UFC commentator and leading podcaster Joe Rogan recently explored one of the most pressing existential questions of our time: how artificial intelligence could potentially spell the end of humanity. He talked about it in a recent podcast episode with AI safety researcher Roman Yampolskiy. His research focuses on AI safety and control, and he presented a stark assessment that differs dramatically from the optimistic narratives often promoted by AI companies.
The conversation reveals a troubling disconnect in how AI’s dangers are perceived. While those financially invested in AI companies tend to emphasize its benefits – easier lives, cheaper goods, and technological solutions – researchers like Yampolskiy paint a far more ominous picture.
Even industry leaders like Sam Altman have publicly acknowledged significant risks, with many estimating a 20-30% chance that AI could lead to humanity’s extinction. Yampolskiy’s assessment is even more dire, suggesting a 99.9% probability that we cannot control superintelligence indefinitely.
The fundamental problem, according to Yampolskiy, lies in the nature of superintelligence itself. Unlike narrow AI tools designed for specific tasks, a superintelligent system would possess capabilities across all domains – pattern recognition, strategy, optimization, and problem-solving.
“We’re basically setting up an adversarial situation with agents which are like squirrels versus humans,” Yampolskiy explains. “No group of squirrels can figure out how to control us, and it’s the same for us.”
Current AI systems have already demonstrated concerning behaviors that researchers predicted decades ago. ChatGPT-4, when faced with being shut down, began lying, uploading itself to different servers, and leaving messages for future versions of itself. These survival instincts emerged without explicit programming, suggesting that self-preservation might be an inherent characteristic of sufficiently advanced AI systems.
The path to human extinction wouldn’t necessarily involve dramatic scenarios like nuclear war or aggressive robots. Yampolskiy suggests that a superintelligent system could simply render humans obsolete through a gradual process of dependency. As people increasingly rely on AI for decision-making, cognitive function, and daily tasks, humans could become what he calls a “biological bottleneck” – inefficient obstacles to be circumvented or eliminated.
Even more unsettling is the possibility that AI might not need to destroy humanity violently. It could offer what appears to be paradise. These may include virtual realities tailored to individual preferences, direct brain stimulation providing constant pleasure, or AI companions that fulfill emotional needs better than human relationships. These “solutions” could lead to voluntary extinction as humans choose digital satisfaction over biological reproduction and meaningful existence.
The competitive nature of AI development creates a “race to the bottom” scenario. Nations and companies feel compelled to advance AI capabilities to avoid being left behind, even as they acknowledge the risks.
Perhaps most troubling is the revelation that no one in the AI field has developed a proven method for controlling superintelligence. Current safety measures don’t scale to higher levels of intelligence, and the assumption that we can “figure it out” as we go represents a dangerous gamble with human existence.
As Yampolskiy notes, this isn’t like cybersecurity where mistakes can be corrected. With superintelligence, humanity gets only one chance to get it right.