The Institute's research is organized around three interconnected pillars. They are not independent silos. The most important questions we ask require moving across all three simultaneously.
This page describes what we work on, why it matters, and what we are building toward. Publications, tools, and datasets will be added here as the work matures.
Quantum computing has moved from theoretical possibility to physical reality, but the gap between what quantum hardware can do in a controlled laboratory and what it can do reliably in a deployed system remains wide. Closing that gap requires work at multiple levels simultaneously: at the level of the physics, the error correction schemes, the hybrid architectures that combine quantum and classical processing, and the algorithms designed to run on them.
Our work in this pillar focuses on hybrid quantum-classical systems, architectures that do not attempt to replace classical computing but to augment it with quantum components where those components offer genuine advantage. We are particularly interested in the design of such systems for learning tasks, where the interaction between quantum and classical components raises difficult questions about what the system has actually learned and how one would know.
We also work on the security implications of quantum computing, specifically on the post-quantum cryptographic primitives that will be necessary as quantum hardware becomes capable of breaking current standards. This work is both theoretical and practical: we want to understand the mathematics and to build tools that implement it correctly.
Underlying all of this is experimental physics. Mads leads our engagement with quantum hardware directly, and we believe that researchers working on quantum AI need to understand the physical systems they are building on, not just the abstractions layered above them.
The field of AI has produced systems of remarkable capability whose internal workings remain largely opaque. This opacity is not merely an aesthetic problem. It creates serious risks in deployment, limits the ability to debug and improve models, and makes it difficult to assign responsibility when systems behave badly. We believe interpretability is one of the central unsolved problems of the field, and one of the most important.
Our work in interpretability is both empirical and theoretical. We study the representations that models learn: what is encoded, where it is encoded, and how it is used in computation. We develop tools that make these representations legible to researchers and to the practitioners who deploy AI systems. And we think carefully about what it would mean for a system to be interpretable in a rigorous sense, rather than merely explainable in a post-hoc, approximate way.
We also build investigative platforms: secure, open tools for researchers who need to study AI systems seriously, including their failure modes. We believe that the infrastructure for AI safety research is itself an important contribution, and that making such tools open-source is essential to the kind of collaborative scrutiny the field requires.
Maia leads this pillar, bringing to it a systems architecture perspective that treats interpretability not as an add-on but as a design constraint from the start.
Modern AI systems are built on mathematical structures that most of the people who deploy them do not fully understand. This is not an accusation but a practical consequence of how fast the field has moved and how poorly its foundations have been taught. We believe this gap is dangerous and that closing it is one of the most valuable things a research institute can do.
Our work in this pillar focuses on the mathematical and symbolic foundations that underlie both quantum computing and AI. This includes the statistical theory of learning, the algebraic structures that appear in quantum mechanics, and the logical and symbolic frameworks that allow us to reason precisely about the behavior of complex systems.
We are particularly interested in the role of symbolic reasoning in AI: how formal methods, logic, and type theory can be brought to bear on systems that have traditionally resisted formal analysis. This includes work on the verification of learned systems and on the design of architectures that incorporate symbolic structure in ways that make them more interpretable and more robust.
Alongside this research, we invest heavily in pedagogy. We develop frameworks for teaching these foundations in ways that are rigorous without being inaccessible, and we believe that this pedagogical work is itself a form of research. Understanding something well enough to teach it clearly is a higher standard than understanding it well enough to use it.