In this episode, Conor and Bryce chat about the highlights from the recent ISO C++ Committee meeting in St. Louis, their upcoming trip to CppNorth and more! Link to Episode 189 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Twitter ADSP: The Podcast Conor Hoekstra Bryce Adelstein Lelbach Show Notes Date Recorded: 2024-07-03 Date Released: 2024-07-05 CppNorth Reflection for C++26 C++ Senders and Receivers std::inplace_vector Contracts for C++ code::dive Conference Meeting C++ Conference Core C++ Conference Think Parallel ACCU Talk Chains: Exploration of an alternative to Sender/Receiver | Sean Parent | NYC++ Intro Song Info Miss You by Sarah Jansen https://soundcloud.com/sarahjansenmusic Creative Commons - Attribution 3.0 Unported - CC BY 3.0 Free Download / Stream: http://bit.ly/l-miss-you Music promoted by Audio Library https://youtu.be/iYYxnasvfx8
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
BryceWell, uh, two big things, I guess. Um, first is that we uh voted senders, the C asynchrony framework, into C26, and the second is that we design approved uh reflection for uh C26.
ConorWelcome to ADSP the Podcast, episode 189, recorded on July 3rd, 2024. My name is Connor, and today with my co-host Bryce, we chat about the ISO C committee meeting that just took place in St. Louis, as well as our upcoming conference trip to CPP North. Look at that view, folks. He had a nice view before, but it's even better now.
BryceYes, and and and in fact, I've been I've been uh working hard to figure out how to position my my desk in this room so that I can uh show off the beautiful view of Manhattan and also not have my screen be like super backlit. Um but uh it it there's just really no no option for not having a backlit screen here unless I like go in the closet.
ConorSo well it's it's been a while since we've chat. How did the move go? How is the new place? How's the view? The view does look fantastic, folks. You'll have to post a photo for the people, for the listeners.
BryceThe the move is going well.
ConorGoing. It's still going? We talked like a month ago, and you're telling me the.
BryceWe moved in a week before the C St. Louis committee meeting. And so we were here for like six days, and then I went to the St. Louis committee meeting. Um, and then my girlfriend went to uh Florida um to where my parents are to go to play at the tennis academy there.
ConorOoh, fancy.
BryceShe's uh I I told her that uh look, we're both very territorial people, and uh this new apartment needs to be like a shared space. And if one of us is in this new apartment for like a week right after we moved in without the other one, we're gonna make all sorts of executive decisions, and the other person's gonna come back, and there's gonna be no reversing those executive decisions. So I said, either you gotta come with me to St. Louis, I gotta not go to St. Louis, or uh you gotta go somewhere and else and do something else.
BryceAnd she picked the last option. But um couple slight issues. Two of the uh two of the AC units in the apartment uh were dead on arrival. Um but the landlord's been very good about uh getting people out to fix them. But the AC people came right before I was gonna leave for my flight. And um and This is to St.
ConorLouis, I assume.
BryceUh yeah, yeah, to St. Louis. And uh they the circuit breaker box in this apartment um was apparently not labeled correctly when they first put it in there because the uh AC people accidentally turned off the fridge.
ConorOh no.
BryceAnd we came back, and everything in the fridge was spoiled rotten. So uh we came back to no AC and also a slightly stinky apartment. Well, not slightly stinky, it was fairly stinky. And actually, we were both supposed to fly back at the same time on Sunday. I was gonna get into LaGuardia at like 2:30 on Sunday, and my girlfriend was gonna land at like 3.30. Um, but there were thunderstorms in New York and in Florida, and so she ended up being at the airport for like five or six hours on Sunday because it kept being like, oh, we can't leave because there's thunderstorm over Tampa.
BryceAnd then when it was clear over Tampa, it there would be a thunderstorm in New York, and they wouldn't be able to leave and get to New York. And then it just like flipped back and forth all day, and then eventually they were just like, all right, everybody go to sleep, we're gonna fly out the next morning. And in hindsight, it was good that this happened because if we had both arrived home at the same time and she had had to uh uh smell what I smelled, she would not have been happy, and she would have never used that fridge again. We would have just had to buy a new fridge. Uh that's that's still that is still a possibility that uh that that that that we may never get the ick out of the fridge.
