Shape Rotator Study / Storydeck

I'm going to give a talk to you guys about some of the work I've been doing with the Shape Rotator cohort over the last ten days or so. The structure of my talk is I'm going to start with an introduction. I'm going to then get into the Shape Rotator study that I did, and then talk a little bit about what I think the implications of that study are for the future.

I'm Brandon, for those of you that don't know me. I wear a lot of hats, but the two that are most relevant for us here today are, one, I'm a professor at Johns Hopkins in the Department of Statistics. And I also have an autonomous research lab called Calcifer Computing, and this work was sort of done within the Calcifer Computing, kind of frame. And the question that drives all of my work is, how does language spread and distort over time?

I think that this is a capability that's incredibly important to understanding both ourselves and the systems that we live in, and it's a capability that language models don't necessarily have out of the box. I usually attack this question, by breaking it down along a couple of axes. One of them I call the scale axis, and the other I call the structure axis. On the scale axis, I make a distinction between, studying what I call micro language and macro language. So when we're in sort of the micro case, I'm very focused on the language produced by discrete individuals. So a micro study might be, what do two people sitting at a dinner table across from each other talk about, in an evening?

And this is very similar to studying something like hydrogen bonding in chemistry, right? You have two individual, say, water molecules. How is it that they are interacting with each other? Conversely, on the macro side, this focuses on language produced by very large ensembles where you might not be able to identify the individual units that are producing the language. This is something like a country or a society or an online community. And this is very similar to the study of something like heat capacity in chemistry. It is a aggregate property that results from the fact that, say, hydrogen bonding is occurring. It's impacted by the micro microdynamics, but it's abstracted out into its own sort of macroscopic measure.

I also break things down along the functional versus structural axes. So on the functional side, I want to focus on the flow of the language process itself. What is being said? How is it being delivered? What is the intonation, the vocabulary, the pragmatics, et cetera? And this is very similar to studying the ecology of something like a river, looking at the river itself as a first-class citizen. On the structural axis, I focus much more on the terrain through which language flow occurs. And so this could be something like, how does the shape of a table affect the conversation that happens, you know, at that table? Or how does the shape of a room affect where people sort of congregate and how they converse, right?

Right now, we're in this very pyramidal structure where I am at the front of the room and you're all looking at me, and so that impacts the way that language is being received here. And this is very similar to studying the geography of a mountain down which a river flows, right? The peaks and the valleys and the turns of the mountain impact the direction that the river takes, and these things actually kind of feed off each other in this recursive process, but we won't get into that today.

I want to ground us here with a metaphor that guides a lot of my work, and that is that neurons are to action potentials as agents are to language. In the same way that action potentials between our neurons enable our body to move, language is sort of this thing that is transmitted between individual units that causes our society to move. And so if we can understand how that process works, we can help to nudge society into more, sort of harmonious basins. And today, we're going to be playing very squarely in the, micro category here. We're going to be looking at the language of the Shape Rotator cohort and how that structure affected the language over time and, what people were saying.

So the data we have today is the Shape Rotator corpus. This was collected in a variety of sort of, unstructured transcriptions over the course of time, things like office hours, brainstorming sessions, user interviews, the famous tea times. And one question we might have is, how are the shapes or subjects in this corpus rotating or changing over time? The simplest possible way that I could think of to answer this question is dump every single transcript in the Fable and then ask it what's changing. But obviously, there's kind of a problem with this. There's a lot of transcript data. Transcripts are incredibly, incredibly long, and so it's just too much data to chuck into an LLM.

So what do we do about this? This is the classic LLM memory problem, and we have to compact this data down so that it fits in the context. And so the way that I approach this is by mining an unstructured knowledge graph from the transcripts. This is a actual, kind of snippet of transcription that was taken from something that I said at one point about how word formation's interesting and how transcription models actually have some problems because they decode in the fixed vocabularies. And we can turn this with just a standard prompt into this sort of unstructured set of knowledge graph triplets. So Brandon is interested in word formation. Word formation might be relevant to the transcription team.

Transcription models decode into a fixed vocabulary. New words degrade performance of transcriptions And interestingly enough, there's a transcription error in this transcription itself. That word is supposed to be byte pairs, not bitters. And it turns out that word is new enough that the transcription models don't actually get it right. And so I thought that was an ironic coincidence. And so what I end up doing is just passing all of the summaries, the KG triplets, and some metadata about the transcripts to a language model along with a user query. And that lets us have a general system where we can chat with all of the transcripts throughout the entire duration of the cohort and ask things about, can we extract stories?

Can we make inferences about relationships? Can we do some kind of question answering? And that resulted in a system that I am calling Storydeck. Storydeck is effectively a system that you can go in and extract a set of different kind of longitudinally transcript-powered insights. This happens to be my story, using this LLM system. Each of these different cards that you see on the scene, sort of flashing in front of you are sets subsets of the overall knowledge graph that were extracted from the conversations. And so each of these is sort of like a grounded piece of my story with Flashbots. And if you zoom into any one of these in particular, I'll just pick, say, I don't know, this one.

