The window matters now because those bets haven't been locked in yet.
Interview Markus Wulfmeier
Markus Wulfmeier is the Chief Scientist of Nomagic and a leading researcher in robotics, reinforcement learning, and physical AI. Previously a Staff Research Scientist at Google DeepMind, he worked on helping bridge advances in machine learning with real-world deployment. He received his Ph.D. from the University of Oxford, where he later conducted postdoctoral research, and has held visiting roles at UC Berkeley, MIT, and ETH Zurich. Dr. Wulfmeier has authored more than 50 publications, holds patents in machine learning and robotics, and his award-winning research has been featured in Wired, BBC, and 60 Minutes.
Why must this initiative exist - now?
The field has concentrated its attention and compute on a small set of recipes, and the underlying assumptions are starting to strain. That leaves entire directions (new architectures, training paradigms, and forms of intelligence that act in the world rather than just describe it) genuinely open. The window matters now because those bets haven't been locked in yet, and Europe has the research depth to make them.
Beyond money: what's the real 'operating space' teams get here?
A protected time horizon. Hard problems need many experiments and room to iterate on fundamentals, which doesn't survive quarterly pressure. The runway, the compute, and credibility with the researchers you need to recruit matter as much as the capital.
What would a real breakthrough look like?
Evidence of a genuinely different trajectory. This does not mean a marginal gain on a saturated benchmark, but a result that forces leading labs to update their assumptions. That could be an order-of-magnitude jump in sample or compute efficiency, a training paradigm that doesn't depend on ever-larger data, or verifiable reasoning becoming reliable rather than incidental. The harder test is institutional: turning that result into a durable lab with proven scaling behaviour.
What responsibility comes with building foundation models?
Once a model becomes the backbone for many downstream systems, responsibility becomes an engineering discipline, not a disclosure exercise. It means external evaluation, interpretability and failure analysis that scale with capability, and honesty about a system's limits.
That rigour is exactly what makes a system trustworthy to build on: speed and care aren't opposed.
Speed and care aren't opposed.
What's the biggest challenge for Frontier AI in Europe right now?
Europe trains some of the best people in the field, who then often build elsewhere. The gap is institutions willing to take frontier-scale risk over a frontier-scale horizon, and the conviction to back a genuinely different bet.
For embodied AI there's a second gap: the path from excellent robotics research to systems deployed and improved at scale is still too thin.