BryceUm and then uh uh continuing a series of unfortunate events, um I I tried to upgrade CMake on my uh my workstation, which is in Santa Clara yesterday. And um uh Debian wanted to install a new libc, and libc wanted to run this user merge thing. And long story short, my workstation is like completely bricked. Um my uh my dynamic linker, which is a thing that every program needs to run, uh, somehow got overwritten.
BryceThe like the sim link to it got got overwritten. And then like I couldn't run anything. And I could have gotten myself out of this hole because I still had an SSH session into the machine, but uh I was not uh SSH D as root. So I could I eventually could invoke the command that would create the simlink I needed to like restore the system, but uh but I didn't have root access and I needed it to be able to create that simlink. Um and uh and so I had a a a very diligent coworker who was able to go over to the system and um uh live boot it uh with like a you know a live USB um uh and create the simlink.
BryceBut as he's live booting, he's like, you know, this this machine's running like really hot. I think your water cooler is dead. So then he goes to IG to try to find a water cooler, and they don't have they don't have a new cooler. So then eventually he goes and like cannibalizes another system, finally finds a cooler that's like gonna work for my system, gets my system back up. Um I still can't actually get into it. Uh it is like up and running now, but there seems to be some other issue. So it's just not been my week, man.
ConorThere seems to be a theme: AC, fridge, and cooling system. Yeah.
BryceOh, I didn't even think about that. Yeah.
ConorEverything to do with uh temperature is uh is broken. But what so what percentage are you moved? Are you fully moved in?
BryceUm I mean all of our stuff is here. I mean, my my room is unpacked. Uh by my room is here.
ConorIs this a webcam? Is this a webcam? Can you uh pick it up and uh rotate it around?
BryceIt is attached here. Um however I do have controls, so I can't do that. Um this is the second bedroom, which is uh where my office is. You can see it's not you know fully unpacked. Oh, and it does have it does do 360. No, maybe it doesn't go 360. It's called the Insta360, so I would have assumed it would go all the way around, but I guess it gives you 360 degrees, but does not actually rotate all the way around, which kind of makes sense. Um is this a new a new camera?
ConorI've never I've never seen you in such uh in such high fidelity slash uh you've never been able to spin your camera.
BryceThe old camera, um I could there was no way to set the contrast, and um there was no there was no positioning of the camera where I either my face would be very dark or it would be like very backlit. Um and so I needed to get something where I had contrast control. And so I got this thing. It also has like some AI and tracking. And then oh, the the most exciting thing is this building doesn't have a gym, which uh I was able to use as justification to put in this bad boy.
BryceA uh a BoFlex Max 10, Max 16, uh elliptical slash uh stair machine, um, which is very good. Far better than the ellipticals in my old building. I got quite the workout last night.
ConorAnd um That's quite nice because then if you're in a meeting, I've always I've always kind of wanted that. Like if you're in a meeting, you can just like you know do some stairs while you're right at your computer basically.
BryceIt was a big part of the motivation is that this unlocks like an extra two to three hours of meetings for me because it's hard to do meetings in the gym because it would be rude. So um I mean I still like to get out of my bike and actually like go out and ride. Um, but uh especially during the winter, this thing's gonna be clutch for for multitasking. So if I'm out of breath during a meeting, it's probably because I'm I'm on this thing.
ConorI think it's an entirely in like in 2024, it's an entirely reasonable and like it should be encouraged almost. I'm sure there's some study out there that shows exercise elevates brain activity and you know mental focus. And uh I've heard I've heard people on meetings before that they definitely sound like they're on a walk or they're doing some exercise, but like isn't that that's what Steve Jobs used to do, you know? It's like, oh, we're doing a meeting, let's go. We're out the door, we're going for a walk around Cupertino. And uh I think it's good, you know? I think we should we should be encouraging exercise, especially like so many meetings, like 90% of the participants aren't saying anything, right?
