"We are on the brink of an era in which AI systems begin to accelerate their own development and could eventually do so recursively."

Few people have been as deeply involved in the evolution of open-source AI as Andreas Köpf. From co-authoring PyTorch and helping launch Open Assistant to leading new research efforts at Open-Thought, he has spent decades building systems at the frontier of machine learning. As Europe seeks to define its place in the next wave of AI, Köpf argues that the biggest opportunities will come not from scaling existing approaches, but from pursuing fundamentally new ideas.

Andreas Köpf is an AI researcher and engineer based in Münster, Germany with more than two decades of experience in large-scale machine learning and building software companies. He is a co-author of PyTorch (NeurIPS 2019, now cited 70,000+ times) and co-founded Open Assistant, the first major open-source alternative to ChatGPT.

Together with Mark Saroufim he co-founded GPU MODE, today the world's largest GPU and CUDA programming community. He currently leads Open-Thought, whose Reasoning Gym was selected as a Spotlight at NeurIPS 2025 (top ~3%).

He was previously Head of AI at Aleph Alpha and founder of the robotics/computer-vision startup Xamla (2014). His career began in 1998 at PROVISIO, where over twenty years he served as Software Engineer, Chief Software Architect, and CEO/Managing Partner; he remains an indirect co-owner of PROVISIO today.

Why must this initiative exist - now?

We are on the brink of an era in which AI systems begin to accelerate their own development and could eventually do so recursively. If we get there, society's prosperity and competitiveness hinge on its ability to control and leverage synthetic ingenuity. The teams that define the next decade are forming now, and they need a home that takes them seriously.

Europe has a deep base of well-trained, highly motivated researchers and engineers with innovative ideas. SPRIND's NFAI provides room to materialize such ideas at meaningful scale here in Europe – without having to relocate.

Beyond money: what's the real 'operating space' teams get here?

Through SPRIND and the NFAI organizers and the jury, teams get direct access to university chairs, the German Mittelstand, leading AI researchers and deeptech investors. Such a network would otherwise be hard to assemble as a seed-stage startup in Europe. Besides this, SPRIND provides company-building support.

NFAI ties financing to progress and they negotiate contracts with HPC providers, including the EU AI Factories, on behalf of the funded teams. Teams are picked because they want to try something the dominant mainstream path is not — which makes room for moonshots and riskier bets that are less attractive for standard VCs.

"The teams that define the next decade are forming now"

What would a real breakthrough look like?

A real breakthrough opens a capability dimension that is out of reach for brute-force scaling.

Some concrete examples that come to my mind: A system that builds a better version of itself, in a controlled and observable way. Sparse architectures that are drastically more energy- and parameter-efficient than dense transformers. A solution to continual learning and adaptation. Reasoning systems with strong meta-cognition capabilities that explore and invent genuinely novel solutions. Models that are provably correct and meaningfully interpretable. Models that generalize substantially beyond their training distribution. Robots that model the complex physical world, learn from experience operating safely and efficiently in unstructured environments - in the real world. Realtime omni-models processing audio, video, 3D, tactile, actions and emotions. Swarms of agents that conduct long-horizon, autonomous research over weeks or months and stay aligned to the goals human operators set.

What responsibility comes with building foundation models?

The main responsibility is keeping increasingly agentic, increasingly capable systems behaving in line with what their operators and society actually want.

Models need to be genuinely useful, trustworthy, and broadly accessible. As AI becomes part of everyday life, the creators need to deliver on reliability, transparency, safety, and human oversight. Especially in Europe we should ensure that AI works across languages, cultures, and industries.

But the actual responsibility is to unlock the AI opportunities - not build the next slop generators - to accelerate science, improve healthcare and education, and increase productivity across society.

What's the biggest challenge for Frontier AI in Europe right now?

The major challenges are: capital, foreign tech dependency, talent retention, regulation, compute infrastructure, energy, and fragmentation. We need to build the ambition and realize the urgency to shape the next generation AI systems in Europe.

In the best possible future, Europe treats frontier-AI investment with the same seriousness it has historically applied to railroad networks, energy grids, and national defense - sustained public and private capital, deployed at scale. Promising teams stay and build here because the funding, the compute, and the ambition on offer match what they would find elsewhere. Compute infrastructure is European and competitive at every layer of the stack - cloud providers, semiconductor supply chains, and datacenters owned and operated by European companies. The best researchers and engineers stay, or come back, because compensation, compute access, and the seriousness of the work no longer require crossing the Atlantic.

Who should apply - and who shouldn’t?

All people with ambition, technical expertise, and the drive to build the genuinely new next generation of AI systems should consider applying. The form takes a few hours to fill out and the downside is small.

If brute-force scaling or reproductions of existing models is your main thesis, NFAI is not the right venue.

Contact person

Andreas Köpf

AI researcher and engineer, Open-Thought