Empowering people

AI should expand human potential, not reduce humans to supervisors

Today’s AI systems are remarkably capable, yet often unreliable enough to require constant human oversight. Instead of creating, learning and solving meaningful problems, people increasingly find themselves reviewing, correcting and managing machine-generated work.

This is not necessarily the future of human-AI collaboration, but likely a limitation of today’s technology.

We envision AI that works more like an exceptional mentor or intellectual partner – helping people think more deeply, develop their abilities and spend more time doing what they love.

Achieving this vision may require more than scaling existing architectures.

Why Decentralised Intelligence?

We believe advanced intelligence should be able to operate, learn and adapt within its own environment – efficiently, locally and under the control of the people and institutions using it.

Much like the human brain, such systems should continuously adapt to their surroundings without depending on vast external computing infrastructure for every meaningful improvement.

We call this research direction Decentralised Intelligence.

Our ambition is to make advanced AI not only more efficient and accessible, but fundamentally better suited to human flourishing.

Our research

We research new AI architectures and learning algorithms that could make advanced intelligence radically more efficient, transparent and accessible – enabling more researchers, companies and institutions to build and own intelligent systems locally.

Our current focus is on post-Transformer architectures and significantly more effective information routing schemes.

We believe that the only way to preserve healthy society and avoid catastrophic concentration of power is to ensure economic viability of small, specialised AI models that can be deployed on premise.

• [May 2026 – present] – Building on our interpretability results we work on radically more effective post-Transformer information routing schemes for neural networks.
• [July 2025, paper, code, demo] – We provide a unifying view on neural computation and introduce Semantic Pullbacks as an effective explanation method for deep models. Our results suggest that pullbacks are often a more natural and informative backward signal than gradients.
• [March 2024, paper, code] – We introduce Semantic Features as a candidate building block for transparent neural networks that are both self-explainable and adversarially robust.