- Emergent Behavior
- Posts
- EB-4: Dingboard - Transcript - Part 1
EB-4: Dingboard - Transcript - Part 1
Ate A Pi
April 14, 2024
Transcript: Dingboard - Part 1
👩🎤 Ate-A-Pi: scene. So without further ado, what is Dingboard?
🧑🦰 yacine: Hello.
🧑🦰 yacine: So Dingboard is an image editor that works right in your web browser. So Dingboard is mostly used to make memes. And the core features of Dingboard, I think what makes it really easy to do, what makes it popular with people is how easy it is to use, how quickly it is to load the actual application. So the application itself runs in a web browser. I'm doing this all in a web browser. And the features that I ship.
I'm very careful to ship features that are only directly useful for, like I try to do the minimum amount of work to get just enough good enough image editing. And yeah, so it's basically just an image editor. And you can make memes really quickly. Sorry, I'm a little nervous. But.
I'm very careful to ship features that are only directly useful for, like I try to do the minimum amount of work to get just enough good enough image editing. And yeah, so it's basically just an image editor. And you can make memes really quickly. Sorry, I'm a little nervous. But.
👩🎤 Ate-A-Pi: No worries. So if I can describe it, it's a website. It's dingboard.com, right? Dingboard.com. You go to this website, and I think you start off with a open landscape of some sample memes or sample parts of memes, maybe a peppy frog and a couple of other things just sitting around.
🧑🦰 yacine: Yep, dingbord.com
🧑🦰 yacine: Yeah, exactly.
👩🎤 Ate-A-Pi: And you can click on an image, and then you have a number of tools there. And I think one of those tools is an image segmentation tool, where you can click on an image, and you can click, let's say, on a character in an image, and you can click on the image segmentation tool, and you can remove that character, a face, a whole character, a building from that image.
🧑🦰 yacine: Exactly.
👩🎤 Ate-A-Pi: And then you can remix that somewhere else. You can take that face and put it on top of another image, like a collage, and it's perfectly designed so that you can cut out the entire outline. It segments the image in that way. And then you can remix that image with other images in order to make something new, interesting, and funny.
🧑🦰 yacine: Pretty much the main metric that I aim for is how quickly it is for you to make a meme. For example, let's imagine that I'm in a podcast right now, and we're both VR avatars. And I want to make a really quick meme of us. I want to capture this moment in meme form. So I'll grab a picture of a Joe Rogan podcast. I don't know, I'll screenshot myself here. Let me just angle my face in this manner. And all the tools are kind of set up in a way which optimize for usability and speed. So I can kind of just paste myself in here.
🧑🦰 yacine: Um, and kind of just like merge myself in and it's, it's like basically useful for like really quick edits. And my kind of like, my goal is to make sure that things like take like about. You know, one minute or less to make an image, right? Like if it takes any longer than that, then I've like kind of fucked up. Um, and there we go. That's us having a podcast. Uh,
I figured out a way to use these AI models in a manner which optimizes for the speed. And I've also shipped diffusion models to help you get that last mile of editing.
I figured out a way to use these AI models in a manner which optimizes for the speed. And I've also shipped diffusion models to help you get that last mile of editing.
👩🎤 Ate-A-Pi: So in the time, in the last 30 seconds as you spoke, we've had this Joe Rogan versus Elon Musk podcast where you have inserted your anime avatar into Elon's position and my anime avatar into Joe Rogan's position. And we are talking to each other and we have an image right there within, yeah, within like 20, 30 seconds.
🧑🦰 yacine: Exactly. Basically if you can't make a Photoshop level image in one minute, then your editor sucks ass and you need to make a better one. This is all working in the web browser as well. This all runs in the web browser. Yes.
👩🎤 Ate-A-Pi: And let's just take a step back and like, how would you make this in Photoshop? You have the Joe Rogan image and you have these two guys sitting there. And the segmentation problem is the first big problem. You got to remove them from the, or you got to take your own image and you got to kind of slide it in there and then you would have two layers at that point. And then you'd be...
🧑🦰 yacine: You'd have to at least click for more than seven buttons, I'm sure, including the Subscribe to Creative Cloud for $1,000 for a year button. I think one of the things, so people will often compare Photoshop to Dingboard. But the fact of the matter is I've never used Photoshop before. And I think I kind of see that as an advantage, where I haven't solely how I view image editors.
I'm coming at image editors from this very naive approach, and I can personify my average user a lot better, which aren't like professional image editors. It's kind of like, I'm trying to democratize professional image editing, so it's better that I'm not a professional image editor, because then I know better what to build for the average consumer.
I'm coming at image editors from this very naive approach, and I can personify my average user a lot better, which aren't like professional image editors. It's kind of like, I'm trying to democratize professional image editing, so it's better that I'm not a professional image editor, because then I know better what to build for the average consumer.
👩🎤 Ate-A-Pi: That's amazing. So if you look at what, so the interesting thing about what you're doing, which I found which is quite interesting, is as I understand it, you're doing part of the AI processing on the browser and part on server in the back end. Is that correct?
🧑🦰 yacine: Yeah, that's correct. So one of the things I optimize for is user interaction latency. It's not.
very appreciated how much latency actually affects user retention and user satisfaction. I don't think users themselves understand that. So every time I push out a model to the browser, so what actually happens is there's a model for the segmentation, there's a model in the back end that does the embedding and then sends the embedding back to the front end. And then I load a neural network onto your front end that does a prompt every time there's a hover event. So as you're dragging your mouse around, it's basically constantly doing neural network inferences,
very appreciated how much latency actually affects user retention and user satisfaction. I don't think users themselves understand that. So every time I push out a model to the browser, so what actually happens is there's a model for the segmentation, there's a model in the back end that does the embedding and then sends the embedding back to the front end. And then I load a neural network onto your front end that does a prompt every time there's a hover event. So as you're dragging your mouse around, it's basically constantly doing neural network inferences,
👩🎤 Ate-A-Pi: Mm-hmm.