ConorSo if you've got a stationary bike or an elliptical or even like a I've I've heard of these little like walking, like un unelectric walking treadmills that are just like you start walking and then the belt starts moving. Yeah. That kind of stuff.
BryceIt is tough because I I like to take minutes in pretty much every meeting that I'm running. Um, and those are the meetings that I have the control over the schedule. So uh it is a little bit tricky, but maybe I'll try maybe I'll try like AI transcriptions or something like that. Oh, and you know what? We we should have had I just realized we should have had this recording session an hour later, because right after this, I'm going to pick up my chiming clock. It's gonna drive you crazy that is now finally repaired.
ConorI'm sure there's a future several hundred episodes where where Henry the Clock or is it Henri Henri the Clock?
BryceHenry the clock, yeah.
ConorUh he will make his uh debut. Anyways, we got uh we got some things to chat about. We we uh mentioned the uh ISO C committee meeting in um St.
BryceLouis, which Yeah, reflection, baby.
ConorYou want to give a recap on uh we'll call this episode, it'll be uh will moving get mentioned, we're not sure, but we'll call it, you know, the St. Louis uh ISO committee.
BryceI think Connor, I you know, I'm I'm such a boisterous and brash character that you you it would be a huge mistake to not include some of the little content where I'm just having a bad week, where just like I'm cursed. Like I feel like the listener deserves.
ConorThat's all staying in the episode. That's all staying in the episode. The question is, does it impact the title of the episode? That's what we were talking about there. Because we kind of it needs to be clickbaity, right? You know, it's like people care about the committee meetings and the progress of C. And then we're also gonna briefly mention in what is it? It's July 2nd, I believe, in 20, in T minus 20 days. Bad news though. Actually, maybe we'll talk about this first. Bad news. Uh CPP North is happening, and the bad news is I'm gonna miss your talk, Bryce. Um because I don't land back in Toronto from a bachelor party I'm gonna be at in uh Mexico until 2 p.m.
Conoron Monday.
BryceWait, and your talk is that you're not gonna be around for July 21st, my birthday. I don't care about you missing my talk. You missing my birthday. Well, you're gonna have to figure out how to make that up to me.
ConorDidn't we already have dinner plans or something like that? I don't think that was for well, if it was for the Sunday, uh yes, bad news.
BryceUh actually it may have been. Let me let me check.
ConorUm wait, we already have reservations?
BryceUh, pretty sure.
ConorI didn't realize that.
BryceI'd be surprised if we didn't.
ConorI would be unsurprised. I feel like I would have known about these reservations.
BryceUm I thought I made reservations at Allo. I don't see it here though.
ConorI think we we definitely talked about it because you correct me if I'm wrong, are going to Niagara Falls or Niagara on the Lakes.
BryceWe have reservations on July 23rd at Allo.
ConorSo we do? We've already been there, man. You've been there like three times. Alright, folks. Uh oh. That's true. That's true. Well, yeah. TBD, where we go. Uh point is CPP North is happening 21st to 23rd. I think, are those the dates? We should get these dates right. I assume the the 23rd is the Wednesday, correct?
BryceYeah, I believe so. Yeah, yeah. It's gotta be the Wednesday.
ConorTuesday. Yeah, 21st to 24th. Actually, no. So that means that the 23rd is the first one. Tuesday, yeah, Tuesday. The Tuesday. Uh why am I looking at August right? Because I shouldn't have done that. So yeah. It says it starts on the 21st, because I think on the 21st there is an evening social where people pick up their badges.
BryceAnd celebrate Bryce's birthday, of course.
ConorAnd yes, I guess maybe they should just rebrand that to the uh pick up your badge. Actually, it should be the Bryce birthday and pick up your badge celebration. And then the Monday to Wednesday is when the conference is happening. Oh, Bryce. And yes, the reason we bring it up is because we're both gonna be there.
ConorLast year, uh I had to replace you with Ben Dean. That's true. Um this year, you will not be replaced as you will be there. Ben Dean will be there. We'll probably talk to him. He's gotta talk. You've got to talk. Um anyways, we'll circle back to CPP North at the end. Give us the update about the uh ISO committee meeting.
