Agi Meta-Objective: a Derivation
Event details
| Date | 06.05.2026 |
| Hour | 12:15 › 14:00 |
| Speaker | Clément Hongler |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Machines long learned to master games such as go and chess. But these games were not designed to make machines more cognitively able. Do such games exist and can machines create them to self-improve?
We consider the problem of building an artificial general intelligence by self-play on self-creared sequences of games, and of the associated problem of automated relevant skill discovery. We provide a concrete substrate where to formulate the problem and argue that if there exists a greedy optimization algorithm that achieves this, then the form of the meta-objective satisfies gauge symmetries that lead to an explicit formula.
Practical information
- General public
- Free
Organizer
- João Penedones
Contact
- Corinne Weibel