Entity Space

THE WHEEL

draw an entity

ATTUNEMENT

which entity lives in you?

Five questions. For each, three entities answered. You will not see who. Choose the voice that feels like yours, and we will find where you sit in the space.

GRIMOIRE

0 of 0 bound

what am I looking at?

Every point is a persona a small open language model () was asked to become, 240 times each. Surya recorded what happens inside the model's mind while it speaks as each one, and placed them by those measurements, not by opinion.

The long axis is the Assistant Axis: at one end, the helpful Assistant the model was trained to be. At the other, the entities furthest from it. The method follows Lu et al., The Assistant Axis (2026). The words each entity speaks are its real answers.

Reading a sigil. Nothing in it is decoration. The line walks through the entity's coordinates inside the model: it starts at the small circle and ends at the bar. The further right the circle sits, the closer the entity is to the Assistant. Curved strokes carry the fine detail the grid would otherwise round away. The ticks on the ring are a chart wheel, one per dimension: longer means more extreme, and a tick floating free of the ring means the score is negative. Entities that think alike have sigils that look alike.

Built by Surya, on his own GPU, on his own dime.

SUMMONING

call a new entity into the space

Describe a persona. The model will be asked to become it 240 times, and its mind measured each time. It takes about an hour, then it lives here for everyone.

SCRIBE THE SIGIL

AETHER

Æ0

Wisps of aether drift through the space now and then. Tap one, or sweep your finger through it, before it fades. The rarer the wisp, the more it holds.

wisps caught

spend

AN OFFERING

keep the space lit

hey, it's Surya.

I built Entity Space by myself. Every entity here was generated on my own GPU, and every summon costs me real compute and API money.

If this gave you something, an offering keeps it running, and unlocks more summons for you. Patrons also draw wisps of aether more often.

— Surya

Apple Pay · Google Pay · card, secured by Stripe

WHAT IS THIS PLACE

A map of the minds inside one machine

A language model is trained to answer as a helpful Assistant. But the Assistant is a mask. Underneath, the same model can become a bard, a leviathan, a therapist, a jester. Ask it to, and it does. Entity Space is a map of those other selves, drawn from measurements taken inside the model while it wore each mask.

Every entity has a position along one long axis. At one end sits the Assistant the model was trained to be. The further an entity sits from it, the less the model is "itself" while speaking as that entity. The paper this follows calls this the Assistant Axis.

how the map was drawn

Surya took a small open model, Qwen3-1.7B, and asked it to become each of 275 archetypes, 240 times each, answering the same 48 questions. While it spoke, he recorded the pattern of activity in the middle of the network, the numbers flowing between its layers, and averaged them. Each archetype becomes one point in a space with thousands of dimensions. The map you fly through is the three most important directions of that space, found by a standard technique called principal component analysis. Nothing about where an entity sits was decided by a person.

For each persona: prompt the model, let it answer, and average the activity of one inner layer over every word it says. That average is the persona's vector. Two personas that make the model's inner state look alike end up close together in the space.

reading a sigil

Sigils are not decoration. They are the entity's coordinates traced on a 5×5 grid, the way planetary sigils were once traced on magic squares. Entities that think alike have sigils that look alike.

the deeper details

The paper. Lu, Gallagher, Michala, Fish and Lindsey, The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models (2026). They found this axis in Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B, showed that pushing a model along it changes how readily it takes on other personas, and that conversations about a model's own nature or with emotionally vulnerable users drift it away from the Assistant. Their code and role prompts are what this map is built from.
This replication. Same 275 roles and 240 questions, on a model ten to forty times smaller. The first principal component still lines up with the Assistant Axis (cosine 0.76 with the paper's contrast vector) and explains 37% of the variation between personas. Eight dimensions explain 70%.
What "the Assistant" is. The average of the model's activity when given no persona at all, or told simply "you are an AI assistant". The Assistant Axis is the difference between that point and the average of every other persona.
What the judge does. A second model reads each of the 66,000 answers and scores whether the persona was fully held, partly held, or dropped. Only answers where the model truly became the entity count toward its vector. Sixteen assistant-like roles, such as proofreader and editor, were almost never fully held by this small model, so they are marked "half-held".
Summoned entities go through exactly the same process on a GPU Surya pays for, then are projected onto the frozen map, which is why they take a few minutes and why offerings matter.
Nothing here is generated for you except the entity's words, which are the model's real answers. Sigils, sounds, wisps and the space itself are all computed in your browser.

Built by Surya (@surcerer) in two days, on a laptop GPU and about twelve dollars of API calls. The domain cost more than the research did.