Ruslan Salakhutdinov, an AI researcher from Carnegie Mellon, discusses autonomous agents that could work independently. Current models hit success rates around 45-50%, and jumping from 80% to 90% accuracy is extremely difficult. An unrestricted agent once erased an entire database in seconds, raising safety concerns as systems gain more autonomy.
Again, a lot of these things I think in the future will just be automated. And we're seeing success rates of existing models hitting, you know, maybe like 45, 50%, which is very impressive.
One of my favorite questions on the show is asking where you still need a human in the loop. So far, no guest has been brave enough to say, let's just let the AI agents run wild. But there's a definite trend towards autonomy. Today, I want to know how we're progressing towards this goal of autonomy through better reasoning, running tasks for longer, using more tools.
From 80%, people went to 90%, was very hard, and going beyond 90% just became impossible. It's hard for me to see that it's gonna, you know, hit 100%, because there's a fundamental limitation of these systems.
Our guest is Russ Salikhoudinov, a professor at Carnegie Mellon University. Russ had a prestigious start to his AI research career as one of Geoffrey Hinton's postdocs. He's also spent time as an executive at Apple and Meta before returning to his academic roots.
Somebody gave full access to some of these agentic systems and the agent just deleted the entire database in 8 seconds. Let's learn about the latest AI agent research.