645: More Frosting

43:01That was such a, you know, a sobering story to so many developers. But like the reality is like, yeah, things like ideas get copied and free usually wins over paid. like that's i believe me i in a similar era i had a paid app with big free competition and i learned that lesson too it sucked but that's the market you know and and in the market of ideas no one cares who did it first no one cares who ripped off who and no one cares who's doing it better because threes is a better app but that doesn't matter either totally but yeah exactly like that that doesn't matter like no one cares if you invented something first except you no one cares if you do something better except you look i mean i do the same thing now like with overcast like a lot of my audio features i was like the first one to do something or i was one of the first ones or i did it in a new way or something and who who even knows that no one and who cares i do but no one else does so you got to do that stuff for yourself not the market we are sponsored this episode by delete me delete me makes it easy quick and safe to remove your personal data from hundreds of data brokers online at a time when surveillance and data breaches are common enough to make everyone vulnerable.

44:15So here's, you know, if you go and like search for your name on the web, you're going to find a whole bunch of websites that give anybody quick access to your, you know, your full name, your family members' names, your home address, your phone number, their phone numbers. Like it's kind of messed up and creepy. And so the reason they're doing this is that there's all these data brokers out there who buy and sell your data. What Delete.me does, Delete.me knows your privacy is worth protecting. So you sign up for Delete.me, you give them the information that you want deleted, and their experts will take it from there.

44:44You're done at that point. And then what they will do is they will go to hundreds of data brokers and will file opt-outs and takedowns to get your information deleted from their databases. And this just makes your stuff show up in way fewer places online. They will send you regular personalized reports showing what they found, where, and what they were able to remove. and this is always working for you. Not just a one-time thing. They're constantly monitoring and removing information from new places wherever it pops up that you don't want out there on the internet. So it's a great service. Check it out if you want something like this.

45:14I was looking around for something like this a few years back and before they were a sponsor, I went and found DeleteMe and I signed up and I did it and it was great. Take control of your data and keep your private life private by signing up for DeleteMe now at a special discount for our listeners. Get 20 % off your DeleteMe plan when you go to joindeleteme.com slash ATP and use promo code ATP at checkout. The only way to get 20 % off is to go to joindeliteme.com slash ATP and enter code ATP at checkout. That's joindeliteme.com slash ATP code ATP. Delete Me does all the hard work of wiping you and your family's personal information from data broker websites.

45:50Thank you so much to Delete Me for sponsoring our show.

45:56We're going to get cranky for a second, so if that's not your bag, that's fine. Well, no, I'm trying to be more positive in general. Sometimes I'm succeeding, sometimes I'm not. You're predicting all of our responses to this. You don't know what we're going to say about this first item. Sure, John. Okay, sure. Apparently, from Joaquim, I think, forgive me if I'm pronouncing that incorrectly, but they write, when uploading a cyst diagnose or probably any other attachments to feedback, you get the usual privacy notice that there's likely a lot of private and other sensitive info in those log files.

46:28So, you know, you're making a bug report to Apple, It'll capture what they call sysdiagnose, which is a bunch of information about the system that you were using. And you might have attachments or whatever. And you're uploading this to Apple as part of a bug report. It's not a great feeling, but I mostly trust the folks at Apple to treat it with respect. And I trust the logging system to redact the most serious bits. However, when filing a feedback today, this was a couple of weeks ago, I think, I noticed a new addition to the privacy notice. Quote, by submitting, you blah, blah, blah, blah, blah, agree that Apple may use your submission to train Apple intelligence models and other machine learning models.

47:03I'm sorry, what? I don't think I love this. I'm really trying to be positive, but I don't think I love this. So my one complaint is that there is no way to opt out of the training part because basically if you don't agree to this, then you just don't get to submit a feedback, which seems pretty harsh because this is a new addition. Before, again, SysDiagnosed captures tons of stuff about your system. Every app that you're running, everything that's installed, like just tons of stuff and uh what this was referring to about like the logging system has a way to redact private information programmers have to use it correctly in their apps not to leak private information but presumably apple does a good job with that anyway you're already sending a bunch of private information but you're doing it as like as a part of the bug report like you have a problem you want to send up like here's the system that i'm running here's all the apps that i'm running here's all the kernel extensions here's the versions of all my things like you're trying to help them solve whatever bug you're having so that's part of the bug report but now it's like also uh we might train apple intelligence models uh and other machine learning models on the stuff that you submit um and there's no way to say okay i'll do the submission but please don't train your models on it because i feel like training the models doesn't really have anything to do with addressing whatever bug you know the sys diagnose stuff does have to do with the bug it's telling you know apple potentially how to reproduce the bug or what's involved in you know crash reports logs stuff like that right but the training the models that's a whole separate thing it's like okay maybe that's the thing they want to do and it's great that they're telling us about it but they should be asking us hey do you want to submit the thing anyway but also tell us not to train and that's not an option and that's crappy and that i think is actually my only complaint about this that it's basically like don't submit feedback at all or allow us to train i understand why people don't like it and like this is one of those things where i think a lot of people will blindly click through this and not think about it but i'll try to give people something to think about a lot of the ongoing court cases on which we're about to touch on involve people saying oh you trained on you know uh my book or new york times as you trained on our articles and the reason i know that is if i ask a question in a certain way i see paragraphs of my book spit back at me i see paragraphs from a new york times article spit back at me either almost verbatim or 100 verbatim and so it's like so i can tell you trained on my stuff because look i can get it out of your thing if i ask the right question um this is true of your source code too if you submit stuff in a feedback and they train their model on it let's we first of all we don't have no idea what they're training because they don't say but let's say they're training swift assist or one of those things that helps you in xcode it could be that someone somewhere type something and like auto completes to a paragraph of code that you wrote and you might not think about that maybe that paragraph of code that you wrote has like a snarky comment or a variable name or like things that are personal to you that would you wouldn't want being out there in the world like when you're submitting a feedback report you're not thinking someone somewhere someday will uh ask some apple lm for something and my code will come back with my comments and my variable names and also obviously proprietary stuff like if you have some secret algorithm in your app that does some fancy thing now people can autocomplete that without knowing where it came from without giving any credit to you or whatever so maybe you want to think twice about that the reason i'm personally not concerned about it is because i never upload my real source code to feedback i'm always making sample apps and the sample apps have nothing that i care about in them they're just bare bones sample apps with code that is as far as i'm concerned could be public and if radar slash feedback had a way to market as public i would do that but that's not true of everybody if you're like adobe and you're like we're having a problem in some part of photoshop and you upload that portion of photoshop i don't think adobe wants someone auto completing their like filter kernel in some xcode project a year from now because apple trained on it uh and the final thing that i'll notice and a lot of these like oh by the where we might train your stuff your we're trying machine learning on your stuff this cycle has happened multiple times already with other companies and a lot of the time it's just you know lawyers getting in there and saying just i'll write this language so that we can do anything we want and then the company swears up and down oh we didn't mean it to be that broad we just wanted to use it for this very narrow thing but the lawyers made us put this broad thing and so we'll change it to make a more narrow thing that works if a company is i was going to say like it is uh employs humans who are responsive to feedback in real time i would say that does not characterize apple when it comes to this type of thing so i'm not optimistic that apple is going to remedy this language to clarify what they mean in fact this story is weeks old and i don't think i've heard anything about it but just fyi this is a thing apple wants you to agree to which you may or may not care about i personally don't because of the way i use feedback but lots of people legitimately could because of the way they need to use feedback and like i said the only option is to not use feedback so congratulations apple you put up one more barrier to discourage people from sending feedbacks all right so with the disclaimer that i know this is going to make people mad i think this is not that big of a deal see casey you're over two i know i'm i'm shooketh over here i cannot believe my ears my goodness i think john is right about like you know for the most part what you should submit if you want your feedback report to actually get fixed.

