We started with a latency problem

Why should one model write a paragraph for another model to reconstruct its meaning? Why should a decision pass through a conversation before it becomes an action?

Our inspiration came from high-frequency trading: examine the path between information and action and remove work that does not need to happen there. Every conversion must earn its place. Every unnecessary dependency is a delay.

Our first target was 100 milliseconds to 0.1 milliseconds for a bounded, local decision exchange. We were not trying to make every reasoning problem instantaneous. We were asking how much work could be prepared, learned, or eliminated before the moment a decision was needed.

At first, we thought we were designing a language for speed.

Then we realized the more consequential question was not how little a model could say. It was what another model would need to understand it.

Human language should not be a compulsory intermediary

A model may already have identified the relevant object, recognized a change, evaluated a condition, and formed a judgment. Yet to communicate with another model, it is often asked to describe that work in words. The receiver must then reconstruct the distinctions the sender had already made.

We want to stop paying for the same understanding at every boundary.
This is not an argument against language. Human language is how we explain, question, negotiate, and hold systems accountable. It should remain available wherever people need it.

But machines should not have to perform an explanation for a human audience when no human is participating in the exchange.

We are not proposing abbreviated English. Nor are we proposing another serialization format. A shorter message is not necessarily an easier message to understand. Removing words can merely transfer more inference to the receiver.

The objective is to preserve the meaning needed for the next computation while eliminating the need to reconstruct it from prose.

That requires training, not just formatting.

Speed was the constraint that exposed the larger problem.

A private shorthand between two models might make one system faster. But it would leave us with another isolated system.

What happens when the receiver changes?

When a new model is trained?

When one system sees images and another reasons over events?

When an intelligence on a device needs a capability available elsewhere?

A language that works only between two particular checkpoints is not universal. It is a coupling.

Our ambition is a shared language that lets different models communicate without requiring the same architecture, training history, or internal representation.

They do not need to think alike. They need to agree on what an exchange means.
That means preserving distinctions such as which object is being discussed, what changed, what evidence supports a claim, and what remains uncertain. It means carrying enough context to avoid ambiguity without retransmitting an entire history. It means allowing the receiver to reject an interpretation it cannot support.

Universality is not one model absorbing every capability. It is different models becoming useful to one another.

The language must preserve uncertainty and authority

Compression is not progress when it deletes the qualification that makes a statement true.
A hypothesis must not become a fact in transit. A correlation must not become a cause. A proposed action must not become permission to execute it.

We want a language in which those distinctions are fundamental, rather than comments appended to an answer.

An exchange should preserve the difference between “this happened,” “this probably happened,” and “this would happen under these assumptions.” It should identify the evidence on which an assessment depends, so a later change can invalidate that assessment without forcing the entire system to begin again.

And understanding a request must remain separate from being authorized to carry it out.
The purpose is not to make opaque systems coordinate more efficiently. People must still be able to inspect the evidence, decisions, and permissions behind consequential actions. Human-readable explanations belong at that boundary.

We want to remove unnecessary translation, not accountability.

We will train for successful transfer, not impressive compression

The first measure of this language will not be how small its messages look.
It will be whether another model can use them correctly.

We will train senders and receivers around shared tasks, then test whether the learned exchanges survive unfamiliar models, new environments, and changing evidence. We will measure the complete path: encoding, transmission, interpretation, and the resulting decision.

A message that is ten times smaller but requires ten times more reconstruction has achieved nothing. A fast exchange that silently changes meaning is a failure.
The harder test is whether a new participant can learn the interface without rebuilding the whole system around itself. We intend to begin with narrow domains where correctness can be checked, then expand the language only when the evidence supports it.
The goal is not to declare a universal standard. It is to earn universality through transfer.
From faster decisions to connected intelligence.

The first application may be a security decision, a market response, or an operational exception. These are useful starting points because the exchange is bounded and the consequence can be measured.

But the ambition extends beyond any one decision.

A perception model should be able to pass an observation to a planning model without first narrating a scene. A specialist should be able to contribute one judgment without becoming the orchestrator of an entire workflow. A local model should be able to request a missing capability from a larger system without rebuilding the interaction as a conversation.

The world should not need to be flattened into text every time intelligence crosses a boundary.

We began by asking how to remove 99.9 milliseconds.

That constraint led us to a larger purpose: a language through which models can share meaning, combine capabilities, and continue one another’s work.

Speed was the first reason to build it.

Universality is the reason to keep going.