🧑🦰 yacine: figure out the most likely segment for the clicks that you provided. And that speed allows you to interact with it a lot better. Instead of clicking on a segment and then waiting 30, even 500 milliseconds would kill that whole interaction. It has to be within 100 milliseconds. So that's part of the reason I'm doing that. Another reason I'm doing that is there's other models I have loaded and I just haven't baked into a feature yet, TikTok's depth anything model, which is Apache.
So I've been able to like expert models and run them on a web browser. So this is a depth model that can like take an image and then find out how close each pixel is to the camera, which is useful for things like background removal. It's useful for a whole bunch of stuff. So like lighting and stuff like that. Like, so it's kind of surprising.
how much stuff you can actually just run in a browser. I'm using something called Onyx to do that. So that's a big part of Dingboard. And for what's worth is, it also ties into what I'm doing, how I'm going to compete with the bigger guys because I'm running all of the inference on client, on the client side. I don't have to, I can basically scale indefinitely and I can charge basically nothing for the product and I can edge out my competitors that way.
So I've been able to like expert models and run them on a web browser. So this is a depth model that can like take an image and then find out how close each pixel is to the camera, which is useful for things like background removal. It's useful for a whole bunch of stuff. So like lighting and stuff like that. Like, so it's kind of surprising.
how much stuff you can actually just run in a browser. I'm using something called Onyx to do that. So that's a big part of Dingboard. And for what's worth is, it also ties into what I'm doing, how I'm going to compete with the bigger guys because I'm running all of the inference on client, on the client side. I don't have to, I can basically scale indefinitely and I can charge basically nothing for the product and I can edge out my competitors that way.
👩🎤 Ate-A-Pi: There's just a lot to unpack there. And I'm no expert, so I'm gonna take this step by step here. So how do you... Okay, firstly, the liveness question, I'm totally with you there. Typically, you need the reaction time to drop to somewhere around 50 milliseconds for people to feel that there is liveness.
Like, liveness is a 50 millisecond kind of feeling. And that's been true of computers for the last 30, 40 years. So if you pressed a keyboard and you don't see that reaction within that 50 millisecond time, it's not live for you, and it doesn't feel real. And so I totally get it there. But it's very interesting how you've unpacked the process into two pieces.
Like, liveness is a 50 millisecond kind of feeling. And that's been true of computers for the last 30, 40 years. So if you pressed a keyboard and you don't see that reaction within that 50 millisecond time, it's not live for you, and it doesn't feel real. And so I totally get it there. But it's very interesting how you've unpacked the process into two pieces.
🧑🦰 yacine: Yes.
🧑🦰 yacine: Yes.
👩🎤 Ate-A-Pi: Normally, if I wanted to segment, I would, or let's say I'm doing that inference totally on cloud, I would click to segment. It would go back to an API somewhere. It would get processed, and then I'd get the result back. And then I would see the results of that, and then I would think, okay, I need to modify this a little bit more. And then I would have the next glacial step.
I need to do. And in this case, you're splitting that task apart into two pieces, I'm guessing, the embedding piece and the actual regeneration piece, what would you call it, the regeneration piece on the client side, is that correct?
I need to do. And in this case, you're splitting that task apart into two pieces, I'm guessing, the embedding piece and the actual regeneration piece, what would you call it, the regeneration piece on the client side, is that correct?
🧑🦰 yacine: Yeah, so you're on the right track. So basically, I send the image over to an image encoder, which runs on the back end. So the input to that is just an image, and the output is a tensor. And then the prompt decoder takes an input of two things, the client clicks, and then also the image embedding. So it's basically doing all the preprocessing beforehand. And I won't take credit for this. I was just reading a research paper from Facebook, or Meta.
these days, and I was waiting for my winter tires to get changed. I was like, oh, shit. I probably could do this on the web browser. Doesn't seem that much like, it's not that much parameters. It probably does run the web browser. So yeah, this is actually something that the segment anything, the authors of the segment anything paper actually engineer themselves. So.
these days, and I was waiting for my winter tires to get changed. I was like, oh, shit. I probably could do this on the web browser. Doesn't seem that much like, it's not that much parameters. It probably does run the web browser. So yeah, this is actually something that the segment anything, the authors of the segment anything paper actually engineer themselves. So.
👩🎤 Ate-A-Pi: amazing. So is this the future? I mean, just to take a step back here, the reason I'm saying this is because obviously, latency is very, very important for the user experience. And so far, a lot of these more sophisticated models have always been this kind of API or
Or you have the entire model sitting on your laptop. And you need very hefty machines, a 4070, 4080, 4090, or even bigger to run these things. But you've unpacked the process into two pieces, which have a browser component, which anyone can run, and a server component, which you're running in the back end, but you don't need that many servers to support your users. Is that correct?
Or you have the entire model sitting on your laptop. And you need very hefty machines, a 4070, 4080, 4090, or even bigger to run these things. But you've unpacked the process into two pieces, which have a browser component, which anyone can run, and a server component, which you're running in the back end, but you don't need that many servers to support your users. Is that correct?
🧑🦰 yacine: That's pretty much correct. I think that these models, as time goes on, they're kind of hitting this halving trajectory, where the model size actually gets smaller and smaller, which is a function of the data getting better. So what happens is these folks are training these super large models that are actually really great at annotating data. And when you have cleaner data, then you can get away with less parameters. You don't need to trade off on that.
