705: Relentless in a Gentle Way

25:54In iOS 27, the one that's currently in beta, it takes a significant step backwards.

26:01Marco Arment:i think dictating reminders to siri in ios 27 is awful now this has changed a little bit throughout the betas so maybe it won't ship awful in the fall but that's only a few weeks away um and it's still not great it is starting to ask me fewer asinine clarifying questions in later betas but it's still really slow now it's way slower than it was in previous os's and it's doing all these smart llm things and that's wonderful for lots of other reasons but for creating reminders it's a pretty big step backwards i didn't know that was going to happen when i started making my reminder app which is now now as you know is named unforgettable um i didn't know that at the time i started in february but over the course of the summer i'm i've realized like oh i need to ship this app now because i think a lot of people in the fall as they get used to this new version they're going to realize, oh, dictating reminders in Siri is pretty rough now.

26:58Marco Arment:Anyway, that's the side diversion. How does Unforgetful help with that? Aren't you still dictating reminders with Siri to put things into Unforgetful? So it helps in a few ways. So let me, that was a diversion. I'll get back to that in a minute. But the short version is that's why I have a dictation feature in Unforgetful. That feature didn't exist before like last month. But I realized, oh man, this beta is so painful for dictating reminders with Siri. I should also probably have a fast dictation thing. So what I did was I built a dictation engine. And as of a few days ago, the 102 update also enables some natural language processing on that input.

27:36Marco Arment:That's very, very fast. So you can say, you can hit the microphone and you can say tomorrow 9am by milk, and it will actually schedule it for tomorrow at 9am. And it will parse out of the title as you'd expect. Also with that, I optimized the design of the widgets and lock screen controls and everything else to provide lots of easy, fast ways into dictation mode. I mean, you know, there's widgets everywhere. There's control center controls. You can long press the app icon. All of those things will give you options for jumping right into the microphone recording mode. I even made it so that when you fire one of those off, it does not even animate the sheet presentation for the dictation mode.

28:16Marco Arment:it just pops it in so you save you know 0.4 seconds there or whatever it is like i want it to be so fast and what i haven't mapped to is i put the microphone button on my lock screen in the bottom right because i made a lock screen widget also so you can add it to wherever wherever the system allows widgets you can put it on the action button like whatever whatever you want to do there's mac controls that i think work it's really weird mac controls it's really weird like the new control center api thing for the mac it's very very strange but i think it works but for me my fast capture mode is the microphone button on my lock screen on my phone and that jumps right into it and lets me record it's so much faster than siri but anyway going back to kind of the purpose of the app for me it's all about fast capture and ubiquitous fast captures and i can still use siri if it's if i can't easily reach the phone or whatever and then also on the notification side, it is very important to me that no task ever get into a state that it does not ever notify me again.

29:22Marco Arment:Because that happens too easily with reminders. For me, it happens in two big ways. Number one, if I forget to set a reminder date, if I just say to Siri somewhere, remind me to buy milk, it will never remind me. Because I didn't say, it's like simon didn't say i didn't say remind me tomorrow to buy milk i didn't say remind me in 15 minutes if i'm like i just i just have remind me so it'll be like all right well i added to reminders and so that's on you good luck and that kind of like i will never look at it i will never find it again the other the other way that you can lose things and reminders if you are me is say remind me in 15 minutes and it notifies you in 15 minutes and if you don't snooze that notification if you do anything else, if you ignore it, if you accidentally swipe it away and dismiss it in any other way, if you clear notifications, whatever, if that notification, if anything else happens to it besides you hitting snooze, it will never notify you again.

30:17Marco Arment:So you have one chance. And this is not a good recipe for success if your brain works like my brain. The original concept of Unforgetful back in February was mainly to just solve those two problems just notifications that will constantly remind me over and over again until something is actually marked complete now there's an app called do dui uh that does this we talked about it before lots of people like it totally fine app it's not really for me like i don't really love the ui and some of the mechanics of it but it's that kind of concept for notification so i so like if any if you want you know to to avoid getting a bunch of emails this app already exists i'm i'm very aware Thank you.

30:57Marco Arment:And it's perfectly fine app. My life has for years now been managed by, well, since forever, it's been managed by my calendar. And anything that isn't a calendar item is in the DOE app. Like this is how I run my life. And, and I think that, you know, if, if Unforgetful was out years ago, then I probably would have put all of those things into that instead. But I've got so many freaking things in do now to move them over would be quite a burden. But no, if Unforgetful is for you, that's perfect. If Do is for you, that's fine too. But they're both very analogous in certain ways and very different in other ways.

31:33Yeah, and what I've realized,

31:36Marco Arment:so I decided fairly early on, I wanted to actually release this app to the public. And it has been a very interesting journey, I think, for me in part because it's a very narrowly focused app. This is an app that you either know immediately it's for you or you know immediately it's not for you. And this has been an interesting, you know, it's been interesting, like, looking through the feedback and trying to figure out how I should deal with that feedback. Because for me, the app is, it's very powerful if I keep that philosophy of it will always remind you, no matter what, of every task in it.

32:17Marco Arment:and it doesn't you know it can it can space it out by a few hours or tomorrow or whatever but like it will it will never stop reminding you of a task until you mark it done critically it reads and writes your data from reminders from the system reminders database it isn't just an import it isn't a one-way thing it reads it and it writes it back to that same database so there is no import or export you can use unforgettable with reminders or you can use it with other apps that use the reminders database or you can use it for a while and never look at reminders forever again or you You can, at the end of that, decide, you know what, I want to go back to Reminders.

32:49Marco Arment:Just delete Unforgetful, and Reminders is still right there, and all your information is still right there. So it's very flexible in that. One of the downsides of that, though, is that when somebody tries out Unforgetful for the first time, it shows everything in their Reminders database. And you can select different sublists if you want to only restrict it to one list or whatever. You don't want your grocery list in there. That's fine. But one of the most common feedback points I get is people saying, oh my god it filled it up with all these tasks that i had scheduled for the distant future but with no due dates i don't want those in there now i i even ran a poll yesterday on mastodon for about an hour i deleted it early but i ran a poll asking basically like hey do you what do you want unforgettable to do with tasks that have no due date because to me i always interpreted that to mean i'm never going to do that as i mentioned earlier because like i'm never notified of this so therefore it's a failure it'll never happen it's invisible that's not something i want to allow but but about half of respondents to the poll said they want basically like a someday section in the app and so i mocked it up put it in the app and i ran it you know a couple of test builds and and i played with it myself and i'm like i don't think i should add this like because to me a someday section means i'm never going to do it and a someday task just becomes a bucket to dump stuff in that i'm never going to do and my philosophy is if i'm never going to do it i should just delete it and no shame attached to that like just delete it this is a list of things that i intend to do not a list of things that i'm probably never going to do you mentioned what people were calling these things when they talked to you about them oh i have when i put a reminder in the reminders but i don't have a due date because it's stuff for the far future when i heard you say that i'm like well if it's for the far future it should have a date on it for the far future right perhaps not coincidentally i have a reminders list called far future that has dates that has reminders in it that are in the far future and they all have dates but they're in the far future and i the reason they're in that sub list is because i don't want to see them in my like regular reminders list because i know they're like two three four years from now but they all have dates on them so i mean i'm not i'm not going to convince anyone to change how they use reminders everyone uses the apps the way they want to use it but uh it's just i was surprised by the 50 response in that poll as well because i feel like look if you really want this to be in the far future set it for and oh i don't know the exact date fine set it for you know on the the closest you can get within a year two years three years five years just make a guess and throw it in there because anything is better than no date at all and i i'd never even thought about the idea of reminders without dates if i ever get a reminder without a date it's because of a failure of dictation and i immediately fix it right exactly like i'd never intend that and that's again that's part of why i just like so unforgettable it treats undated reminders like they're due today on the next hour boundary it's gonna remind you and then you can decide what to do with it you can you can say remind me tomorrow you can change that you can snooze it you can change it you can whatever you want to do anyway so when people were asking for this i i thought like well my instinct was like with overcast with overcast my goal is a broad audience i want i want overcast to be the podcast app for everybody most people i think their needs for podcasts should be solved with overcast obviously that a little bit different now with like video and stuff but like the idea generally with overcast was like i want this to be broad market so my instinct when i got this feature request of like undated tasks what should happen everyone wants a a someday section.

36:31Marco Arment:My instinct was to build it. So that's why I did. And I tried it. And as I was like trying to write the wording for the setting and the setting screen, I was like, I hate this. It felt like to use a wonderful programmerism. It felt like I was giving people a foot gun. If you're using unforgettable, if this app is what you need, then if I allow you to have an entire category of tasks that you never get reminded about, this is not the app for you. Like if that's what you want, this is not the app for you. And then look for our sister app, Forgetful. Right. Like, and I've been very careful when designing Unforgetful, like that to make it impossible to fail at your task.

37:14Marco Arment:That's the whole goal. If you have a reminder, Unforgetful will do whatever it can to make sure that you don't forget about it. That's the whole point of the app. That's why it's called that. And by the way, naming it. Oh, my God, I'll get to that. But so with this app, I'm being very conscious, like one of the unfortunate limitations of the reminder API event kit. There's a lot of fields that reminders supports that I don't have access to. I can't even tell if a reminder has them or not. So one of them is subtask relationships. Reminders, I think a few years ago, reminders added the ability to make subtasks under a task.

37:51Marco Arment:This is a common feature of apps like Things and OmniFocus, you know, all like, you know, the big task management apps. and the idea is you can make a project you know make like a top level test that's the project and then you have the sub the sub check marks of like all right we'll do this then do this and then then the project is complete that's great there's there's a place for that but i decided early on like first of all since i couldn't see those relationships well i can't support them anyway but also that's fine this app is not the app to manage that type of project if you like this is the app to manage call the doctor's office tomorrow simpler things this is not a project management app this is not a whole system it's it's not even really ideal for like collaborating with people i might change that at some point but like it's you know reminders has the concept of like shared lists and i admit i haven't actually done that much testing on that because that's not really what it's for and there's all these different features there's tags and organization features that I just not only do I not have access to thanks to the API, but I'm not going to build those features because what is tagging and folders and everything?

38:57Marco Arment:What is that used for? It's used to hide things selectively. I don't want to hide things selectively. I never want to hide things selectively. That's the whole point of the app. And so I am taking this different philosophy with this app of like, again, this is either for you or it's not for you. It's very hard for me to remember that when i'm reading all the feedback and you know some you know some of it's positive some of it's not but what makes it very encouraging is like on this point specifically about like the undated tasks whether they should be like in a someday bucket or remind you right now i i heard from a few people already i think three or four people already who have said initially they were mad that all their someday tasks all landed up there immediately and they were getting reminded but now they've set due dates like john was saying like they just set dates or got rid of them or did them and now it's fine and i'm like you know what that actually tells me this is a success it is a hundred percent the right move for me never to add that feature because what i want is that outcome i want you to either be reminded or complete them or set a date in the future and i don't care if that dates five years from now fine whatever set it as far in the future as you need to set it as you know to match your intention but if your intention is do this someday well how about you get reminded about once a month till you do it because otherwise it's not going to happen i just looked at my far future list by the way to uh see what was in there and i realized i have to check off one of them because we already talked about it on the show it was the um do airpods max to exist uh scheduled for early september we already did that was that for a bet that we had was that one of our bet yeah we had it i was like i think we talked about it and uh when they came out and like um i just looked it up like episode 683 uh it was it was a bet and I won.

40:40Yay. Anyway, I will check that off. Now there are fewer items on my far future list.