52:15First of all, you're rolling the dice anyway. Good luck. Odds are you're wasting your time. But if any human being ever actually looks at it, which is, again, very unlikely, what you need to attach to it is a very, very simple sample project that doesn't really contain any of your code. It's just like, here is me using this API in this way and when I do these four lines of code, it doesn't work. That's the kind of thing they will actually respond to. It should fit on one screen. Yeah, if you're uploading your actual source code to your actual functioning app of thousands of lines at least of code, that's probably not going to even get your bug fixed because it's too complex.

52:54And if you can't figure out a very small amount of code that shows the problem, they won't even bother looking at it either. Ideally, you're not really submitting anything of significant value or exclusive knowledge to you. that being said I think broadly speaking and please forgive me everyone I think no one's going to care about this kind of stuff in like two years wait define this stuff training on training AIs on quote my content like what you know I think people are largely not going to care about this in the long term I think the AI tools provide a huge amount of value and we're all starting to use them in all sorts of ways.

53:39And so as that value goes up, first of all, I do think it weakens the argument that a lot of people make that like, well, you ripped off my content and you're creating value just for yourself. Because the reality is most of us over the next few years, if we aren't already using AI tools to help us out, it's going to just be in everything, in everywhere. We're all going to use AI tools in some way. Even if you just do Google searches now, you're using AI, whether you like to or not. It's everywhere and it's getting everywhere even more so because these tools are so useful and they have very high utility to so many different contexts.

54:18You're using AI everywhere. Whether you intend to or not, it's just getting everywhere. There are always going to be people who resist it and who are against it for various reasons. but I think the numbers of those people are going to dwindle over time and it's going to become marginalized the same way that people freaked out about a lot about the internet when the internet was new and over time for the most part most people have gotten past that or those people just got more and more marginalized who wouldn't who wouldn't let go of those objections because there was so much value being created in so many other ways that it's helped so many other people in so many ways, everyone kind of moved on.

54:56I think that's happening with AI. I think that will continue to happen over time. And so people who were super against AI on principle are just going to keep getting more and more marginalized. And if you are one of those people, this probably sucks to hear and you're probably really mad at me. And I'm sorry, because it's not, it isn't super clear cut. I can't say this is a universal good with no downsides. It's not, but it is creating a ton of value for a ton of people and it's just going everywhere now. So there is no fight against it. When technology develops a way to do something that is very valuable, it goes everywhere.

55:31You can't stop it. And this is one of those techniques. So I think in the long term, people will stop caring about what AI trained on and why. Because it's just being trained on everything, everywhere, all the time. The same way web search engines crawl everything everywhere as much as they can. They index that data for you to make use of it later. Like that's where we are now. That's where we are headed. It's just more of AI ubiquity and all the different products that we use and most people just not caring. Young people already don't care. Young people who are coming into the workforce now or in college now, they've been using ChatGPT for a few years now.

56:11And that's just what they have. It's just available to them. That's their technology world. You can be a part of it and you can embrace it now or you can wait for it to be forced upon you by all of your platform vendors over the next few years. but I think embracing it now is probably the way to go. So I think this is largely a temporary objection of it even mattering at any meaningful scale and in the long term it won't. And again, if you are on the other side of this you're probably really mad at me and I don't disagree with you about why you should be mad at me but I think this is the reality of where this is going.

56:44There's only so much we can fight it. And I think once you are using AI tools even if you didn't mean to But hey, you know what? Oh, there's now there's there's new button in Google apps or whatever. Or like, even if you are if you are using AI tools, I do think it then becomes a little bit problematic to say, well, I'm going to happily take the value of this tool, but I'm not going to allow it to train on anything that I make. that's a little bit of a tricky place to be um so i just think about that like again i'm not i don't think i'm gonna change anybody's mind with any of this but but i think just this is what's happening like this wave is crashing over all of this stuff on all of our tech world it's only going to keep increasing because again there is a lot of value being created and a lot of utility being created and whether you use it or not everyone else is using it so i don't think it's a big deal i don't think it's a tricky position to be in to say i'll use this thing but i don't want it to train on my data that's there's plenty of other analogous situations in the world where you will accept a service made by somebody else but you don't want to contribute to that service for whatever reason i feel like that should still be your option which is i don't want to contribute to this thing so don't contribute to the thing but you could still use it because it's filled with other people who presumably consented to their contributions let's say in the in the most ideal scenario so i don't i don't think that's a big deal at all but that well as your other points I think we will touch on them as we go down the line of other items in this very long topic.

