Real actions, not just answers
Execution inside real business systems.
The next curve
AI that keeps learning after it ships, on the hardware it ships with. We build the mechanisms for it, measure every claim against a declared scope, and start where models already act: real business systems.
For a decade, the answer to every gap was more scale. Many now argue that pre-training gains are flattening and that human-written data is finite. And a model that has already shipped cannot be scaled after the fact. Most deployed AI is frozen the day it leaves the lab.
We work on the other curve: learning after deployment, inside a fixed compute and memory budget, without losing what was learned before. It is an open problem, and we treat it as one. We publish what we measure, with its scope attached.
Where that work meets real systems today. One gateway between your models and the systems they act on: it turns intent into an action you can trust, and holds that trust across an entire task.
Execution inside real business systems.
Checks before anything irreversible.
Identity and state that last the whole task.
Our research program studies how a model can keep learning within a fixed resource budget: what it writes, what it keeps, and what that costs. Results are published with their code, and patent applications are filed in Korea.
Attractor Dynamics Inc. was founded in Seoul in June 2026 by Jeonghoon Lee, who leads its research and product work. Before founding the company he spent fifteen years in B2B sales and started a company in Japan. He works in Korean, English, and Japanese.
We publish our research with its code, file patents on what we build, and say plainly what each result does and does not show.
If you build models that have to keep working after they ship, or agents that need to act in real systems, we would be glad to hear from you.