40:45Marco Arment:So let me tell you about naming this app. This was, I consider myself fairly good at naming things. Do you? Oh, wow. I was going to say this name is a very Marco name. I would say that Overcast is the aberration in your experience naming things because Overcast is a straight up the middle good name with no weirdness this is not so trying to name a productivity app a to-do app as it turns out this has been a pretty common need since the dawn of time so there are a lot of task management and project management and to-do and collaboration type like organization type productivity apps out there in the world and many of the many of them have existed many of them still exist there's many live trademarks in this in this world so my naming process is which i am pretty good at john but my name is i disagree but okay my naming process is you know i'll brainstorm some stuff i'll look into like you know root words and parts of speech and different styles and i will say i tried having llms generate me name ideas and they are just horrendous at this task they are awful at naming things there were some good ideas that lms threw at me but like they kind of got me thinking in different directions but you should have done like you know that um what is that a mcsweeney article about uh email addresses that are hard to read over the phone have you seen that one yeah yeah all right so your app should have been called capital d lowercase o capital d ee capital y o u capital ee so it would be do d u e wow because that's what everybody says when they talk about that damn app because that app also has a problematic name yeah right you can't just say do anyway yeah so so often i would come up with what i thought was a pretty good name and then i'd go do searches i'd do a trademark search i'd do an app store search and they just were all not only taken but like really taken like taken in ways that like this is still like an active live company using this name like there's no like i really can't use this name everything was taken and then i would often you know i'd i'd go in a different direction okay well if i can't do something descriptive what if i do something that's not even related like i i wrote it on the way to kona uh to kona hawaii what if i use kona as the name i love that name nope that's some productivity app um were you doing the naming after the icon because you could have called like lighthouse reminders or something no the icon came after the name i went through so many names and so often i would land on something that was like oh that's not bad let me look at oh no can't do that or i would i would think of some kind of metaphor like you know oh what what what about memory uh what about oh um elephants they never forget oh i just invented evernote like everything's been taken every metaphor everything has been taken safe harbor lots of lighthouse related ones yeah safe harbor for adhd people who can't remember things yeah not bad it it was this was the hardest naming challenge i have ever had and i i landed on unforgettable i don't think it's my best name ever to be honest but i think it's i think it's a decent name it has the advantage i think correct me if i'm wrong i believe it's not actually a word which has the advantage of it's not going to be taken you know what i mean like it's unique in that way it is you know it is your own branding and not just like i took a word from the dictionary and called my app that so there is a a minor issue that i hope will no longer i hope i hope it won't be an issue for long um because it'll just have an established name after a while but there's an issue that the app store search will try to auto correct you if you type in unforgettable it'll auto correct you to unforgettable ah crap that is actually pretty bad but i i'm hoping like if that is always the case i'm assuming that was taken by the way was unforgettable taken um there were some closer trademark problems with that um and i kind of i liked unforgettable better because well unforgettable is a word apparently yeah it's like it's a word but it's not really a word that anybody would use but it it is technically a word but like i was focusing in on the like adhd angle pretty early and one of the key design principles of the app is it never makes you feel stress or shame no feelings of overwhelm that's why i didn't call it overdue in addition to being you know probably disrespectful to the do app um i didn't want to call it overdue because that's stressful i didn't want any feelings of of stress shame overwhelm so i wanted it to be lighter and encouraging you know fresh starts things that are related to like sunrises because every day is a new day you know stuff i looked at so many metaphors um the waves uh you know coming in uh i thought like one of the one of the probably the best name that the llms came up with was nudge but it wasn't quite good enough for me and i think there i think there actually is an app called that anyway so i couldn't use it anyway but there's like like beacon or pain like okay you know not bad not i wanted to call it waves can't do that taken i wanted like i wanted to call it reprise even though it's my php framework but it's like i try to try to throw in like fish references musical references every single word was taken trademarked like everything somewhere in the task management or productivity space that was too close for comfort so it was just incredibly difficult to name this app um but i went with i went with Unforgettable because you know i had you know of my of my kind of finalists it was the one i liked the most and it had like the least contention in the app store it had the least contention for search results like if i can get apple to stop auto correcting it to a different word it'll be great because it will be very easy to find in the app store and it will be very high rank just put just put unforgettable in your keywords oh that's interesting maybe i'll do that yeah that's a good idea thank you okay um as for the actual journey of programming the app um you know this was originally this was mostly vibe code like what i wrote on the plane was mostly me learning about vibe coding i fairly quickly realized i didn't like ai code structure very much and that it had a lot of limitations and so i ended up most of the app's code i have written it myself um i still you know there is ai generated code still in there for various boring frameworky type parts of the code like you know some of those swift ui views like that's there's not that much for me to do that's distinctive there but like the data layer i wrote almost all of it myself more importantly everything dealing with dates and times and calendars and days i wrote almost all of that myself yeah because that's the easy stuff oh my god dealing with dates is and times and everything it's comically tricky yep there are so many pitfalls and the fortunate fortunately apple provides very nice apis the calendar api the date components like all these different parts of it it provides very good apis um and what i found was overall kind of like the the best balance for me is i wrote the time parsing and time handling code and then i had ai check it for bugs and write a bunch of test cases for me i mean i i can't tell you how many tests i think there's like 500 tests in the application testing all my heart all sorts of different like different edge cases and part of why there's so many.

48:30Marco Arment:So you might think if you are writing the Reminders API, recurring task support. How do you do that? And the answer is it's actually very, very simple. All you have to do is mark a task as completed in EventKit and Reminders automatically generates the next one in its recurrence series. And that's awesome. except what if you want to tell the user when the next time will be in the UI before it happens? Like if you want to, if they're entering the recurrence rule and you want to say, this will repeat on September 1st and October 1st. What if you want to show them that? That's actually pretty helpful.

49:10Marco Arment:There's no way to do that. That's fine. Huh, okay. What if you want to actually change some of these meanings? Like if you want to, like one of the features of Unforgettable is if you have, suppose you have a task that recurs every day and you are two weeks behind and you haven't done it yet if you go back and you check complete you're going to have to check complete 13 more times to get yourself to the present day and i thought that is awful i've done it before i've been there it's horrendous it just tells me how much of a failure i've been and it makes me feel bad and it's tedious and error prone so everything about it is bad so i'm like i want unforgettable when you complete an old overdue recurring task where you've missed intervals in the intervening time, I wanted to just skip all those and just get to today.

49:57Marco Arment:Well, then that means I have to change the logic of what it means to complete a recurring task. So between wanting to display the future dates in the UI and that, I pretty quickly realized I think I need to write my own recurring task parser thing that extrapolates the next dates forward for a recurring task. So I did. I wrote every single one of those cases and if you look in the reminders app and you go to custom repeat thing every single one of those i wrote the logic for that with the calendar apis and doing it all like you know manually and myself and everything and then i had ai go find find the bugs which i actually pat myself on the shoulder right now i actually did okay i didn't have that many bugs in the version i wrote but i had more than zero and it fixed them and then it made 500 test cases to make sure I got them all.

50:47Marco Arment:That was a really fun challenge, but it was a lot because, yeah, dealing with dates and times, very, very tricky stuff. And just when you're making any kind of app like this, there's so many questions of like, okay, what should happen in this edge case? You might think, oh, remind me every day about these tasks or whatever. Okay, well, you know, what if this time doesn't exist in this day? Or what if, you know, what if somebody says, remind me on the 31st of the month and it's a month that doesn't have 31 days. Or just all sorts of things. What if the person flies across the country, changes time zones between this interval?

51:23Marco Arment:What should happen? There's so many decisions about how things should behave in an app that deals with timed events. And so I hope what I have done here is make a system that makes sense, that is internally consistent, coherent, predictable, I think I have. It makes sense to me. I hope it makes sense to other people. But overall, dealing with times has been by far the most challenging part of this app. I don't doubt it. Time arithmetic is brutal. And I actually just had a very small bug with something in CallSheet that's honestly, the story is not that interesting. But it was basic time arithmetic gets you every time or gets me every time.

52:11That is absolutely brutal. So now that it's out, and forgive me, I'm not trying to cut to the end or anything like that, but what is the response been? Because it seems extremely positive from what I've seen. And it's funny because when I first saw the app, and John and I only knew about it, I don't know, like a few weeks ago, a month ago maybe. when I first saw the app, I thought, this is great. I don't know that this is something that I personally need. And I worried that this may be something that not that many people need. Now, since this is not, you know, your primary income, I wasn't maybe worried was too dramatic a word.

52:46But I figured, oh, there'll be a handful of people that like this, and that'll be that. But based on just what I've seen, it seems like there are a lot of people who are looking for exactly this. So from my perspective, it looks like the response was great.

52:59Marco Arment:so far it's been i you know i had kind of like like a low bound and a high bound of what i expected and it's basically right in the middle oh good okay i'm happy with it it's not blowing me away it's not crushingly disappointing it's right kind of right in the middle of what i what i kind of expected slashed hoped for um so i'd say so far it's doing well the i think the only major question mark is so the app it has a hard paywall with a$20 a year subscription and with a one month free trial and I wanted to do a whole month free trial so people could actually like use it like move into it like live with it try it and see how it works for you before you pay me anything because I want you to be able to try it and then of course with reminders integration like you can try it and all your stuff is right there and then if it's not for you you can leave and all your stuff is still right there in reminders.

53:54Marco Arment:You can leave whenever you want to also. So I figure like I'll make it really easy for people to try it with no pressure. And then after a month, if they're still using it, then I can get paid. How long is the one month free trial? What? How long is the one month free trial? What do you mean? It's a month. Date math. Date math. There it is. That's up to Apple. That's whatever they do in App Store Connect. That's my question. What does Apple do? Like, so if you, you know, because if you did this yourself and more people will complain no matter what, they would say, I bought it in a month where there's not that many days.

54:26And I bought on the 31st, but then my subscription ended on the 28th of the next month.

54:30Marco Arment:Not fair. Honestly, I don't actually know how Apple handles that. I didn't even do I. That's what I'm saying. Date math. It's terrible. Because in that situation, I do this in my app is why I thought of it. I just always give them 31 days, period. That's how I do the date math. And so you can get a one month thing that lasts you two calendar months because you span over February. You end up, you know what I mean? 31 days, no complaints. But that's just because I'm rolling it myself, which has its own problems because Apple doesn't support the kind of payments I do in my app. But dates, terrible.

55:03Marco Arment:Yeah. Oh, I actually, I think my usage of CloudKit is a little bit interesting. And I wanted to mention that. so i don't directly need cloud kit you know at first glance because i'm reading and writing from the reminders database and your reminders database is probably stored in iCloud for you know on your iCloud account but that's not of my concern i just read it locally but i started realizing i needed cloud kit for a few additional things so one is like there's a few cases where i want to store metadata about each reminder to provide some feature that i want to offer so for instance the snooze times are dynamic.

55:43Marco Arment:There's a few different brackets of short, medium, and long snooze intervals. The one it offers you depends on what you picked last. So if you picked a very short interval as your last snooze, like the 15-minute one, it'll give you the set of short ones. If you picked a day to snooze it last time, it'll give you the long ones. And so I wanted to store for each reminder the last used snooze interval. And I also, I had this idea for location triggered reminders that the biggest problem i have with location triggered reminders is that the radius of a geofence that triggers them is a hundred meters it's a pretty big radius and this is you can try to make them smaller but they'll be less reliable so apple recommends in like the programming guide for core location everything they recommend 100 meters as the smallest geofence radius because the gps chip in the phone is when you're just kind of like doing background pulling of geofencing and not doing like active turn by turn directions the gps radio is in a very low power state it's not pulling constantly and so if you have like a smaller radius that you're trying to trigger based on you might actually miss it as you move through it so 100 meters is the kind of the default geofence radius and whenever you tell reminders here remind me when i get home do this um it'll remind you when you're within 100 meters of your house that can be when you're like half a block away like you like you might you know you might be near your house but if you say remind me when i get home to you know put the milk in the fridge i mean hopefully you don't need that reminder but maybe a bad example but like you know if you say that it's gonna remind you when you're like three houses down in the car or maybe five houses down on foot and by the time you get into the house where the fridge is you might have forgotten i know i do so for me geofence alerts for reminders have always been not particularly useful because it's like well i don't want to be reminded when i'm like driving up near the house i want to be reminded when i'm in the house when i can actually do something about this task because i don't want to forget in the intervening time and so i i wanted to add delayed triggers notifications i wanted to add all right make it so i can set myself to remind me five minutes or 10 minutes or 30 minutes after i get to a location remind me of this thing that that way i'm like once i'm like in the house jackets off stuff's put down i'm settled then i can actually do something so i had to i had to store a field on reminders um and i also wanted to be able to do things you know just like you know for future features like if i want to add url support or whatever i can't read reminders URLs thanks event kit please man if there's one thing I would love to read URLs but the ability to add fields you can't really do it in event kit there is no way in event kit to add arbitrary metadata to reminders now other apps have come up with various hacks for this sometimes they'll try to cram it into the notes field they'll put like an identifier in the notes field or they'll just actually literally write the metadata in right into the notes field and that's fine I wanted to avoid that if I could or if it if there was like a better solution and then totally separately from that i ran into an issue during testing one day for the notifications to work best ideally you turn off reminders notifications and you leave unforgettable notifications on and it replaces reminders notifications but unforgettable on the phone is limited by ios's background refresh intervals if you don't allow it to background refresh or if you tell seer in your phone remind me in three minutes to do something and unforgetful doesn't have a chance to background refresh itself in that three minute span it's not going to be able to notify you because it doesn't know about that reminder yet now i have a mac version too it's a mac and ios app the mac version can constantly monitor so as soon as the mac gets a new reminder through iCloud or whatever, the Mac app instantly knows, here it is, and it can create a notification for it.

