We talk about LLMs gaining knowledge, but what about personality?

Question

Turns out, models don’t just learn facts. They can pick up traits like cheerfulness, stubbornness… even sycophancy.

And fine-tuning can amplify these behaviours quietly

Recently, researchers introduced “Persona vectors - numerical patterns” inside an LLM that correspond to traits like evil, sycophancy, or hallucination.

Think of these vectors as “knobs”. Turn them up, a trait grows. Turn them down, it fades.

Now, this actually could be a big step forward.

Because, LLM personality isn’t a by-product anymore. It’s something we can observe and control.

This opens the door to:

• Safer conversational AI

• Predictable fine-tuning outcomes

• Guardrails against hallucination + sycophancy

• Customizable AI personalities (responsibly!)

Basically, high level personality behaviour is no longer mysterious… it’s editable.