BryceWell, uh, two big things, I guess. Um, first is that we uh voted uh senders, the C asynchrony framework, into C26. And the second is that we design approved uh reflection for uh C26. Um and uh both of those are pretty big features. Um I mean, I think reflection is you know uh uh a much bigger deal. Reflection's a huge game-changing feature. Um Herb said that he thinks it's gonna have the biggest impact on C since uh any of any feature since C 98.
BryceUm I I I think that that's probably fair. I think it it may compete with a couple others. Um modules impacts the language in ways outside of just the source code, right? It it it impacts how we structure code, and it's gonna still be years before we start to see the impact of modules. Um but it's it's still gonna be years? I think so, okay, because I because it takes a while for the ecosystem to adopt things. Um you could argue that maybe some some set of C11 things have had as big of an impact, like I don't know, maybe one could argue move semantics.
BryceI'm not even sure I would go that far. Um constexpr maybe. Constexper is maybe the thing because it unlocked this whole new um field of uh compile-time uh metaprogramming, but you could kind of do that before. Um reflection is is really very game-changing. It um it unlocks uh uh a variety of things that weren't previously possible, but also it gives us uh new ways to evolve the language.
BryceUm things that we previously had to build as library features we could build as uh or the previous things we previously had to build as language features, we can now build as library features.
ConorUm Spaceship one of those?
BryceYeah, could have done um Spa well spaceship had a syntactic um component, right? Like there was a new operator, but um if you had just decided to call it like a you know, just to call it like a member function, you know, just pick some some protocol name, you know. Like uh like call like a compare, um, then you could have done the rest, yeah. Um so yeah, it's a big deal. I'm I'm very confident it'll get in at this point.
BryceUm my uh our colleague, actually, Andre Alexandrescu, has been working on uh some follow-on things, um, helping to drive uh some of the uh injection part of it. How do you put things back in? How do you generate new stuff based on reflection? Um, you know, how do we generate new functions? How do we like wrap functions and members and classes? Um and how do you manipulate uh uh the things that you've reflected and and the model that we've settled upon is this thing called token sequences.
BryceUm there's a couple different approaches we explored. We explored like a spec approach where like you could define like a new function or type by you know, you'd call like define class, and then you'd pass in this struct that uh describes all the members of the class and the the types and stuff like that. And we looked into that too for functions where you define the signature and and whatnot.
BryceBut one of the properties we wanted is that we wanted when you're writing the in the injective code, when you're writing new code that's going to be generated at compile time, we wanted it to look like C code. And this spec approach doesn't really look like C ⁇ code and it gets very verbose, very fast. And there's a couple different variations of that approach. There's one that's um uh I think in sort of the same broad category, but it's more AST-based, where you're like manipulating the AST in a very similar way, where you're sort of doing it programmatically.
BryceAnd I think it had all the same downsides, but it also had the additional downside that um exposing up the AST in that way makes it harder to evolve the language. Um because, like, you know, if for example, if we ever standardized uh uh statement expressions, um then like the the AST nodes uh for expressions um uh might change in type.
BryceThey might now not be statement nodes, and and so it might limit how we'd be able to evolve the language going forward. Um and then we also looked at like a string-based manipulation approach, where the way you injected code is you just you know gave it a string and said, here, here's some string of C stuff, um, inject this into the program. Um but there were some some concerns that that's not really um not going to really be robust, that it could be brittle, that diagnostics could be not great, um uh that it would sort of not be hygienic, um and uh you know you wouldn't get any syntax highlighting on any of that code in the string, which would be a little unfortunate.
BryceUm but that was sort of kind of the run runner-up approach. And then the token sequence approach is that instead of manipulating strings, you that you'd manipulate these sequences of tokens. Um uh and um this is sort of information that pretty much all the compilers carry around. Um and so this would be a little bit cleaner. Um, and we even imagine that we could build a hygienic macro facility on top of this. So in some ways it's similar to, I think it's somewhat similar to Rust macros um and how they work, the this token sequence stuff.
