Semantic opposites are not opposite in AI embedding space. Hot and cold point toward each other. They cluster together. They have a cosine similarity of 0.894, nearly identical, because they almost always appear in the same sentences, the same topics, the same contexts. "It was hot outside." "It was cold outside." Same structure, same co-occurrence patterns, same neighborhood in the vector space.

 This is not a bug. It tells us something profound about what language models actually encode. And it is the finding that reframed everything I thought I knew.

 Let me back up.

 In April 2026, I started running independent experiments on a question that would not leave me alone: Can meaning be represented as position in a geometric space? Not statistically approximated across billions of parameters. Actually, encoded as geometry.

I called the project KAIA, Knowledge Architecture for Intelligent Agents. The original motivation was practical: build a semantic reasoning system that runs on CPU hardware anyone can own, not just research labs with GPU clusters or individuals who can pay for ChatGPT and Claude subscriptions. The equity dimension was the starting point.

I ran 27 experiments in the language track, then a mathematical track, and then opened a geometry investigation. No prior knowledge of academic literature. I was building and running code, following the results wherever they led.

As the conclusions began to take shape, I searched the literature and found three independent bodies of research converging on the same structural ideas from completely different directions. Peter Gärdenfors at Lund University has been developing the Conceptual Spaces theory since 2000, arguing that meaning is fundamentally geometric. In January 2026, a neuroscience team published findings showing that the human hippocampus organizes word meaning along stable geometric axes, the parallelogram structure of analogies found in actual neurons. In October 2025, LLM interpretability researchers found that large language models accidentally learn geometric structure as a side effect of next-token prediction at scale.

I arrived at the same conclusion through experiments. That independent convergence is part of what convinced me the direction is real.

But the most interesting finding was not what I expected. It was what the data kept showing me that contradicted my assumptions.

The First Surprise: Opposites Cluster Together

When I ran a systematic map of how well standard Euclidean geometry preserves different types of semantic relationships, the opposition result stopped me cold.

I tested 62 word pairs that should be opposites: physical pairs like hot/cold and fast/slow; mathematical pairs like addition/subtraction and finite/infinite; and abstract human pairs like love/hate and good/bad.

Every single pair had positive cosine similarity. Not one pointed away from the other.

Physical opposites: mean cosine 0.83. Mathematical opposites: 0.72. Abstract human opposites: 0.67.

GloVe, the word embedding technology used throughout this research, is working exactly as designed. It encodes distributional co-occurrence: words that appear in similar contexts get similar vectors. And semantic opposites appear overwhelmingly in the same contexts. They share a topic domain. Hot and cold are both about temperature. Addition and subtraction are both about arithmetic.

GloVe encodes topic membership, not conceptual opposition.

This reframed everything. The 13-axis semantic space I had been building was not finding directions between conceptual poles. It was finding directions within topic clusters, the dimensions along which temperature words differ from each other, inside the temperature cluster. Real structure. Useful structure. But different from what I assumed.

The Second Surprise: Hierarchy Runs Backward

If opposites cluster together, what happens to the containment hierarchy?

I tested 8 chains: algebraic structures (field inside ring inside group inside algebra), number types (integer inside rational inside real inside complex), living things (poodle inside dog inside mammal inside animal inside entity), and five others.

Every single chain was inverted. Mean Spearman correlation of -0.82 between hierarchy depth and cosine distance from the root concept. The most specific terms were geometrically further from the root than the general terms. The hierarchy ran backward.

The mechanism is the same as for opposition. Poodle appears in narrow, specific contexts, such as sentences about dog breeds. Entity appears in broad, diverse contexts spanning philosophy, computer science, legal documents, and everyday speech. Different distributional profile, different position in vector space, large cosine distance, even though poodle is a subtype of entity.

The co-occurrence signal runs exactly backward from the ontological hierarchy.

I tested five approaches to recovering the hierarchy through geometric transformation. All five failed. The diagnostic that explains everything: the correlation between GloVe norm and hierarchy depth is 0.08, essentially zero. The signal is not there. No geometric transformation can recover information that was never encoded.


The Third Surprise: Change the Training Signal, Change the Geometry

Here is where it gets interesting.

I trained a Poincaré hyperbolic embedding, not on co-occurrence statistics, but on explicit (child, parent) relationship pairs. 39,242 pairs from WordNet. 60 curated mathematical hierarchy pairs. The training signal was the hierarchy itself, not the order in which words appear in sentences.

Mean Spearman jumped from -0.82 to +0.43. Seven of eight chains showed positive correlation. Three chains were fully recovered above the 0.70 target.

The geometry was right all along. Hyperbolic space is the correct geometry for hierarchy. The exponential volume growth of the Poincaré disk naturally accommodates the exponential branching of hierarchical structures. What was wrong was not the geometry. It was the training signal.

Co-occurrence encodes discourse patterns. Supervision encodes relationships. These are fundamentally different things, and they require different geometric foundations.


The Geometry-Forcing Finding

One more result. I tested whether the domain of a concept affects the geometric quality of its representation. The hypothesis is that mathematics, in which structure is determined by logic rather than cultural convention, should produce cleaner geometric axes than language does.

It does.

After optimization, the mathematical axes achieve a Gram matrix mean of 0.0673, below the 0.08 target. The language domain, starting from a more entangled baseline, requires more optimization steps to reach a comparable result. Both confirm what the theory predicts: the domain of a concept shapes the geometric quality of its representation, and logically forced domains start cleaner and stay cleaner than culturally contingent ones.

Temperature is geometry. Mathematical operations are geometry. Moral evaluation is messier because culture shapes it differently across corpora, communities, and time, and the distributional signal reflects that variation.

What This Actually Means

The picture that emerges from two months of experiments is this: meaning has at least two distinct geometric structures.

Within-cluster structure, the dimensions that differentiate concepts within the same topic domain, are captured well by Euclidean distributional embeddings. This is what makes king minus man plus woman equals queen work. Both pairs are within the same human/royalty topic cluster, and the direction from man to woman is parallel to the direction from king to queen inside that cluster.

Containment structure, whose concepts are subtypes of which, whose knowledge is prerequisite to which, is not in the co-occurrence data at all. It requires supervision. Explicit relationship pairs. And it lives naturally in hyperbolic space, not Euclidean space.

Current large language models learn from co-occurrence at an enormous scale. They acquire within-cluster structure implicitly, as a side effect of predicting the next token across billions of examples. This is why analogy arithmetic works in these models.

But the containment structure, the hierarchy, the "field is a ring is a group" relationships that underlie mathematical reasoning, the "poodle is a dog is a mammal" relationships that underlie commonsense inference, these are not reliably in the co-occurrence signal. They must be put in deliberately.

This research started as a question about hardware access: can we build AI that runs on hardware anyone owns? Two months of experiments have led to something more fundamental.

Current AI systems are not built on geometry. They are built on statistics. The geometric structure that interpretability researchers find inside large language models is emergent, accidental, and a side effect of doing next-token prediction at an enormous scale. Nobody designed it in. It shows up uninvited, and researchers find it afterward from the outside.

What I am trying to do is build the geometry deliberately, from the start. Not discovering it inside a statistical model after training. Design a system where geometric structure is the foundation, not the side effect. The experiments have already shown it is possible. The geometric structure is real, measurable, and behaves exactly as the theory predicts. The open question is whether it scales to something useful, full language coverage, real-world tasks, and generation quality, at a fraction of the cost of statistical approaches. That is what the research is building toward.