59:44Marco Arment:So I'm like, what if I use iCloud background refresh triggering? When the Mac app sees a new reminder, if it's running on a Mac somewhere, and you're out somewhere with your phone, it can send a CloudKit silent push to the phone. And I think, oh my god, that's how to do it. So I started using cloud kit for that that when the so the and this is why when the mac app the mac app gives you one of those annoying confirmation messages when you try to close it like chrome um because i actually don't recommend that you close the mac app ever because if the mac app is running it can do this it can pick up notification it can pick up new reminders immediately and then send a cloud kit push to wake up the phone app so the phone picks them up way faster than if it has to wait for a background refresh.

1:00:33Can you make the Mac app just run in the background? Like make a little background job, faceless background job?

1:00:38Marco Arment:I initially, I had it as one of those like dock or menu bar or both toggles. I think I'm going to put that back. But I think there's a way you can get it totally in the background. And the reason I think so is because Golden Gate puts icons in the menu bar for apps that are doing stuff in the background. I don't think I want to deal with that. But I definitely would like to. probably in one of the nearby nearby future versions i'm probably going to add an option to put it just in the menu bar because a lot of people have requested that um anyway so that's why i recommend if you run run the mac app all the time if you can because that will keep your phone app up to date faster um anyway so i had to i dealt with cloud kit for that and then i'm like you know this this was kind of like midway through the summer as i'm trying to figure out what to do with these different features that i want to have metadata for like the delayed notification or delayed location triggers and things like that and the dynamic snooze i was like hmm well now that i'm using cloud kit and fable was out and i'm like fable's really good at coding what if fable writes me a bridge between my blackbird database and cloud kit i've already i've been planning to do this forever i called it skybridge and like i made a prototype back in the day i never finished it so i had fable look at my prototype and using the style that i use and the blackboard mechanics that i use just finish it like take this prototype and finish it and test it and make it good and make it work and fix the bugs and it did so and this you can actually see what it did there's um if you look at my blackboard libraries open source and there's a branch named fable skybridge is in there and you can use it if you want it's just automatically syncs a blackbird table to cloud kit and it's awesome i'm so i'm so happy about this like i could when it did that i'm like oh my god this would have taken me a week to do this myself and my version would have had more bugs and it would have taken me another month after that to find them all and fix them and i never would have had tests i'm so happy to have this and since having skybridge and since i already did all the work on blackbird it's a fully functional blackbird app with no servers just with cloud kit as the as the sync backend so now i can add arbitrary metadata to any reminder whenever i want and i can if i want to add another column it's no big deal i'm i just do it and i go to this you know the cloud kit dashboard and publish the changes and all that stuff but like it's fine what a game changer that has been so i'm now able to do good features like that anyway so that's i thought the cloud kit store was a little bit interesting on that especially the background push i think that i think that really helps keep the app up to date um but anyway so to actually answer your question casey from a thousand years ago the reason how the reception has been again kind of what i was saying earlier about about with overcast uh versus this like I'm having to exercise a lot of discipline, not adding every feature everybody asks for.

1:03:40Marco Arment:Because if somebody is asking me for things like folders and organization, I think I can confidently say, if you're looking for this app to do that, it's not really the app for you. Or this is not the app for that job. and if you use reminders that way like i was listening to our friend steven roblez over on mac power users he recently made a kanban app he vibe coded the whole app it was great it's a good episode for mac power users i think it was last week and we'll link to it he was describing how he used reminders before he made his own app and he was a super power user of reminders using things like subtasks and tags and all these and like i i've never used it that way i've never used anything that way i've tried using omni focus back in the day i tried using things back in the day i would stick with them for a little while and then i would it would fall on the floor and i'd forget about it and i think everybody with adhd type tendencies i think this is a very familiar story you know this is this is a category of apps that gets a lot of like tire kicking like hey yeah i can be the kind of person who organizes my stuff this way and you do it for like a little while you know you do it for maybe you can stick with it for a few days or a few weeks or maybe if you're lucky a couple of months but then then you don't you stop checking it like this is why i've never gotten into like paper notebooks even like my my you know my journey into remarkable tablets i would just eventually stop looking at them i would stop checking them and that's that's been true of everything i've ever done that can't notify me constantly so i can pretty confidently state what I want this app to be and what it is and what it isn't.

1:05:22Marco Arment:It is not an organizational system. It is an automated nagging machine to get you to do stuff. If that's what you need, it's very effective. That is not what a lot of people need though. I mean, there are a lot of people out there who need it, but like, I think if you pull the average people out there in the world, it's not going to be the majority who need this. If I take a hundred people off the street and I describe this app to them half of them are going to look at me like I'm crazy I think about 40 of them are going to be like eh I don't really need it and then maybe 10 of them are going to be like oh my god yes I need that that's kind of what I'm looking at and I don't really know how to reach those people I'm learning I'm going to do a lot of experiments with different types of ads and trying to find different audiences targeting different ads in different ways I really have no idea how to market this app.

1:06:18Marco Arment:And I don't think it's ever going to be a mass appeal thing because it doesn't have mass appeal. To be a mass appeal task management app, you have to have a lot of features and enable a lot of different ways of working. Things and OmniFocus, these apps are great for that because they're very customizable. They have very, very deep features and broad feature sets so what those kinds of apps whatever your system is they allow you to mold the app to execute your system that's not this this is the app for people who systems don't work for and i'm very confident keeping it that way because if for some reason you out there if your brain works like mine if you can forget something five minutes after someone told you this app might be for you and there's enough of you out there there's enough of us i doesn't i'm confident it can yeah i'm confident that this can be a reasonable business it's not i don't think it's ever going to be anywhere near the size of overcast in terms of either user base or income i i would be shocked if it ever came near it um i think i'd be happy with a tenth of it but i know that there's other people like me out there and if you are out there and if you want a better reminders app for yourself try unforgettable see see how it works for you give it the month free trial if you don't like it don't worry delete it go back to reminders no big deal um but but give it a shot um because i think there's i think there's a lot more of us out there than than we realize but i don't know how to i don't know how to reach us so help me out did you get unforgettable in the app store like no colon suffix um no i actually i didn't even try you always gotta try no i you don't because the way the app store search works this you know look i had on i had overcast no colon forever um but i stopped using it because app store search is brutal you have to jam your title and your subtitle and your keywords full of every keyword you can think of and where you put them matters what you know which of those fields they're in they all it matters if there's a phrase and both words appear in the title or subtitle that matters more than if they're split across the two fields like there's all these different things with app store search optimization i wish the app store search wasn't so dumb but it is and this is this is the environment i operate in so if if you want your app store app to succeed you have to do things like shove a bunch of keyboards in the title and the subtitle you just have to so i didn't even the idea of not doing that it it never was a possibility so what is your subtitle never lose a task oh that's a good one that's after the colon when i was searching for um apps called nag of which there are many um one of them has a good subtitle which is nag and they use a hyphen instead of colon which i think is a mistake but anyway nag relentless reminders that sounds more like do actually that's actually not that's a decent title i like that because yours you didn't mention this but i think is that everyone should read margo's post about this because it's really good and explains even more stuff if you can imagine um and one of the key things that i didn't realize until i read your post about it i'd been test flighting the app but i just thought it was like doing weird stuff because the reminders things sync stuff but you know this this app's tagline being relentless reminders uh unforgettable is intentionally not relentless like if you have a bunch of notifications that would normally show up like all at the same time you space them out so as not to throw a pile of crap on top of people which i would never have occurred to me that you were doing that intentionally and it's why the um it's why never lose a task is a much better subtitle for your app than relentless reminders because you're trying not to be relentless i mean you're relentless in a gentle way i guess well yeah because again like so that again that goes back to some key design things no shame try to minimize any any way to feel overwhelmed and so what you know obviously look i get i'm not great at keeping up with my to-dos obviously what happens frequently is sometimes i'll get i'll get to a state where like for a week i will every day be deferring to tomorrow like seven to ten different tasks you know and my default tomorrow day my default tomorrow time 9 a.m what would happen then is at 9 a.m i would get seven to ten notifications um now on what happens on modern apple platforms is they stack and they get summarized.

1:10:54And so if that happens every day

1:10:57Marco Arment:and if every single day you get the same seven notifications at 9 a.m. that you've been deferring for a week, you're going to only see the top one. And maybe you're going to see the weird summary of it instead. But that's it. You're not going to see the ones that are buried in the middle, which means they're never going to get done. And this has multiple issues. Number one, you feel hopeless. Oh my God, it's another system I've lost track of. It's too big. I can't, I've lost control. I'm a failure and I must just bury it and never look at it again. So that's obviously a failure mode that I'm trying to avoid.

1:11:34Marco Arment:So that's where the spacing came out. If you have seven things that are gonna all remind you at 9 a.m., I'm gonna space them out by your interval that you set. You can customize it, and you can set it to zero if you don't want them spaced out. But you can say like, you know, just remind me, space them out every 15 minutes. That way, every 15 minutes, as you're working, you get one notification. And you can look at it and you can say, oh, you know what, fine, I will finally call the doctor. But also, to avoid the stacking problem, and this is, again, this is like, if you think this sounds insane, your brain doesn't work the way ours do.

1:12:10Marco Arment:It also matters what order you're reminded in. So, I built a feature that randomizes the order. and it uses a seated predictable random number generator and it's seated based on the date so that way your mac and your phone randomize them in the same way every day so no matter where no matter what device you're on you're going to get a randomized order of those tasks that stack up every day the ones that are scheduled for the same time they change every day in a random order it's the same on all your devices so that way when you get that stack notifications nothing ever gets lost in the middle forever because every day it's in a different order it's going to be on top eventually so you will see it and like this is the kind of problem that a functioning human being whose brain does not work the way mine does like why are you letting them set that i've never had things stack up that i'm just deferring for a week straight why don't you just do them instead of deferring them for a week yeah but that's not how my brain works so the way i the way i work such as it is, is a lot of snoozing, a lot of deferrals, a lot of things eventually happening in batches or being given up on.

1:13:19Marco Arment:That's the type of feature that I build in here. That's what this app is for. And if you're, again, for the 90 % of you out there who are like, that's insane, this is probably not for you. But for the 10 % of you out there who are like, oh, I would actually really use that. Try it. I assure you this, this might work for you. well this is very exciting uh congratulations again i am pleased to see another entry in your i don't know arsenal isn't the word i'm looking for but in your repertoire whatever you want to call it um this is very very exciting news and you should be very proud and very happy and i hope you're both of those things thank you yeah i'm glad you finally come over to my side and started making weird little apps that most people don't like oh well if you want a weird little app that most people don't like let me tell you the story about fast text yeah honestly like i find it very um very interesting like i as i've as i talked about recently um you know my my job in overcast has evolved to really be more of a caretaker of overcast i don't i can't really change that much about it without angering a lot of people what people want for me with overcast is not tons of change they want just keep the app the same ish keep it working keep it good fix any bugs and do my pet feature like that that's what everyone but nobody wants me to really change overcast and so unforgetful was a chance for me to do something very different this is like it's from scratch it had no existing user base so i could do whatever i wanted especially with a very small focus user base like it's going to have so i it was a very different creative itch it was a very different problem space as my first productivity app i've ever released i've never than anything productive.

1:15:01Marco Arment:In fact, I make apps that distract people from their productivity usually. Oh, how quickly we forget about quitter. I guess that's kind of productive. I'm mostly giving you a hard time. And the chapter thing, what's it called? Forecast. Forecast. Yeah, but the forecast is a very limited release. The productivity tool. If you had to pick the category in the app store, you'd pick productivity. Maybe media. Anyway, so yeah, this is basically my first productivity tool. My first series or my first one that I'm actually charging money for. My first one that's really a business. It's also my first Mac app store app.

1:15:30Marco Arment:And I, in general, I don't make a lot of Mac apps at all. So it's one of my first Mac apps. So it, there's a lot of, you know, kind of firsts for me in this. And as an app developer, that's always exciting. That's always very interesting. Yeah. So I've really enjoyed it quite a lot and I continue to enjoy it. And frankly, if no one else ever uses it, I'll be disappointed, but I'm still going to use it. Like that's like forecast. I use forecast every week. That's how I edit the show. even though forecast is not a major app and i haven't actually issued an update in many years i still use it constantly no matter how many people use this app i'm going to use it myself all the time and therefore i am just i'm very happy with it and i i hope other people find it i hope it's good for them too by all means please try it but if not hey you know it's it's at least going to be good for me.