BryceUm so not not clear whether we'll see token sequences in C26. It's not part of the core reflection proposal that's been approved. Um and uh it's not even clear how much of injection we'll see in C26, but we're definitely still working on it, and we've still got a good bit of time. So we're pretty hopeful that we're gonna get some of this follow-on work done. Um we also landed in-place vector at uh this meeting, which is uh you know, a vector where you can which has a fixed capacity where you can just give it a block of memory that's you know on the stack or been statically allocated, and then it can grow dynamically within that capacity.
BryceIt'll never dynamically allocate memory, but it can grow in its size within that fixed capacity. Um so that's been something we've been working on for a while. Um the the std SIMD proposals made a lot of progress. We're gonna likely see that get advanced into C26 at the next um uh meeting. Those are sort of the two big library features that I think we can expect for um 26 are gonna be uh SIM D and uh senders and receivers.
ConorI know that uh so that's I mean that's pretty exciting because uh std execution and reflection are else, obviously, both big. I know that uh I can't remember whose plan it was, but there's some C26, you know, large plan um pattern matching was on that list. Is there any update on pattern matching or no?
BryceUm not that I'm aware of. Um it yeah, I it it could well have been discussed last week. I I don't know.
ConorYeah, there's a lot of there's a lot of room.
BryceSo wait, wait there was some some time was spent talking about trivial relocation. Um and uh contracts has been another big discussion. I do not think contracts will make it into C twenty six. Um I think it's gonna probably go to uh a T S or something. There's still a lot of a lot of open questions there.
ConorThat's that's uh uh interesting update. Because I think the last thing I heard about contracts, because I think um doesn't teamur co-host of CPP Cast. Uh he's I don't think he's is he the chair or vice chair? If he's not the chair or vice chair, he's one of the uh regular attendees of that working group. And I think he's been pretty hopeful on the episodes of CPP Cast when he mentions it that it'll uh get in for C26.
BryceBut uh I I I I just think it's unlikely because there's still a lot of uncertainty about the design and it would pretty much need to make it in at the next meeting or maybe the meeting after that, but really it would need to it'd need to be settled by the next meeting. Um Where's the next meeting? It's in Poland in Wrocław.
ConorWho's hosting that?
BryceUh uh Nokia, I believe. And my uh my Polish girlfriend and my mother will both be uh be there.
ConorThat's exciting. When is it? That's uh November. September. November. Why so far? I thought there were three meetings in there.
BryceThere are there are, but it's always um it's always um like February, March, uh, early summer, and then November.
ConorAnd is it in Vrotslav or is it in Warsaw? No, it's in Wroclav. Okay. So that is it gonna be coinciding with the codedive conference then?
BryceUm it is. The coded conference is like the the week after in in Warsaw. I don't think I'll actually be able to go to Codedive this year, unfortunately, because I'm going to Core C. Um, and so we'll be uh we'll be flying from Warsaw, I think, on the the 25th, because Core C starts on the 26th, and um uh codedive is on the 25th, so I'm not certain if it's gonna work out.
ConorAre you going to meeting C as well? The one in Berlin?
BryceI'm going to meet in C and I'm doing a little bit of a tour of uh German uh C user groups and HBC centers and people who like GPUs.
ConorNice.
BryceWell what uh what's the talks that you're gonna be given in November? It's gonna be my same talk, the ThinkParol talk. Um I'm probably going to give a slightly updated version of it that focuses on some different algorithms than those the scan algorithms that have been in the versions that I've given thus far. So I may I may talk about some reduction algorithms.
ConorOkay. And yeah, and then I guess that full circle, full circle, half a circle, who knows what portion of a circle brings us back to CPP North. Uh which are you giving us?
BryceI'm giving my think parallel scans talk.
ConorOkay, because I was gonna say I've seen the ACCU talks start to come out online. I have not seen yours yet because I've been waiting to update the link from whatever episode we included that in the show notes, and I guess now I have to include it again in this show notes. Uh but it'll be uh an unlinked link because the the recording, at least as of yesterday or two days ago, um, is not out yet. Um but you'll be giving that same talk.
