000Initialising
Orrery

Orrery / computational core, build 4.2

A machine that keeps rearranging its mind.

  • Movesteer the core
  • Hoverattract its nodes
  • Clickopen the inner layers
  • Scrollseparate the shells

Layer 1 — Identity

We build instruments for looking inside thinking machines.

Orrery is an independent research studio. We make the tools that let people see what a model is doing while it does it — not dashboards after the fact, but live, navigable views of computation.

Everything on this page runs on the same idea: a system you can steer, perturb and open up, and that answers back.

Founded
2019
Researchers
11
Open tools
23

Layer 2 — Capabilities

Five shells, five disciplines.

  • Interpretability tooling

    L0 · activations

    Probes, feature maps and circuit tracing, packaged for teams who ship models.

  • Agent simulation

    L1 · rollouts

    Thousands of agents in reproducible worlds, instrumented down to the token.

  • Realtime visualisation

    L2 · WebGL / GPU

    Million-point views that stay at 60 fps on a laptop.

  • Inference systems

    L3 · latency

    Serving stacks tuned for observability, not just throughput.

  • Interface research

    L4 · humans

    How people should touch, question and correct a model.

Layer 3 — Projects

Selected work.

Layer 4 — Experiments

Running on the bench.

Run 0412Live

Self-sorting shells

Particles negotiate their own lattice without a target.

Run 0398Live

Gradient weather

Loss landscapes rendered as moving pressure fronts.

Run 0351Paused

Echo prompts

What a model remembers of the question it just answered.

Run 0277Archived

Sparse chorus

Sonifying 512 features as a slowly tuning choir.

Layer 5 — Contact

Bring us a system you can't see into.

hello@orrery.studio