And if you have less parameters, then you can run on lighter and lighter compute. So I actually think that's going to be kind of the way forward for a lot of this stuff. So diffusion models are still a bit too big. They're getting really cheap computationally, but they're still too large in size. So
I don't even think it's the trend is splitting the models up. I think the trend is actually going to be the whole ass model just gets a lot smaller. So I have a branch, a test branch, running on Dingboard where I do the whole segmentation, the encoding, and the decoding all in the front end. So the whole model. So it's called efficient SAM. And what these folks are basically doing is they're just distilling the model down. You have a really great model that you just train a lot of parameters and you distill it down. It's like.
I mean, I'm pretty sure that's how depth anything works as well. I can't remember reading the paper. So I think just generally smaller models are going to become more and more common. And what's really great about this is that you can, the more narrow your task, the smaller your model can be. The less generality you need, the smaller your model can be. So you can imagine like.
You could probably cover 50% of the use cases with fidelity as good as mid-journey if you just have a bag of five really great fine-tuned, stable diffusion models. You'd lose a generality, but it is what it is. And users would rather have that in the speed that comes with it. And also, it's just like it makes sense from a perspective of it's very cheap to run. Business-wise, it makes a lot of sense to go in this direction.
And if you have less parameters, then you can run on lighter and lighter compute. So I actually think that's going to be kind of the way forward for a lot of this stuff. So diffusion models are still a bit too big. They're getting really cheap computationally, but they're still too large in size. So
I don't even think it's the trend is splitting the models up. I think the trend is actually going to be the whole ass model just gets a lot smaller. So I have a branch, a test branch, running on Dingboard where I do the whole segmentation, the encoding, and the decoding all in the front end. So the whole model. So it's called efficient SAM. And what these folks are basically doing is they're just distilling the model down. You have a really great model that you just train a lot of parameters and you distill it down. It's like.
I mean, I'm pretty sure that's how depth anything works as well. I can't remember reading the paper. So I think just generally smaller models are going to become more and more common. And what's really great about this is that you can, the more narrow your task, the smaller your model can be. The less generality you need, the smaller your model can be. So you can imagine like.
You could probably cover 50% of the use cases with fidelity as good as mid-journey if you just have a bag of five really great fine-tuned, stable diffusion models. You'd lose a generality, but it is what it is. And users would rather have that in the speed that comes with it. And also, it's just like it makes sense from a perspective of it's very cheap to run. Business-wise, it makes a lot of sense to go in this direction.
👩🎤 Ate-A-Pi: Absolutely. What's been the user growth at DingBoard Bin? It's been, I think, seven months, you said, seven months since you kicked off the project?
🧑🦰 yacine: Yeah, it's been seven months. The last time I published my like MRR numbers, I was like around eight grand or something, eight grand Canadian dollars. So that's like, not quite American, but if you say, if you measure everything by Canadian dollars, then your MRR is basically 30% more and the number is more impressive. So that's why I do that. But my user, I actually like very recently ran a DAU count, was like 400 in the order of like hundreds of DAUs, pretty stable too, DAUs. I have like.
👩🎤 Ate-A-Pi: Mm-hmm.
🧑🦰 yacine: dedicated users, so users from a very long time ago, stick around. I think that's a pretty good metric to keep an eye on. And I've been getting a pretty steady growth of like, it depends, if someone big will tweet a meme made with Ding Board, I'll get a big rush, like a big bum rush of users. But right now, I could check the charts right now, but it's probably like 50 to 100 new users a day. I think the count of total users signed up is probably around 15 grand.
At this point.
At this point.
👩🎤 Ate-A-Pi: It's, you know, those that's often seen as the initial signs of product market fit. You know, the initial signs that you have dedicated users who are using you after a significant amount of time. And that's, you know, the maxim is always find dedicated fans and then figure out how to serve them better, right? So it's pretty amazing that you have that, yeah.
🧑🦰 yacine: Yeah, and it's just like symbiotic relationship, too, where I got really great users very early. So they helped me out quite a lot by just telling me what they wanted.
and what was annoying them. So I just kept on sending patches. And basically, a ding board is like 80% built by the users because they tell me what's broken, what they want, and I just do the thing that they ask me to do. And I keep on doing the things that make them quite happy. So without my true, I guess, the true ding board fans, ding board basically wouldn't exist in the form that it exists today.
and what was annoying them. So I just kept on sending patches. And basically, a ding board is like 80% built by the users because they tell me what's broken, what they want, and I just do the thing that they ask me to do. And I keep on doing the things that make them quite happy. So without my true, I guess, the true ding board fans, ding board basically wouldn't exist in the form that it exists today.
👩🎤 Ate-A-Pi: Indeed. Maybe just to take a step back and to give a sense of scale, I think one of the things that I noted was the cursor crown. So the Atlantic published a article on Mark Zuckerberg. As usual, I think it was not a very salutary article.
and they used a crown made of mouse cursors. And I think to a lot of people, that looked like a digging board creation.
and they used a crown made of mouse cursors. And I think to a lot of people, that looked like a digging board creation.
🧑🦰 yacine: Yeah, so if you give people really easy tools to edit images, the site itself becomes an asset generator where people will create really cool assets very quickly. So I'm pretty clever, and I understand how, I have a very keen understanding of how memes kind of proliferate. And I saw a.
opportunity there where you can kind of play off of negative sentiment and flip it and turn into positive sentiment. So, and I also noticed that it was a clever, uh, PFP virality hack. So if you encourage people to add it onto their PFPs or profile pictures on Twitter, they'll see what it is and they'll, and they'll want to get it themselves. And then, um, well, guess what the place, the easiest place to do that on is it's going to be ding board.
So it was kind of like a, I observed it, you know, no one, I bet you no one read the article. And I don't think it's kind of funny that the Atlantic didn't like actually mean to make it look so goddamn cool, the crown. So yeah, I just like grabbed it really quickly in like a moment. And I think like the actual software itself kind of enables this because it's so easy to make, like the fact that it takes me like 10 seconds to just do that, like just like R sync an asset, like grab the picture, like remix it on DingBoard and then just like R sync it up and then done.
I think that also enables the virality. Because a lot of virality is based on recency. So if you just see the Atlantic and then you see, like, oh, you see he's goofing on them, you're a lot more likely. You're going to find that kind of funny. So I definitely thought it was quite funny. I was laughing quite a lot.
opportunity there where you can kind of play off of negative sentiment and flip it and turn into positive sentiment. So, and I also noticed that it was a clever, uh, PFP virality hack. So if you encourage people to add it onto their PFPs or profile pictures on Twitter, they'll see what it is and they'll, and they'll want to get it themselves. And then, um, well, guess what the place, the easiest place to do that on is it's going to be ding board.
So it was kind of like a, I observed it, you know, no one, I bet you no one read the article. And I don't think it's kind of funny that the Atlantic didn't like actually mean to make it look so goddamn cool, the crown. So yeah, I just like grabbed it really quickly in like a moment. And I think like the actual software itself kind of enables this because it's so easy to make, like the fact that it takes me like 10 seconds to just do that, like just like R sync an asset, like grab the picture, like remix it on DingBoard and then just like R sync it up and then done.
I think that also enables the virality. Because a lot of virality is based on recency. So if you just see the Atlantic and then you see, like, oh, you see he's goofing on them, you're a lot more likely. You're going to find that kind of funny. So I definitely thought it was quite funny. I was laughing quite a lot.
👩🎤 Ate-A-Pi: Uh, what, what does it mean?
🧑🦰 yacine: Um, I think a meme is, so there's a lot of definitions of the meme. So the origin definition comes from Richard Dawkins. Um, but I think a meme is a piece of shared cultural context. And.
it's something that like spreads. So a lot of it is a meme can be an inside joke with your high school friends. If you're hanging out with your high school friends and you guys just like latch onto some term, it's like an inside group joke. It's like some shared cultural context that you guys can use to signal and relate to each other. It's also like a way to spread information. Humans kind of basically just like evolved to use these symbols to communicate with each other. And you know, what they say is a picture is worth a thousand words. And I do think that's quite true.
And I think the thing that gives memes a lot of power is the shared cultural context, right? Like there's a lot of like, I think sort of the part of Twitter that me and you reside in is extremely internet addicted folks. So a lot of our memes are memes that are ..
images that are relatable for people who are very internet addicted or are callouts to old forum culture from like 2000, like the 2000s. So I think that's what a meme really is. It's one of those things where if the more you think about it, the harder it gets to define. But you definitely know a meme when you see it, right? And there are good memes and there are bad memes. So
it's something that like spreads. So a lot of it is a meme can be an inside joke with your high school friends. If you're hanging out with your high school friends and you guys just like latch onto some term, it's like an inside group joke. It's like some shared cultural context that you guys can use to signal and relate to each other. It's also like a way to spread information. Humans kind of basically just like evolved to use these symbols to communicate with each other. And you know, what they say is a picture is worth a thousand words. And I do think that's quite true.
And I think the thing that gives memes a lot of power is the shared cultural context, right? Like there's a lot of like, I think sort of the part of Twitter that me and you reside in is extremely internet addicted folks. So a lot of our memes are memes that are ..
images that are relatable for people who are very internet addicted or are callouts to old forum culture from like 2000, like the 2000s. So I think that's what a meme really is. It's one of those things where if the more you think about it, the harder it gets to define. But you definitely know a meme when you see it, right? And there are good memes and there are bad memes. So
👩🎤 Ate-A-Pi: Yeah, I play Twitter like a game. So I regard Twitter as kind of like this kind of like verbal visual kind of game. And so for me, I use, it's almost like this kind of like cultural unlock of a surface because I can, you know, because during peak times I can sense who else is online.
kind of I have in the back of my mind, like, you know, as I scroll through the feed, I kind of have in the back of my mind who is there. And I kind of also have a sense of like, what is going to work for these people? And as I read through something, I kind of cross relate. So it's, I'm autistic. So I'm running a retrieval augmented generation in the back. I have an embeddings vector.
kind of I have in the back of my mind, like, you know, as I scroll through the feed, I kind of have in the back of my mind who is there. And I kind of also have a sense of like, what is going to work for these people? And as I read through something, I kind of cross relate. So it's, I'm autistic. So I'm running a retrieval augmented generation in the back. I have an embeddings vector.
🧑🦰 yacine: Okay, amazing.
👩🎤 Ate-A-Pi: I have somewhat of an aedetic memory for text and images. So I'm running that. And as I see something that picks up that the embedding vectors kind of cross and there's some cosine similarity, and I'm like, okay, I have these two things that kind of not really related to each other, but the embedding vector is showing some similarity there. How do I write something that relates the two in an interesting way? And if I can get that kind of like,
a Sudoku-like unlock, I write it down, I make a meme, or I do whatever, and then that kind of goes on the stream, and then I see the reaction to that to see whether or not it works or not. And if it works, then I go downstream further. I update or I do something else. So for me, it's a game. It's a kind of streaming game that you kind of play.
a Sudoku-like unlock, I write it down, I make a meme, or I do whatever, and then that kind of goes on the stream, and then I see the reaction to that to see whether or not it works or not. And if it works, then I go downstream further. I update or I do something else. So for me, it's a game. It's a kind of streaming game that you kind of play.
👩🎤 Ate-A-Pi: I have my own metrics, right? I used to have this metric, you know, views to followers metric that I would look at. And now I've kind of like diverged from there because as your follower count goes up, those metrics start to be less meaningful and you start looking at other things.
