AI has come a long way, but I would argue that the current most popular direction in the field, scaling with human generated data, is misguided. If we really want our agents to scale, we need to focus on how they will learn from their own data. This means creating reinforcement learning (RL) agents that are efficient enough to learn from a single, embodied stream of data.
Outline
0:00 – Intro
1:22 – The problem with the field
3:13 – The runtime learning perspective
7:24 – Point 1
9:31 – Point 2
12:26 – Point 3
14:30 – My research
Social Media
YouTube – https://youtube.com/c/EdanMeyer
X – https://X.com/ejmejm1
Sources:
Andy Barto’s talk: https://www.youtube.com/watch?v=-gQNM7rAWP0
source