58:12I'm stunned. And I think I mostly agree with you guys. But I think the thing that chaps my bottom the most is that basically, hey, you've done all this free work for us, but we would like to use the things you send us to do more free labor. Okay, thanks. I mean, look, there's lots of reasons to be mad at Apple for the terrible feedback system that just at every possible – in every possible way, they dump the entire cost on us in every way. and they expect us to do all sorts of free work so that one of their bug filter people three months later can mass close thousands of bugs by saying, does it still work in the current beta?

58:53And they never even bothered running your sample project. Like that's the reality of the system. The reality of the system is it's wasting our time for lots of other reasons. This is a small slap in the face in that context. But when taken in absolute terms, do I object to this particular behavior? eh, it's not great, but there are worse things. Yep, I think I agree with that. And I also wholeheartedly agree with what John said at the beginning of the segment. There should be a way to say, yeah, I understand you're getting sensitive data, and that's fine, but I do not want to use it to feed your never-ending, never-satiated LLM.

59:31But here we are. All right, speaking of LLMs and AIs and things like that, Meta has some thoughts on the value of AI training data. This is reading from an article in Vanity Fair. This is how Meta AI staffers deemed more than 7 million books to have no, quote, economic value. Reading from the Post. Lawyers for Meta are invoking a fair use defense in a copyright suit that's been winding. I think it says winding here. Winding is a word. Okay. Today I learned. Wending its way, not winding, folks. I didn't know either. See, thank you. Okay. Wending its way through the Northern District of California legal system for nearly two years.

1:00:09The reams of confidential Meta communications newly in the public record as exhibits for the plaintiffs offers an unprecedented look at the internal maneuverings behind the company's decision to train its model on a database containing more than 7 million pirated books. Meta's attorneys argue that under precedent, it does not matter whether Meta downloaded datasets containing pirated books from a third party who lacked authorization to distribute them or borrowed used books from the library and scanned them by hand to achieve the same result. but their defense also hinges on the argument that the individual books themselves are essentially worthless one expert witness for meta describes the influence of a single book in llm pertaining quote adjusted its performance by less than 0.06 percent on industry standard benchmarks a meaningless change no different from noise that first of all that actually sounds kind of high it's like okay so you're also gonna say that it might have left some zeros off of that number so you only need a few hundred books to train your entire model to be amazing?

1:01:04Well, maybe it's kind of the type of thing of like, you know, you change one bit in something and like the MD5 changes radically, right? So maybe even just omitting one book can change it by 0.06%. Continuing from Vanity Fair, furthermore, Meta says that while the company has invested hundreds of millions of dollars in LLM development, they see no market in paying authors to license their books because, quote, for there to be a market, there must be something of value to exchange. but none of the plaintiff's works has economic value individually as training data, end quote. An argument essential to fair use, but that also sounds like a scaled-up version of a scenario in which the New York Philharmonic Board argues against paying individual members of the orchestra because the organization spent a lot of money on the upkeep of the David Geffen Hall.

1:01:45And also, a solo bassoon cannot play every part in the Rite of Spring. I love this quote that I bolded it here in the notes of saying, you know, in order for there to be a market, there must be something of value to exchange, but none of the plaintiff's works has economic value individually as training data they're basically saying all the books in the world are worth billions of dollars to our company but your one book is worth zero so if we add up billions of zeros we get billions of dollars of value it's an interesting math trick to say well our product is worthless without your data but your data is also worthless but if we add up all the worthless data in the world it equals billions and billions of dollars worth of value so good luck to their lawyers about this like i get what they're saying they're like well what is one book like what they're basically making the argument is like look if we had to pay for all these books it would cost us so much money there's lots of other articles you'll find here uh similar things so basically saying if we had to pay that for the data that we train ais on it would bankrupt us and you couldn't have ai so obviously we can't do that if i had to pay for all the ferraris in the world i couldn't have all these ferraris so obviously we can't do that you should just give them to me for free um but anyway they're literally making the argument in court your one book is worth zero i'm just rounding now it's worth zero dollars and it's like okay great well so if these things are worthless we'll just take all them away it's like well if you take them all away we don't have a product so we can't have that but we're just saying your one book is worth zero uh i mean it's any kind of collective bargaining you know any individual worker is worthless but if you unionize suddenly the workers on mass have value and now you have to negotiate with them so i mean people We'll make all sorts of arguments in court cases.

1:03:19This one I don't think has wrapped up yet, although this is an older article. But yeah, this is one way to go, which is to say we need your stuff, but we don't want to pay for it because we had to pay for it to be expensive and your stuff is worthless anyway. What a crazy argument. Like despite what I just said about defending the value of AI, like, I mean, Meta's a terrible company. They just do ridiculously unethical, terrible things all the time. And this is a terrible argument. This is like saying, we don't think that i mean this is it's basically the napster argument it's like we think you are overpriced therefore we will take it for free and that is reasonable it's great that you brought that up because that is the first thing that sprung to mind when you were saying before how it's like well there's no fighting this because it's just going to happen i feel like it's very much like the napster days where it's like suddenly we have the technology to copy music everywhere for free and it's like look there's no fighting that you can't stop this technology everyone's going to be able to do it and i think all these court cases as terrible as they may be as a way to get to this all these court cases and thoughts around it it gets back to what i said very early on on this topic months ago we need to come up with some kind of economic system that incorporates the new technology like acknowledges its existence and it's going to be there but nevertheless gives us some kind of path forward and the adamant of napster and mp3s or whatever essentially decimated the recording industry as it existed.

