

02:12:41   I loved the i4, I said that while just ignoring the fact that the kidneys took up the entire front

02:12:48   of the car. This is better. It's a little weird, but it's better. I don't think I like the back at

02:12:55   all. And the interior, I am not here for the vertical braces on the steering wheel. You're just going to

02:13:01   have to see to understand. And I'm not here for the trapezoidal center screen thing. However, on the

02:13:08   whole, this looks really good. Well, you mentioned this being the, this being spoiled yesterday or

02:13:12   whatever. Well, the way BMW has rolled this out has been, I mean, you know, obviously they had the

02:13:18   original like concept cars, the Neue class of whatever, however many years ago, those came out.

02:13:23   And then obviously the i3 is out and you can buy and is the first car on this platform. Although

02:13:29   it's not a sedan, obviously, but this car, this exact car, the i3, um, they've been doing the thing

02:13:35   for, I think maybe a year, at least six months or so where, um, they've been allowing journalists to

02:13:41   see and show the car quote unquote disguised. And if you're in the car world, you know, the normal way

02:13:47   you disguise cars is you put a bunch of things on the car that use like black and white, like dazzle

02:13:53   camouflage, but just all sorts of weird swirly black and white patterns to make it. So you can't get a

02:13:58   read for what shape the car really is. Often there'll be kind of like a bra over the front and rear. So you

02:14:04   can't see what the front grill looks like. And sometimes they put other panels on it of like

02:14:08   cardboard or wood or whatever to just hide the shape of the car. But this car, the i3 has been

02:14:14   shown to journalists with essentially a wrap on it. Meaning like if you put a red wrap on your car,

02:14:19   people think your car is painted red, just a complete, just a wrap. And, but instead of the

02:14:24   wrap being red, the wrap is dazzle candle. So it's like, how much of this car are you hiding when you

02:14:30   just essentially put a wrap on it where you can see the whole car and yet it's maybe a little bit

02:14:34   harder to see what shape it is, but there's the car. There's nothing disguising it except

02:14:38   an ugly paint job essentially. But still I was holding out hope that when I saw the finished

02:14:43   version of it, it would not be as homely as the camouflaged cars they've been showing to the

02:14:49   press. I mean, they would let them drive them in the snow and do all this stuff, but that's the car.

02:14:53   And anyway, uh, there's not much that can be done to save this. I was hoping that they could

02:14:58   de-emphasize those giant, uh, you know, headlights on the front, uh, with styling that was not apparent

02:15:04   with the wrap on it, but nope, they're there. They're big. I don't like them. And also in general,

02:15:09   even though they said this car emphasized horizontality or whatever, it looks tall,

02:15:13   especially compared to the existing I four. I think the existing I four, like, do they make a coupe of

02:15:20   this? Whatever, whatever I four I've seen on the road, I've taken pictures of it before the current I four

02:15:24   body shape and back looks so good. And so just pretty and athletic and nice. And then you get to

02:15:33   the front and it's got big beaver teeth. I understand that. I'm not a fan of the front, but if you never

02:15:37   see the front and you just see it from the side and the back, it looks amazing. This car looks tall and

02:15:42   squat from the side. This car has ugly taillights and an ugly rear. And this car's front, I think is not

02:15:48   actually an improvement over the buck teeth. It's just ugly in a different way. You are so wrong, sir.

02:15:53   As for Casey's steering wheel that he doesn't like, I think I posted this on our little Slack

02:15:56   channel. I think if you get the M sport steering wheel, those spokes are hard. I put a picture of

02:16:01   it into our Slack. Those spokes are horizontal, not vertical, which helps it a little bit. But I am

02:16:05   also dubious about the comfort of the steering wheel because of the way that they've, it's not like a tube

02:16:11   around it. They've really done these big cutouts and like these strange sort of, you have to look to

02:16:15   see it. It's, it's a strange steering wheel. I'll give them credit for making the squircle a little bit

02:16:19   more Urkel-y than squircle. I mean, it's not, it's not a circle because you can't put circles on the,

02:16:25   on car wheels anymore because it's against the law, but it's not as bad as like, say the Corvette wheel,

02:16:30   which is just really weird hexagon thing. Um, I do like that they've updated their architecture.

02:16:35   This is basically the same thing, same architecture as the iX3. So it's good. Uh, I think they went to

02:16:40   800 volt with this. So I'm sure it'll be a great car, but just the styling problems happening over

02:16:47   there in Germany continue to be dire. I mean, not all in Germany is I feel like Porsche styling has

02:16:52   still been pretty good. And I'm mostly on board with it and Mercedes styling and BMW styling. They've

02:16:57   been, they've been in a rut for what, two or three decades now. Uh, I don't think it's near as bad.

02:17:04   Certainly this doesn't look near as bad to me as it does to you. And I don't think it's been two

02:17:09   decades. It's probably been one ish. Oh, no. You, you starts with the bangle, but that's the nineties,

02:17:14   right? Or is that 2000? No, that was early to mid 2000s, I believe. And even, I don't know. And it

02:17:19   wasn't all of the cars. It was some of the cars. I don't know. Like the three series has mostly been

02:17:24   good with like the, the, the 90 that I had was good. The F10 was most, or no, that's the, the five

02:17:30   series is an F30 is what I'm thinking of. F30 was all right. And then it got ugly after that.