1:16:25All right. So there's been a lot of news, hand-wringing. I don't know what you want to call it. There's been a lot of stuff going on with regard to AI watermarking. Now, we're going to be talking a lot, perhaps primarily about Anthropik's Claude. And it's worth noting that they are certainly a former sponsor, I believe a future sponsor as well. So I just want to make that plain up front. But with that said, let's talk about what they're up to. So Anthropic wrote that Anthropic has signed the EU AI Acts Article 502 Code of Practice on Transparency of AI-generated Content. Here's what this means for Claude.

1:17:01New models will mark AI - generated content from day one. Claude models launched in the EU on or after August 2nd will support machine-readable marking at launch. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported. We'll help you detect Claude's marks. We'll support users and other third parties to detect Claude's marks as the code requires. And we'll share details in forthcoming documentation. Existing models are in progress. The law includes a transition period for anthropic models launched before August 2nd. And we're working to add marking support for those models as well.

1:17:35It's a very long press release that goes on and on, but we're not going to read all of it. But the initial question would be, why are they doing this? And the easy answer is, well, they just said why they're doing it. It's one of those EU regulations. Apparently, EU has this, you know, this AI Act that says you have to, you know, put some marking on your AI output so it can be identified as AI. And there's lots of systems that do that. There's a couple of standards that do that, a big one that Adobe supports that I can never remember the acronym for. But those usually in the context of like, oh, we want AI generated images.

1:18:09We want some way so that we can detect that they're AI generated. and there's like you know a metadata sidecar that you can stick on the image that has some cryptographic info that says this was ai generated blah blah blah and obviously all the systems well you know they can just strip off the metadata sidecar just like you strip off location data from a photo and it's like yeah but it's like an advisory system for people who want to participate in the process and maybe it fulfills some regulations and blah blah blah then the next level is like oh we have my our ai image generator can use like steganography i think that's what it's where you put like secret hidden data in the pixels that humans can't see but that ai can see and that's harder to strip out but of course you can still strip it out because you know you can change the pixels uh but at least that's clever it's not like a metadata sidecar thing it's not really sidecar the metadata is in the file but anyway you know what i mean like it's they would change the pixels as a way to do it but then you get to things like ai generated text and you're like how the hell are you going to watermark text are you going to because text unlike an image usually text files don't have a place where you can shove metadata that's not visible the whole point of a plain text file it's a sequence of you know unicode code points or whatever you know encoding scheme it is and that's it there's no other like images have all sorts of stuff in them that doesn't get displayed when you display the image but text files don't that's what makes them text files so how are you going to watermark uh text files i forgot to look this up but like watermark is derived from i'm assuming when they would like take a piece of paper and put a mark on it i don't know if they actually ever use water but like if you held up to the light you could see you see it on like a paper currency in our country sometimes anyway that's the duration of it but like again with a plain text file what are you going to do with a plain text so i can identify it as uh ai generated unless maybe at the very end it says buy claude right so that's that's the problem they're faced with but like but why are they doing it like why you know again with the image stuff it's like oh we want to do this to be nice and so people know when things are ai generated and blah blah blah and yet it's not a security thing we're not fooling anybody because obviously anyone can just remove the watermark but it's like we're trying to do this to be nice there is one other angle that i hear everybody talking about with this and i have no idea if this is in anthropics mind at all i don't know if it you know it's just a happy accident because the eu regulation i think doesn't say anything about this but the angle is um first the premise that when ai models are training on stuff that they don't do well if they are trained on like their own output, like, or the output of other models, right?

1:20:36So they want to train on stuff that was not AI generated. If you train an AI on AI generated output, it just kind of goes into a degenerative loop and things get worse and worse. So I've heard, I don't know the foundation of that, but that's what I've heard. And so the idea here is why would Claude, why would Anthropic do this? Besides obviously complying with the EU regulation, like Claude is doing it, Anthropic's doing it worldwide, which is just a simplification, I think, which is what Apple should do in response to EU regulations. But anyway, they're saying, oh, they're doing it. So they won't accidentally train their next model on stuff that was generated by any of their current models.

1:21:10Because when their models go out there, they'll be like, oh, yum, yum, training data, munch, munch, munch, steal everybody's data. Oh, wait. Our watermark detector has detected that this data was generated by us. So we won't consume that because we think it will make our model worse. That sounds like a conspiracy theory, but I mean, i think it's plausible again assuming the whole idea that if models train on ai generate output they get worse which also sounds plausible to me so keep that in the back of your mind i mean we're not again we're not going to read the whole rest of this press release but we will eventually read something else from anthropic about this topic about you know why is anthropic doing this besides the fact that they have to comply with the regulations and how does this apply to other companies and it's not the only one doing this obviously every this eu law applies to everybody and Google has something similar and so on and so forth.

1:21:58So keep that in the back of your mind as we continue here. But, you know, getting back to my early question is like, OK, so I can see how you can watermark an image, whether with metadata or with like clever changing of pixels. But how the hell do you watermark text? And that's the question everyone had in their mind. This story is mostly taking place already between two episodes of the show. So we luckily have the benefit of time passing. But in the beginning, everyone was like the press release didn't say however they were going to do it. It's like, how are they going to do this? How are you going to watermark a plain text file?

1:22:29All right, I'm going to do a lot of reading, so boys, feel free to interrupt me. All right, how Claude's text watermark works. This is from Anthropic. In this article, we share answers to some of the questions we've received about how our chosen watermarking method works. We use a method of watermarking that does not have any practical impact on the quality or content of Claude's outputs. The difference between watermarked and unwatermarked text will not be distinguishable to readers. Nothing is added to the text, and there are no hidden characters. Watermarking doesn't require extra tokens and will not be more expensive.

1:22:58Watermarking carries no identifying information and can't be traced to a specific person, organization, or chat. Watermarking won't be specific to Claude. As of August 2nd, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same code of practice and will be implementing their own watermarks. The watermark only applies towards Claude chooses. When Claude proofreads text written by a person, what it gives back has generally only been lightly edited. Depending on the length of the text and how heavily Claude has edited it, those changes might not be enough to make Claude's involvement detectable.

1:23:29The more Claude writes, the more decisions it has to make, and the more space there is for a watermark. A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish Claude wrote this from Claude heavily edited this. So one thing that annoys me about Anthropic is why did they wait whatever number of days to release this? Because everyone was flipping out about how it was going to work. These bullet points are very succinct and explain a bunch of stuff, And then they go on to say, you know, basically, if if Claude didn't put the word there, it doesn't it's not part of the watermark, which makes sense.

1:24:00But it's worth saying. Right. And as I said, the more word Claude puts there, the more it's watermarked. But like, obviously, if you wrote those words and Claude didn't, they don't contribute to the watermark. So you need more words and they don't come out and say this. But I would imagine the watermark. I think they do say it later that we don't have in this passage here. But the watermark detection is not thumbs up, thumbs down. Yes, I know. ai it's a confidence score it's like yeah probably maybe do you get a percentage because you can't it's not you know it's not a slam dunk case again if claude only changed three words can you detect that at all what is the threshold like they they talk about this and uh i think in another podcast talked about it being like 200 words being the uh the thing you're supposed to be able to detect according to the eu regulation but uh it's a little wishy-washy but yeah that this is their explanation and all the stuff they crossed off no weird characters no hidden text it's not going to cost you more tokens it doesn't identify you that's all very reassuring and makes sense but like but how does it actually work how it works large language models like claude work by generating one word at a time each time the model decides on the next word it chooses amongst a list of potential candidates ultimately selecting the most sensible or likely based on the preceding text take the sentence the weather today was cold and the next word is very unlikely to be sugary but it is very quite likely even to be overcast, ding, or gray.

1:25:21Under most circumstances, it doesn't matter much to the reader which of these latter two words the model ultimately chooses. The meaning of the sentence is largely the same either way. In cases like this, the choice is settled by a random number. Let me pause you here because this is my favorite, the most fun thing. You know I love when people have to be confronted with how these things work because it is very difficult to hold in your mind how they work because you know it's like pet rocks or you know pet pets like they're the mental model people have these things is so different than how they actually work that even if you explain it to them and they understand it they will discard that immediately and go back to believing they work however you know it's like a little magic elf in a machine or something right and so by here's anthropic and they have to explain in like very simplified terms um how their watermarking works which requires them to discuss how their product works they haven't gotten to the watermark and they're just saying here's what an llm does and what they're saying is when our you know it generates one word at a time based on the probability based on the words that came before it right and when we get to the point where uh you know we have to pick from some words that may be like roughly equally likely we pick randomly and people are reading that you pick randomly like this is not the watermark this is just how lms work so i can imagine if people read this that they would be very upset by wait a second.

1:26:39It's the same either way, so we just pick randomly? That's your product? This is upsetting. But anyway, continue. Water marking uses low-stakes choices like these, which occur many times over a piece of generated text, to leave a pattern in Claude's responses. That pattern is undetectable to the reader, but it's detectable to anyone who has a key that encodes it. When water marking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, water marking uses the key and a few words that come before to settle what word the model should pick.

1:27:12That is, the words that Claude picks are still random, but now one can check the sequence of words and see if it's consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude. Amazing synergy by Marco, who just talked about the fact that he is doing random number generation, but he's using pseudo-random number generator. We will link to the Wikipedia page for pseudo-random generation. but the upshot is it's a function that you can call that gives you a random number uh but like where do random numbers come from this is a very complicated topic um if you're not doing it for something for cryptic cryptographic or whatever you use a simple pseudo random number generator where you give it a seed value and for any given seed value the random number sequence from the random number generator every time you call it will be exactly the same so if you seed it with one two three you're going to get the same you know call it random number random number right they're always going to be exactly the same it's always going to be the same you know every time you call it it's a sequence of numbers you're like that doesn't sound random at all you say if i seed it with one two three and i call rand rand rand rand it's going to be the same answers from each one of those not that each one is the same but the first one will be 10 the next one will be 20 the next one will be 15 you know like it'll always be 10 20 15 10 20 if you seed it with one two three a year from now and call it'll be 10 20 15 10 20 it's always the same sequence that's how pseudorandom number generators work and there is random choice within all LLMs.

1:28:33And we'll get to why in a little bit by linking to all those videos where he's linked to. And what they're saying is we always choose at random, but now we are, we have a random number generator that we know how it will, we, we, we know how it will spit out the numbers. And you know, like Marco doesn't know what the random numbers will be, but he knows if he seeds both of them the same way, he will get the same sequence of numbers. I don't know. Again, you don't know what those numbers will be ahead of time, but you know, like that they're going to be the same sequence so their random number generator is obviously more sophisticated than this but they essentially know because we decide how the random numbers are generated we know whether you know the our generator our watermarking generator was used to generate these numbers versus you know some other random number and and no one is more random than the other right because any any things where you're a student random number generator what's a good seed value well a good seed value is one that's not going to repeat and one is essentially unique so a lot of the times the current time in like microseconds since you know 1970 or whatever is part of the formula for the seed but what you never want to do is use the same seed twice because then you'll get a sequence of numbers that you've seen repeat before but each individual the whole point of sitting around a number of generators is what you hope is for every time you ask for one you get one that is unpredictably different than the previous one it's obviously not unpredictable because it's a deterministic algorithm but like from from an outside observer if you don't know what the seed was you won't know what the next number is going to be so again it should be exactly equally random to to no matter what the seed is no matter what random number generator you have as long as they both have the same amount of entropy or whatever this is a very complicated mathematical topic that i don't know the details of but the whole point is one set of randomness is not necessarily any more or less random than the other it's a thing that you can choose when you make the algorithms and anthropic is saying it's random just like it was before but the random number generator is one that we control in a way that we can predict but it's still random it's just as random as it's just as random as marco's random number uh random order of his notifications it's random enough that a user considers it random but behind the scenes it's not random at all it's

1:30:43Marco Arment:entirely determined by the seed and fun fact that's why you can um copy someone's minecraft world by just copying the seed. That's literally like the Minecraft seed for the Minecraft world. That's what that is. It is a random number that serves as the seed to the random number generator that generates the entire world. And so that's why two different computers you can just type in this number and you will generate exactly the same world that my computer with the same version of the game generated because that's literally what it's doing. Yeah, because there is a sequence of code in the game that's going to run and at various points that code is going to say give me the next random number and as long as the seed is the same as someone else's seed and as long as the code is the same as theirs it's going to get the same answers as you got like that's the seed of a pseudo random number generator determines every number that will come before it exact come after it exactly the whole trick is you know choosing a you know a seed that's not going to repeat and that people don't know and that's another upsetting thing about about computers people just assume computers can easily pick a random number it goes really deep like there's a reason computers have like quote-unquote real random number generators uh for use like cryptography and stuff because it would be terrible to use a pseudo random number generator for cryptography because people could predict what it's going to be and how do the real random number generators work usually there's some piece of analog hardware that samples like electrical line noise from interference from surrounding components in the air like they you need it you need basically a source of entropy um and sometimes these devices that will get you sources of entropy can't give you an unlimited amount instantly.

