INDEPENDENT RESEARCH & SOFTWARE
For what
comes next.
New ways to think. Better ways to build.
We explore the frontiers of AI—and put the tools in your hands.
01 / OUR SYSTEMS
One lab.
Many possibilities.
From the mathematics inside a model to the worlds on your screen. Our work connects research, creative software, and infrastructure.
5376 features
learned operator
5376 features
Axon Studio
Build neural networks as naturally as you build an idea.
Explore Axon StudioAether
An ambitious engine for building worlds, systems, and experiences.
Explore AetherModel J Cloud
Our vision for simpler compute, storage, and application delivery.
Explore Model J Cloud02 / THE RESEARCH QUESTION
Can intelligence
have a shorter
description?
We study whether complex neural transformations can be approximated by compact mathematical operators. The goal is useful intelligence with less machinery.
Read our researchA RICH RESPONSE.
Local approximation results. End-to-end behavioral preservation remains an open question.
03 / LATEST ANALYSIS
What we’re
looking into.
Current questions. Unfinished ideas. The next experiment that might change how we build.
Explore the analysisA smaller operator. The same behavior?
Our next question is whether compact hidden-state approximations survive inside the complete model.
Can rational updates buy more expression per parameter?
Continued fractions give us a different primitive to explore. Stability and fair comparisons come first.
Grow the graph. Hold the active budget.
A graph that stores more capability without turning every training step into an irregular, expensive traversal.
THE READING DESK / 24–30 SEPTEMBER 2026
Research news this week.
And our take.
Small controllers, large generative models.
A fresh control perspective on image generation—and a useful question for modular AI tools.
Read the original researchA small learned control module around an existing model fits a direction we want to explore in Axon: compose a powerful backbone with an inspectable, task-specific operator. Our interpretation is about modularity; this work does not establish that our transformer surrogates preserve behavior.
The finding & our analysisCoherence needs a memory of the world.
Long-form generation raises a familiar creative-tool problem: keeping one world consistent across many decisions.
Read the original researchFor Aether, the interesting idea is persistent world state: identities, relationships, and scene rules that remain inspectable across creative actions. Our interpretation is that generation should participate in a shared scene model instead of producing disconnected assets.
The finding & our analysis04 / PRODUCTS & FEATURES COMING SOON
The next
possibilities.
The studio, on your desktop.
A native workflow for visual AI development, connected to your Model J membership.
In developmentFrom editor to playable world.
Connect scene authoring, visual logic, and an integrated game runtime.
In developmentA home for what you build.
A simpler path from an application or experiment to running infrastructure.
In developmentIn the making. Explore milestones and dependencies on the roadmap; launch dates are not yet announced.
05 / FROM THE LAB
Ideas in progress.
The search for a smaller description of intelligence.
What structured operators reveal—and what a good hidden-state score cannot tell us.
Powerful tools should invite you in.
Why visual workflows, accessible experiments, and honest capability boundaries matter.
A network that grows where learning asks it to.
A proposal for rational graph cells, useful connections, and responsible growth.
The tools to imagine
should belong to everyone.
Explore for free. Build deeper with a membership. Help us make the next generation of tools possible.