The many masks LLMs wear - Kai Williams

Fantastic article in Timothy B. Lee's *Understanding AI* newsletter (also fantastic) that goes deep on all the ways LLM personas go off the rails, why it might happen, and what the real world consequences are, both first- and second-order.

As any economist will tell you, everything comes down to trade-offs (just as computer scientists might tell you there’s no free lunch). Although those phrases don’t appear anywhere in the article, the entire history of model alignment is an exercise in turning one dial in the ‘good’ direction, only to have some other dial, perhaps one we didn’t previously know about, turn in the ‘bad’ direction.

Really recommended reading.