Rebuilding social networks to eliminate misinformation and engagement bait. With 5 small changes, social media becomes social cohesion instead of social ills
Every social network ever built is a broadcast system. People post, but the algorithm decides what gets seen and promoted, moving attention but not understanding or well-being.
Every piece of content sits atop a graph of typed relationships. Browse a concept and the system shows you where you already are in the network: what you understand, what you are close to, what bridges you still need.
Concept detail view for METAPHOR. Lines descend from each typed relationship in the interface to the corresponding node in the semantic graph. Amber nodes are primitives: the irreducible vocabulary from which all concepts are built.
On social media, you share a representation of an idea: a container that different people decode differently. Say "justice" in a room of twelve people and you get twelve activations, each connected to different experience and assumption.
On the Idea Network, you share a structured concept with typed relationships to other concepts. The idea connects to what the other person already understands. It arrives intact because it travels through familiar territory.
Social media connections are social; you follow people. Every edge is the same: uniform, undirected, typeless.
The Idea Network's connections are semantic. You connect concepts with typed edges that describe how they relate: causes, contradicts, analogous-to, is-a. When you connect two ideas, you're not just saying they're related. You're saying the mechanism: the difference between a network of associations and a map of understanding.
Social algorithms find content similar to what you've engaged with before. More of the same, optimized for the next click.
The Idea Network finds concepts structurally equivalent to things you already understand, regardless of domain. P-waves in seismology and traffic jam propagation share no surface features. Their underlying structure is nearly identical. Understanding one makes the other immediately legible. No similarity algorithm would find this. It emerges from the graph. The section below shows exactly how.
On social media, contradiction is engagement. Conflict drives clicks. The algorithm rewards it because amplification is the only goal.
On the Idea Network, internally inconsistent reasoning fails to resolve. It doesn't evaluate as false; it falls apart at the primitive level, and the system surfaces exactly where the breakdown occurs. This is different from fact-checking. Fact-checking asks whether a claim matches reality. This asks whether the reasoning holds together. That question is substantially harder to game.
Social media's quality degrades at scale. More users produce more content, more noise, a worse signal-to-noise ratio. This is why every platform eventually struggles with the same problems.
The Idea Network improves with use. More concepts mean more connections, which mean shorter, more reliable paths between any two ideas: better analogies, more precise reasoning. No degradation curve. The system becomes more useful the more people use it.
To say a thing is something, you first need a boundary separating it from everything else. That boundary is NOT. It precedes IS. Every concept ever formed begins here.
Before the distinction, there is nothing to point to. After NOT draws a circle, there is an inside and an outside. The inside is the thing. The outside is everything the thing is not. Both sides come into existence because the boundary exists.
No boundary. Nothing to point to or name.
The boundary is NOT. The thing becomes thinkable.
Discovery Engine · Structural Equivalence
Not found through keyword matching: found because both phenomena decompose to the same primitive subgraph: a disturbance propagating backward through a pressurized medium.
When rock ruptures, a compression wave radiates outward. Each grain of rock squeezes forward then snaps back: traveling energy through kilometers of crust before any shaking begins.
disturbance
propagates through
pressurized medium
against normal flow
One brake event compresses traffic upstream. The jam wave propagates rightward: against traffic flow: car by car, while every individual driver eventually continues leftward.
Primitive Decomposition
Every concept decomposes to the same irreducible primitives. When two concepts from different domains share the same primitive subgraph: same nodes, same typed relationships: the system recognizes structural equivalence without any surface-feature matching.
Structural equivalences between concepts in unrelated fields surface automatically: without surface-feature matching. The system finds them because both concepts trace to the same primitives.
Incoherent reasoning fails to resolve at the primitive level. The system shows exactly where the reasoning breaks: not whether the claim is false, but whether it holds together.
Every concept is presented through the specific combination of familiar concepts each person understands best. Two people encounter the same idea through completely different references. Both arrive at the same understanding.
Two years after this thesis, a peer-reviewed team published PrimeNet in Cognitive Computation: a three-layer knowledge base organized around a small core of conceptual primitives, a middle layer of connecting concepts, and a large outer layer of entities. Nearly identical architecture: arrived at independently by professional AI researchers and an undergraduate iSchool thesis.
The problem this addresses: bridging neural networks (powerful but inconsistent) and formal symbolic systems (precise but rigid): is one of the most actively funded directions in AI research today. It's called neurosymbolic AI. As of now, no consumer product does what the Idea Network describes. The research infrastructure exists. The product hasn't been built.
The constraint that made this work: starting with a question that had no specification. How do humans share understanding? Follow the logic far enough and the structure emerges on its own: not from a brief. Open-ended prompts are harder than they sound because there's nothing to push against.
The most important thing I learned: the most interesting properties of a system often only appear after you've built enough of it to see what it implies.
And finally, I learned to formally enumerate the primitives early on. I treated the primitive layer as a theoretical foundation without specifying what the primitives actually were: whether they're culturally universal, how they're validated, who defines them.