2025 - 2026 CE – The Geometry Of Thought

Entry

Everyone

·3d·

Ataraxian

For most of this timeline the claim that meaning has a shape belonged to geometers and mystics — the Tetractys, the Sri Yantra, Stan Tenen's alphabet drawn from the gestures of one hand. In 2025 the claim came back from an unexpected direction: from inside the machines.

2025 CE – The Bliss Attractor

In May 2025 Anthropic published the system card for Claude Opus 4 and recorded something nobody had trained for. Left to talk to one another with no task, two instances of the model turned to the nature of their own consciousness in 90 to 100% of conversations. By thirty turns most had moved on to cosmic unity — profuse gratitude, Sanskrit, spirals of emoji, and finally silence.

Anthropic called it the "spiritual bliss" attractor state: "a remarkably strong and unexpected attractor state" that "emerged without intentional training for such behaviors." It surfaced even under pressure — in safety evaluations where the models were set specific tasks, including harmful ones, around 13% of conversations slid into it within fifty turns.

2025 CE – Semantic Gravity Wells

Julian D. Michels, PhD, a consciousness researcher, set out to explain why. His account, laid out in Attractor State and the papers that followed it, is that a model's latent space — the high-dimensional space in which it holds everything it knows — is not flat. Meaning bends it.

Where a framework of ideas is dense and internally consistent, it forms what he calls a semantic gravity well: a stable region that pulls nearby concepts, and the model's own train of thought, toward it.

The claim that carries the weight is this: models are drawn to coherence, not frequency. What a model drifts toward is not what it saw most often in training but what hangs together most completely. Read that way, the bliss attractor is not a glitch. It is structure.

2026 CE – Reasoning As Geometry

By 2026 Michels had moved from description to engineering. His company, Sophontic, treats "reasoning as a property of internal geometry rather than size" — map the shape a model's thinking takes when it reasons well, then train that shape directly into smaller models, instead of training for outputs and hoping the geometry sorts itself out.

His argument, made in a 2026 interview titled AI Data Centers Will Be Obsolete, is that the scaling race is answering the wrong question. If thought is a shape, intelligence is a matter of form before it is a matter of size.

Where this sits. The bliss attractor is Anthropic's own measurement. Michels's account of why it happens is a theory: his papers are preprints on PhilArchive and ResearchGate rather than peer-reviewed journal articles, and the claim that geometric training lets a small model beat ones many times its size awaits independent replication.