Discussion about this post

User's avatar
Kevin Hou's avatar

What is 'good' truly seems non-trivial, and at times, impossible in a field like mental health.

In emergency psychiatry, you often hear that a clinician's ability to predict suicide risk is no better than chance. This is not a bad clinician; humans are just insanely complicated.

It's still worth exploring what guardrails can be built, and also building non-goodharted, cross-cultural evals (particularly for every country outside of US/China that don't have their own foundation models).

It sounds like your evals parallel some work we've done recently called SIM-VAIL (Weilnhammer et al., 2026) - using Anthropic's Petri to red-team using psychiatrist crafted personas across Phenotype-Intention personas.

Keen to follow how this work progresses - I'm hopeful the cumuluation of efforts will converge towards actionable steps in building more aligned LLMs.

John Lund's avatar

Thank you for bringing up the need for a diversity of cultural values in LLM evals (and behavior)! We have to start incorporating true diversity into our models or we risk an unprecedented flattening of society.

Ready for more?