We build foundation models that reason about interventions.
Untiring Systems is an RSI lab. We work on scientific discovery, cybersecurity, and autonomous research.
Windmill is our model family.
Windmill predicts what an intervention will do when the mechanism underneath is only partly known. Often the intervention is its own.
Training mixes language modeling with causal modeling, long-context attention, and sparse compute, then adds agentic post-training.
Recursive self-improvement is the goal, and most of our research is a piece of it. A model that improves itself has to tell cause from correlation, run experiments that separate two explanations fitting the same data, and recover when a hypothesis fails. It has to stay coherent across long horizons. Compute is the binding constraint, so attention must not spend equal effort on every token, and sparse architectures have to buy capacity without buying flops.
The same reasoning shows up regardless of subject. Planning an experiment in a lab, tracing the root cause of a vulnerability, and predicting which training change will help the next model are versions of one problem. That is why a single model family can serve scientific discovery and cybersecurity at once.
We treat the architecture as a variable.
The standard transformer is where we start. When a conventional choice costs performance or compute, we replace it.
Every change has to earn its keep.
- Attention
- Routing
- Training objectives
- Context handling
- Causal supervision
- Expert specialization
- Development workflows
Answering questions is the easy case. The models we care about act on the world and learn from what comes back.
Windmill is the base. Around it we are building the machinery — training methods, evaluations, feedback loops — that makes each model cheaper to build than the one before it.