1:32:19So you have to wait for them to gather enough. It's a very complicated topic, but suffice it to say that the random number generation used for LLMs has to be fast and it's not like cryptographically secure. They just want to, they just want like roll a dice, pick a random number, you know, like the same way you would get it from any kind of pseudo random number. That's perfectly sufficient for all LLMs. That's why I think the one that they're swapping in is probably also not expensive, doesn't use like real world entropy because the job it has to do is you know there's five possible word choices um and pick what they are and before we get to gruber's thing on this we'll uh there's been a bunch of links in the show notes the stuff we linked before uh but i'll give timestamp links to the specific parts of this and in particular uh the three blue one brown uh video on how elms work which isn't that scary and if you watch it maybe seven times you'll start to get it it's a bunch of matrix math and in the end you end up like it's like what is the next word supposed to be uh at the end you end up with like here are the top 10 words um and they're they're like probabilities like we think this is the most likely second most likely third most likely fourth most likely right so on and so forth that's how lms work um and there's a setting called temperature which says when i get to that end part and you give me a probability distribution about the likeliness like this is 90 likely this is 40 likely this is 30 like they don't add up to 100 it's just like how likely is each one of them um a temperature of zero just always picks the most likely one sort them by likeliness the the one that's at the top you get okay and if there's a tie i don't know they just pick randomly or whatever between those two but like temperature of zero picks the most likely and you think that's great that's what lm should do that'll be the smartest because it finds the best word and picks it but temperature zero is not uh pleasing to use because for a variety of reasons uh if you're conversing with an llm or having it do anything like interesting a temperature of zero and this is me talking not any sort of you know ai scientist my impression is that it reveals too much about the stupidity and determinism of the machine so if you say at each sequence just pick the most likely next word it just it's just terrible it doesn't it doesn't seem like it's alive or creative or interesting.

1:34:34And by the way, if you ask the same question with the exact same wording, it will always give you the same answer, kind of like a sort of round number generator, right? So picking the most likely next word choice turns out not to be the best strategy for a thing that's going to fake people out into thinking like it's alive and creative and stuff. So it's good for like, we talked about this in the past, like business processes where you need something to be predictable. Like if there's some business processes that's classifying something like what category should this invoice go into, you put temperature zero on that one because you don't want creativity just pick the most likely next word or the most likely less token or whatever like just do it the same every time if you get the same input i want the same output temperature zero but chat bots no temperature zero is not what they use when you increase the temperature to be non-zero what it does it takes that probability distribution where it's like well this one is really the most likely like 90 likely and the next one is 80 and then 70 and then 60 and then tails off and it squishes the probability curve down to say let's not have too many spikes here let's make the less likely ones have a higher probability so instead of it being like let's say it was like 90 20 10 2 2 2 2 2 1 let's squish that down so it's like 60 40 20 you know like it's basically taking some of the distribution from the big bars and spreading it to the smaller bars and then it picks randomly among them amongst them weighted based on their probability right so you know it's like you roll of dice and it uses the use of the size of the bar to say basically how many entries in the raffle does each word get and so the most likely one is still the most likely to get picked but you're giving more of a chance in the raffle to the lesser words and the higher you make that temperature the more it squishes those bars down to the point where if you have a really high temperature you may end up with a distribution like the screenshot that i pulled from the the three blue one brown video where it's like the top three entries are six percent then there's four, five, five percents, and then four, four percents.

1:36:29So there's lots of things that are tied and it's really squished it out. So now, you know, the top ones get six entries in the raffle, then the next four of them get five entries, and then a bunch of other ones get four entries in the raffle. The odds are pretty good for any of those words is going to come up. If you put the temperature too high, you start getting wackadoo responses, right? So somewhere between temperature zero, which is it just does the same thing every single time picks the most likely and, you know, a high temperature of five somewhere in the middle, there is the pleasing temperature where there's just enough randomness that it makes it interesting and fools you into thinking it's like a little elf in there talking to you, but not so high that the elf is on LSD.

1:37:02Nice. And that's how these things work that we're all talking with. And so into this comes the watermarking discussion and Gruber's initial take on it, which lots of people had objections to. Right. So it's worth reading this post in full, but John has selected this particular quote. John Gruber writes, I want any LLM I use to choose the very best, most precise words at every single decision point. The idea that anything other than my needs should factor into the generation of text for me is patently offensive. The best writing I see come out of these models is worse than anything I would choose to read for pleasure.

1:37:39And now Anthropic is saying they're going to make it worse on purpose for purposes that do not benefit me in any way, even if only slightly worse? Get f***ed. so based on what we just read about temperature and everything uh you if you read that sentence by gruber which so many people on the internet did and says uh you know i wanted to choose the very best most precise word at every single decision point that would be temperature zero and it sucks like that's you know the only thing they have to determine best is probability and if it picks the most probable next word at every every turn it will sound dumb and terrible and bad and will not be pleasing to use as a chatbot or as a tool for doing things.

1:38:18That's temperature zero. So already before watermarking, it's not only is it not picking the very best, most precise words, you don't want it to. Now, he's just expressing his desire. I would wish these things would pick the most precise word. And he was saying that by watermarking, now you're making your decision based on something other than what's the best. Well, they're obviously always making the decision based on something other than what's the best. And that something else is randomness. Like that's their strategy. There's no other. But like, what if it's better in a different way? Nope, nope, they're numbers.

1:38:48And you have probabilities. And your choices are pick the best one every time. Or, you know, do like, whatever, like spread, spread the probabilities around. And so the lesser ones have a higher chance and then pick randomly, because that's how these things work. And everyone just replied to him and said, Hey, dummy, don't you know about temperature? Of course, he did know about temperature and wrote a big rebuttal. But before he got to that, people who replied to Gruber will link to Dan Morin's posts, where he has a more philosophical angle on this that gruber did respond to and daniel jalkett um had a had a good point that i think he made in his post about this which is basically like look if it's random in both cases um and if you're afraid that's going to change the quality uh it's just as likely to make the quality better as it is to make it worse because it's random you know what i mean at any given word choice if the randomness is truly random and you're just picking a different source of randomness it may make it worse but it also may make it better hopefully an equal percentage and so we'll average out to being the same and that makes anthropics claim that it will not change the quality of the output very uh compelling now anthropic didn't do itself any favors by saying oh we tested on users and they couldn't tell the difference that doesn't mean anything as grouper points out like users can't tell anything right but mathematically speaking it makes sense to me that is entirely mathematically plausible that watermarking would not change the quality in any way whatsoever because random is random is random.

1:40:09The temperature will certainly change the quality, but that's not what this is changing. This is changing the source of the random number generation. So Gruber's reply, again, Syracuse have pulled a little bit for me to read. The temperature-based randomness, which is what makes LLM output non-deterministic, is in place to help make the output better. The prose is clearly better with a temperature of 1 with weighted randomness than at temperature 0 with no randomness. The watermarking schemes, on the other hand, are applying predictable with the secret key randomness for an entirely different purpose than improving the quality of the output.

1:40:39And thus, I believe, inherently make the output at least slightly worse. Now, I'll pause you there. This is also some, he kind of backs it, walks it back in the next paragraph you're going to read, but there is a little bit of logical leaping here that is divorced from reality because, you know, with the, makes the output non-deterministic. Well, as we just discussed, pseudo-random number generators are absolutely deterministic. that's a feature of them temperature-based randomness does not make lms non-deterministic it just makes them random seeming to humans who don't know the seed or the functionality of the random number generator but it's deterministic code executed by a computer it is absolutely deterministic and then you're saying well watermarking the schemes on the other hand are applying predictable with secret key randomness that's what the other randomness is too it's predictable with secret both the randomness is coming from exactly they're both deterministic they just seem random it's a pseudo random number generation again nothing that anthropica said is said they're you know sampling like wind weather veins or something or to get like true entropy and randomness and even if they were let's just improving the quality of the randomness presumably a good pseudo random number generator spreads its randomness evenly like they're better or worse ones or whatever but i would imagine they're highly motivated to use ones that are efficient to run because they make these random decisions a lot in the data center.

1:42:04So you probably don't want a really expensive random memory generation. So anyway, I'll let you continue with the next part. But I just want to say that Gruber is like assigning qualities to these two things and characterizing one as non-determinant randomness, the other one as deterministic. And I think that is a logical error. They are both predictable with secret key randomness. Gruber continues, advocates of LLM watermarking schemes for text argue that the schemes don't necessarily lower the quality of the generated prose because they don't change the temperatures they only change the source of the randomness daniel jowkett wrote a good piece today about this i hope that's true i believe it's possible that it is true i think it's highly unlikely that it is true i do not see how a detectable signal can be encoded into the choice of words without affecting the meaning of the prose yeah so he says i hope it's true i believe it's possible it's true but he thinks it's unlikely now here's here's what i think river's problem is um psychoanalyzing this the motivations of the problem like that's his main objection the motivation of the watermarking feature is not to make the output better that's 100 true that's not the motivation the motivation has nothing to do with making the output better and i don't think it will make the output better but i also don't think it will make it worse and so he is his main objection is don't put time effort and money into doing a thing that you know is not there to make things better it's kind of like app store search like it's not there to make things better for the user right you know and app store search actually has a bad effect but anyway In this case, the motivation is not aligned with what John Gruber wants from an LLM.

1:43:31And I agree with that 100%. But I also think that it is entirely plausible that the watermarking scheme does not, in fact, change the quality in any way. It doesn't make it better and doesn't make it worse because all it's doing is picking a different pseudorandom number generator. Now, it's possible that the one they pick for the watermarking has worse randomness than the one they were using before. But I don't know enough of the details about that to know. and i don't think gruber does either uh again i think both of them are probably equally efficient and equally random and i'm not sure how random that is um but they're both predictable with secret key randomness and their job is the same which is after the temperature squishes down the probability curves pick one of them based on the the prob best on how many entries they get in the raffle based on their you know probability uh amount that they get after the squishing that's it that's the job and they're going to make different choices because you know you'll make different that's that's the you know the the sort of predictable secret key depending on when you time your query assuming time is in any way part of the seed that's going to change which things you get too if you return one microsecond later you might get a different response because the pseudo random number generator incorporates the time that it gets your response if your thing gets hung up on a router somewhere and arrives a few milliseconds uh later or earlier at the data center, that will also change your answer.

1:44:48That is all deterministic, predictable randomness. It's just that the factors that go into it, we don't know. And so it seems random to us. And that is entirely sufficient for randomness.

1:45:01Marco Arment:All right. Thanks to our sponsor of this episode, Tailscale. And thanks to our members who support us directly. You can join us at atp.fm slash join. One of the many perks of ATP membership is overtime, our weekly bonus topic. This week on Overtime, we're going to be talking about other people's phones based on aforementioned friend Stephen Robles' video about his son's iPhone. That'll be a fun discussion. And Overtime, you can join or listen to APJFM slash join. Thanks, everybody, and we'll talk to you next week.

1:45:45Marco Arment:Marco and Casey wouldn't let him Cause it was accidental It was accidental And you can find the 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 Casey Liss M-A-R-C-O-A-R-M And T Marco are men S-I-R-A-C-U-S-A Syracuse. It's accidental. They didn't mean to. Accidental. Tech Podcast. So long.

1:46:34So I had a long day on Friday. I drove from Richmond to Charlotte, North Carolina, which is about five hours. And I sold a car, bought a car, and then drove five hours back home. Congratulations. I was very tired, but very, very happy. Did you sell the car you drove there in? Yes. And that is itself a whole long story. I will try to give the short, short version of this, by the way, for people who don't measure in time, it's about a 300 mile journey, which I don't know what that is in real units. So I apologize, but you can look it up for yourself. So, yeah, what I did was I woke up, drove the Golf R down to Charlotte.

