What it is
Study of open-endedness: systems that keep generating novel, learnable challenges and solutions without a fixed end goal. Autocurricula from multi-agent interaction and the case that open-ended search is the route to superhuman systems. The formal branch (UED, replay-guided curricula) has its own page: Unsupervised Environment Design.
Notes from the ICLR 2025 invited talk Open-Endedness, World Models, and the Automation of Innovation:
- legends of the field who explore this idea
- this is what we are doing as a company - enabling research that helps us get here
- it’s interesting that this seems to be a broader philosophical property
From Michael Dennis’s UED intro talk:
- other ideas I should read on
- these are ideas where we train models on real2sim and then use that to train sim2real
- Genie is an interesting area
- check Michael Dennis’s work for more ideas
Key sources
- Open-Endedness is Essential for Artificial Superhuman Intelligence — the definition to use (novel and learnable to an observer); argues FMs + open-ended search is the route to ASI
- Autocurricula and the Emergence of Innovation from Social Interaction — multi-agent origin of the idea; the problem problem and autocurricula from competition/cooperation
Related
- Unsupervised Environment Design — the formal branch, split out 2026-09-10
- World Models
- Reinforcement Learning - RL
- Reading lists not in this repo: World Models × Open-Endedness, UED Core Reading
To ingest
Talks (not arxiv)