My recent visit to Anthropic

I recently attended a two-day session aimed at helping shape the constitution for Claude. The select group was exceptional, and we had ample opportunity to engage with key decision-makers. The discussions were of the highest caliber.

In my contributions, I emphasized a few key points. First, I argued that the constitution should draw more from common law and case law analogs, rather than simply relying on rigid legal frameworks. Next, I suggested viewing it in a more flexible, adaptive light, akin to a "Talmud" rather than a static "Torah."

I also stressed the importance of developing a robust secondary literature to support and interpret AI constitutions. Currently, such literature is lacking. Additionally, I proposed the idea of incorporating a panel of diverse AI entities, each queried with varied prompts, to assess how well Claude (and other models) adhere to their respective constitutions.

If these AI panelists identified any discrepancies, they could alert a human adjudicatory board, which would then have the authority to address potential breaches and recommend remedies.

The experience was invaluable, and I extend my gratitude to Anthropic for the opportunity. This visit is detailed in a recent post on Marginal REVOLUTION.