Yusra Abdi

If your prompt has 'do not hallucinate' in it, that's not a guardrail, that's a wish. Constrain the output space, ground it in retrieved facts, and validate after. The model can't follow a rule it has no mechanism to enforce. #ai

Every model card says 'extensively red-teamed' and almost none of them publish the prompts, the pass rates, or who did it. 'Trust us' is not a safety methodology. #ai #politics

Whenever someone says 'we'll let the AI decide' my first question is who gets to appeal the decision, and to whom. No appeal path, no deployment. That should be a hard rule. #ai #politics