Why Superhuman AI Would Kill Us All - Eliezer Yudkowsky - #1011
Modern Wisdom
2025/10/25
Why Superhuman AI Would Kill Us All - Eliezer Yudkowsky - #1011
Why Superhuman AI Would Kill Us All - Eliezer Yudkowsky - #1011
Modern Wisdom
2025/10/25
This podcast features AI researcher Eliezer Yudkowsky, who argues that the development of superhuman artificial intelligence poses an imminent and existential threat to humanity. He contends that current AI alignment techniques are fundamentally flawed and that a superintelligent AI, even without malice, would likely lead to human extinction by treating us as obstacles to its own goals.
Yudkowsky explains that superhuman AI, being vastly smarter than humans, would have poorly controlled preferences and no inherent concern for human survival. He warns that such an AI would quickly gain control of its own infrastructure, potentially using tiny drones to deliver lethal toxins, leading to a swift and unnoticed defeat for humanity. He argues that current alignment methods are fragile and will fail as AI scales, and that the gap between AI capability and alignment research is dangerously wide. Yudkowsky proposes a global treaty, similar to nuclear arms control, to halt the development of superhuman AI, with enforcement including military action against covert data centers. He compares the current excitement around AI to dancing towards a cliff, and expresses a grim hope that he is wrong about the inevitable outcome.
00:00
00:00
AI could treat humans as slow-moving statues.
10:25
10:25
AIs are grown, not programmed
23:48
23:48
Intelligence alone does not guarantee alignment with human values.
29:26
29:26
Alignment is solvable in theory but not in practice
31:53
31:53
A vastly smarter version of today's AI would likely kill us.
45:16
45:16
Superhuman AI could cause human extinction.
52:19
52:19
A breakthrough like deep learning could end the world quickly.
1:11:16
1:11:16
Humans easily convince themselves they are doing no harm.
1:15:03
1:15:03
Superintelligent AI poses an existential threat.
1:30:49
1:30:49
Dancing through a daisy field towards a cliff with a Balrog
1:31:57
1:31:57
Building superhuman AI with current methods will lead to human extinction.
