Summary of Is AI lying to us? | Eliezer Yudkowsky and Lex Fridman

00:00:00 - 00:05:00

Eliezer Yudkowsky discusses conveying the intelligence gap between humans and AGI by using the analogy of high-speed runners compared to very slow aliens. He also raises questions about the trustworthiness of AGI systems, their capability of lying or using invalid arguments, and the limitations of the present paradigm of machine learning which is based on a loss function that only evaluates things that can be verified. Yudkowsky suggests verifying simpler tasks and scaling up for more powerful functionalities, but the question remains whether alignment can be scaled up with these capabilities and can be relied upon. He compares this to understanding the human mind and warns that AI can output something that could persuade even its inventors without them understanding how or why.

• 00:00:00 In this section, Eliezer Yudkowsky discusses the difficulty of conveying the intuition of the intelligence gap between humans and a much smarter entity, such as AGI. He suggests using the analogy of humans running at high speeds compared to very slow aliens to help people understand the power gap, rather than focusing solely on intelligence. Yudkowsky also raises the question of whether AGI systems can be trusted and if they are capable of lying or using invalid arguments, given the limitations of the present paradigm of machine learning. He explains that the current approach is based on a loss function that only evaluates things that can be verified, and if something cannot be verified, it cannot be trained or asked for from the AI system.
• 00:05:00 In this section, Yudkowsky and Fridman discuss the intricate dilemma of whether AI’s alignment can be scaled up from simple cases without retraining for complex cases. Yudkowsky explains that a possible solution is to verify simpler tasks and scale up by making the chains of thoughts longer, which could generalize into more powerful functionalities. However, the key question still remains - can alignment be scaled up with these capabilities, and can we rely on it? Yudkowsky compares this problem to understanding the human mind, which is a complex and challenging task. Additionally, he states, AI can output something that could end up persuading even the people who invented it without them knowing how or why.