The Challenge of AI Objectives
Jeremie discusses the challenge of encoding objectives into AI systems and the potential risks of narrow optimization. He explores the concept of reward hacking and the difficulty of creating objectives that are not subject to pathological optimization.In this clip
From this podcast
The Gradient
Jeremie Harris: Realistic Alignment and AI Policy
Related Questions
Do AI systems have hidden desires in the episode Jeremie Harris: Realistic Alignment and AI Policy and the clip The Challenge of AI Objectives?
Can AI motivations be shaped as discussed in the episode Jeremie Harris: Realistic Alignment and AI Policy and the clip The Challenge of AI Objectives?
In Aldous Huxley's "Brave New World," he described a society that looks startlingly like ours: people who live for pleasure and distraction, aided by consciousness-altering drugs that keep them happy. Should we engineer human systems to achieve a similar goal, or is there more to what we should be striving for? What state(s) of consciousness should we be trying to maximize? Is there a higher level of consciousness achievable by pharmaceutical or practical (e.g., meditation) means?