State of the art model architecture

We’ve come a long way since GPT-3 introduced LLMs to the world. The core of Avoriana Labs is a principled system for building out recursively self-improving LLMs with the latest model architecture innovations, for maximum bang-per-parameter.

Small to medium parameter count

Frontier models need expensive, exotic server hardware to run. More modestly sized models — 5bn to 500bn parameters — are runnable on consumer equipment today, and we believe will lag the largest frontier models in usefulness by no more than six months.

Intelligence density

Success at Avoriana Labs is defined by increasing the intelligence delivered in an equal-compute-minute. We build our models to deliver outcomes without leaning on excessive time and compute in thinking loops.

Recursive self-improvement

2nd generation AI labs lean on the ability of AI to quickly improve existing AI systems, accelerating every step of the model training pipeline to enable rapid productisation of the latest research advances.