Building an open ecosystem for AI research

Connecting the dots

He opened with a decades-old vision of a book-sized computer that took about 25 years and a whole ecosystem to become today's laptop. Trustworthy, open AI is the same: build the individual parts now, so the ingredients exist in ten years. He told four stories.

1. OpenML

OpenML lets anyone share a dataset with one line of code, with machine-readable tasks and results that upload automatically, so experiments are comparable and reproducible. Lessons from ten years: in-person hackathons built the community, simple APIs and a shared vocabulary mattered, and open platforms struggle to fund engineers and upkeep.

2. Rewarding good data

He co-founded the NeurIPS Datasets and Benchmarks track because people won't build good datasets unless it counts for their careers. The track judges data papers on documentation, accessibility, maintenance and ethics rather than on novelty, and now requires data to be hosted with proper metadata. His lesson: reward people in the currency they value.

3. Croissant

Croissant is a shared metadata standard for ML datasets, built with the major dataset platforms. It describes a dataset in layers (basic details, files, record structure, meaning), so any tool can find and load it. Communities have added extensions for responsible-AI documentation, Earth observation and biomedical data. Lesson: build bridges, not another platform.

4. Fully open models

Many "open" models share their weights but not their training data or recipes. OpenEuroLLM is building fully open models for all EU languages, publishing data and methods as it goes. Some practical points:

• Train in phases and save the best data for the end (annealing).
• Predict performance before spending big compute, by fitting a scaling law on a ladder of small training runs:

• Build safety in at every step: respect opt-outs, scrub toxic and personal data, limit memorisation, test, then red-team.

From his slide: "The hard part isn't the architecture. It's engineering, data cleaning, and knowing exactly what data went in and exactly how you measured what came out."

What's next

He's building infrastructure to find any dataset and link it to evaluations and papers, including an MCP server that lets AI agents search for and load datasets. In his words: "What we need now is more people."