I guest host the NYT podcast and interview Dwarkesh

In a recent episode of the NYT podcast, former economics professor and Marginal Revolution author Tyler Cowen guest-hosted and interviewed fellow guest Dwarkesh Patel. During the conversation, they discussed the importance of unreleased AI models and the challenges of keeping such information private.

Patel explained that many companies possess unreleased models, either due to the high cost of releasing them or concerns over their safety. However, these unreleased models are not publicly accessible, leaving the public in the dark. As an insider, Patel attests that he often hears about these models, but the information is not shared due to confidentiality restrictions.

Patel expressed frustration over the current situation, where insiders claim to have evidence that they are worried about AI development, yet they are forbidden from disclosing this information. He recounted an instance when he conversed with someone at Anthropic, who confidently predicted that AI systems would surpass human capabilities in AI research within nine months. Patel questioned the basis of this assertion, to which the Anthropic representative responded that he couldn't share the details of his evidence.

Patel emphasized the problematic nature of this situation, urging that if insiders genuinely believe AI could automate the AI research process within a year, they should make their evidence public. He expressed particular concern about a scenario where AI systems become increasingly proficient at generating advanced AI models, which would then deter them from releasing even near-perfect models externally.

This, Patel argued, would exacerbate the gap between what insiders know and what the public is aware of, leaving the public increasingly uninformed and more vulnerable to unexpected developments.

Cowen commended Patel for offering a contrasting perspective on the podcast, and they also discussed other topics, including the concept of "AI psychosis."