BryceSo I guess uh it's good news for I know they're I know they're working on it because they asked me if they could give it a longer title because Think Parallel is not a very good title for SEO. And so they wanted they wanted the YouTube video to have a longer title.
ConorYeah, that's an interesting uh what are your thoughts on that?
BryceBecause I think my CPP North talk from last year, its title was Composition Intuition, but they uh ended up renaming it at least in the YouTube video, like uh function composition in I I I told them I they asked whether they could rename it, and I told them that I would prefer not to because I dislike um talks that have a subtitle in the name. Um like they wanted to do like think parallel call and like something, something, something.
BryceAnd I just am never a fan of that. Um and I think I'm happy to have you know some keywords put in the description and uh you know peep people find it if they're gonna find it. You know, I I trust the YouTube algorithm is what I'm saying.
ConorYou know. Yeah, I will say, I will say, oh my god, I well, a little tangent here. Just a small tangent, folks. I love the YouTube recommendation algorithm. Whatever, folks, if you're a listener or you're friends of someone that works on the YouTube algorithm, give them a handshake next time. Literally, with like I think it was 37 views, the Sean Parent meetup talk from New York C, which I actually don't know. Maybe I am subscribed and that's why, but it was it was not in my subscription pane, it was on my homepage with the Chains talk that we saw back in March.
ConorIt had 37 views, and YouTube recommended that talk to me. Like, let's go. It's the most esoteric thing.
BryceDo you think that YouTube's recommendation algorithm does content analysis? Do you think that they recommend solely based on title and description?
ConorOr do you think that they No, they do it because the reason I would say that it definitely does content analysis is because when you do a Google search these days, sometimes it'll give you the timestamp within a vi a YouTube video. It'll say like minute thir 337 discusses like if you say how to, you know, do something, it'll give you a YouTube video and then YouTube videos you can now include chapters in the YouTube video.
BryceSo you can say like minute zero to minute one is the intro, minute one to minute three is you know some other topic. I don't know if you've ever seen that on a YouTube video, but it might be that.
ConorBut um but I wonder whether it could just be the keywords too as well that you attach to YouTube videos, right?
BryceUm Yeah, but that wouldn't give you the timestamp. But I I I I would bet that if they're not doing content analysis for recommendations now, they will they will be sometime in the in the near future. Um you I mean, because they're already doing AI transcription, right?
ConorYes, yes. They've been doing transcripts for a while.
BryceSo they they have So they they have the Chow Screen transcript. So that's that's just text. That's pretty low cost for them to use in ranking and recommendations, right? Like if if they're if they're already doing the thing that that generates you know text from the video, which is the expensive operation, I would assume, and they're already doing it because there's already a use case for that transcript, then they've already got that data, they may as well use it. Um they're almost certainly already doing um like things like sentiment analysis to see like to analyze like whether somebody's really angry or there's like violence in a video because they have to do that for content moderation, right?
BryceLike if someone like uploads a video of like, you know, like crime or like like stuff like that or like violence, they they need to be able to detect that um so that they can either moderate it away or that they can put a you know graphic violence warning on it. Um uh, you know, that's that's probably AI. I mean it maybe maybe there's somebody who just watches all the videos. I assume some AI is involved. Um and then they also they also have to do content analysis for um uh you know copyright infringement.
BryceUm uh so like they're already doing a lot of stuff. I I'd be I'd be shocked if they're not feeding um uh like data from from what you're actually talking about into your um into the uh the recommendation algorithm. So so maybe the talk title doesn't really matter.
ConorNo, I I definitely think it does. I mean, talk title and thumbnail, I think, are the two most significant things. I think for a conference talk it doesn't matter as much, but I think for like some viral video or just some video where you're trying to get into the video. But like I don't I don't think I need I don't think I need to put the word C into the title. I mean, as a talk title, you definitely don't. For optimizing on YouTube, I actually don't know.
BryceUm let me put it another would it help? Almost certainly it would help. Do I think that YouTube, that YouTube's algorithm knows that my talk is about C even without me putting a keyword into it? Yeah, I think it does. I think it's got some smarts there. I think it's done some analysis on the transcripts, I think it's done some rankings, and I think it knows this is a thing about C. And if somebody searches for C, you know, parallel programming that my talk will come up.