1:04:46But today, people still pay money for music. Spotify is terrible. Record labels are terrible. They've artists get paid less, they have to go on tour and sell teachers, blah, blah, blah. But what didn't happen is now all music is free forever for everybody. That's not a thing that happened. It is more free than it used to be. But still, we have come up with models, business models and markets, however imperfect and however impoverished compared to the old ones that have lots of downsides versus the old ones, but also some upsides that we're trying to find a way forward and that happened through lawsuits and through new laws being passed and through new business models being created and you know people who like the old record system will tell you that the new one is bad but there are good aspects to it as well like i feel like we need that for ai training and we are at the very early stages we're at the metallica suing like people in their basement stage of ai training it's like you can't fight the technology there's no going back from like oh it's trivial to distribute music online it's it's oh it's still trivial to distribute music and it's always going to be but the innovation was like you know itunes what we will we will compete with piracy by being more convenient uh we will let you pay a monthly fee for essentially all the music in the world and it's like well how does that work how can artists make money like well it's better than everyone getting all the music for free so let's try to work something out and i feel like you know i'm not saying we've reached a successful end point and that the music industry should be the model for the rest of the world especially since the sort of gatekeepers in the record label still managed to somehow cling to power and be a thorn in our side to this day and spotify is trying to become the new version modern version of that so i'm not saying it's all roses what i am saying is that this whole thing looks familiar and i feel like what you said marco about like there is no way to fight this is true but also what we should trying to do is figure out a way forward without sort of like holding back the tide or boiling the ocean or whatever analogy you want to say it's like you can't you're not going to stop this by refusing to use ai you're not going to outlaw ai you're not going to do any of the things that they tried to do to stop napster but we should also not say it's a free for all all data is free for anybody at all times uh we have to figure out something in the middle and as this meta case emphasizes it can't be like okay we'll pay everybody a dollar for their book because that also seems like it wouldn't work because it would be a lot of dollars and even so that's not an ongoing thing so i don't know again i don't know what the answer is but that's what these cases are hopefully moving us toward is some kind of market for creative work that allows ai to exist and be useful but also allows people to still somehow make money for creating things it's it's tough because there's no clear answer and i don't know where to go from here uh related but a new topic uh the new york times first generative ai deal is with amazon a reading from the verge the new york times has struck a multi-year ai licensing deal with amazon that will bring its editorial content to a variety of amazon customer experiences the outlet announced on some recent thursday under the agreement amazon will include summaries and short excerpts of the times content and products like alexa and will also use times articles to help train its AI models.

1:08:00The deal comes over a year after the Times sued Microsoft and OpenAI for copyright infringement, accusing the companies of copying and using millions of its articles to train their AI models, while depriving the publication of subscription, licensing, advertising, and affiliate revenue. Several other outlets have sued OpenAI on similar grounds, including The Intercept, Raw Story, CBC, Radio Canada, the owner of IGN and CNET. Other publishers, like the Atlantic News Corp and the Verge parent company Vox Media, have struck AI licensing deals. In addition to using content from the Times, Amazon will also draw from the outlet's sports-focused outlet, The Athletic, and its recipe hub, New York Times Cooking.

1:08:36Amazon launched its AI-upgraded Alexa Plus in early access earlier this year and claims hundreds of thousands of customers have tried the assistant. The financial terms of the deal were not disclosed. So this has been going on for a while. This is like rallying off a list of people who have cut deals. This is New York Times is cutting a deal. This is, it makes sense that this is happening. and in some ways it certainly makes sense for the ai companies they're like we really want that new york times data we have a lot of money can we just give you a deal that like we give you this amount of money or whatever the financial terms are and then you let us train on your on your data because we're doing it anyway and you're trying to sue us and let's let's work something out right but this is not i think a system that we it's not a a good system because what it means is big rich companies they're going to afford to sue open ai like disney and the new york times and like, you know, the Atlantic and News Corp and Vox Media, they will strike deals where they get paid to allow, you know, they'll be paid by everyone who wants to train on their data.

1:09:33They'll strike a deal, whatever those deals are, they'll figure out the terms like this makes sense to us. You can train our data and in exchange, we get some part of the riches that you're getting from our content. But the only people who can do that are these big companies with lots of money and lawyers who can strike these deals. And it's basically what that basically means is if you're a peon, all these AI companies get your content for free and you have no recourse. If you're a big company you get paid and that doesn't seem to be to be a great system at least in the napster days everyone's music could be pirated you know like it wasn't just some people who got paid for it and it does have some precedence of like the record company staying around and still taking all the money and a few streaming companies but you know dominating in the current era but i don't like the shape of this but on the other hand i do like the idea of essentially setting a precedent that if you want to train on our content then pay us something and presumably the deal is not like you pay us a dollar per word like it's obviously a deal that they can do it's i'm sure it's like spotify it's like well if you have to pay you know 250 000 for every infringement or whatever the hell it is for like each song uh what what would it look like for a service where you can get access to all the songs in the world that must be like prohibitively expensive it's like no it's a couple bucks a month and you can get access to all the songs in the world it's like how does that make any sense that's like a fraction of a fraction of a fraction of a fraction of a cent for each song how does that make any sense it's like well it's not great but it's still better than zero and people still can get some money from it so like we're still working that out but i don't like the idea of like if you are a big company you get paid and everyone else doesn't yeah yeah again there's no clear good answers but it'll work itself out one or the other uh reading from engadget this time it turns out you can train ai models without copyrighted material it's just a pain in the tuchus uh so reading from the article ai companies claim their tools couldn't exist without training on copyrighted material it turns out they could it's just really hard to prove it ai researchers trained a new model that's less powerful but much more ethical that's because the llm's data set uses only public domain and openly licensed material the paper via the washington post was a collaboration between 14 different institutions the authors represent various universities and non-profits the group built an eight terabyte ethically sourced data set.

1:11:45Among the data was a set of 130 ,000 books in the Library of Congress. After inputting the material, they trained a 7 billion parameter large language model, or LLM, on that data. The results? It performed about as well as meta's similarly sized LLAMA-27B from 2023. The team didn't publish benchmarks comparing its results to today's top models. Performance comparable to a two-year-old model wasn't the only downside. The process of putting it all together was also a grind. Much of the data couldn't be read by machines, so humans had to sift through it. Quote, we use automated tools, but all our stuff was manually annotated.

