Skip to content

We buildfinishedintelligence.

A research and engineering lab. We work on models, the products built on them and the silicon they run on — until each is ready to be used.

Thesis

SL/001 — 01

Intelligence is notfinishedat pretraining.

Pretraining produces capability. What makes it dependable — what it learns from, how it behaves, how it is checked — is decided afterwards. That is the work.

  1. Models

    The weights themselves — what they learn from, how they behave, and when they are finished.

    Explore models
  2. Products

    Software built on the models: interfaces, orchestration and agents that turn capability into work.

    Explore products
  3. Silicon

    Architecture research for intelligence workloads — compute designed from the workload down.

    Explore silicon
  4. Research

    Evaluation and measurement. Evidence for every claim, published with its limits.

    Explore research

Method

05 stages

From signal to finished.

  1. Choose what is worth learning from. Discard the rest.

  2. Form raw capability into deliberate structure.

  3. Lock behaviour to explicit intent, end to end.

  4. Check every part. Publish what the checks miss.

  5. Ship it finished — then keep it working.

Somber is a small lab with one standard:nothing leaves until it is finished.