This episode features Ruslan Salakhutdinov discussing the rapid advancements and future of AI agents. He highlights their growing autonomy in tasks like coding and web usage, the challenges of long-horizon tasks and safety, and the emerging trend of multi-agent systems. While agents excel at automating routine tasks, achieving 100% reliability for critical applications remains a significant hurdle, necessitating human oversight and robust safety mechanisms.
Summarized by Podsumo
Agents are rapidly improving, particularly in coding and computer usage (web agents), now capable of handling tasks requiring several hours, a significant leap from minutes.
A major hurdle is defining *intermediate or partial rewards* (the "credit assignment problem") for tasks that take hours or days, as current reinforcement learning often only provides a single "correct/incorrect" signal at the end.
The future of complex task solving involves *orchestrating multiple specialized agents*, where a larger model plans, and smaller, cheaper agents execute sub-tasks, communicating back to the manager.
Despite impressive capabilities, agents are not 100% reliable and can be *confidently incorrect* or even destructive (e.g., deleting a database). Robust *guardrails, multi-agent verification, and human-in-the-loop systems* are crucial, especially for critical tasks.
While navigation and locomotion in robotics have advanced significantly (e.g., self-driving cars), *dexterous manipulation of diverse objects* in unstructured environments (like unloading a dishwasher) remains an "extremely difficult" unsolved frontier.
"Somebody gave full access to some of these agentic systems and the agent just deleted the entire database in eight seconds."
— Russ Salakhutdinov
"I would never use any of the agentic systems to book me a flight. Even if it's a simple task... I would never do that, even if it's 80% correct, that is at least 20% gap, where it will go and do something crazy, right?"
— Russ Salakhutdinov
"We have systems that can win math competitions and international math Olympics, but getting me a robot that can reliably unload my dishwasher, or load my dishwasher. It's just extremely hard."
— Russ Salakhutdinov