1:12:16And at the end of the day, at the end of the day and checked by people, quote, co-author Stella Biderman told the Washington Post, and that's just really hard. Figuring out the legal details also made the process hard. The team had to determine which license applied to each website they scanned. It's a lot of work. Yeah. So, I mean, this is about, you know, text training on books and things that are probably available. I think Adobe has already shown that you can make a useful, viable product because uh without training on essentially stolen data because according to adobe all of their photoshop ai image generation features are trained on content that adobe already owns or licensed uh that is a very conservative safe way to do this pay for or already own the content you want to train on and then train on it and if it's content you already own fine you've already paid for it if it's content you want to pay for and someone will license it to you knowing that you're going to train your model on it fine like that you can make useful things out of that the the place where it gets like untenable is like okay but that's great for a thing that makes images in photoshop or some tiny model that with seven billion parameters that it runs on your phone or something but what about the big models they need all the knowledge in the world what are you going to do pay everybody in the world for their stuff you can't even figure out who to pay or how much to pay it's like it doesn't make any sense you either you either can have a model like the the biggest chat gpt for or whatever or you can pay people but you can't do both um but it is worth pointing out that this doesn't mean that okay well anybody who's ever training a model there's no way you can do without stealing everything you can you totally can yes it is more hard but also it is essentially insurance against like depending on which way court cases go like adobe's butt is covered like we're training on stuff that we own or license for this purpose what everyone is getting you know these are all legal agreements they're all sound there's nothing weird about it licensing data is a thing that happens all the time as long as you put the terms in the license that's we're going to train like they're fine no matter what it's a great hedge for them to say we're we're putting out ai products we're telling you photoshop can do image generation photoshop can do this ai stuff and we don't have to worry whatever happens in all the laws unless they try to like outlaw ai which i don't think anybody's going to successfully do at this point um but just putting it out there and more people trying to do it it is possible to do this if you are very careful and yes it is a pain but there are dividends for you doing this there are dividends for you and there are dividends for your customers because it's just they don't have to worry about suddenly your product becoming like useless because all of a sudden it becomes it turns out it's violating tons of laws that just got passed or something finally uh for tonight and this was breaking as of just yesterday i believe anthropic wins a major fair use victory for ai but it's still in trouble for stealing books reading from the verge a federal judge has sided with anthropic in an ai copyright case ruling that training and only training its ai models on legally purchased books without author's permission is fair use.

1:15:04It's a first-of-its-kind ruling in favor of the AI industry, but it's importantly limited specifically to physical books Anthropic purchased and digitized. Judge William Alsip of the Northern District of California also says in his decision that the company must face a separate trial for pirating millions of books from the internet. The decision also does not address whether the outputs of an AI model infringe copyrights, which is at issue in other related cases. So a minor victory for Anthropic here, but I feel like not ruling on the outputs like I'm not a lawyer but my brief understanding from the story that just broke yesterday is what they're saying is like okay it's it's fair use to essentially you know if you pay for a book and you scan it you're it's like taking it and like making an archive backup copy and compressing it into a model and yada yada yada but it says nothing about the quote-unquote outputs which I think what they're saying is okay you can do that and it's fine but we're not saying anything about you charging someone$20 a month to talk to your model that's an output so now all of a sudden you have a product that you're selling whose value is derived from the books that you scan so yeah you can train it and you can talk to it yourself just like you can like the whole sort of quote-unquote backup thing that people use like pirate roms or whatever if you bought a game you have a legal right to dump the rom for your own personal quote unquote backup purpose just because you own the rom you bought it you paid for it you can make a digital copy of it stick it in your attic whatever like it's you can do that what you can't do is probably according to the this type of analogy is uh you know uh start a store where you uh rent out time on people paying you money so they can play your game or whatever unless you have a you know like it's like it's like buying a vhs i'm trying to think of an analogy that's relevant to modern people if you bought a vhs tape you could watch it at your house but blockbuster did not pay the same price as you for that same vhs tape because they are going to run a business on it by renting it out again and again and they had a very different price than you for that copy of diehard on vhs anyway um not addressing the outputs it's like well how useful is training a model if you're not going to charge people money to use it i guess and even giving it away for free it's like what if i did this i i bought your book i scanned it and now i'm letting the entire world talk to the lm that i trained and oops there goes passages from your book in the output um for whatever reason the lawyers in this case like didn't contest that angle didn't talk about it at all like they were not trying to you know they basically conceded that like the training of it is all they're concerned about and so yeah they won that part you were allowed to buy a book uh and scan it and train a model on it but then it's like okay but they're not just training a model on it and putting it in a drawer somewhere something happens after that that involves lots of money and people using it and that was not part of this case so stay tuned but i feel like this is at least one legal decision on the side of the ai people which is like that one little piece because you paid for the book like you literally bought the book they bought copies of these individual books and then they cut the pages out of them and ran into a digital scanner because you did that yeah it's your book you can train your ai model on it it's just like you know again buying a computer thing and taking it apart and dumping the roms it's your thing you can do whatever you want with it tune in later to see whatever the all the court cases that talk about, okay, but can I charge people$20 a month to talk to the thing that I just made without paying the authors anything?

1:18:20Unclear. All right. Do you want to do a little Ask ATP? Yeah, I haven't done it in ages. Yes. All right. Samuel Pollay writes, I noticed that Marco's car is one of the handful of models that works with Apple CarKey. Do he or his wife use the feature or still carry around fobs for convenience sake? Are there any frustrations with BMW's implementation of CarKey that have caused him to stop using it? My partner and I are considering a new vehicle, and the Kia Niro we're considering also supports CarKey. So I'm curious if there's any noteworthy limitations we should be aware of. I do use CarKey. It is one of the reasons why I picked the iX, actually, because BMW is seemingly a pretty close partner whenever Apple deploys new stuff, with the notable exception of CarPlay Ultra, which I don't know if they're going to ever do that.

1:19:02Allegedly not. That broke earlier today, actually. I don't know if you saw that. Oh, I didn't see that, no. But for the most part, whenever Apple deploys some other kind of technology around cars, BMW is usually one of the first and often one of the best to implement it. So that's one of the reasons I went back to that direction. So car key works great. I use it most of the time. I do have – so most modern cars that support this kind of thing, usually you also have some kind of key card. It's like an RFID card. I keep that card in my wallet just in case I ever need it. I think I had to use it once and I've had my car for almost a year.

