Banning Superintelligence: Are We Building Humanity's Replacement? with Roman Yampolskiy

Roman Yampolskiy (0:00)

If we built general superintelligence, we would not be able to control it. You are trying to establish a perpetual safety machine, which will never ever have a single slip up, no matter how advanced the AI gets. If it interacts with malevolent actors, if it gets hacked, I don't think that's a property of complex software systems. And it's definitely not gonna happen for something smarter than all of us combined. The pause is not enough. It has to be a permanent ban. You never create general superintelligence. I like technology, I like science, but don't build gods.

Nate Hagens (0:36)

You're listening to The Great Simplification. I'm Nate Hagens. On this show, we describe how energy, the economy, the environment and human behavior all fit together and what it might mean for our future. By sharing insights from global thinkers, we hope to inform and inspire more humans to play emergent roles in the coming Great Simplification.

Today, I'm pleased to be joined by AI safety expert and the actual originator of that field, Roman Yampolskiy for a dive into his 12 years of research on the threat that artificial super intelligence poses to humanity and the biosphere, and how this risk reflects our larger, non-systemic approach to technology, progress, and governance. Roman Yampolskiy is a tenured AI expert at the University of Louisville, boasting over 100 publications in AI safety, digital forensics, and cyber security. As the founding director of the Cyber Security Lab, Yampolskiy's expertise spans AI safety, behavioral biometrics, cybersecurity, and more, with broad impact in academic and media circles. In this conversation, Roman does not beat around the bush regarding the severity of the threat that artificial superintelligence, as opposed to just simple AI, poses to our society. He unpacks how the recent acceleration of AI development has shifted his outlook on humanity's future, and shares findings from recent research testing AI capabilities, research that continues to reveal how little control we actually have over the systems that are evolving and we continue to build. While I personally believe AI is just one piece of the larger puzzle of the interconnected crises we face, I do increasingly see it being a primary hurdle for more benign human futures. And beyond that, it is increasingly clear that the way we're approaching the existential risk of general superintelligence is itself a microcosm of the fundamental governance issues embedded within the human superorganism. With that, please welcome Roman Yampolskiy.

Roman Yampolskiy, welcome to TGS.

Roman Yampolskiy (3:11)

Thank you for inviting me.

Nate Hagens (3:12)

As it happens, this week I am preparing a frankly which is my Friday monologue on humanity's Icarus moment, the 20 risks from AI. One of the risks is the jailbreak humans losing control scenario, which is you are a world expert on that, but there's a lot of risks.

Before I get into it, I just find myself in this paradoxical position that I actually like using AI. Within limits, and I simultaneously wish it had never been invented. What are your thoughts on that?

Roman Yampolskiy (3:54)

We use the term AI to mean very different technologies. If you like using AI tools, which are narrow and helpful, I'm with you.

If you like general super intelligent agents to replace humanity, I'm not with you. It depends on how we use the term.

Nate Hagens (4:09)

Okay, well said.

So I am to understand you are one of the earliest researchers in the field of what is today called AI safety, which I believe you're credited with coming up with that term 15 years ago. You spent much of your career trying to understand how we might make increasingly intelligent machines safer. And as you've continued your work, you have arrived at a conclusion that our attempts to create boundaries and security mindsets around AI are actually in reality pretty limited, and that some of the properties we would need to make these models safer, like predictability and explainability, verification, control, and the like, may actually be impossible for us to do. So when did you start to realize that this might go beyond an engineering problem, and what implications has that had on your own research trajectory on all this?

Roman Yampolskiy (5:16)

Actually, it surprisingly took me almost a decade to realize that no, I'm not going to control godlike super machines. That's a lot of hubris right there. I don't know why it wasn't obvious. It should be, it's just common sense. But I went through the technical arguments and showing individual impossibilities of different tools we would need to get to control before this, aha, obviously not going to happen.

Nate Hagens (5:40)

Well, actually, that's a lot of wisdom in your one sentence there. Because that actually is kind of a microcosm of what our whole culture is going through, perhaps.