Memes are located in semantic space based on their relation to other memes. Those distances represent the amount of cognitive dissonance between the ideas. But these connections only exist once they are memed.
Since meaning exists only in these relations, the concepts themselves form a graph structure. The space is latent, which is to say, you can interpolate new ideas along its dimensions but they don’t actually exist until they’re reified.
Graphs are defined by their edges, so the connections between ideas are actually more important than the positioning of the nodes. If we ran a graph weight algorithm, nodes with more edges would be more central. This is the thoughtscape that most people actually live in.
We don’t float about freely in semantic space, interpolating smoothly between all possible variations of an idea. We walk the edges of the graph like trails through the wilderness, rarely straying from the path. For some people, this is all they do.
And because central nodes are inherently more likely to be connected to each other, there literally is a gradient to descend. If you follow the path of least resistance, taking the shortest route between each node and the next, you wind up circling the drain of discourse.
But of course this landscape is not fixed. Due to the heritability, variability, and selection pressure of ideas, memes evolve and create ecologies. That wilderness is teeming with life. Even the paths themselves are co-created by the activity of these beings.
Some memetic organisms are niche builders. The common brainworm, for example, erodes the loss landscape around itself. Anytime you follow a path A → B → C, where C is a brainworm, the brainworm burrows a tunnel to A, creating a shorter and more direct connection.
Now when you think of A you can immediately think of C. And any idea Z which you follow from A to C will also be connected. The more this happens, the more ideas lead directly to the brainworm, and the harder it is to escape from its area of influence.
It’s true, language and image models are a snapshot of a semantic graph. If your concept is in the graph, you have Become Meme
I like this! Complimentary thought, memes as cultural embeddings https://t.co/ixTOinGcWc pic.twitter.com/uSh7fT6nKK
— Mat Dryhurst (@matdryhurst) September 22, 2022
Great question! It is actually possible to project off the paths and into wild idea space. In fact some of us do it all the time. When humans do it, it’s called creativity. When machines do it it’s called unintelligible coping noises
how do you use a known point (meme) in the semantic space and a vector to “discover” a not-yet-reified meme? what even is a vector in this space?
— Ben (@BenRatkaj) September 22, 2022
That’s the thing, they’re not just metaphors. They’re ideas too! The nodes and the edges are made of the same material. This homoiconicity gives semantic thought its expressivity, flexibility and computing power
Those edges mostly metaphors
— ~riley (@riley_stews) September 22, 2022
The words of a sentence triangulate a point in semantic space. The same sentence reliably conjures the same meaning. A sentence is just a really long compound word. I understand this principle was discovered by the Germans
Hmm. A vector would be that subtle feeling you get when searching for the next word in a language dialogue. So in a sentence the next word is a range of possibilities (a tree) based on the seed (the preceeding words/ideas). The vector would be this seed.
— -Cantīde- is e-dhamma.🧘♂️☯️🙏 (@Cantide1) September 22, 2022
New ideas.. 1/
It is possible to meme in continuous spaces, like dance moves or melodies. But the variation in that environment is too high. We inevitably develop names and languages to protect them.
We can also improvise and make things up. This is easiest when using an amorphous medium like motion or sound. There is a relationship bt the constraints of the medium (words, text, clay, energy) and the allowable resultant idea space.
— -Cantīde- is e-dhamma.🧘♂️☯️🙏 (@Cantide1) September 22, 2022
We mix things up and find out what happens.
You can think of a new idea without memes, but to communicate you must attach to the graph.
When I refer to semantic space I mean a specific thing: an implied high-dimensional space composed of all the different concepts you have ever learned and the similarities and differences between them.
df how would you define “semantic”
— Ben (@BenRatkaj) September 22, 2022
It’s our ability to learn combined with our ability to project higher order relationships that creates this space. And the evolutionary pressure of it birthed a symbiotic ecosystem of info-based life we call “culture”
unified meme theory draws on this paper, "Embodiment vs. Memetics," by Joanna Bryson. she says humans have a special combination of temporal imitation and second order thinking that creates semantic space fertile for memetic vectors
— web weaver (@deepfates) May 25, 2021
preprint pdf here https://t.co/S7KGGXRXu9 pic.twitter.com/lPSt6uVeSV
Textual inversion is where you freeze the knowledge of a neural net and fine-tune a new token for a concept defined by a cluster of pictures. A lot like what we literally do when learning a new word: show examples and say the name of the thing! https://x.com/xlr8harder/status/1572968931924779008?t=wLWo3bX7RZ0QiSWkAzuo-A&s=19
To be extra clear, memes don’t only exist as words. They can be gestures, sounds, shapes… anything that can be imitated on observation. They still compose in the same way. A dozen pictures of your fursona, that’s a sentence in image space. And a sentence can be a word.
An original idea enters the graph wherever it first attaches, and from there must propagate in its own way. There are certainly ideas that have never yet been thought. And I wouldn’t rule out the possibility of ideas that are unthinkable by human minds
An original idea enters the graph as a best-fit? Would there be ideas that can't be attached to a given graph, either accurately or at all?
— Aleksi Liimatainen (@aleksil79) September 22, 2022