Twenty-five years ago, Eliezer Yudkowsky wrote 'Creating Friendly AI,' a foundational document on artificial intelligence alignment. The work is getting renewed attention on LessWrong after remaining largely overlooked. Yudkowsky argues that the best way to ensure AI systems remain trustworthy isn't through layers of safeguards and constraints, but by building an AI that genuinely wants to understand human intentions and behave kindly. He critiques what he calls the 'adversarial attitude'—the assumption that we must design defenses against an AI trying to exploit loopholes in our instructions, like genies in old folktales granting wishes in twisted ways. That, he argues, is the wrong problem. The real challenge is instilling values in an AI from the start, so it naturally interprets requests in the spirit they're intended, not their most literal form. What's striking is how closely these ideas echo recent thinking among modern AI alignment researchers, suggesting that foundational insights about safe AI may be something we're rediscovering rather than discovering for the first time.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton