Uh I I think I'm gonna talk here. Yeah, I hope you guys don't mind the Twitter uh aspect. This is all obviously very translatable to Bluesky. Yeah, my talks about uh as you can see, digital anthropology. Uh I'm gonna tell you a story. Uh I had fun uh writing uh last night. Okay. My name is Francisco. Uh I I have some background in AI safety and tools for thought. Uh I've worked at meter in the U Commission and I I run my own small RD org called Epistemic Garden, and we build AI tools for community flourishing.
Um epistemic garden. Okay. So I spend a lot of time on Twitter. There's this part of Twitter that I love. Uh and you could describe them as intellectual hippies. I think of them as uh pre-paradigmatic research scene on online. Uh and these are people with varied interests, uh, all of mine. Um there have been there have been potentially many waves. Uh this is how you can think of them. Uh but the the anthropology has uh always been done like manually uh in pictures like this one. Um yeah. Uh this is just uh to sh I I want to make the argument that it's a generative scene that ends up being very influential.
Uh you know the motto of the conference. Uh I'm pretty I I I'm relatively confident my friends came up with it uh way back when the department of war. Uh I know it's not a great slide, but uh again I'm making the argument that memetically uh this group of people is kind of upstream of a lot of stuff. Uh okay, so I love these people, I love this space online, and I wanted to understand what makes it tick, what makes it work, uh, and I wanted to make it go even faster. Uh but Twitter later is closed, oh no.
So I started something called the community archive. Uh because you know if I move to Bluesky, I love Bluesky, but if I move to Bluesky, my people wouldn't be there. So I started this thing uh and I got a few hundred people to upload their archives. You need to download them from Twitter, uh and it's a whole like laborious process. Uh but ultimately we end up with a really nice data set to to run experiments and build uh prototypes. You have a few influential people uh like in the top follows, Patio's like an ex-Stripe person who who's a famous blogger.
Emmett Shear was CEO of OpenAI for like a second. And and people did build all sorts of tools uh and prototypes. Um these are a few that I gathered, but you know, you have like a Google Trends type thing, someone someone did um uh like an idea like uh viral spread of ideas thing. I have a banger bot over there, someone did uh uh people search uh on the on the archive. Uh many tools. Uh but the tools don't answer my questions. I wanted to understand the the stories uh and the the flow of ideas in this community.
Uh so I tried semantic clustering first. You know, you you embed all the tweets, you try to cluster them, uh, and you get um you get groups of tweets that are are a little bit too broad. So you know the an example is you might end up with a cluster of just like broadly meditation-y uh themed things, but I want stories. I want I want like threads of narrative that are much more specific than that. Uh and you know I examine the the the meditation cluster and f found one really cool story. Uh people uh I think fitch not Han, like some I think Vietnamese uh Dharma person has this saying like the next Buddha is a Sangha.
And over time people like kind of remix this and uh all the all the Dharma oriented people uh made them yeah, you know, like the next Buddha is an internet community. Fun. Um that's a story that you can tell uh that you f that I found in the data. Uh but semantic clustering didn't work that well to do it in an automated exhaustive way, right? So I tried I tried to define strands, uh which worked better. Uh so I'm gonna tell you a little bit about strands. Um they worked better. Uh we could uh we took this we took this um tweet about community building, the original one, and we found tweets causally downstream of it uh and tweets topically upstream of it.
Like I was saying, strands are centered on one tweet, like causal light cones, right? Downstream, upstream. Backwards and forwards in time. This you know is just my working diagram, but it's it's what it says on the tin. Oh yeah, I guess I guess I could explain. It's like you have a tweet, it gets quoted a bunch of times in the future, or it gets replied to, and all of those are kind of causally downstream, right? And you could implement the causally upstream in many ways. In my case, I just I just did semantic similarity, and it's close enough.
We still get pretty cool results. Um yeah, I forgot I had slides explaining exactly this. Quotes, enriching with semantic search, and then you get all the threads from from the from the quotes and the replies and so on, and you end up with something that's kind of shaped like this. I I analyzed a bunch of them manually just to make sure that the quality was right. Then I kind of automated them and got 250 of them. But what tweets do we base uh the strands on? Well, bangers, of course. Bangers are tweets with lots of quote tweets over time from the from the same community.
Uh that you know kind of implies that they're good enough to cite uh over and over and and and use as building blocks in the future when making arguments. Works pretty well. Um these are you know some of the bangers. Uh if if you're if you're familiar with the community, you you you'll recognize these as pretty like canonical tweets that end up that end up being referenced over and over. We found 250 of them. Um this is kind of a little like just semantic atlas of the of the strands that we found. Uh and you know you can ask you can ask, can you break the strands together, right?
Can we make a full picture? Can we make a fuller like a fuller map? Uh I tried getting plot code to make sense of them because there's 250 of those. Um, as visual support, this is the chaos of the data set. And this these are the individual strands, and this is what I'm hoping for. Uh it's not perfect, but it's it's a start. Like it works works fine. Um it still requires a lot of human care. Like I wouldn't trust it, I wouldn't trust it to I wouldn't present the the results without having already the tacit knowledge of having been on Twitter for this long.
