High Variance with Danny Buerkli Building an Open LLM – with Antoine Bosselut
Nov 5, 2025
In this discussion, Antoine Bosselut, an assistant professor at EPFL and lead creator of the Apertus open LLM, shares insights into his groundbreaking work in transparent AI. He argues for public funding to support open models, explaining how accessibility enhances scientific research. Antoine dives into the cost-effectiveness achieved through public supercomputing and reveals the challenges of data compliance and talent acquisition in AI development. He also discusses the importance of community collaboration and the future of synthetic data in creating robust language models.
AI Snips
Chapters
Transcript
Episode notes
Open Weights Enable Scientific Audit
- Apertus is a large open-weight LLM trained on ~15 trillion tokens with released weights, checkpoints, and training data for full transparency.
- This level of openness enables true scientific auditability and customizable reuse for research and industry.
Public Money Should Fund Open Models
- Use public funding to build open, compliant foundation models that private firms avoid because they lack immediate profit incentives.
- Public investment creates foundational infrastructure and skilled people that enable downstream innovation and sovereign capability.
Public Compute Lowers Training Costs
- Training Apertus consumed ~10 million GPU hours and student/staff salaries around $3M, but public supercomputing cut compute costs to $5–7M.
- Publicly funded infrastructure can make otherwise unaffordable large-model training economically feasible.