1:19:41The only time I really like I will occasionally have like a glitch where like the Bluetooth on the phone won't actually unlock the door for like a couple of seconds as I'm sitting there holding the handle waiting for it to open. This only got me like kind of wet in the rain more than it had to be once again in a year. That's really not bad. um so and the convenience of just having my phone be the key and not having some giant key fob in my pocket all the time which sucks like that's very very nice um if i'm going like on a trip like so whenever i have like my backpack with me like my main backpack that i bring everywhere i do keep one of the key fobs in the backpack so just in case i'm like out in the middle of nowhere somewhere and for some reason it's not working i can use the key fob it'll be fine but I have not needed to yet in a year.

1:20:33And it seems like whatever the downside or slowness of it sometimes is, it sure seems like it's Apple's bug, not BMW's. Everything with iPhones and cars and Bluetooth works really well like 97 % of the time but then there's that 3 % that it doesn't. And because I've seen this across multiple cars from multiple brands, it seems like that's an Apple thing, not a BMW thing. Um, so, you know, there is, there is that, you know, occasional unreliability. Uh, also when I get into the car and start driving, it has wireless CarPlay. Sometimes the CarPlay doesn't connect for like the first 20 or 30 seconds of the drive.

1:21:16Again, not frequently, but sometimes. And again, that feels like an Apple bug, not a BMW bug. Um, but for the most part, the rest of the integration really is, is quite solid. and I use it again. That is the main way I use my car is I walk near it with my phone in my pocket and it opens up and I get in and I sit down and I just start driving. It's great. And at this point, I don't think I would consider a car that didn't have that functionality. Now imagine this, but having a always updated central infotainment system, which you do have now, as you said, this is why people are such nut jobs, including me, hello, about CarPlay.

1:21:54Like, I don't think I really need my phone to unlock my car, although maybe I'm saying that only because I'm ignorant and I don't realize how amazing it is. But I tell you, I am not going to buy a car without CarPlay. I'm just not. It's not going to happen. This is an interesting confluence of technology development related to getting into your car with the sort of the rise of the smartphone coinciding with car manufacturers moving to, you know, having a primitive keyless entry and moving up the ladder. So very old sort of keyless entry, you know, key fob thing that you can use as radio to unlock your car if you press a button and then eventually uses radio to unlock your car when you get close to it.

1:22:33In the beginning, those use tech that would like go off the like the what is it? The I forgot the number thing, but the battery that's in an AirTag. What is that one? Twenty thirty two. Yeah. Anyway, like like a little cell battery and it would last for like years and years. Like eventually your thing would die if you own your car for a long time. All right. um and those you know they were cheap to make they weren't particularly technically sophisticated they did the job a lot of them were not particularly secure but you know whatever um and then smartphones started rising but then the car manufacturers like well this technology exists now like these new standards bluetooth rf id my car is like a little computer all that stuff we were doing with those little coin batteries and those primitive like rf transmitter receivers our key fob should become a little miniature computer now instead and so what happened is key fobs started to become gigantic because i don't know like there's a little computer in there that has bluetooth and other like i don't i don't understand why they got so big i assume it's because they stopped being like a little radio shack radio frequency things and started being tiny little computers and they just got so big that it would lead you to say well the second i don't have to carry this rock in my pocket i can just use my phone which i'm carrying anyway i'm just going to throw this thing away and not use it and like key fob inflation like it goes right alongside like the size of cars in america some of the key fobs i see in these car reviews i'm like that can't be real like that cannot be the size of the key fob like they're they're the size of a submarine sandwich like who who can even fit that in their pocket what is in there like uh and then there are of course they're ugly ones the ionic series has a horrendous uh key fob that's also kind of big but anyway this is leading people i think to it will eventually like once this sort trickles down from the fancy BMWs to other cars as soon as regular people can just use their phone as their car key assuming this works in any way reliably it will be such a relief not to bring those giant key fobs along with you in the same way that we had that brief period where there's a relief not to have to stick a metal key into a thing the one they used to have the ones where the keys like closed down like a switchblade so if you needed it was there or whatever then they're like yeah just abandon it there's no actual physical key it's just the fob but then the fobs got giant so now car makers are essentially driving people towards car key and other solutions if those things become reliable i'm glad the bmw one is reliable but i'm not confident that every car brand will be as reliable but all i know is i don't want my key fobs to get any bigger and that's been happening because i don't yet have any fancy cars but even just in the the series of hundas i buy the key fobs keep getting bigger i'm like easy guys back off my pockets are only so big i mean and for the for the record i first of all i think my favorite key fob of any car I ever had was actually the Tesla Model S key fob.

1:25:14It was basically a little tiny black Tesla. It was shaped like the Model S, and you'd click the front for the frunk to open the front, click the back to open the back, and click the middle, the roof of it, like a little matchbox car to lock and unlock. And it was small and sleek, and because it was curved like the car, it actually didn't make a huge bulge in your pocket. It was nice. but also for the record like the the car key situation uh rivian i don't know if they support the actual car i think they might i think they i think their version 2 supports car key better my r1s was like the the first version and it had it used bluetooth in some different way to basically be a car key and that also worked very reliably the vast majority of the time the actual apple car key does work better um it's more reliable more of the time than the Rivian one was.

1:26:06But the Rivian one was reliable enough that that's what I used, again, the vast majority of the time. I never used the key fob in that car. And I think, if I recall correctly, I think even my Land Rover Defender before that also had some kind of app Bluetooth key thing. And I think that also worked just fine. But again, the car key in the BMW has been the best so far of all of them. I also wanted to briefly address, listener Steve Stutz wrote in earlier yesterday to ask just generally how I'm liking the BMW so far. And when I was looking for a replacement for the Rivian, I'd mentioned that I didn't want to go with a Lucid or a Fisker or some other EV startup.

1:26:43I wanted to go with a big established company because I was tired of it. And Steve wanted to know why, basically, and how that's working out for me. So, so far, I love the iX. It's great. It is such a terrible looking car. with respect can confirm yeah it's it is not attractive but it is really nice to like once you are once you're no longer outside of it and you once you get inside of it it's a very nice vehicle um and so and what i what i was tired of with the ev startups was with tesla i was tired of i mean first of all the politics of the company pretty significantly changed over the time but but that was actually mostly after i was already off of it but um with tesla i was mostly irritated by the software constantly redesigning and moving controls in weird places and that was kind of annoying but overall it was a great ownership experience that really was not that big of a problem i mainly moved off of tesla's because i needed something off-road when i was living on the beach um and of course now i wouldn't go back because the politics but if not for the politics I might go back because their cars really were very nice to own.

