self‑improving systems

Anything that can improve itself stops being linear.

Nonlinearity is an early‑stage AI company. We build systems that feed their own output back into their own input — and we think that loop reshapes far more than the models running inside it.

01 / the shape of feedback

A straight line, until it touches itself.

Effort alone buys you a straight line. Let the result feed back into the thing that produced it and the same clock buys something else entirely.

ẏ = a + βy  ·  a = 1, y(0) = 0
feedback β0.25

with feedbackno feedback
02 / scope

The loop is not only about AI improving AI.

Anything that lives in computer and language space — anything that can be written down, run and checked — can be placed inside it.

01 / science

Discovery

Hypothesis, experiment and revision as a closed loop that keeps its own notes and sharpens its own next question.

02 / models & agents

Systems that rewrite themselves

Models and agents that read their own traces, repair their own tools and return stronger than the version that started the run.

03 / language

Everything writable

Code, proofs, specifications, plans. If it can be expressed in language, it can be generated, executed and refined.

03 / about

Early, on purpose.

We are a small AI company at the very beginning of its work. For now we are building and testing prototypes rather than selling them, and we would rather show something real than describe something imaginary.

More here soon.

If it can improve itself, it will not stay linear.

nonlinearity · 2026 · in progress