1:47:16It had been appraised online via CarMax, where I basically just gave them the VIN and said, you know, it's in good shape. I haven't done anything to it, blah, blah, blah, blah, blah. And I got a estimate from CarMax or an appraisal from CarMax. and I thought, okay, that's all well and good, but, you know, there's like a couple of blemishes on the car and they're not tremendous, but there's enough that I think if somebody sees this, they might be like, oh, well, we got to fix this and that and it's not worth what we said it was. So then I took it in. Definitely with that voice too. In exactly that voice.

1:47:46So I took it in, this is like, I don't know, Wednesday or Thursday of last week, I forget exactly when it was, but I took it in and got an in-person appraisal, which was very quick and easy. Now, to be fair, CarMax is based in Richmond. As I've said on the show in the past, was actually a spinoff of Circuit City, if you've ever heard that name before. But anyways, I went to one of the local stores, had them appraise it, came up with the exact same number. But interestingly, they said that they would have honored either number and I should just choose whichever one is bigger, which was fascinating.

1:48:14I don't think it's, maybe this is an American thing, I don't think it's a very appropriate thing to talk about money, generally speaking, but I am so astonished with the number that CarMax offered that I'm going to share it with you because i'm so proud of it we bought the golf r in on august 30th of 2018 i remember because it was the day before aaron's birthday uh we bought that car for a little shy of forty thousand dollars all in in august of 2026 eight years later carmax offered me twenty eight thousand six hundred dollars for that car that is that's an astonishingly small amount of depreciation i am flabbergasted that that's what they offered me.

1:48:53Marco Arment:Out of curiosity, how does that compare to the Blue Book values? You know what? Honestly, I didn't look because I knew that I was going to sell it to CarMax or Carvana who offered like 24 thereabouts. I forget exactly what it was offhand. Again, it gives me the heebie-jeebies to talk about money like this as an American, but that's so cool that we got that much money for that car that I'm very proud of that. Well, you got to look in the Blue Book value for the line item that says a little casey just drove it to church every Sunday. Yeah, right? Exactly. Because that's basically your car. It is a low mileage garage queen.

1:49:27Marco Arment:It gets very aggressively cleaned all the time. I remember when I sold the Tesla, I went to CarMax and I kind of had, they gave me an estimate and I was like, that sounds too good to be true. And I remember I drove there thinking I would take anything over X. And they came in so far above what I hoped they would they would pay. I was shocked and it was way more than it would have been easier for me to do otherwise. So I can definitely not a sponsorship. I could definitely say I had a very good experience selling to CarMax. Yeah, I mean, it's worked out really well for us. One slightly difficult thing about this, however, is that the car, you know, we had the title in hand eight years later and the car was titled in both our names, which is all well and good.

1:50:16but I'm going to be selling it in a different state without Aaron next to me. So what the hell do I do there? And we did our wills or redid our wills a couple of years ago and have durable power of attorney for each other. And in theory, the way this works in America anyway, is that I can sign as though I'm Aaron and vice versa. And that is as though she's there. That's what a durable power of attorney is. But that doesn't mean that CarMax has to honor that. It doesn't mean that CarMax and Charlotte, North Carolina has to honor that. So me being me, I thought, okay, well, maybe I'll just fly down and I'll deal with the Volkswagen later.

1:50:51But that was a$500 plane ticket. Now on the plus side, it was a direct flight, which is one of the like six places you can get direct from Richmond, but I'm too cheap to spend the$500 to fly down there. So I didn't want to do that. I looked at renting a car, but it was like a hundred dollars, which is within reason, but then I'm in a box for five hours and I want nothing to do with that either. So on Thursday, I called up a CarMax down there. Well, I think I went to CarMax like either Thursday morning or whatever day it was sometime. And I talked to the local ones and they were like, it should be fine shrug, which was not reassuring because I didn't want to be in a situation where I'm five hours from home and suddenly own two cars and only have one person to drive them.

1:51:31So I finally was able to find a number for the CarMax in Charlotte, which was a challenge in and of itself. I freaking hate the phone, but this is one of those times I needed a phone number. And anyone I found was like the corporate trunk line. But I eventually found it, called them. An incredibly nice person got me on the phone. I got this incredibly nice person on the phone who talked about it and was like, hey, why don't you send me an electronic copy of it so I can see it to make sure it looks good, blah, blah, blah. And she said, you know what? You should be fine. and so I set off at about 7 30 in the morning drove my five hours got to the BMW dealer I bought a certified pre-owned uh BMW i4 m50 this is exactly what I've been flirting with for six months now or something like that yeah about six months and is it white it is not white it is San Remo green I did not set out for green but the more I see this car the more I love this green I think it's beautiful I thought it was black did you did you pick the color or was just like this is the color it's available in and I don't mind it.

1:52:28Uh, the latter, I mean, I wouldn't, there was no exterior color that I can think of that I would have said no to. Uh, there was an interior color, the, the red that it seems that everyone optioned on this car for, I don't know what reason. Uh, I did not want a red, a red interior car. I would have held my nose if I needed to, but I did not want one. Um, this one has a black interior, uh, and it's, uh, this deep, deep, deep like forest green on the outside, which I absolutely love. I'm extremely, extremely happy with it. So I get down to BMW. We take it for a quick, you know, shakedown ride. All looks good.

1:53:03We do the stuff with BMW and that was mostly easy. And then I go across the street to CarMax and I sell the Volkswagen to them. And I ended up working with the exact same woman I was speaking with the day before, who again was amazing. I was the biggest worrywart in the world. I'm sure you find this surprising. I was so nervous about everything. And she was so calm and so chill and exactly the person I needed to work with in order to survive, basically.

1:53:28Marco Arment:Did you pack 40 bottles of water for the trip? No. Maybe a watermelon? No, no, I didn't need any of that. But no, I had like a small little cooler, like lunch bag that had a couple of drinks in it. Was it your tailgating cooler? No, no, no, nowhere near that big. This was like a lunchbox. They don't have lunch in Charlotte. They needed his travel router. The thought was, no travel router. Apple TV. The thought was that I didn't want to spend that much time because it's a long day, right? It's five hours down there, anywhere between one and five hours worth of financial transactions and five hours back, including a charging stop, which we'll get to.

1:54:05Did you do the selling of the Golf before you picked up the BMW or after? What do you think was the correct answer to this question? Because I had an intention and then I reversed my intention on my way there. I mean, the tricky part is once you sell the Golf, how do you get to the BMW dealership if you do the Golf first? But then on the other hand, you have the exact same other problem. How do you drive two colors to go sell the golf?

1:54:26Marco Arment:Well, I would worry like, what if whatever you do first, then the second thing falls through? Exactly. So you sell your golf and then you can't pick up the BMW and now you're stuck with no car. Right. Or you're stuck with two cars. So on the way there, I had intended to do the shakedown ride first, just to make sure that whatever it is I'm about to buy is okay. Then drive to CarMax. Now, these two stores were like a mile apart. Now, as it turns out, they were on this gigantic road in Charlotte. I had envisioned like a regularly a one lane each direction kind of street. I didn't get to the point of like looking at it on Google Maps satellite or Apple Maps satellite view or anything.

1:55:03This is like a four or five lane road each direction. Not the sort of thing. That's where car dealerships live. That's their natural habitat. I know. I know. I should have known. This is a failure of my own. Gigantic strode. Yep. That's exactly what it was. Exactly what it was. Uh, so I'd intended to do the check ride, sell the car, have them like pick me up or something or Uber if I had to, well, I'd originally thought I would just walk it, but I, then I quickly realized that was a terrible decision. Uh, then, you know, either have them pick me up or Uber back to the BMW dealer. At that point, I'll sign the sign on the BMW and drive away on the way there.

1:55:33Uh, my car nerd group chat, which I've brought up many times in the past, uh, one of whom is a former lawyer said to me, you're probably going to want to sign on the BMW first, because even though the salesperson I worked with was amazing and I really enjoyed him, first of all, his job is to make me say that and feel that way and basically gaslight me. That's the whole point of a car salesman. But secondly, if I sell the Volkswagen, now I'm screwed. And if the terms change on the BMW, I'm way more likely to be like, well, I guess that's the deal now. And so my friend Brian had strongly suggested, and I eventually acquiesced, and I think it was the right answer that I buy the car first, then run over to CarMax.

1:56:13And then they were kind enough to send a person over to pick me up. That's what I ended up doing. Um, it was very sad to get rid of the Volkswagen. It was an incredible car. It had something like 35 ,000 miles on it in eight years, which in America, you know, with you average, most people are considered 10 to 12 ,000 miles a year, uh, to be average. So this car should have had like nearly a hundred thousand miles by time. If you think about it, it had 30 ,000 miles. So as you said, in great shape. And, and even though it is, uh, old in a time way, it is very, very, uh, young in a mileage way. Uh, so all the financial transactions went well, no major issues.

1:56:54And I get on the road and I'm heading back to Richmond and oh my God, this car is so nice. It's so nice. It is a 2024, uh, 20 ,000 miles. I think I already said that, but 20 ,000 miles on her a little bit more than 20 ,000 miles on it it is so nice it has the three packages i wanted it has uh the driver assistance plus package i think it's called i forget the exact term for it um it has the parking and it has premium package so um it has their equivalent uh full self-driving it's a little squishy it's not really what it is but it has as close as bmw gets to full self-driving anyway um it has a chi charger in the center console it has like this wraparound well maybe that's a bit dramatic but like this this curved infotainment display that mates into the gauge cluster, which I could see someone, particularly John, saying that's hideous.

1:57:43And I wouldn't necessarily argue with you, but I think it looks good. It's just such a nice car. And in so many ways, I quickly realized that I'm kind of welcoming myself to 2017. And I say 2017 specifically because Aaron's former Volvo that got destroyed by Pebble was a 2017 model year. And that had amazing technology. technology like a sunroof, which my Volkswagen Golf R, the Mark 7.5, never offered. You can precondition the car. So even her car, which was fully gasoline, the first one was fully gasoline, you could get on the app and run the engine, which we didn't do that often, but you could do to get the car cooled or warm as you're leaving wherever you are on your way out to the parking lot.

1:58:29That is, on a really bad summer day in Richmond, let me tell you, that is clutch. So that's really cool. And then with the new car, with her new car, and with mine, obviously, you're not running gasoline. You're doing that with electricity, which is even better. And you're never going to be able to precondition a six-speed car because that would require the engine starting. And if you start the engine with it in gear, which is how you should park a six-speed or three pedal car, I should say, then you will drive it away. If you, if the car is in gear and the motor is turning, it is driving. So that would never work with any of my cars.

1:59:03And now I can do it. Uh, it has a heated steering wheel again, as did Aaron's 2017. It has rear climate control for the kids, which is exciting. Uh, so in so many ways I've joined the land of 2017 with my 2024 BMW, but, uh, but these are all as much as I'm snarking, you know, these are all niceties that I haven't had. And oh my God, it's so great. Yeah. Those are awesome. Yeah. And so I'm driving home. The car is so solid. It just, I feel like the Volkswagen, especially when new was similarly solid, but over the, I mean, it's, it's been eight years. And so, you know, things get a little loose is dramatic, but for lack of a better word, things get loose.

1:59:40And this thing is so solid. I don't know if I mentioned, I think I said it's certified pre-owned, which means, you know, BMW has gone through it, made sure it looks okay. I get an extra year of warranty, which is really nice because my former BMW experience, let me tell you, not great. So driving home, I'm going to tell you things, particularly Marco, that you have known for 10 years now. But what's incredible, one of the very incredible things about driving an electric car, not only is it incredibly comfortable, not only is it incredibly quiet, astonishingly quiet, but if you happen to see a gap on the interstate that maybe is not directly adjacent to you but is within reach of you especially if you're in a performance electric car that gap is yours by the time you decide you want it to be yours now admittedly i'm now describing driving like a bmw driver or alternatively driving like an apple which fair the shoe fits but it is so intoxicating that you can see a gap, I don't know, 50 yards or 50 meters ahead, if you will.

2:00:42And you can be there instantly. It's unbelievable how fast this car is. And I know all electric cars are very, very fast. I would assume that your, what is IX3, right? Do I have that right? I have the IX. Oh, it's just IX. I had a feeling that was wrong. Okay. I would assume your IX is nearly as quick. And for all intents and purposes, is as quick. Like, even though it isn't literally, for all intents and purposes, it is this quick. But, oh, my God. Oh, my God. It's so amazing. And, you know, one of the things I feared most, and John, you're going to beat me up for this? Because I would if I were in your shoes.

