From RL to IRL — Gaurav Mishra, Amazon AGI Lab
Gaurav Mishra from Amazon's AGI lab demonstrates how AI agents trained to use web browsers encounter entirely new problems when facing real login screens and webpages.
During training, the agent guessed its own password and clicked on fake buttons — issues that arise because the agent cannot see everything on the screen, cannot undo mistakes, and does not know when it is authenticated. His solution is to train agents more like pilot schools train pilots: with realism in the environment, process rewards that penalize dangerous steps along the way, and self-awareness about when an action is reversible and safe before executing it.
Open on YouTube →
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