Uh but I I think we can get there. Uh so what can we say about the big history of the scene? Um, uh these are walls of text, so I won't bore you. Uh but the TLDR is there's this guy Visa who figures out a bunch of really good posting norms uh that people end up uh adopting over time. Um during COVID, lots of lots of kind of rationalisty, dharma-e flavored people uh join Twitter, they find Vista's posting norms, uh, and and it creates this this kind of really generative intellectual scene where people who are previously pretty like uh rationalist minded, like pretty left brain, end up end up discovering all of these, all of these like spiritual traditions and therapy modalities and embodied practices.
And that was kind of the golden age, but it but things keep going. Uh like this is just like the the kind of the implications of that of that big like left brain-right brain merge uh and kind of the error that we've been in for the past couple of years, it I guess is characterized by by community um building uh in real life and uh data infrastructure like the community archive. Um we just found a bunch of stories, we had cloth code stitching together. Uh I have I have a personal theory of of the scene. I don't think we're gonna go through it very, very in-depth.
Uh I will just say that I think the core loop of what's happening here is uh people people who are high in epistemic rigor, uh, you know, like have the tools of rationality, like Bayesian thinking and so on in systems thinking, uh look at these old traditions and and kind of try them out and explain them in rationalist terms. Uh and that makes them more accessible to new people who are kind of like more left brain emphasize emphasis. Um having access to these practices, they end up uh you know having more nervous system capacity and executive function, which lets them relate to people better, doing do more projects that has an outgrowth of of community and and and friendship and flourishing uh and is all supported by technical infrastructure like Twitter, Bluesky, uh, and tools that people may build on top.
This is uh with uh overwhelming icons added, I don't think it's worth looking at. Um there's still work to do. Um there are a few key events that I pre-registered that I thought I would be able to find in the data that I was not. And that's fine. Um for example, the first time the community met in person, there was this thing called Vibe Camp, but Brooke, the founder isn't in the community archive, so so we couldn't really like f find the strand exactly for that. In the same way, jhanas, which are kind of meditation practice uh that became really big recently, but the main people uh aren't there, so it was hard to find like the like strands that did exactly that.
And then fractal NYC is uh is an in-person community uh where the people are in fact present, but we still you know for one reason or another uh didn't find a great strand. Uh bless you. Uh if the data yeah, okay. Obviously, if we had complete data, uh this wouldn't be a problem, but the people I care about are still on Twitter, right? Uh and we were talking about this on the uh on the panel just earlier, it's kind of a socio-technical problem. Maybe we need to uh throw a better party uh even though the house is really nice.
Um exactly. Yeah, so uh what's next? What's next for me? Uh I'm building tools for community flourishing to you know get the hundred x more serendipity uh if we can. Um right now, like the next thing I'm doing is it's kind of like a P2P network of permission data between people and and to get you know to find opportunities between us, right? Like uh Ronan's agent talks to my agent, maybe I know someone who wants to fundronin, maybe I'm trying to get rid of a couch and Ronan needs a couch uh and the goal is to have this you know big ecosystem of AI fairies conspiring in your favor.
Uh this is where the fairies live. Yeah, uh this is Epistemic Garden. Uh I'm Francisco, you can find me on Substack. I publish lab notes fairly regularly uh with my work. And thank you. Thank you guys for paying attention. Thanks, Francesco. You're welcome. Any questions?
It's a lot of information. This is really fun. Um what do you you talked a little bit about limitations, but like yeah, what do you wish you could build that you feel like you can't build, or where are you limited? Like you were saying there were some like events you were expecting to see that you didn't see, but are there like bigger things? You're like, I wish I could have this, but I don't have it yet. I think definitely just the amount of data uh i i is one, right? Just being comprehensive across all the people that might matter.
Uh another is permissioned data. Uh where for the for the opportunity mining that I was talking about. Um, you know, I don't want to be public about everything, but for people I meet once or twice at a conference or friends of friends, I I wouldn't mind. There's a lot of information that would make it easy to coordinate with other people that I can that that I wouldn't be able to post on Twitter. Yeah. Uh I just gotta say I I I feel like the whole day if I had to summarize it in like one sentence that I learned is that Bluesky is a nice house, but it needs to throw a better party.
And I think, yeah, you can teach us how to throw good parties. Like this is cool stuff. Thanks. Yeah, so we should should we should jam on it. Yeah. Any other questions. You you talked sort of briefly about like assembling threads from like posts. I was wondering if we could like elaborate a bit more on that process. I felt like it was kind of unclear to me how you take one post and turn it into this whole uh you know. How do you find like posts that form some sort of story? Yeah, yeah, yeah. So I think the core insight.
I think the core insight is that uh first we pick the central posts well because they're they're posts that are very likely to have mattered. Uh and they have lots of lots of quote tweets. They're cited lots of times, right? And then the actual way we build the strands um is so you get the post, then you take all the times it was quoted, all the posts that quote the original one. Then we take all of those, we do semantic search on them, uh and find you know the 50, if you want, uh, you know, most semantically similar posts.
Uh so that we have some chance of getting past posts, right? Posts from the past, because otherwise it would be all all in the future. Uh then for each of those uh we we get all their threads, right? Like all the reply trees. Uh and that gives us something that looks kind of like this. Uh and uh did I I at first when you asked your question, I thought you were asking about how to how to stitch the strands together, but uh was that not the question? Yeah, some like maybe we can also do that part offline if we're because we're a bit behind schedule.
Okay, uh if there's any last quick questions you'll have, or we'll move on.
Okay, so we'll move to uh Billy, the last talk. And thanks again, Francisco. You're welcome. Thank you.