λ
"Theory can provide insights and confidence that experience alone cannot."

— Terence Tao, Fields Medalist

Latrix Rena

A research-driven extension for AI fundamental science.

An open-source initiative dedicated to transforming AI from "alchemy" back to "science".

Scroll to explore

π

The "Alchemy" of Our Time

The Dilemma of Current AI Research and Development

Success is Hard to Replicate

Dependent on accidental hyperparameters, massive data, and inexhaustible computing power. A model's success often stems from countless trials and enormous resource investments, rather than clear theoretical guidance. This "alchemical" development paradigm makes AI progress full of uncertainty.

Failure Cannot Be Explained

The internal logic of black-box models remains a mystery to us. When models underperform, we cannot accurately pinpoint the problem. Is it data quality? Architecture design? Or training strategy? Without theoretical foundation, debugging becomes like blind men touching an elephant.

Trust Cannot Be Established

Can we entrust the future of civilization to an intelligence we do not fully understand? Without solid theoretical support, the reliability, safety, and controllability of AI systems all face fundamental questions. This is not just a technical issue, but a major challenge for human civilization.

Latrix Rena

Theory-Driven, Reproducible AI Science Platform

Rena is not just another "model training framework." It's a new paradigm: starting from theory, compiling theory into executable verification systems, and packaging the entire process as reproducible "digital papers."

1

Describe Theory

In Theory Description Language (TDL)

Using Theory Description Language (TDL), express your AI theories, hypotheses, and inductive biases directly as code. TDL is not low-level Python scripts, but a high-level, declarative "language of ideas."

TDL: Transform Thoughts into Code
2

Compile Theory

Into Executable Reality

The Rena compiler automatically translates your TDL theory into two parts: 1) Test Suites: experiments to validate theoretical predictions; 2) Loss Components: embedding theoretical assumptions into training.

Rena Compiler: Theory Experiments
3

Reproduce Experiments

With One Click

All experimental results, theory code, and training configurations are packaged as .ltx "digital papers." Anyone can reproduce your experiments with one click and verify your theory. This is the essence of science: reproducibility.

.ltx: One-Click "Knowledge Containers"
λ
π

Become a "Lamplighter" of the Next Scientific Revolution.

Join the Latrix Rena community and build theory-driven AI science together

Rena is an open-source community of researchers, engineers, and thinkers. We believe that the future of AI should not be monopolized by a few laboratories, but should be built on an open, transparent, and reproducible scientific foundation.

Join Rena Researcher Forum

Exchange ideas, share theories, and discuss frontier problems with global AI theory researchers

Contribute Your First "Theory Suite"

Transform your research theory into TDL code for the world to verify and reproduce

Apply for Rena Scholar Program

Receive resource support to deeply explore theory-driven AI research directions

Latrix Rena: Bringing AI from Alchemy back to Science