02:17:34   We just didn't know what was going to happen. So you look back in the F30 and you're like, wow,

02:17:38   that was nice. Yeah. That's fair. And it's got dump store handle. It's got basically Tesla door

02:17:42   handles in 2026, 2027 model year. You're going to do Tesla. They're not exactly the same because

02:17:47   there is a mechanical pull on them, I believe, but I think they might still have to redo them for

02:17:51   China. You know, we did that story about China banning, uh, electronic door handles. I think they

02:17:55   might still have to redo them for China just because the electronic mechanism makes it come out

02:17:59   like the model S handles. So it doesn't matter that there's a mechanical pull. If you can't

02:18:03   actually get your fingers behind the thing, cause they made it like, why, why, as you noted, Casey,

02:18:07   the I four has flush door handles that you just put your fingers under and pull. And I don't even

02:18:12   know if those are mechanical. Those might be metal. Uh, what do you call it? Uh, electronic switches.

02:18:16   But the point is they could be mechanical because you don't need anything to happen. You just walk up

02:18:20   to the car, stick your fingers underneath the handle and pull. Like it's, I sent you the video of like

02:18:25   the Honda Prelude thing, whatever that was like 1995 Honda Prelude flush door handles,

02:18:31   a hundred percent mechanical. You just put your fingers underneath it and you pull and the door

02:18:34   opens, but no, not on the BMW i3. They have to electronically come out like it's 2012 and you're

02:18:41   a model S. So dumb. With the exception of the door handles, I think both the iX3 and the new i3

02:18:49   look probably like incredible cars. Like, yeah, I agree. Yeah. I think these are going to be hits.

02:18:55   I think they're good cars. I just don't like the styling and I don't like the shape of the steering

02:18:59   wheel. I also don't like the trapezoid screen, but I do like the, the, the, uh, the like distant

02:19:03   dashboard, that stripe thing. I think that's actually a good idea. Lots of other cars have done that. I

02:19:08   mean, Toyota Prius famously started this trend, I think of like putting things farther away than you

02:19:12   think they would be. And I think it does work. Um, yeah, and it might be a little tight in the backseat.

02:19:18   I'll have to wait till I see some tall YouTubers get into this thing, but, um, it's nice that they

02:19:22   updated their platform. I think they have a bright future because they're, they're there. As you've

02:19:25   noted with the i4, their previous platform, which was hacked into their gas car platform was already

02:19:30   pretty good. And presumably they've learned things since then. So I think this, this car will be a

02:19:35   great car as long as you can, you know, stay inside it. So you don't have to see the outside.

02:19:41   Honestly, I, the, the i3 looks incredible to me overall. Like I don't share how much you hate it.

02:19:48   I think it looks great and I would be very happy to drive that car around. I think the, like, as,

02:19:55   as I think of like, you know, what will my next car be? Like my, my current lease is, you know,

02:19:59   it has about a year and a half on it. I think I'm going to look very hard at the i3 and the iX3.

02:20:04   The i3, it still doesn't have a hatchback, right? Isn't it? It's still like a regular trunk.

02:20:09   Although they apparently showed a touring version, which to us is a wagon.

02:20:13   Probably not going to be in this country. Yeah. Probably not here, but, but it will exist

02:20:17   somewhere allegedly. Yeah. They always get better models in Europe, but yeah, like I,

02:20:20   the, the iX3 looks just smaller and sleeker enough compared to my iX that it might be nice.

02:20:29   Um, but it still has like the utility of the, the big liftback trunk, which I, I do like and use all

02:20:37   the time. I think even, even though I would love to have like one of these regular sedan i3s to drive

02:20:43   around most of the time, I think I would really hate not having a liftback trunk now. Cause I got so

02:20:49   like ever since the Model S, like I'm so spoiled by having the, a big trunk opening. It's so useful.

02:20:55   It's so good. I don't think I can really have a car without it anymore. And I'm, I'm hoping that

02:21:01   like whatever the next gen of the i4, maybe we start to see what that starts to look like. Um,

02:21:06   cause that could be a good future car for me as well. But I think this thing looks great.

02:21:09   Yeah. I don't know why more people don't copy what, uh, Tesla did with the S just like the liftback

02:21:14   on a sedan is such a good idea. Everyone should be doing it. I think, uh, who does it? Um,

02:21:18   Audi does it. Audi, Audi does it. Um, but Porsche has, well, they have the weird, the ugly, uh,

02:21:24   Taycan with the wagon. I mean, the i4 was, the i4 was a liftback, you know, cause it's an i4 grand

02:21:30   coupe. I think they mostly do it cause a NVH, like it's harder to like isolate the cabin and people want

02:21:36   an electric car and yada, yada, but it can be done. I just think it's people, more people should copy it.

02:21:40   Yeah. It's, it's a massive advantage. Don't copy the door handles. Copy the liftback.

02:21:44   Right, exactly.