Since May of this year, I’ve been running a series of experiments in my free time that might point to something surprising:
That meaning itself may have an underlying mathematical structure — one that can be detected, tested, and potentially engineered.
This work began as a curiosity about whether two agents, without sharing tokens or language, could converge on the same “truths” purely from the structure of their knowledge graphs.
Over dozens of iterations, I’ve found consistent signals that there is a structural law of semantics — a kind of “math of meaning” — that persists even when we try to break it.
I’ve been exploring this almost entirely alone.
So far, I haven’t met anyone who both understands the mix of mathematics and philosophy involved and is willing to take the idea seriously enough to engage with it in depth.
That’s part of why I’m posting here: to find people willing to stress-test, falsify, and expand on what I’ve found.
What I’ve been using:
- Tools: Python, NetworkX for graph theory, statistical validation with SciPy, and embedding models for semantic similarity.
- Data: Public knowledge graph datasets (e.g., ConceptNet) as a testing ground, with both positive and hard-negative examples to stress the system.
- Method: Iterative agent-based simulations where multiple independent “reasoners” attempt to build contradiction-free semantic graphs from noisy and incomplete input.
What I’ve been able to show so far:
- Structural Coherence: Even with noisy or deliberately “wrong” data, the underlying law manifests — when we violate it, the violations are specific and measurable.
- Contradiction-Free Memory Graphs: It’s possible to curate a representation of knowledge that is internally consistent and grounded in real-world truth, and the process seems to be generalizable.
- Cross-Agent Convergence: Multiple agents, starting from different inputs, can align structurally without direct symbol sharing, suggesting that semantics can be independent of surface language.
Why this might matter:
If this holds under more rigorous and large-scale testing, it could mean:
- A path toward language-agnostic AI reasoning that is falsifiable, interpretable, and robust.
- A principled way to detect and remove contradictions from knowledge bases.
- A foundation for AI alignment grounded in structure, not just statistical correlations.
I’m sharing this here because LessWrong has a history of stress-testing unusual claims and poking holes in them.
If this idea is wrong, I’d like to know exactly why — if it’s right, then we might be looking at a core property of reality’s “semantic substrate” that deserves serious investigation.
I’ve documented experimental runs, statistical tests, and the code used to generate them.
Happy to share specifics or rerun experiments with suggested parameters.
If you suspected you had found a physical law governing meaning, what would be your next falsification tests?
And if it survived those, how would you prove it’s not just an artifact of the dataset or method?