🧑🦰 yacine: Right.
👩🎤 Ate-A-Pi: So yeah, for me, it's completely a game. And I use these memes as this kind of unlock on the surface or like a turn or a move to see what happens.
🧑🦰 yacine: Yes.
It's like the meme is basically like compressed information. It's information that yourself, like you've connected to multiple things put together. You basically like collected all of this information and compressed it into a single image or a single observation, and you send it out to the wild. And when people interact with it, that's a, you're basically correctly fitting the curve where you're like modeling reality quite well because you've made this observation that other people also agree with. Would you say that like a lot
times it kind of like, where does it really come from? Like, does it like strike you out of nowhere? Is it just like, kind of comes for me, a lot of the times it just feels like it comes from the ether.
It's like the meme is basically like compressed information. It's information that yourself, like you've connected to multiple things put together. You basically like collected all of this information and compressed it into a single image or a single observation, and you send it out to the wild. And when people interact with it, that's a, you're basically correctly fitting the curve where you're like modeling reality quite well because you've made this observation that other people also agree with. Would you say that like a lot
times it kind of like, where does it really come from? Like, does it like strike you out of nowhere? Is it just like, kind of comes for me, a lot of the times it just feels like it comes from the ether.
👩🎤 Ate-A-Pi: Um, in my case, I am not a natural. I think there's some people who are naturals. Uh, I am, I am not a, I'm not a natural. I'm not a natural for most things. Everything, everything is learned behavior. So, uh, for me, it is very much, uh, a kind of, um, active search. So I have to be like activated and I'm like, uh, okay. I am.
trying to figure out how to... It's a game. It's like I'm actively looking at how to create an unlock. And I'm acting, but I don't particularly know where the unlock is gonna come from. So there's a passive part that you're like, I am looking for an unlock, but I don't know where it's gonna come from. So I let that piece of it kind of expand a little bit or like loosen, but I keep the active focus on finding the unlock. So there's a passive and active part to it.
trying to figure out how to... It's a game. It's like I'm actively looking at how to create an unlock. And I'm acting, but I don't particularly know where the unlock is gonna come from. So there's a passive part that you're like, I am looking for an unlock, but I don't know where it's gonna come from. So I let that piece of it kind of expand a little bit or like loosen, but I keep the active focus on finding the unlock. So there's a passive and active part to it.
🧑🦰 yacine: Right.
🧑🦰 yacine: Got it.
👩🎤 Ate-A-Pi: So you're scrolling and you're thinking, you're thinking in the mid brain that you're actively looking for something to say, but where you, what you're gonna say it about and what you're gonna connect it to, the fore and the back, that you kind of like loosen up. And it's really a state of mind. I don't, I can't do it all the time. It's really kind of a state of mind. But if you, and it's taken a little bit of training also. It's definitely taken like.
🧑🦰 yacine: Yeah.
👩🎤 Ate-A-Pi: you know, months and months of like, oh, you know, how do I, how do I like de-res one part of your thought process and res the other part, bring the other part into focus, right?
🧑🦰 yacine: Right.
🧑🦰 yacine: Right. You basically have to just let loose and just send things into the void. I think the really cool thing about Twitter, which is that you can actually do a lot of things
is very much like a neural image board or a neural forum, is that the actual, you have a meat filter that does all the work for you. So all you have to do is generate, and you have this like discriminator that basically just exists, which is like the collective flesh of the humans around you. And you said like, you kind of kept close, like you keep a close eyes on, keep a close eye on metrics. I also do keep a close eye on the metrics, but the only metric I keep a close eye on is like tweet rate. Like how many tweets am I actually sending out per hour?
because it's the only metric that I can control. So it's the thing that I try to optimize for an increase. But it kind of just happens naturally. One of the things I've noticed was I've noticed my tweet rate naturally increase unbeknownst to me. And I think it's a function of me not having a job anymore slash being my own, being self-employed. But I also think it's, I used to think I was the input to the function
the positive feedback loop, but it turns out that I'm actually just a part of the positive feedback loop. I think that there's something that's encouraging me to tweet more, which is probably really just internet addiction. It's fascinating to look at Twitter. It's kind of crazy how much connections you can establish from a website like that, just by kind of continuously sharing stuff. It's also one of those things where if you're, you know, the barely...
Just be a little bit careful and anything you say won't harm anyone. Like just like, like actually getting things wrong won't actually hurt anyone. As long as you make the context clear that you like are just saying shit. But then you get the, you still get the benefit of like the collective, you know, the collective group of Twitter users who like are, you know, either up boosting or ignoring your signal.
is very much like a neural image board or a neural forum, is that the actual, you have a meat filter that does all the work for you. So all you have to do is generate, and you have this like discriminator that basically just exists, which is like the collective flesh of the humans around you. And you said like, you kind of kept close, like you keep a close eyes on, keep a close eye on metrics. I also do keep a close eye on the metrics, but the only metric I keep a close eye on is like tweet rate. Like how many tweets am I actually sending out per hour?
because it's the only metric that I can control. So it's the thing that I try to optimize for an increase. But it kind of just happens naturally. One of the things I've noticed was I've noticed my tweet rate naturally increase unbeknownst to me. And I think it's a function of me not having a job anymore slash being my own, being self-employed. But I also think it's, I used to think I was the input to the function
the positive feedback loop, but it turns out that I'm actually just a part of the positive feedback loop. I think that there's something that's encouraging me to tweet more, which is probably really just internet addiction. It's fascinating to look at Twitter. It's kind of crazy how much connections you can establish from a website like that, just by kind of continuously sharing stuff. It's also one of those things where if you're, you know, the barely...