ConorYeah, like if yeah, exactly. Like I I'm sure putting code report in my talk title instead of my actual name, because that's what I'm known by on YouTube, that would help, but I'm not gonna do that. Although actually I've seen that um uh one conference actually did. They put my name and then they put my handle next to it. But anyways, the name of the talk, they switched composition intuition to function composition in programming languages. Was that a good decision? I mean, it is one of the top five talks on the YouTube channel, the CPV North YouTube channel. So I don't know, maybe they do know what they're doing.
ConorWho knows? Um I from a uh what's the word, like a a purity point of view, you know, uh it upsets me that like it's not the name of the talk, you know. It's the name of the talk is composition intuition, and it matches my other talk, algorithm intuition. So to change the title in the YouTube thing, uh, you know.
BryceSo I I want to dig into this a little bit more. I wanna I want to think about um I want to think about what are the expensive compute workloads in processing a YouTube video. So generating the transcript isn't that bad. And I and I say that because um audio data is substantially smaller than video data, right? Um you you've got this audio data, you're um all I know about audio processing is there's a lot of Fourier transforms.
BryceSo I assume that the way that they generate the transcript is if they take the audio data, something something Fourier and uh some AI and neural nets these days, and then it it produces words out of it. Um things that are gonna be more expensive, I would imagine, are doing things like OCR, like like recognizing like text that's on the video content, and also doing things like um uh object you know detection or like I I I don't know whether they have a thing that searches for like copyrighted logos or or or stuff like that, you know, like copyrighted images or things of that nature.
BryceUm but uh I would presume that there's something like that. But like if you have like a you know a 30-minute YouTube video at like 4K and 60 FPS and you want to search through like every frame of that video to see if there's any of a wide range of things that may be in that video, you know, a logo or you know, uh some graphic content.
BryceUm like that's that's gotta be pretty expensive, right? You know?
ConorYeah, I have no idea, but I would imagine that they have some kind of like similar to like uh what do they call it, like a Monte Carlo tree search, if you're familiar with the MCTS, where it's like they choose that was like uh the algorithm underpinning Alpha Go and Alpha Zero. Actually, I'm not sure about Alpha Zero, but definitely Alpha Go, is that instead of exploring all possible paths, they would do a couple iterations, figure out the best options, and then continue to explore those. So every single few iterations they would prune for the top ones and then keep going.
ConorSo I imagine they have some similar like filtering mechanism where like they've got some you know relatively inexpensive way of checking, you know, does this pass all of these things? And then it it narrows it down to the 5% of videos that look suspect, and then even the step they do with the 5% of frames, right?
BryceLike, you know, they could take that 4K video, they could downsample it to like a very low resolution, and then they could do an initial pass to see like, you know, okay, which which parts of this do we think are uh sketchy and maybe need to to you know to be analyzed more.
ConorYeah. Yeah. I'm I imagine it's something like that. I would also assume that like they profile uh channels, especially ones with a lot of subscribers, and those ones immediately they get hit with the full. So if you've if you've got a million plus or you know, even half a million plus subscribers, uh especially if you have a history of posting videos with questionable content, you immediately get hit with you know all of the uh all the checking versus like the almost guaranteed the hardest things to check for are these new YouTube channels that are brand new and don't have any like you know for for my channel, I'm sure that they you know they don't actually check a lot of they they do a couple like oh verification, verification, but like all my stuff is just programming content.
ConorSo as long as they can verify that it's oh it's another programming content video, you know, it goes down some separate tier and gets a green check mark, whereas some other bigger YouTubers that are covering like news or something like that and potentially are gonna be violating some um you know code of conduct thing or it's gonna get put some label on it, those immediately are gonna be, you know, put through some other pipeline. I imagine there's a bunch of pipelines and it's not just like, oh, we treat every video equal. Like that that seems like it would be a recipe for false positives or yeah, yeah, no, no, no.
