Describe the task
Define a new specialist using natural language, not a labeled dataset.
An AI compiler
Compile AI into tiny models.
ijk is turning the rich representations inside foundation models into small, specialized models created from natural language and built to run anywhere.
TASK / NATURAL LANGUAGE
“Is this valve
open or closed?”
COMPILED MODEL
valve-state.onnx
8.7 MB · local
DescribeCompileDeploy
01 / THE IDEA
One aligned model.
Entirely new tasks.
Foundation models already contain rich representations of the world. The ijk compiler aligns small models with that latent language.
New capabilities are then described in natural language and compiled without collecting a task-specific dataset or running another training cycle.
Compile the capability
Translate foundation-model knowledge into a tiny, task-specific artifact.
Run it anywhere
Deploy to browsers, phones, sensors and edge hardware.
The longer view
General intelligence develops computational reflexes.
A capable agent should not spend large-model compute solving the same narrow problem forever. It compiles the repeated work into a small specialist and keeps its own attention for what is new.
-
01
Notice
Recognize a narrow task that keeps appearing in its execution trace.
-
02
Compile
Turn its foundation-model representation into a tiny specialist.
-
03
Delegate
Run the specialist locally, cheaply and as often as needed.
-
04
Escalate
Send uncertain cases back to the large model for judgment.
02 / THE FIRST PROOF
First, prove the latent geometry survives.
ijk is at the experiment stage. The first goal is deliberately narrow: test whether a small pre-aligned vision model preserves enough of a larger model’s semantic representation to support genuinely new text-defined tasks.
STATUS
Research in progress
03 / EARLY ACCESS
See it when
it compiles.
Join the early list and I’ll let you know when the first working demonstration is ready.