July 7th, 2024 Open AGI Summit Brussels
Chi Jin, Electrical and Computer Engineering Professor, Princeton University
Full Session Recording
Talk Notes
The Present and Past:
- The current generation of AI depends on human-generated data
- Models sizes have increased significantly, and with this, the quantity of human-generated data in creating these models has also increased significantly
- As depicted in this chart from Epoch AI, this means that models will soon approach the total quantity of human-generated data.
Future: Self-improving AI
-
The first tool you can use to create self-improving AI is self-evaluation. You can create a two-agent system where:
- One is a teacher that gives rewards and tells you how to improve
- The other is a student agent that learns from this reward system
- This is the kind of system behind Generative Adversarial Networks
-
The second tool you can use is self-play, adversarial training
- The model learns corner cases and learns to be robust
-
Only through multi-agent learning can you achieve superhuman performance
- This has already been showing in chess, go, and strategic games like Starcraft
-
Still, there is a lot of room to improve in areas like mathematical reasoning and coding tasks
How do we achieve improvement through multi-agent learning?
-
We look at the solution concepts that we would like to find:
- One concept is equilibrium. Finding equilibrium is at the core of game theory.
- Still, there are concepts beyond equilibrium. One of these is rationalization. For example, you don’t want to play clearly dominated actions.
- In equal games, you may also seek an equal share.
-
Game design is also essential to developing multi-agent systems
- In a GAN, for example, you do self-critiques
- You can have a lot of agents cooperating or competing with each other.
-
Finally, you need benchmarks to evaluate multi-agent systems:
