Thank you for sharing John - brilliant insights here, lots of food for thought! I do fear L4 and L5 are contingent on the LLM being able to hold and maintain a divergent view, even in the face of challenge - I’ve not seen much evidence of AI achieving this (yet) and welcome any examples. My experience of it has been mostly sycophancy - even very well-disguised sycophancy where on the surface it may seem like it’s disagreeing. I would really like to follow your and Louise’s experiments with this, and wondering also how we might think to implement this on our own advisory board.
Really thought-provoking. Thank you.
Thank you for sharing John - brilliant insights here, lots of food for thought! I do fear L4 and L5 are contingent on the LLM being able to hold and maintain a divergent view, even in the face of challenge - I’ve not seen much evidence of AI achieving this (yet) and welcome any examples. My experience of it has been mostly sycophancy - even very well-disguised sycophancy where on the surface it may seem like it’s disagreeing. I would really like to follow your and Louise’s experiments with this, and wondering also how we might think to implement this on our own advisory board.