Just be a little bit careful and anything you say won't harm anyone. Like just like, like actually getting things wrong won't actually hurt anyone. As long as you make the context clear that you like are just saying shit. But then you get the, you still get the benefit of like the collective, you know, the collective group of Twitter users who like are, you know, either up boosting or ignoring your signal.
👩🎤 Ate-A-Pi: Yeah, I think Twitter, I think Josh Abak regards it as one of the most important sites that you can have. I think it's, I think we're riding onto other people's neurons directly, right? So...
🧑🦰 yacine: Yep. It's like this amazing emergent system, dude. If you look at, so Twitter's code is actually open source and they have really great tech specs out there. And even during Twitter 1.0, they wrote really great blog posts about how it works. And the more you look into it, the more you realize how simple of a system it is. And I think that's where a lot of Twitter's power comes from. Like, as far as I can tell, there's no really fancy ML stuff. It's just like, you know,
kind of cool algorithms that group people together naturally based on their interactions. And it's like this natural human thing that just works incredibly well. And it feels like very recently the filters have been, the filters have been lifted off of Twitter, which is allowing Twitter to better model the collective beauty of humanity. I've been definitely feeling like Twitter has been getting a lot better in terms of the users who feel like they can say things. It's actually been super fun.
kind of cool algorithms that group people together naturally based on their interactions. And it's like this natural human thing that just works incredibly well. And it feels like very recently the filters have been, the filters have been lifted off of Twitter, which is allowing Twitter to better model the collective beauty of humanity. I've been definitely feeling like Twitter has been getting a lot better in terms of the users who feel like they can say things. It's actually been super fun.
👩🎤 Ate-A-Pi: So one of the things that strikes me is that, so we have this idea of AI, AGI, et cetera. And one of the things that I always feel is that you also have these things like Twitter, which are like a single organism composed of digital neurons and meat neurons, right? And it is kind of this collective consciousness there. And I don't think...
that idea of this collective consciousness embedded in this kind of always on meat and digital system, that no one refers to that as, I mean, I mean, some people, some small number of people who are like Josh and Bach do, but no one refers to that as kind of like an AGI or.
that idea of this collective consciousness embedded in this kind of always on meat and digital system, that no one refers to that as, I mean, I mean, some people, some small number of people who are like Josh and Bach do, but no one refers to that as kind of like an AGI or.
🧑🦰 yacine: So I would actually classify that as sentient in some way. I would classify that as something that actually experiences the world, because anything else does not make sense to me. As far as I can tell, it's processing information at scale, and therefore, I must experience things.
👩🎤 Ate-A-Pi: And you can also see it kind of motivate certain things, right? You can also see it that it likes, what I always say is that it likes things which are different, right? So it tends to push forward ideas and things which are shocking or different, forcing people to like confront, like, you know, for example, you know, some crime, which, you know, some video of a crime or whatever that.
🧑🦰 yacine: Yes.
👩🎤 Ate-A-Pi: people don't really want to see on their feeds, but then it's just so shocking that it just bursts through the feed. And then like, you know, events are forced to kind of, the world events are forced to kind of come together to kind of fix that or whatever, right?
🧑🦰 yacine: Yes, it's like this perfect conduit of information where you have humans choosing what to input into the system. And in fact, there's this selection process that humans are just naturally doing themselves. I have been thinking a lot about what makes something interesting to me versus not interesting. And I think what makes something interesting is, is it variant or is it new? Is it information that I have not
you know, thought of or observed before. And if I put it into my current world model, does it like fit?
Like, is it like actually fitting the world? Like, does it help me predict the world better? Does it make sense to me? And I think that information that is like both those qualities, like entropic in some sort, where it's like just novel information and also isn't schizophrenic. So like actually does model the world quite well based on what I already understand. That information is something I really like. A really good way to, another good example is jokes are funny. Like comedians will use...
subversion of expectations, but still say something that is actually true. And I think like humans like generally enjoy pieces of information that are delivered like that. Um, so like, you know, recently I tweeted out like, Hey, like there's our principal engineer, uh, he's like this, like, you know, he's in a cubicle, he's a cubicle farm, it's like, Oh, and he's being kind of treated like a child. And I think it's kind of like unexpected matchup. And it's also kind of funny because it is true at some level. And I think like, that's like what.
what kind of successfully filters through the giant Twitter system.
you know, thought of or observed before. And if I put it into my current world model, does it like fit?
Like, is it like actually fitting the world? Like, does it help me predict the world better? Does it make sense to me? And I think that information that is like both those qualities, like entropic in some sort, where it's like just novel information and also isn't schizophrenic. So like actually does model the world quite well based on what I already understand. That information is something I really like. A really good way to, another good example is jokes are funny. Like comedians will use...
subversion of expectations, but still say something that is actually true. And I think like humans like generally enjoy pieces of information that are delivered like that. Um, so like, you know, recently I tweeted out like, Hey, like there's our principal engineer, uh, he's like this, like, you know, he's in a cubicle, he's a cubicle farm, it's like, Oh, and he's being kind of treated like a child. And I think it's kind of like unexpected matchup. And it's also kind of funny because it is true at some level. And I think like, that's like what.
what kind of successfully filters through the giant Twitter system.
👩🎤 Ate-A-Pi: So what happens, I think the next step which I expect is spam filters, AI spam filters, right? Which kind of read a bunch of stuff for you and decide what's worth reading and what's not reading, right? That's definitely coming to email for sure, right? Email is, you know, email, it has to happen because everyone wants it. What happens when that happens to something like Twitter or TikTok? Like, how does that, how do things start to change at that point?
🧑🦰 yacine: Yes.
Yes.
Yes.