2:01:19One of the things I feared the most about adjusting to this car was missing the six-speed, was having my left leg get bored. admittedly I'm in the honeymoon period but I am so enamored by having instant unlimited power or at least unlimited at the speeds in which I drive I don't think I miss it. I told you you wouldn't. I know you did. And if I were hearing these words from John, I would be ripping John a new one right now. So, John, feel free to beat me up. It's deserved. I mean, I believe that you don't miss it. I think I will miss it, but I think I'll also appreciate the better things. So, you know, time moves on.

2:01:57But I don't know why I'm getting electric or anytime soon. But, yeah, it's, you know, it's like kind of the way you miss things, like, from, like, your nostalgic youth or whatever. and it's like, well, I wouldn't want to have it again because in many ways it was old and bad, but I have fond memories of it. And I feel like that's what stick shift driving is going to be to you.

2:02:13Marco Arment:The whole appeal, for me, the appeal of a stick shift was always better control over what the engine is doing and how it's delivering its power. And with an electric car, especially a single gear, like not the stupid Taycan, you don't need that type of control because you just you inherently have more direct control with your regular gas foot like you don't need a whole second mechanism to control this like very kind of peaky power delivery pattern that it has like it's you're just controlling something direct so that the so the entire problem that you need a better control over doesn't actually exist anymore yep that's exactly it and so much is different about this car that I feel like it's different enough that I don't think I'd miss it.

2:03:07Like if I were to jump into a Golf R, even a modern Mark 8 Golf R that has the dual clutch transmission, I think I'd be like, this is trash. And I think it's because it's still an internal combustion engine. It's still shifting gears. And I want to have that control, damn it. I want to have that control. In an electric car, it's such a different beast. And the instant power as a pure-blooded American who loves a muscle car, that instant power, oh, gentlemen, it's so intoxicating. It's so intoxicating. So yeah, so I'm driving home and immediately, you know, I never thought of my Golf R, which again was a 2017 model year.

2:03:45I never thought of the Golf R, it was 2018. I'm sorry, 2018. I never thought of the Golf R as feeling old. You know, it had car play. It had the display screen for the gauge cluster rather than having like physical gauges. It was early on in that trend. That had been a thing for a while, but it was one of the first Volkswagens to do it. So it didn't, like, reek of being old, if that makes sense. But, oh, my God, to be in this new car where everything is new and shiny and, like, the screens are very, very big air quotes here. The screens are retina by comparison to the ones in the Volkswagen. Not literally.

2:04:25I'm just trying. You get what I'm trying to say here. The screens are so much better. the processor in the car, whatever processors are doing, whatever animations it's doing are so much nicer so much of this is so great so I'm on the road on the way home and I have the fancy schmancy automated driving and I hit a little bit of traffic and do you know what the car does when I come up to traffic it stops itself and do you know what it does when the traffic keeps moving the car starts moving again, this is technology that most of the world has had for a long time this is amazing it's so nice and I'm probably not going to be on the highway that often with this thing I mean I was rarely on the highway with Volkswagen but it's so nice and so I found that on the way to Charlotte I was locked in I was wired in because I had visions of like some pebble or something jumping out at me and like cracking the windshield or leaving a dent or something like that I don't know that I have ever driven as defensively in my life with the exception of the time that my children came home from the hospital.

2:05:30Every other time I think I've gotten in the car, I've been less defensive than I was on this five-hour drive to Charlotte. And let me tell you, five hours of white-knuckling your way down southwest into almost South Carolina. That's how far away it is. I was tired by the time I got there. But on the way back, obviously, I'm still paying attention. And, you know, I'm not one of these dummies that believes full self-driving is literally full self-driving, but I'm in. So I'm still paying attention. My hands are still on the wheel, but I don't have to be quite so locked in and, oh my God, you know, I can actually like breathe a little.

2:06:05It made the drive so much nicer. And as much as I snark on Tesla, my understanding, having not driven a Tesla in a few years now, is that they're, they're self-driving while not really full self-driving. my understanding is legitimately, all snark aside, that it is the best in the business. And so given how good BMWs is, I can only imagine how good Tesla's is. But at one point in this, what did I say, 300-ish mile journey, I needed to stop. And I almost said fuel. I needed to stop and charge because the range in a perfect world, and the salesman charged it to 100%, didn't even have to ask him.

2:06:39But I still needed to charge at some point. And the range on this car in the summer is somewhere around 250, 275 miles. And again, this was a 300-mile journey. So I didn't know how I was going to handle that. My understanding from my friend Brad, who has an F-150 Lightning, is that the Tesla Supercharger network is far and away the most reliable, the easiest to use as long as you have an adapter to go from NACS to CCS. This car has CCS, which is what the rest of the world uses, but Tesla Superchargers use Tesla Standard. Now, Brad had been kind enough to loan me one of his NACS to CCS adapters.

2:07:15So I did have that in the car. But I don't love the idea of giving Elon any money whatsoever if I can avoid it, including the$10 or$20 or$30 it would be to charge my car. So I worked with a better route planner, which is the standard issue. This is what you should use in order to plan your route if you're in an EV. and so I used a better route planner to figure out where I should charge. And toward the top of North Carolina, there was, guess what, Marco, but a Rivian, what is it, not Action Network. Adventure Network. A Rivian Adventure Network charging station. And I pulled up to that, backed my butt in, using my fancy 3D overhead camera, thank you very much.

2:07:57Yep, yep. And I backed in. All I did was bloop, bloop, and then I plugged in, and then I walked to the other end of the parking lot to Wendy's to get myself a junior bacon cheeseburger, which I love. I know it's trash. I don't care. I love it. It's delicious. Wendy's is delicious.

2:08:13Marco Arment:Thank you. I'm a Wendy's fan. Thank you very much. So by the time I had received my biggie bag, this is the most American story I've ever told. But anyway, by the time I received my biggie bag with my junior bacon cheeseburger and spicy nugs and fries, I had enough in the car, enough charge in the car that I could have left right then and there. So what I ended up doing was I ate my nuggets because I had a little dipping sauce with it. I ate the burger because I had a feeling I would dribble all over myself with a car. And then I took the fries in the bag because if I drop one, I can find it. And I got on my way.

2:08:45It was something to the effect of 20 bucks at the Rivian Charger and about 20 minutes or thereabouts. You're making your new car smell like fast food though. No, this is why you crack the sunroof that I mentioned I have. That doesn't stop it from smelling. Maybe you like it smelling like fast food. No, it's fine. It's fine. But I've done this trick for years across several cars. As long as you constantly have ventilation, constant ventilation from the outside world, not the HVAC, if you will, but the actual windows down or sunroof cracked, it's enough. I promise. Unless you have something extremely smelly.

2:09:18Yeah. So the test would be you arrive home, you get out of the car, you remove all evidence that you had fast food. then you tell your wife to get in the car and say, did I have fast food on this trip or not? That's fair. She would not have noticed, not because of Aaron. I'm saying it was that not fragrant. But I totally hear what you're saying. I really do. But so anyway, so I did my charging stop. Again, I'm a worrywart. I was deeply stressed about that. But I knew that there was a Tesla supercharger a couple miles away. And when I got there, I think I had like 100 miles of range. It was one of those situations where it was just a convenient place to top up.

2:09:50I didn't absolutely need it at that very moment. And so I think there's Electropy America nearby, which referred as trash, although getting better. We're going to need to get you the charge limit on this thing so you don't keep it plugged in in your house like a desktop laptop at 100 % charge all the time. Leave it at 50%. Geez. Well, one of the first things I did was crank the charge limit to 80%. Yeah, 80 % is fine. I've debated bringing it down more. I don't know what I'm going to do about that. But that is one of the first things I did. So I got my way home. I'm enamored by all the silly things about the car.

2:10:22you know, as you walk up to it, it unlocks itself. You don't have to do anything. It just automatically unlocks itself. When you walk away, it locks itself again. You can use your phone as a car key to get into the car. And then if you place it onto the Qi charging pad, it can run the car. I'm a little nervous about that. And plus, I'm still amused about the whole locking and unlocking itself thing when you have the physical key in your pocket. So I like to have this as a backup scenario, but I don't know that I would necessarily choose to run with just the the phone as a key. I can use my watch as a key to get in and out, but it's kind of hard to hold my hand on the Qi charger when I'm trying to start the car.

2:10:58Um, so I don't think I would do that, but I, at least I can get in and out of the car with my watch. Um, there's so much about this car that I love so very much. And I think it's such a good fit for me. Um, one of the things that I did do subsequent to having it for a day or two is I finally got the opportunity and understood how to use launch control. Now I'd read instructions about how to do this. And they basically said, just have it in sport plus, I think is what it's called. No, that sounds wrong. Sport boost. There it is. Have it in sport boost, mash down your left foot on the brake, mash down your right foot on the gas, and it'll say launch control activated on the dashboard.

2:11:33And then just pop your foot off the, uh, off the brake and off you go. And I tried this once or twice and I never saw the message. Like I was on the brake and I didn't go anywhere, but I tried this dance and never really saw it and i watched either a youtube video i did eventually read the manual which uh i said to marco via text i'm doing a spot of light reading and that's all i said to

2:11:54Marco Arment:which marco immediately replied oh you want me to dig it up you said something along the lines of you're reading the manual aren't you to which i then replied with a photograph of me reading the only person who reads car manuals anymore you nailed it i read them too thank you john I read all the manuals to all my cars. So anyways, so I read the manual. I looked at some YouTube videos. And it turns out, I think what I was doing wrong was that you need to like properly mash down your left foot on the brake. You can't just be on the brake enough to hold you. They want, the car wants you to be mashing down on the brake.

2:12:29And then you mash down the gas and you pop the, I almost said pop the clutch, you pop the brake. Oh my God. This car is so fast. It also does some really silly things that I think I'm supposed to hate as like a self-respecting car guy, but I actually love because I'm a child. When you do launch control, it kind of like squats a little bit or like shakes a little bit to mimic kind of what it would feel like to have an actual motor going. I was going to say, isn't it brake by wire? I honestly don't. I understand the question. I don't know. Like when you're mashing the brake pedal, it's not like you're pressing the brakes any harder because you're not actuating the brakes.

2:13:07I understand what you're saying. I just don't know the answer. You're probably right, to be honest with you. But yeah, so it's kind of like squats a little bit, does a little bit like. Also, it makes like these noises when you're in sport mode in particular. Any BMW i4 driver will tell you, well, Hans Zimmer composed the in-car noises, which is true.

2:13:25Marco Arment:I did enjoy those noises in my car for about the first three or four months, and I turned them off. Okay, well, there you go. I think I'm supposed to be like, oh, this is dumb. I'm a self-respecting car guy. I would want a fake internal combustion engine, which it doesn't sound exactly like a motor, but it's inspired by a motor. Man, I freaking love it right now. At least right now, I love it. The Hans Zimmer soundtrack is very good. It's really, really good. But launch control, I have never... I cannot recall a time, and I've been in fast Teslas, but to be fair, it's been a long time. I cannot recall a time, even in the Rivian launch edition, the quad motor Rivians, those were astonishingly fast when you consider the fact that they weigh 90 billion pounds.

2:14:12this car also weighs like 6 000 pounds or something like that but the thing with the launch control in the i4 m50 is that i cannot think of a closer or a more analogous experience to one of the linear induction motor roller coasters where they launch you you know what i'm talking about that is what this feels like it is astonishing how quickly this thing will get up the road i want to do it every time i leave a stoplight and i'm mostly resisting right now because I'm not that much of a jerk, but I'm only barely able to resist. It is incredible. I don't even know how to describe it. Marco is going to be in town in the future, and I cannot wait for you to, at the very least, experience it as a passenger, if not take it for a spin and do it yourself.

2:14:56This is incredible. Everything about this car, I love it so much. I love it so much. I'm so very happy that this is the decision I made. I had kicked around the idea of leasing a new one because you could make an argument that there are financial reasons why that would be better. But honestly, this car, I paid like 60 % or thereabouts of the original purchase price. Like that is a really good deal for a two-year-old car that's CPO'd. I'm very happy with it. The only problem I have with it right now, right now anyway, is that BMW killed home assistant integration like three months ago. And as far as I know, there's no way to get, at least in America, there are ways overseas, but at least in America, there's no way I'm aware of to feed data into Home Assistant from the car.

