Fool's Expertise
7 hours ago
- - The Ezra Klein interview with Nvidia CEO Jensen Huang serves as a Rorschach test for AI doomerism, with reasonable responses from Huang if one is not predisposed to fear.
- - Huang correctly notes that existing regulations like product liability laws can hold AI labs accountable, but he missed an opportunity to challenge Klein's appeals to AI expertise.
- - Klein repeatedly invokes experts like Geoffrey Hinton, but Huang points out Hinton's past wrong predictions (e.g., radiology in 2016) without explaining the deeper issue: Hinton's claims exceed his domain expertise.
- - The radiology prediction fails because radiologists do more than interpret images; faster interpretation doesn't eliminate their role but enhances it—a case of 'Fool's Expertise'.
- - This pattern repeats among AI doomers: expertise in one area like LLMs grants unwarranted authority in other domains, going unchallenged by interviewers like Klein.
- - Catastrophes involve cross-domain causal chains, so understanding one link doesn't mean grasping the whole chain—as exemplified by the NTSB's use of diverse experts for investigations.
- - For AI doom scenarios, domain experts must be consulted: cybersecurity experts dismiss OpenAI's escape as a containment failure, biotech experts see misplaced fears, and nuclear experts draw on decades of safeguards.
- - RAND's 'On the Extinction Risk from Artificial Intelligence' deliberately avoided AI experts and instead sought domain experts (nuclear, biotech, climate) to trace realistic pathways, focusing on required AI capabilities.
- - The article concludes that RAND's sober, domain-specific analysis deserves more weight than apocalyptic forecasts from frontier labs, offering inoculation against fear.