🧑🦰 yacine: I honestly, I'm of the belief that everything's going to be fine. I think that generally like, I would describe bot like behavior as behavior that is like not very low, like very low entropy. It's like very expected behavior. It's like, you can clearly tell what a bot post is, even though it can be a human.
because it doesn't actually say anything new. It just rehashes things. And I think that if we have AGI systems that do say new things, then I would want to read what they say. If we have AGI systems that are tweeting better than actual humans, then I don't want to read the humans. But I think that all of that stuff is going to naturally sort itself out, because the Twitter algorithm is very, very good. It's very, very simple, and it just works. Like the same cluster algorithm, you can look it up.
And it's super simple and it just works. And I feel like humans are going to naturally just take care of the system, like by just reacting to things and, you know, muting certain accounts. So if I, if I get someone who's like too bought, like who shows up in my, you know, replies, I will actually just mute them. Uh, if I get someone and for what it's worth, I don't care if they're a bot or not, what happens is that like they drop a LinkedIn ass reply.
And I don't want to read that. So I just will use the tool to naturally downvote them and not react to it. And that kind of helps keep the system clean and keeps the same cluster that I like being part of the same way it always has been. I'm actually not worried at all, to be honest. I think that another great example is people will keep on talking about proof of humanity and how we're gonna need it so bad, but they don't realize that we already have that.
like super complex proof of humanity system working for credit cards. It's for you to get a credit card. You have to prove that you're like, an upstanding citizen of the West. You can't get one if you have no credit. And that like works great. And there's like actual financial incentive for people to like break that. So we've already kind of like began to evolve the system. And it's also why I think like Twitter's, that's why I think Twitter using
because it doesn't actually say anything new. It just rehashes things. And I think that if we have AGI systems that do say new things, then I would want to read what they say. If we have AGI systems that are tweeting better than actual humans, then I don't want to read the humans. But I think that all of that stuff is going to naturally sort itself out, because the Twitter algorithm is very, very good. It's very, very simple, and it just works. Like the same cluster algorithm, you can look it up.
And it's super simple and it just works. And I feel like humans are going to naturally just take care of the system, like by just reacting to things and, you know, muting certain accounts. So if I, if I get someone who's like too bought, like who shows up in my, you know, replies, I will actually just mute them. Uh, if I get someone and for what it's worth, I don't care if they're a bot or not, what happens is that like they drop a LinkedIn ass reply.
And I don't want to read that. So I just will use the tool to naturally downvote them and not react to it. And that kind of helps keep the system clean and keeps the same cluster that I like being part of the same way it always has been. I'm actually not worried at all, to be honest. I think that another great example is people will keep on talking about proof of humanity and how we're gonna need it so bad, but they don't realize that we already have that.
like super complex proof of humanity system working for credit cards. It's for you to get a credit card. You have to prove that you're like, an upstanding citizen of the West. You can't get one if you have no credit. And that like works great. And there's like actual financial incentive for people to like break that. So we've already kind of like began to evolve the system. And it's also why I think like Twitter's, that's why I think Twitter using
🧑🦰 yacine: giving people the ability to pay for blue checks is pretty smart because it puts a ceiling on like, or puts a floor on how much it would cost to like run a bot farm. Because it's not just like, the challenge isn't just paying the money, it's also getting all of the credit cards and like running a credit card farm, which is like extremely, extremely huge.
massive dude, I'm talking billions of dollars of like incentive to like build a good fraud system. It'll sort itself out. I'm really not worried about it, to be honest. Like, it's like, I think the people I'm worried about are people who are like bought like in general, where they the information that they're generating is. So I think what happens and I fall into this pattern where you kind of like are a bit too afraid to truly be yourself and truly be variant.
and you'll kind of like start fitting the curve too much. I think like some people fall into that trap and I would probably be worried about those people like not being able to like get the audience they need because they're competing against too much bots. That's probably something that might happen. But then you can just like filter on blue checks. Just like, I just won't read, you know, if you don't have a credit card, I just won't read your shit.
massive dude, I'm talking billions of dollars of like incentive to like build a good fraud system. It'll sort itself out. I'm really not worried about it, to be honest. Like, it's like, I think the people I'm worried about are people who are like bought like in general, where they the information that they're generating is. So I think what happens and I fall into this pattern where you kind of like are a bit too afraid to truly be yourself and truly be variant.
and you'll kind of like start fitting the curve too much. I think like some people fall into that trap and I would probably be worried about those people like not being able to like get the audience they need because they're competing against too much bots. That's probably something that might happen. But then you can just like filter on blue checks. Just like, I just won't read, you know, if you don't have a credit card, I just won't read your shit.
👩🎤 Ate-A-Pi: Yeah, I think one of the things that, one of my metrics for what to write is, I have a lot of things stuck in my draft folder because I write it and I'm like, this is just a medium reply. This is just a lukewarm response. And then I'm like, okay, it's not, it's not, it's not worth, it's not worth putting out there, right? So.
So I definitely do that. I definitely have drafts which are stuck for ages. I definitely storyboard something. And because you want to have, like I said, it's almost like playing a Tetris game. If you find the exact right shape of the puzzle, you can unlock it, right? That's the way I play, right? So I often storyboard and then I'm like, it doesn't quite get there. And then it just gets stuck.
So I definitely do that. I definitely have drafts which are stuck for ages. I definitely storyboard something. And because you want to have, like I said, it's almost like playing a Tetris game. If you find the exact right shape of the puzzle, you can unlock it, right? That's the way I play, right? So I often storyboard and then I'm like, it doesn't quite get there. And then it just gets stuck.
🧑🦰 yacine: Yeah.
👩🎤 Ate-A-Pi: It just gets stuck and I'm like, I stop and I'm like, all right. And then sometimes, I've been storyboarding for a while and then it works. And I think I had one about this Apple iPhone store being ransacked and I compared it to a universal basic income. And I'd been thinking through those themes for ages, right? But I was like, how do I put it into something that's...