1:27:57It is a different beast now that they've scaled a lot of stuff up, and I think quality has suffered a little bit from what I've heard, but I don't follow it that closely. Rivian, that was very much a version one car for a version one car company, and the service situation was atrocious. That's really what got me. It needed a bunch of service on minor things because it was a version one car, and the service situation was just terrible. Like having to drive it very far away on a very unpleasant drive to even get service and not being able to get service for literally 11 weeks when I needed it, that was quite a poor experience.

1:28:32But when the car worked, it was great, but I was so tired of dealing with bugs. BMW, it's really mature. It's a very low drama ownership experience for a very nice car. And the car itself, the iX is the nicest EV I've ever driven in so many ways. most notably what you notice in that car is that is a luxury car when you are inside of it it feels like a luxury car it sounds like a luxury car like it's so it's it is the smoothest nicest ride i have ever felt in any car like it is so smooth you feel like you're just driving a cloud down the street and it's so quiet inside because they like they use like sound dampening well compared to Rivian and Tesla were both horrendous at that.

1:29:22It's a quiet, smooth, nice ride. It has infinite speed. I didn't even get the super fast model, but it's very, very fast. The range is great. It's way better than they advertise in practice. It's been amazing. So I'm very happy with it. And if I ever do need service, which so far in a year I have not, because more than I can say about Rivian, but if I ever do need service, the service like tiff has a bmw she got service recently she she caught up like you know the day before and got her car right in like it's so easy uh because they're a big brand with an established dealer network and established supply chain and their service people know what they're doing and been doing it for years and so it's just a much nicer easier experience buying the car was easier i drove to a dealership test drove one and started a lease they they have their paperwork game together you don't have to do weird games with like oh they forgot to pay the sales tax or they They forget to cancel the lease, Tesla.

1:30:16There's so many problems when you're dealing with young car companies that have nothing to do with how well the car drives. And you have to deal with none of those when you're at one of the big companies. So overall, I'm very, very happy with it. And whenever this car – I have a three-year lease. I have two years left on it. In two years, I'll look around again. But it wouldn't surprise me if I buy the exact same car again or lease the exact same car again. I'm glad it's going well. Getting back to key fobs, by the way. the winner of the key fob game in the current generation of cars is ferrari their key fob is big and ridiculous but it is beautiful and there's a special place in the car where you put it and it looks like a little piece of art of course i'm surprised you didn't say honda to be fair so point their key fobs are terrible uh david campbell writes sometimes i import photos from my big camera to my mac and sometimes to my ipad can john explain when mac os wants me to why and when mac os wants me to eject the card before removing it while ios doesn't care it It always feels wrong to just rip the SD card out of the iPad.

1:31:14Yeah, this is the eternal struggle from Mac users versus PC users. Mac users need to unmount or reject their things, and PC users just yank the thing out when the light's not on and hope for the best. It's all the same technical situation under the covers, which is that operating systems do not write the data to your removable storage immediately. Instead, they build it up in buffers and dump those buffers on a schedule that they feel like as a complicated part of their I.O. system and a cache hierarchy, and it's all to say it's like the computer is more complicated than you think so we don't have drive lights anymore you can't tell when the data has all been written even we did have drive lights oh there could be one more blink you thought it was done writing that floppy but oh but you yanked it up or you opened the drive door or whatever um so yeah mac os uh is trying to stop you from removing the disk before all the data that you think is written to it has been written to it um most of the time you can do it and you'll be fine because you're probably not doing a lot of i.o and that buffer probably did get flushed but that's why ios wants you to do it and what could go wrong if you have a fancy like journaling file system where it's or any kind of file system that's always in an inconsistent state you may be missing some data but you won't like corrupt the data structures in the file system making the entire thing unreadable but if it's a crappy old file system that doesn't have those safeguards like is often used on a cheap removable media and you yank it out not only may the data that you thought was supposed to be on the card not be there but maybe it was in the middle of updating some important data structure or directory entry which will cause tons of other data on the drive to become gone or quote-unquote unreadable and you have to use one of those undelete programs to find this scavenge the data and repair the damaged directory entry to pull your stuff back off which is why there's a million tools for recovering or quote-unquote undeleting data from sd cards and stuff so my advice is to unmount the stuff before you yank it out of your computer if possible and why does ios not do that maybe they have a different buffering system.

1:33:04Maybe there's just no way to tell iOS to unmount it. Maybe someday iPadOS will get one and will become a big boy computer. But for now, I mean, if there's no way to unmount, then you just yank it out when you feel it's safe. But on the Mac, you do have an option, so don't do that. Well, also on iOS, there are very few ways that you can write data to an SD card in iOS. It's almost always just reading it. So if it knows that, like, so, you If you are importing a picture, which is the example David gives here, importing a picture on iOS, that is a read-only operation. That is never going to write anything back to the card.

1:33:39When you mount that card on a Mac, it mounts as a removable volume. And there could be some app in the background that starts doing whatever it wants on that volume. I can tell you exactly what app is going to be in the background doing whatever to it. We all know it, too. MDS. MDS Store. Spotlight. It's going to be indexing it. And it's going to tell you, oh, sorry, I couldn't eject that because some other process is using it or because the device is busy. And if you know how to use PS or even activity monitor or LSOF, you'll find out, oh, look, it's MDS or MDS store that is now doing stuff. And it's like, stop, stop, Spotlight.

1:34:10I just want to eject my thing. And so you kill MDS and you unmount your thing. Right. But iOS is not doing that. iOS knows what SD cards are and aren't. And it has very limited uses for them. It doesn't let any app just access them freely. i think the only way to write to it from any app i think would be probably the files app if you're like manually dragging a file onto it but for the most part no apps can touch it and that's what i imagine people be doing like if you're copying things in the files app onto your sd card and you're like well i'm done and you yank it out but it wasn't really done yeah but like if it's just in there and some app is going to read the photos off of it probably apple um then they they can control like they know exactly what is going to write to it and and almost nothing is so that way they can say like all right yeah you can just pull it out like obviously if you're in the middle of a read operation you pull it out that operation will fail but it's not going to corrupt the data on the card because no rights were taking place if you're lucky that operation will fail in a sensible way instead of giving you a half corrupted image that might crash the crash they have every time you try to view it hopefully we'll just not write the image at all but error handling is tricky and that's exactly where you find like edge cases and obligations where like what does happen if i'm importing photos from this sd card and i yank it out in the middle does it just like okay all All the ones successfully written are there and all the ones not successfully written aren't there?

