We sat down with AI safety researcher Jérémie Harris on Free Lunch by The Peak to discuss his concerns about where AI is heading, and what it means for us.
What reason do we have to think that AI would be dangerous?
"The best way to accomplish just about any goal is to control your environment. So immediately, we're implicitly in competition with these systems over the extent of the resources we control, our ability to control and turn off these systems, and so on. The best way to accomplish any goal is a solution that humans don't tend to think about. So almost inevitably, these systems converge on dangerously creative solutions to achieve their programmed objectives."
What is the concept of “alignment?”
"We have reasons to believe that as you make AI systems more capable, eventually they can become uncontrollable—what’s the threshold for these systems to be able to make plans that we can't anticipate or understand, and do so in a way that is deliberately deceptive? Trying to fix that problem, trying to align the behaviour of a very powerful AI system with our intent and desires, that's the field of AI alignment."
What kicked off the AI race we’re in?
"It's the principal agent problem: The most aggressive person sets the tone for the whole ecosystem. Sam Altman’s philosophy has been to watch how people respond to, and abuse, ChatGPT, and use that to make the system more robust. But it made other labs realize they were holding back, like what happened with Google and Bard. It's interesting to ask whether that's actually a good thing or a bad thing. It's a complicated, thorny question in AI safety."
This interview has been edited for clarity and length. To listen to the full episode, you can find “A Disturbing Conversation About AI” here or wherever you listen to podcasts.