BryceI mean, but but but what what you're saying makes sense. And and it makes sense that they would use the the history of a channel. And I what you just said about like categorization, I bet that that plays into it too. I bet one of the first things that they do is that they have some relatively cheap pass that um that categorizes what kind of video it is. Um and like if it's if it's a video about you know news or something, um, or like current events, that maybe they run some different set of filters.
BryceOr like if it's if it's about politics, you know, they run some other set of filters.
ConorYeah. Already you have to well you don't have to, but m it helps if you do. And actually you might have to, because I I actually don't know. Uh you have to s you there's a combo box for specifying what category of video it is. So all of mine are like education or technology or something like that. But they definitely have like new sports. I imagine sports is like a pretty low one, right? Like TSN has a YouTube channel or ESPN has a YouTube channel, they're always uploading their you know, sports highlights. Uh they probably get that stuff green checkmarked immediately, and like the biggest problem is copycat channels that are illegally posting copyrighted uh videos.
BryceYeah, I don't know. I I I think sports actually um is probably the more intense ones because like all of the because uh sports broadcasting is like pretty strict on who gets to broadcast what. Like I every every tennis grand slam my girlfriend and I go through the same song and dance of figuring out like where can she actually watch you know Wimbledon or the French Open. Uh and it's because you know they they're not just like available to to stream for free. Um there's like a ton of people who stream like live commentary of them, but don't show the the video.
BryceLike I'm I'm I'm sure that sports is is actually uh uh uh a tricky one. Um I bet I bet like the the ones that are um I don't I don't know what like the easy ones would be. There's there's like a lot there's a lot of things to think about there uh because there's like so much there's you know it's not just like recommendation that they gotta do, but they also have to tackle this whole like content moderation space, and like that gets just real complicated. Um Yeah.
ConorEspecially when you consider the I can't remember if it's billions of hours of content that get uploaded every hour, every day, or something like that. Yeah. It's some it's some absurd number that's like unfathomable. And they're like, oh, how did your moderation team not catch this? And they're like, moderation team? Like we have a moderation team, but they they manually look at a fraction of a fraction of the of the content.
BryceUh I didn't even think about this, but but but like language has also got to be a uh a fairly big thing too, right? Like you need to I don't know. Is it is that when you submit a video, do you say what language it's in? I don't think you do.
ConorNo, but it I think it auto-detects that stuff like that. Yeah, yeah, it's gotta be pretty easy. I think that part is pretty like you're not allowed to swear in the first 10 seconds or something, but like that I imagine is not actually a hard problem these days. Really noneows to swear in the first 10 seconds. I I I I think you are allowed to, but if you do, you get like some kind of not safe for something label. Um I don't I don't know because I don't swear in my YouTube videos, but I think uh there's certain YouTube channels that have like a part of their like uh catchphrase is like you know, says some word that's on some you know not allowed list.
ConorStay tuned, listener. We've got Kevlin Henney almost lined up for next week. And uh we'll have a couple episodes of that, and then we'll probably be live from CBP North. If you're there, come say hi. All right, we'll we'll chat next time.
BryceAll right, I gotta get back to figuring out how Python decorators work.
ConorBe sure to check these show notes either in your podcast app or at adspthepodcast.com for links to anything we mentioned in today's episode, as well as a link to a GitHub discussion where you can leave thoughts, comments, and questions. Thanks for listening. We hope you enjoyed and have a great day.
BryceLow quality, high quality, that is the tagline of our podcast.
ConorThat's not the tagline. Our tagline is chaos with sprinkles of information.
In this episode, Conor and Bryce chat with Zach Laine about C++26 Reflection vs CRTP, AI, kids & more. Link to Episode 313 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP:...
In this episode, Conor and Bryce chat with Zach Laine about the C++26 contracts fiasco. Link to Episode 312 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...
In this episode, Conor and Bryce talk with Zach Laine about his career and the career changes he has made. Link to Episode 311 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials...
In this episode, Conor and Bryce talk about Bryce's next chapter and where he is leaving NVIDIA for. Link to Episode 310 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The...
In this episode, Conor and Ben chat with Jason Turner about default construction in C++. Link to Episode 309 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...