1:35:25Or is there a half-written thing that every time I scroll past it, crashes the Photos app? Fingers crossed. All right. Thank you to our sponsors this episode, Notion and Delete Me. And thanks to our listeners who support us directly as members. You can join us at atp.fm slash join. One of the many perks of membership is ATP Overtime, our weekly bonus topic. Every episode, we do bonus content that only members get to hear. Plus, as we mentioned earlier, we do specials and all sorts of other fun stuff. To hear Overtime, go to atp.com slash join. This week on Overtime, we're talking about the overlap between app development and game development.

1:36:01Once again, thank you so much for listening. We'll talk to you next week.

1:36:20search marco and casey wouldn't let him because 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 list m-a-r-c-o-a-r-m-e-n-t marco are meant S-I-R-A-C-U-S-A Syracuse. It's accidental. Accidental. They didn't mean to. Accidental. Accidental. Tech podcast. So long. All right. So there's been some TV-related news. I have to recap. Marco was thinking about buying some new TVs. I was giving him advice, but I said, you know, during the course of the year, it takes a while for all of the TVs to be released and so it's hard to give a recommendation maybe you should wait until everyone everyone has put their cards on the table for 2025 and so we could say finally all the 2020-25 TVs are out they've been reviewed and now I can give better assessments so I gave some tentative recommendations last time without having all the information in particular Sony's Bravia 8 II the stupidly named Bravia 8 II was not out then it had just been announced but not reviewed and I'm like well that's let's hedge our bets because maybe that'll be great maybe it'll stink but we don't know so let's wait until that's out um the all the reviews are not for this tv in particular my favorite reviewer has not yet to review this tv but enough people have reviewed this tv that i trust that i'm ready to make a call on what i think marco should do so first of all we'll put some links in the show notes to uh reviews of the bravi 8 too uh in particular there's a good one from artings they usually do a pretty good job and they continue to improve their methodology with each passing year so i continue to recommend that site um they're maybe not as in the weeds as my favorite uh reviewer uh but they do like i said they do a good job and the upshot is the bravi 8 2 fulfills this role as not being the top tv in sony's marketing lineup because its number is lower like it's not it's probably 8 2 instead of 9 or x or whatever right it is not as good as when their qd oled was the top model which in the previous naming screen scheme was the a95l and it shows it is not as carefully calibrated as the a95l was it is not as good in some respects as the a95l and if i've i believe it's a three-year-old television now sony didn't bother revising it because even in its second and third year was still one of the best tvs you can buy guess what the a95l which you can still buy if you can find it somewhere and it's real cheap because you know get them while they last is still probably a better tv than the bra v8 too if what you care about is ultimate picture fidelity and you're probably going to watch in a dark room the bra v8 too is brighter in some ways but in other ways it's not brighter than the a95l which is shocking because it uses a three-year newer panel that supposedly is brighter watch the artings video for more details so marco if you're waiting to hear the deal with the bravi a2 is yeah the answer is they it's not great like it's there's a second video from rting saying bravi a2 versus a95l which should you get and they basically say get the a95l of those two tvs the a95l is still better it's amazing this tv is like undefeatable i don't think they're making them anymore except for maybe in 77 inch so you can get them cheap like not cheap cheap but you know like what like for half the price they originally sold for you can get an A95L so cheap that I was thinking should I get an A95L because I have the predecessor of the A95L the A95K and I'm like no don't don't buy because don't buy a three-year-old TV just because I just wonder if they'll ever make a TV as good as the A95L again it's depressing but anyway for your purposes the LG G5 is brighter than the Bravia 8.2 it's brighter than the A95L it is a true 2025 5 tv the lg g5 did have a problem where they had some contouring issues with with uh hdr 10 output but they fixed that in a firmware update while we were waiting for the braviate 2 reviews lg updated its firmware and they fixed that bug so i'll put a bunch of links in show notes to reviews of the lg g5 that is my recommendation for you uh for all situations where you want an oled in the year 2025 the lg g5 it's very bright uh even in terms of its peak brightness may not be as bright as you think but like in whole like we went over this last time but in looking at the whole scene a mini led tvs we're gonna have to turn down their backlight to stop blooming the lg g5 doesn't have to so there are situations where the g5 is even brighter than a mini led tv which is usually unheard of it's great tv the lg is good with their software you're used to lg that's your answer 2025 it's the lg g5 if you want an oled television final answer yeah i mean the bribe a2 is not a bad tv the fact that it gets beat at all in anything by a three-year-old anna5l is just embarrassing for sony sony get your act together and why shouldn't i just get the anna5l it's not as bright as the g5 like you have your bright room situation like the i feel like the anna5l is still like the most accurate best tv if you're going to watch in a dark room so if you have that situation fine but for a sunny room the g5 gets so much brighter than the anna5l in so many different conditions and it's not that much worse like the lg's processing is better uh this year uh their their color volume it's not as good as the anna file but you'll never notice it like the lg g5 is the better all-around tv and it supports hgig which helps with the stupid switch to hdr brightness issue like i just think the g5 is going to fit into your life a lot better for the super duper video file i would still recommend the 895l but that's not you so if you really want to get a smaller 895L at bargain prices to put in like a bedroom or someplace that doesn't have bright sunlight if you want the absolute best in color fidelity and you're never going to game on it yeah it's still a great tv and for me if i had to choose i would pick the 895L over the g5 but for your for the scenarios that you described i think the g5 is a better fit okay john i i'm deeply impressed that was incredibly brisk well done yeah i mean when the when the verdicts are in there you go like like i said my favorite person hasn't come out his review yet it's his hdtv test i'm still waiting for vincent to come out with his review when he does if he totally contradicts our things i'll let you know but it seems like yeah the bravi 8 2 is living up to its number which is eight which is lower than nine