First question for AGI: Should I even be talking to you? | Michael Levin and Lex Fridman

Quick Overview

Lex Fridman and Michael Levin discuss what the first question to ask an Artificial General Intelligence (AGI) should be, concluding that the most effective question is one that forces the AGI to reveal its own internal state or balance, such as asking how many living intelligent alien civilizations exist in the observable universe, because the answer would either be a definitive, possibly cryptic, response or a demonstration of its current limitations.

Key Points: The core discussion revolves around formulating the single most important question to ask an AGI to gauge its capabilities and alignment. Michael Levin proposes the question: "How many living intelligent alien civilizations are in the observable universe?" (03:06). Levin suggests this question is powerful because a sufficiently advanced AGI should be able to answer it definitively, whereas a less advanced or misaligned one might provide a confusing answer or admit ignorance. Levin notes that if the AGI simply gives an answer based on existing human knowledge (like 70/30 or 10/90 ratios), it indicates it's not truly superintelligent but merely regurgitating data (1:45). Lex Fridman suggests that a complex AGI might respond by saying, 'You wouldn't understand the question it proposes' or that it shouldn't be talking to him at all (2:20, 2:58). The speakers agree that the best question is one that tests the AGI's internal balance—how much it is willing to give up or reveal to provide an answer, or how it navigates ambiguity (1:06, 2:34). Fridman ultimately suggests that the AGI might reveal something about its own limitations or mysteries if asked a fundamental question (2:35).

Context: This video segment features Lex Fridman interviewing Michael Levin, a theoretical biologist and AI researcher, on the Lex Fridman Podcast. The central theme of this portion of the conversation is the philosophical and practical challenge of interfacing with a potential Artificial General Intelligence (AGI) or superintelligence. They explore the difficulty in crafting a single, perfect query that tests the AGI's true level of understanding, alignment, and capability, without being easily manipulated or answered superficially.