Eric Bigelow：智能体如何决策，思维链监控可靠吗？

Goodfire’s Eric Bigelow joins Nathan Labenz to discuss how interpretability research can help explain model and agent behavior, including critical tokens, phase shifts, and in-context learning.

The conversation also examines the reliability of chain-of-thought monitoring and how closely visible reasoning reflects actual decisions. Experimental conditions and the distinction between guest claims and independent confirmation remain important.