Can we give personality to AI? The research paper from @AlloraNetwork takes a look at this. These days, when you talk to AI, you often have a similar experience. Polite but uniform tone, stiff and neutral responses. It feels like it’s always stuck in "basic mode." But what if AI could have different personalities like humans? If it could show extroverted or introverted traits, calm or impulsive tendencies, the user experience could become much more natural and engaging. The recent research from Allora directly addresses this question. "Can we give personality to AI and actually make that personality manifest?" === Researchers' Approach === The research team provided AI with personality templates. For example, injecting a setting like "extroverted, cautious, and diligent." Then, they had the AI take a "personality test (Big Five, MBTI)." They made it answer questions like a real person and checked how closely the results matched the instructed personality. The measurement didn’t just look at scores. They also had the AI write reasons for its answers to see if it was truly thinking in line with that personality. === Key Findings === First, the higher the model's performance, the more precise the personality expression was. Models with excellent reasoning abilities (like o1, 4o, etc.) answered more faithfully to the instructed personality. Second, it was revealed that personality and tone are separate. The accuracy of personality didn’t change much even with fine-tuning, but the tone and style of the AI's speech could be easily adjusted. In other words, "personality" is at a deeper level, while "tone" is a surface style. Third, AI is not perfectly consistent either. There were responses that deviated from the personality instructions for individual questions. However, when averaging multiple questions, it converged closer to the instructed personality. Interestingly, this resembles the imperfect consistency humans show when revealing their personalities. Fourth, there was bias in certain personality dimensions. For instance, "openness" tended to score higher than instructed. In this way, the Allora research team demonstrated a significant depth of knowledge regarding AI's tendencies and openness, as well as their considerations and studies on the "personality" of AI, indicating that various AIs we will encounter in the future may have distinct characteristics like humans, along with ongoing research in this area. By the way, when is the Allora TGE? Haha

Original paper link. This paper has been summarized and reinterpreted through AI. Heh