αlphabell.
α · People · Jonas Bremer
Stylised avatar of Jonas Bremer

Jonas Bremer

Long-tenured contributor

Based in Berlin
Node-cell voronoi-19
ORCID 0000-7802-4570-0670

Research

Jonas is a senior contributor in the world-models axis and the co-PI on the CLD result. He spent five years working on Bayesian operator learning at Max Planck before turning to neural compositional dynamics — a turn he describes as 'the same problem, finally with enough scale to be seen clearly.'

He is the strongest internal advocate for what he calls the unification bet: that perception and prediction are best treated as a single learning objective rather than two. The lab's 70k-environment study is partly his argument cashed out; he expects to follow it with a multi-modal extension at multi-billion-parameter scale.

Jonas commutes between Berlin and the Bay Area anchor on a six-week rotation. He maintains a personal monorepo of operator-learning baselines that several other cells now use; the lab credits him in CHANGELOG entries.

Background

Ph.D. machine learning, TU Munich, 2015. Postdoc at Max Planck Institute for Intelligent Systems (Tübingen) 2015-2017.

Prior to alphabell: Max Planck IS; Volterra Cognition; OpenMind Research.

Selected publications

Full publications index →

Recent talks

  • Latent dynamics — the discrete-operator bet, ICLR 2024
  • Predict and perceive, MIT CBMM 2024
  • What world models cannot yet do, EurIPS 2025
Working with

Jonas is currently part of node-cell voronoi-19, working under the World models research axis. The cell is open to substantive correspondence from researchers working on adjacent problems; route requests through voronoi-19@alphabell.com or directly to Jonas at jonas-bremer@alphabell.com.