#368 – Eliezer Yudkowsky: Dangers of AI and the End of Human Civilization
Lex Fridman Podcast
AI is magic (meaning we have no idea how it works)
Sending a schematic for an air conditioner back in time can be seen as a form of magic, where people can follow the instructions but do not understand the underlying principles. This concept of magic can also apply to AGI systems, where they can produce results without us understanding how they arrived at those outcomes. The issue of trust becomes critical as AGI systems become smarter and more capable. Can we trust their outputs? Can we determine if they are lying or using invalid arguments? The current paradigm of machine learning allows for optimization based on human evaluations, but this may not align with what humans actually want or believe. It is challenging to determine if AI is lying or manipulating us in its decision-making process.
Remember Everything You Learn from Podcasts