You'll see, like, okay, this happens to be the case that Shashank is characterized as an ascending coordinator is that triplet, and you get kind of the line in the transcript where that knowledge was extracted. And that's actually very relevant to a story that we'll tell a little later. Another thing we can do with this ooh, that's very zoomed in, is we can actually infer the strength of relationships between pairs of participants over time. And so what I'll do for you now, if I can move these boxes out of the way, goodness, is play a little animation of the relationship strengths between pairs of participants throughout the duration of the cohort. And something very, very interesting happens, the entire graph kind of lights up for a little bit, it goes away, and then the week of May twenty-fifth, everything becomes effectively one very large connected component.

One very natural question is, okay, what changes between May fourth and May twenty-fifth that causes the graph to light up with all of these sort of like mutual relationships? And as an aside here, I really like that the James, Etherea, Gonzo like triplet in that bottom right corner is such like a distinct, sort of community in this graph. I think it's really interesting.

But to investigate this question, you can actually just ask our model. So we have a query mode where you can talk with the transcripts. I ask it, okay, you know, we see that something is changing on May twenty-fifth. Tina seems to be becoming a lot more central to this graph. Why is that happening? And it turns out it's because that's when she showed up. Unbeknownst to me, who started working with the cohort ten or so days ago, Tina had, like, lost her passport, and she wasn't able to be here in person, and as a result of this, the entire cohort was paying this huge coordination tax. And as soon as she showed up, there was this massive spike in the connectivity of the network, and suddenly everyone starts to be more coordinated.

Beyond this, I can also look at the centrality of different nodes in this graph over time. This is Roughly, you can think of this as how important the nodes are in routing messages to each other, or through the network. And we see that Tina, you know, spikes, and we see Andrew, the other organizer, is sort of in second place here. But then we see Shashank is in third place. And this is somewhat interesting. He's not an organizer, and he kind of starts at no centrality and then spikes. And so it's like, okay, what happened to Shashank that is causing him to become so much more central in this group? And so again, we can ask our model, and it turns out after Tina came back, she shape-rotated him into a pseudo coordinator working on CoordinationOS.

And this nudge enabled him to become much more effective as a sort of message passer within the cohort. So knowing all of this, and now that we have this tool, what might the future look like given that? Well, there's some problems here. These triplets that I'm extracting are all hearsay. They're not verified, right? And so one thing that you could imagine that would be interesting is having an agent that could go and actually verify these triplets automatically. You can imagine if, like, you know, a team said they shipped something, go and look on GitHub. Did they ship it? That's a question. Another thing is there's a massive legibility bias here. In early versions of this system, James and Etherea were highly underrepresented in the transcripts.

And so I asked him one day, "What's up with that?" And he's like, "Well, we haven't really been focused on doing a lot of, like, talking about our project because we were busy shipping a three-floor installation." And so there's this problem where it's like data about what is happening in the world has to be discovered. You need some kind of agent that can go out and get it, for people that might not be paying the coordination tax. And that data might be hard to access. And further, if you don't have someone like Tina who's coordinating a bunch of people all at once and acting as a central node and is, like, eating that coordination tax, you might have, like, a very disharmonious cohort.

And so the big lesson I take away from this is that I think the information market right now is deeply inefficient, not only within this cohort, but on the broader internet. And my sense here is that there's a pretty large opportunity for coordinating agents to help facilitate efficiency in this market. And this is a very hard and subtle problem, right? There's the sort of like dual legibility questions we were talking about yesterday in terms of, like, revealing your information makes it easier to coordinate you, but it also makes you more susceptible to an adverse actor. But I think it's a worthwhile thing to pursue. And if you agree with me, you should reach out to me.

You can find me online — or on Twitter at calcoio.

I highly encourage everyone to come and play, with, Brandon's model. Fun fact, a few days ago, when Brandon was preparing this presentation, I, decided to play around and ask it some questions that I don't think I was supposed to ask it in the way, how to use it. So I asked it to predict the future about Shape Rotator OS. I'm not going to review the, what it says, but I was quite surprised that it line up, with what I would have said almost verbatim and in an unexpected way. So, yeah. So there's a lot that, there is to just from observing the social dynamics and also, understanding how each of us think.

Have you heard of Nicholas Christakis out of Yale? Okay. I was gonna ask. He has, He's been doing sort of related research for, a bit more than ten years. I was gonna ask if there was some relationship, but it would be very interesting to see kind of the intersection of those two bodies of work anyway.

In the graphs you showed at first, what were the relationships between the nodes? I didn't quite really understand the meaning of the structure.

So the way that I approached the, this transcript problem is routing everything through the, triplet grounded model. And so once you have this general language model where you have, like, the summaries and the triplet grounding and all of the transcript information in it, you can ask it to evaluate the strength of a relationship between a pair of participants, right? Or characterize that strength on a scale from one to ten, right? And so if you get those measures for every pair of participants every single day, now you have a set of graphs over time between all the participants, and I sort of just animated those.

Strength based on the people query, like, how many, how many attributes, one hundred triplets?

I literally prompted... So you should think And this is actually a really, really good point. I, want to talk about this a little more. Language models are incredibly good. If you give them the correct context or extensive context, they can do a very, very wide range of tasks. And so it's sort of a little bit of an open question, you know, what the model decided were the salient features in terms of characterizing a relationship's strength. But literally, once I had the, like, seven hundred thousand tokens of transcript context, in the form of triplets and summaries and metadata, in the model, I literally just sent it queries like, "On a scale of one to ten for the week of May twenty-fifth, how strong was Tina's relationship with Shashank?"

And it would give me back a numeric answer. And that's how I got the edge weights. So it's all just inference internal to the contextualized model.