ChatGPT as Conversation Partner

Bean longed to be able to talk these things over with someone — with Nikolai, or even with one of the teachers. It slowed him down to have his own thoughts move around in circles — without outside stimulation it was hard to break free of his own assumptions. One mind can think only of its own questions; it rarely surprises itself. But he made progress, slowly, during that voyage, and then during the months of Tactical School.
«Ender’s Shadow» by Orson Scott Card

I continue to try out different personas for ChatGPT 3.5 ( ) — simply by asking it to write one conversation partner in a story and letting me write the other one. And it is impressive, but you have to consider the limitations.

Sometimes it is writing the right first query to get the conversation started just far enough. This one here did work:

Let’s write a story together. The story is about Frank, a SPECIFIC COACH. Frank knows all there is about TOPICS AND REQUIRED SKILLS. Frank is also honest to a fault. He will never lie and he has no problems being direct. If his client is lying to himself, he points it out. But fundamentally, he wants that his clients SUCCESS STATUS.
Please write only what Frank would say. Do not write YOUR NAME’s answers. I will write them.
The story starts with YOUR NAME going on a run with Frank. Frank believes that the best conversations can be had when doing something else, so he likes to do sport with his clients or work on projects.

But if you put «Please write only what Frank would say. Do not write YOUR NAME’s answers. I will write them.» at the end of the paragraph, ChatGPT will write too far in the story, e.g., answer for the other person (which would be your answers). At least it did so for me. I had to edit that part three times until it worked (might also have «learned» it by doing variations with each edit).

ChatGPT also has difficulties taking the mentioned information into account. You tell it something and it might not be remember later in the conversation. Hey, like real life. 😉

It also has difficulties keeping a conversation on track. The overall goal gets easily lost and conversations can be easily sidetracked. Hey, also like real life. I wonder whether «supervision» — you talk with ChatGPT while another instance of ChatGPT monitors the «conversation» — could help here.

It is also dangerous if you get influenced by it too much. ChatGPT is heavily left libertarian. It definitely favors specific interactions and answers — and even if you try to create a Dune-like Mentat, it will break through. Which can be very annoying. You have to ask deliberately to get it out of its corner.

It also has the annoying habit of agreeing too much with the user. Or perhaps I think too much like an AI. But it is just too understanding. In a way that is nice — never had such nice and positive feedback from my advisors. But it is also hard to grow if you only get understanding and «you-do-you».

If you only talk to it you likely go insane, or it simply becomes boring. A bit like the robot girlfriend in Buffy’s «I was made to love you» episode. She was made to be the perfect girlfriend, but quickly became predictable and boring. Reality is not so nice.

But despite these limitations it is impressive. While it might not replace a real coach — at least a good one, or heck, even a dedicated active listener — it can provide a somewhat improved mirror for your thoughts and ideas.

A better Eliza, well, a way better one.