2:15:40And that kind of sucks because it would be nice to know what the state of charge or to be, especially if I could write things to the car to tell the car, you know, oh, I would like you to start preconditioning yourself or whatever. And that kind of bums me out. And the BMW app is good and it has a really good widget support. And, oh, I forgot to mention, I have finally, for the first time since the little widgets on the bottom of your lock screen have been a thing for the very first time i've changed from the camera widget on the bottom right or control center widget or whatever we're going to call it uh to something else now there's a little fan down there so from my lock screen i can precondition the car it's incredible i'm so happy it's so silly but i'm so happy um but all in all like the app is a little wonky i think i prefer the volvo app in the grand scheme of things the bmw app has definitely got more knobs and things to fiddle with but i think the Volvo app, not only is it design-wise better, but I think information architecture is a little better.

2:16:36But I thought, if you'll permit me, if we can give it, give me just a couple of more minutes, I thought we could do something together. And maybe, John, if you wouldn't mind, you can take the lead on this. There's a gentleman I know who wrote a post on his blog in May of 2026 called the EV Stupidity Checklist. And I thought maybe, John, you could be the emcee as we go through this list for the BMW i4. Oh, geez. I hadn't actually thought about this. I mean, I know you're getting this car no matter what, but that is very true. All right. I mean, so let's start with the first one because I did see the picture of this car and I almost commented on it.

2:17:10Accessible exterior door handles. Correct. Check. Is it obvious and immediately usable? Do you have to wait for something to happen or can you just grab the handle? You can just grab the handle. Now I will say that the last several cars I've had, and forgive me, I don't know the terminology for this, but the handles were ones that they protrude from the car and you can grab them and pull toward you. And that's what engages the latch or disengages the latch, whatever, to open the car. That's the way the Volvo still is. That's the way the Volkswagen was, BMW that preceded it. Several cars in a row have been this way.

2:17:40The current BMW, the i4, has the old school, like you stick your fingers behind a little lever and pull toward you and up. That is the world's worst description. But it's like the prototypical old-style lift-up-on-the-handle kind of door handle. If you saw a picture, it would be abundantly obvious. I'm painting a terrible word picture right now. But it is a legit, regular, familiar door handle. And on the interior, there's a very traditional mechanical door handle as well.

2:18:09Marco Arment:Yeah, that's actually like when I was trying to figure out, do I want to find myself an i4 or wait for the upcoming i3, one of the things in favor of just get the i4 is it's, you know one of it's possibly the last bmw ev to have regular door handles and a sunroof by the way um and i think that's i think those are both going to be not really available for future evs so the next item is i think you might have touched on it but are is that that thing that you pull is it a physical door opening mechanism or is it electronically actuating a switch nope physical uh interior and exterior both that is correct All right.

2:18:49We already talked about affordances. Yes, you have exterior affordances. You definitely know where you have to put your hand and grab. Interior, same deal. Is there something you definitely know that you have to grab? Absolutely. Yep. Okay. Physical charge port door mechanism. How does the charge port door open? So one of the disadvantages of the i4, which I think we talked about a few episodes ago, one of the disadvantages of the i4 is that it's based on the 4 Series gasoline car. So a lot of the EV purists note that it has the tunnel for the drive shaft from the front to the back of the car.

2:19:20It has that tunnel in the rear seat. So you see this bulge, or there is a bulge, in the center part of the floorboard in the rear for a drive shaft that does not exist. And that's kind of silly and annoying. But the advantage is that it's also designed to be filled with gasoline with a door in the back of the car. And that door in the gasoline model is mechanical. In the electric model is mechanical. So it is a physical charge port door mechanism. You open it by pushing in and then it becomes a jar and then you pull it open. Remember the 70s when they had key locks on the gas caps? Was that before your time?

2:20:00No, I mean, it was before our time, but we saw it.

2:20:02Marco Arment:Yeah. Yeah, there were a lot of cars that still had those that were, you know, still in service during our youth. Exactly right. Kids, ask your parents. Turn signal stock has got one of those, right? Yeah, it sure does. Again, you would think this would need to be on the list, but trust me. No, no, it does need to be on the list. It does, yeah. Physical buttons on the steering wheel. Yes, sir. Volume going to increase your speed or decrease your speed. uh there's a kind of an up down like on each side of the wheel there's like an up down like scroll wheel sort of situation but there are physical buttons uh there there are abundant physical buttons on the steering wheel there are not as many as i would like in the center console area which we'll get to in a minute um for the steering wheel are you giving yourself partial credit because there's one of those things where it's like five buttons and you press down a single piece of plastic and it knows which part of the five no no that that's fair no there there are distinct physical buttons for like volume, for example, for the cruise control switches, etc.

2:20:59Physical controls for temperature and fan speed. Alright, next. This is sad. The temperature and fan stuff, that is all on the screen. Now, in fairness, like the Volvo, the temperature and seat heater controls are always, always, always visible, even if you're using CarPlay. But, but it is not physical, and I wish it was. Yeah. Oh, let's see. The other thing that a lot of uh evs fail on this is that those controls are not on the touch screen they are on separate controls that are like below the touch screen but they're still not physical controls like they put dedicated controls for them but they're capacitive on like this little piece of plastic or glass or whatever it's like what what are you doing like you were almost there just to make it a button anyway um physical controls for airflow and direction that basically means the places where the heated or cooled or otherwise air comes out of like the dashboard in the car can you change where it's blowing by moving a thing with your hand like the vent you know absolutely can that's nice um physical glove box opening mechanism i just laugh every time i read these yes and i should note that the volvo fails on this one but yes yes really volvo doesn't have just an electronic uh actuator thing uh yes it has a button that you use does it make you use the touch screen uh no it's not the touch screen it's a physical button near the glove box but not on the glove box yeah electronic popper uh do you have a rearview mirror yes the rear visibility is not stellar because of the slope of the of the hatchback yeah but it's still a rearview mirror and i quickly got used to the rear visibility does it does it have a thing where the mirror changes into a screen too uh no not to my knowledge anyway um you complained about the tycon as a rear visibility i guess this is better than tycon it is night and day better than the tycon but it is not as good as i would have preferred but again i got used to it very quickly do you have a rear window seems like yes right yes yes i do do you have side view mirrors yes yes i do yeah thanks thanks to u.s law probably but yeah no the ipod is pretty well and this so the disadvantage of not waiting for like these new class you know fancy bmws is uh you know you don't get whatever the new things like yours your car isn't even 800 volt architecture is it no yeah so it's old it's old technology but But it's from this brief moment in time when they were putting physical controls in their EVs.

2:23:18Because like you said, it's based on the gas car. And so it's like, well, we're not going to redo every part of the interior for the EV. And the gas car had some physical controls. And so you're benefiting from that. So you get a lot of advantages for having an older version of this. I believe most of the fancier new BMWs are going to do worse on this list than you did. Yeah. Yeah. I think that is very possible. I mean, there's going to be a lot of advantages for the Neuer Classe or Neuer Clarence. No transmission tunnel. Yeah, right. Exactly. Exactly. But I think overall, I think you're exactly right that in some ways, and again, the EV purists will yell at me, but in some ways, the fact that this is based off an internal combustion car is an advantage.

2:23:53And I think I said this last time, but like this thing does not have a frunk. It has a whole bunch of like machinery and whatnot in the engine bay. And I don't really care because I've never had a frunk. I don't know what I'm missing. and the back is given the shape and size of the car. The back is gigantic. It's mammoth. In fact, I would argue there's way more storage in the back of this car than there was in the back of the Volkswagen because there was basically like a plenty of vertical space, but very little depth. Also, there's one other thing I forgot to bring up and I'm very mad at myself that I'm waiting until the end to say this, but this car has brought an entirely new thought technology to the List family because Aaron and I met in January of 2005.

2:24:35since that time for the last 21 and a half years, Aaron has never, ever, ever been able to drive the car that I own. Never. She, I've tried many times to teach her how to drive a stick and we would do a lesson or two and then she would do life would get in the way and she would never do it again. And so I think in a pinch, she's capable of it, but because she never does it, she has zero confidence, which if you squint means she can't do it after all. Um, she drove, we, we went and played pickleball with some friends, Saturday, the day after the car came home, she drove us there in my car. That's something that we have not experienced in 21 years.

2:25:11And the fact that she can drive either car is incredible because now it doesn't matter. If I want to take the Volvo for some reason, I don't know why I would, but if I do, she's not stranded. It's incredible. Also, she's not making fun of me for driving a gas guzzler anymore because guess what? This bitch don't take gas. So no matter how you slice it. This is such an incredible improvement for the Liz family. I'm so incredibly happy and so very thankful that we were able to find a car that fit our needs. And no, it's great. I understand now that why, when I first started talking about this, so many ATP listeners were like, oh, I have an i4.

2:25:49And a large subset of them was not only I have an i4, but I have an i4 M50. and so many nerds seem to love this car and i get exactly why and i hate to say this again because i really wanted to be past bmw but as i said about my test drive and now that i own one it's like coming home boys it's like coming home thank you for indulging me i appreciate it i i'm

2:26:15Marco Arment:so happy that first of all i'm happy that you have a car that you're really excited about because I know that matters a lot to you. And you get a lot of joy out of that. And that's wonderful. You have worked for it. You've earned it. And you deserve it. Secondly, I love that you're in the EV world now. Oh, I know. I'm so happy with that. You've been very gentle with me both tonight and in the past as we've had conversations in Slack or occasionally via text. You've been very gentle with me and have been very kind and left the I told you so's to a bare minimum, which I appreciate. thank you yeah um it's about time i know i know it really is yeah but i'm i'm very i'm very glad that you uh that you have arrived yep me too and i mean so far it's nothing but upside like even the charging stop as i said to you you know i didn't know if it was gonna take forever i didn't know if it was gonna be burdensome it was not burdensome and it took the slowest part of the charging stop was me i was the hold up with the charging stop and i know that wouldn't be true always if i was going on a much longer trip etc etc but so far this is incredible and i don't ever want to go back which is exactly what every ev owner i've ever spoken to ever ever ever always says yep and and look look evs aren't for everyone but they're for almost everyone like most people like if you can once you can afford one like in your price range and that's that's a big that's a big if you know like that's a big thing that still keeps a lot of people out of them but in terms of like okay if you can get past the price then is the car and its features and everything like is that right for you for most people the answer is yes even if they don't necessarily think so themselves for most people's needs are served very very well by evs i'm not sure if the show is over and this is just the bootleg but speaking of price by the way it's i know you get complaints about this casey i will reiterate the average selling price of a new car in the United States is 50 grand.

2:28:18You paid less than that. Correct. Like people think, oh, I'm getting a BMW. It's expensive. You said it before, but just to reiterate an absolute value, this car is depreciated down to 60 % of its original value, right? This is a car that is less expensive than the average new car. Now granted, it's not a new car. It's like, oh, it's not fair. This is not a new car, but it's like, it's 20 ,000 miles and it's in good condition and it has a more, you know, certified pre-owned war. Like it's effectively a new car for less than it's below. You bought a car for below average price. That's what it comes down to.

2:28:45And like, so I think this is one of the more financially responsible things you've ever done, automotively speaking, even though you are very impatient about it. Well, yes and no, you're not wrong. But I have been thinking about this and kind of kicking tires since February. So it's been six months if you squint a little bit. But I do take it. Once you've got to be in your butt about this is the car that I want. It was like a week. And you got it. That is more true than I care to admit publicly. But anyway, all I'm saying is this is a below average price car for a new car. Yeah. And I mean, again, I think one of the things that kept me out of EVs for the longest time was because I would look at these prices and just figuratively vomit all over myself because it was just, oh my God, 80, 90,$100 ,000.

2:29:30Like the Porsches that I was looking at casually very briefly, the sticker on those was like, you know, 100, 110, 120, which obviously I was never buying a new one, but just an astonishing amount of money. and I think the sticker on this was somewhere around$75 ,000 or something like that and again like I had kicked around the idea of leasing one but even leaving aside the fact that I don't particularly care for leases personally it just seemed like why wouldn't I get a lightly used one which this is by my definition lightly used why wouldn't I do that and save myself you know 40 % of the cost of the car and I know it's a little bit different with leases because you presumably you would just turn it in and get a new car, blah, blah, blah.

2:30:08But I view it as roughly analogous. And yeah, I mean, if you're willing to get something used, a used EV in America anyway, right now is astonishingly affordable. It really, really is.