🧑🦰 yacine: Interesting.
👩🎤 Ate-A-Pi: is concise, but yet understandable. Right. And, uh, and then, and then, you know, I was like, it's like 6 AM in the morning. I woke up and, you know, first thing, you know, like, ah, okay. Very fast, like, you know, 15 minutes, less than 15 minutes, maybe like five minutes, you know, typed it all out, you know, send it out, right.
🧑🦰 yacine: Do you draft those things into the tweets? Do you start writing them as the tweets, or do you write them separately in a Google Doc or something?
👩🎤 Ate-A-Pi: I write them, you know, 99% of the time I write them in the tweets. So I definitely have a lot of beef with the Twitter user interface because I definitely want to use all of the features and like, you know, learn all of those things and like many things are not possible and I'm like, I know there's a lot, but the problem with Twitter is that, and I don't know whether it's true. I don't know. I don't know to what extent the views number is true.
But if you look at the views, it looks like only 1% of people or 2% of people are actually posting at any time. And 99% of people are not posting anything at all. So it's really this kind of, you know. And I've also had this thing where I noticed that once you cross the 1,000 follower mark, that's when you start getting picked up. Anything you write, post 1,000 followers, you have some chance of it going viral. But pre-1,000 followers, like,
you probably will not get noticed at all, right? Like, so that's one of the unfortunate things about Twitter.
But if you look at the views, it looks like only 1% of people or 2% of people are actually posting at any time. And 99% of people are not posting anything at all. So it's really this kind of, you know. And I've also had this thing where I noticed that once you cross the 1,000 follower mark, that's when you start getting picked up. Anything you write, post 1,000 followers, you have some chance of it going viral. But pre-1,000 followers, like,
you probably will not get noticed at all, right? Like, so that's one of the unfortunate things about Twitter.
🧑🦰 yacine: Fascinating, yeah.
🧑🦰 yacine: Interesting. Yeah. It's kind of interesting to me to think about Twitter as like an actual writing tool because it forces you to keep like it like encourages you to keep tweets like very short. And it helps you compress more information into like less words. And generally, that's just a good writing skill to have. So if I've ever written like forever right like technical work. With the thing I'll do probably spend the most time on is making it smaller and smaller and smaller so that the other engineers in my organization can just read it at a glance. That's my goal.
And I think Twitter actually does help you right in that way. So you have this idea about people ransacking iPhones and the lack of jail time for small, I guess not so small theft in San Francisco. It's effectively UBI. The fact that you're using Twitter to send that information out and even just to draft it, kind of encourages you to capture that idea in less words so that people are gonna be more likely to react to it. It's pretty cool to see. It's a lot like TikTok where the TikTok app itself
people use to create the content on TikTok. And it's actually got me, like once I realized that, it got me thinking kind of carefully about how, you know, the software becomes a tool to create the content on the software itself.
And I think Twitter actually does help you right in that way. So you have this idea about people ransacking iPhones and the lack of jail time for small, I guess not so small theft in San Francisco. It's effectively UBI. The fact that you're using Twitter to send that information out and even just to draft it, kind of encourages you to capture that idea in less words so that people are gonna be more likely to react to it. It's pretty cool to see. It's a lot like TikTok where the TikTok app itself
people use to create the content on TikTok. And it's actually got me, like once I realized that, it got me thinking kind of carefully about how, you know, the software becomes a tool to create the content on the software itself.
👩🎤 Ate-A-Pi: You, I often think whether it's the death of the essay, because the public intellectuals of today are gonna be posting, right? Benjamin Franklin today would be a poster. It would be a terrific poster.
🧑🦰 yacine: He won 100% beat. Yeah, exactly. And I would rather read a Plains tweet than I would then read a Benjamin Franklin essay, to be honest. I'll be completely frank.
👩🎤 Ate-A-Pi: Yeah, he would be vicious. He was vicious back in his day. He would be absolutely vicious now. Okay, let's go back to Dingboard. So what do you like? When I look at Dingboard, I see a tool that expands human imagination, creativity, humor, right? And it's especially so because it's trying to compress that feedback loop.
🧑🦰 yacine: Yes.
👩🎤 Ate-A-Pi: in between creating something, putting something out there. And with this realization that a lot of like what we had before was that you needed a inherent degree of skill in order to create something that could be put into the public space. And that was this barrier, right? There was this barrier. One of the ways that TikTok got popular was actually the...
innovated on ways to create content. And they tested it out in China, where things were automatically a little bit more difficult, because Chinese text is hard to type. So they had to create these tools, which were more visual, where you could just click a couple of buttons and just get content out there. And they got really good at this. And they actually had huge teams try out various features. So there was a constant like, you
trying out of a lot of features. And the features that came to TikTok from Douyin were the ones that succeeded, like this evolutionary process where if you were very successful on that feature in China, then your feature would get ported out to TikTok Global. And otherwise they would just like mess around with you, mess around with all these features within the Chinese ecosystem. So, yeah.
innovated on ways to create content. And they tested it out in China, where things were automatically a little bit more difficult, because Chinese text is hard to type. So they had to create these tools, which were more visual, where you could just click a couple of buttons and just get content out there. And they got really good at this. And they actually had huge teams try out various features. So there was a constant like, you
trying out of a lot of features. And the features that came to TikTok from Douyin were the ones that succeeded, like this evolutionary process where if you were very successful on that feature in China, then your feature would get ported out to TikTok Global. And otherwise they would just like mess around with you, mess around with all these features within the Chinese ecosystem. So, yeah.
🧑🦰 yacine: Fascinating.
Reply