What connects two people? It’s a big question, and it depends on context. Most people in my life connected through shared interests, but funny enough, we often bonded less over what we liked and more over what we disliked. So it made it easy to exchange ideas and put each other onto new things. It’s such a big question that I’m going to cherry pick one facet of connection: people can connect through an idea.

So what is an idea in network terms? Maybe an idea is a collection of posts. Some ideas are easy to understand; others are harder. Part of what makes things interesting is similarity, but things can also be interesting because they’re completely different, or because of a playful, particularly resonant shift. Let’s call that serendipity (my alchemical gold). 

So how do I imagine ideal connection mechanisms or how do I imagine connecting with people through ideas?

Here’s a naive approach I’ve been experimenting with in my own life. Suppose I have a space. I can run my files through an embedding model (no LLM tho) to produce vectors. If the space is organized by topic, we should see tight clusters, signaling strong relatedness. If I’m disorganized (not judging), we’ll see sparse, distributed vectors, signaling weaker relationships between objects in the space (it looks like noise). Everything is related to everything. 

Now suppose we do this across all my spaces, and maybe across other people’s spaces. I might discover interesting content in someone else’s space because a sample from their space fits within a cluster in mine. If their file fits in my space, I’d maybe want to be put onto it (gimme). We can even trade files: a seed from my space fits a cluster in yours, and vice versa. File for file.

What I’m trying to highlight is that there’s a way to use machine learning, short of what we usually call AI, to connect people through ideas. 

Not by force feeding content, but by subtly making people aware of places where alignment could exist. Little sparks that might light a fire. I wrote about those starting points here: On Agency, Resonance, and Rabbit Holes.

The flow I have so far is:

1. Target a space and run embeddings to expose clusters.
2. Use those clusters to infer: is this space topical, and what is it about?
3. Pick a seed from a cluster.
4. Search for similar seeds in other spaces.
5. Ask two questions: would their seed fit in my space, and would mine fit in theirs?

The above renders decently acceptable results and its not novel. It currently fails at its boundaries: I don’t want someone’s personal nordic foliage space mixing into my mesoamerican foliage space just because the vectors are close. Shit, it’s a start though. What I’m looking for is spontaneity and serendipity, with fine opinions (I’m trying to avoid the word “taste” way too hard) and context awareness. Optimizing for that is hard, and maybe unavoidably a little random.

There was a concept I was playing with Roman back in the day called bands of interest. I still think about it a ton, and it has two paths.

1. Could we take a concept and estimate its boundary (think EQ envelopes)? Through that, we might see where one person’s interests begin and end. There could be an overlap region where one person’s bounds taper off and another’s begin. What lives in that overlap? It sits near the outer bounds of both envelopes, and maybe that’s where horizons expand.
2. The second path was less about the contents and more about the form. Plot the shapes of ideas. Trace their contours. Then strip away what’s inside each shape (the posts, the examples, the obvious semantics) and compare the shapes themselves. At that point, the question changes: not “what is this idea about?” but “what kind of structure does this idea produce?”

Maybe two concepts with completely different content still share the same geometric fingerprint. Maybe the absence of shape is itself a shape. If that’s true, then shape can become signal.

And if shape is signal, we can use it predictively. Instead of searching directly for known intra-space shapes or known overlaps, we can search for those geometric fingerprints under relaxed constraints (not 1 to 1 fingerprint matches), then inspect what semantic or cultural structure those patterns correspond to. Geometry becomes a discovery interface for meaning and ideas.

Why am I interested in this at all? Because I strongly believe connecting people through ideas is core to network coordination: post.

What connects two people? It’s a big question, and it depends on context. Most people in my life connected through shared interests, but funny enough, we often bonded less over what we liked and more over what we disliked. So it made it easy to exchange ideas and put each other onto new things. It’s such a big question that I’m going to cherry pick one facet of connection: people can connect through an idea.

So what is an idea in network terms? Maybe an idea is a collection of posts. Some ideas are easy to understand; others are harder. Part of what makes things interesting is similarity, but things can also be interesting because they’re completely different, or because of a playful, particularly resonant shift. Let’s call that serendipity (my alchemical gold). 

So how do I imagine ideal connection mechanisms or how do I imagine connecting with people through ideas?

Here’s a naive approach I’ve been experimenting with in my own life. Suppose I have a space. I can run my files through an embedding model (no LLM tho) to produce vectors. If the space is organized by topic, we should see tight clusters, signaling strong relatedness. If I’m disorganized (not judging), we’ll see sparse, distributed vectors, signaling weaker relationships between objects in the space (it looks like noise). Everything is related to everything. 

Now suppose we do this across all my spaces, and maybe across other people’s spaces. I might discover interesting content in someone else’s space because a sample from their space fits within a cluster in mine. If their file fits in my space, I’d maybe want to be put onto it (gimme). We can even trade files: a seed from my space fits a cluster in yours, and vice versa. File for file.

What I’m trying to highlight is that there’s a way to use machine learning, short of what we usually call AI, to connect people through ideas. 

Not by force feeding content, but by subtly making people aware of places where alignment could exist. Little sparks that might light a fire. I wrote about those starting points here: On Agency, Resonance, and Rabbit Holes.

The flow I have so far is:

1. Target a space and run embeddings to expose clusters.
2. Use those clusters to infer: is this space topical, and what is it about?
3. Pick a seed from a cluster.
4. Search for similar seeds in other spaces.
5. Ask two questions: would their seed fit in my space, and would mine fit in theirs?

The above renders decently acceptable results and its not novel. It currently fails at its boundaries: I don’t want someone’s personal nordic foliage space mixing into my mesoamerican foliage space just because the vectors are close. Shit, it’s a start though. What I’m looking for is spontaneity and serendipity, with fine opinions (I’m trying to avoid the word “taste” way too hard) and context awareness. Optimizing for that is hard, and maybe unavoidably a little random.

There was a concept I was playing with Roman back in the day called bands of interest. I still think about it a ton, and it has two paths.

1. Could we take a concept and estimate its boundary (think EQ envelopes)? Through that, we might see where one person’s interests begin and end. There could be an overlap region where one person’s bounds taper off and another’s begin. What lives in that overlap? It sits near the outer bounds of both envelopes, and maybe that’s where horizons expand.
2. The second path was less about the contents and more about the form. Plot the shapes of ideas. Trace their contours. Then strip away what’s inside each shape (the posts, the examples, the obvious semantics) and compare the shapes themselves. At that point, the question changes: not “what is this idea about?” but “what kind of structure does this idea produce?”

Maybe two concepts with completely different content still share the same geometric fingerprint. Maybe the absence of shape is itself a shape. If that’s true, then shape can become signal.

And if shape is signal, we can use it predictively. Instead of searching directly for known intra-space shapes or known overlaps, we can search for those geometric fingerprints under relaxed constraints (not 1 to 1 fingerprint matches), then inspect what semantic or cultural structure those patterns correspond to. Geometry becomes a discovery interface for meaning and ideas.

Why am I interested in this at all? Because I strongly believe connecting people through ideas is core to network coordination: post.