Spend compute where the input is new.

Why pay for a full pass on an input the model has already handled?

Most models do a full pass even on inputs close to ones they have already handled. A cache helps only when a request repeats exactly.

We ask whether a model's own history of passes can tell it when to run the full model, a cheaper path, or reuse an earlier result, and whether that adds anything a confidence score does not.

Three routes

1. New
   Runs the full model.
2. Familiar
   Runs a light path.
3. Already seen
   Reuses the earlier result.

The gate has to be right, not just fast. The failure that matters is an input wrongly called familiar, so that is the error we measure.

We publish the protocol before we run anything.

Every experiment starts with a written protocol: the workloads, the baselines, and the rule for what counts as a pass, fixed before the first run. We report every result, including the ones that fail, and we invite researchers to break the claim.

The stage A protocol will be published here and on GitHub before the first run.

Models that change how much work they do per input.

Our flagship hypothesis, FANN (Familiarity-Aware Neural Networks), asks whether familiarity, read from a model's own history, is a better reason to skip work than difficulty alone. Around it, we study the ways a network can vary its compute.

Research directions

• Dynamic neural networks

  Depth, width, or path chosen per input.

• Early-exit networks

  Stop once the answer is reliable enough.

• Mixture-of-experts

  A few experts active per token, and whether familiarity can guide the routing.

• Looped and depth-recurrent networks

  Reuse the same layers for as many steps as the input needs.

• Memory of past passes

  Remember what was processed, and reuse it when that is safe.

Stages

1. Stage A

   Policy

   A gate over frozen open models, tested on public workloads. This is the work in progress.

2. Stage B

   Architecture

   The behaviour trained into a model. Only if A holds.

3. Stage C

   Capability

   A system that improves with experience. A direction, not a claim.

Quantization, distillation, and pruning are tools we use to build the light path. We do not sell optimization services, a platform, or a dashboard.