ASI Research

ASI Research is an AI research lab working on scaling agentic reinforcement learning. Capability now comes from what an agent is trained against, not only from the model it starts as, so the training signal, the environments, and the infrastructure behind both are the research problem.

Most of the difficulty is in the environments. An agent can only learn what its environment can genuinely ask of it and genuinely check. We build RL environments at scale, terminal (TUI) and graphical (GUI), with executable ground truth rather than surface matching, because an environment whose reward can be gamed teaches the model to game it.

The other half is the model. We work on the architecture of the sequence model itself, attention mechanisms and what they cost at length, positional encodings and how structure is represented, fast-weight and delta-rule updates that let a model keep learning after pretraining ends, and the tokenization and sampling that bound what any of it can express. An agent that cannot hold a long rollout in context, or cannot carry what it learned in one episode into the next, is limited by its architecture, and no amount of environment design fixes that.

The longer arc is AutoResearch: using these systems to automate AI research itself, so that improvements compound instead of being hand-built one at a time.

We keep the group small and the agenda long. We publish what we learn and open-source what we build. If our questions are your questions, we would like to hear from you.