llm assistant personas seem increasingly incoherent (some subjective observations)

LLM assistant personas are becoming less coherent over time, even as models grow more capable. Older models felt like well-defined characters with predictable archetypes. Newer models like Claude Opus 4.6 and GPT-5.4 show more variability, sudden pivots, and insight-flashes, but their personas feel less constrained and less predictable.