We want to understand how intelligence works.
Through mathematics, experiments, and working code, we explore the foundations of learning, reasoning, and computation.
Explore our researchCan a parameterized quantum circuit learn context-dependent next-token distributions with better parameter efficiency or computational scaling than a comparably sized classical model?
We’re building the pieces that let us test this directly: the simulator that runs the circuits, the engine that runs the classical baselines, and the transformer we compare against.
How we’re approaching itWhat we’re curious about
Our questions start small and get more precise through experimentation. These are the areas guiding our current work.
Learning and representation
How neural networks organize information, and how their architecture shapes what they learn.
Reasoning and generalization
Why models succeed at some problems, fail at others, and what actually accounts for the difference.
Foundations of computation
Classical and quantum systems, and the mathematical limits and possibilities of each.
From the notebook
Experiments and the thinking behind them, newest first. Every entry is labeled Plan, In progress, or Result.
How close can a from-scratch autograd engine get to PyTorch?
A fair micro-benchmark of aaronnet, our in-house CPU autograd engine, against PyTorch on a two-layer MLP.
Taking the transformer apart
Building a small causal transformer on aaronnet and running one controlled ablation to see what each component contributes.
A state-vector simulator, built from first principles
Working through the mathematics of quantum states, gates, and measurement by implementing a small simulator in Rust.
An open record of work in progress.
We believe good research should make its questions, methods, and limitations visible, not just its conclusions. This site is where we publish ours as the work happens.
About the lab