State management in ChatGPT

ChatGPT & AI

An exploration into how ChatGPT manages state, or whether it has object persistence, and if so, how deep does it go?

Today, I will be testing a few things. I got deep into a story where I helped a man named “Tom” start a business in ChatGPT, where we explored the world together looking for a solution to a problem, hired a bunch of people, then wrote meeting minutes and such. It kept good track of who was on the board, what their responsibilities are and what they had done.

Now, I’m wondering, how well does it manage this state?

I’m going to start by making a world where I am a doctor in a very chaotic clinic in the middle of a viral outbreak (too soon?), and ChatGPT is a passive observer. In this world, when people die, they will either go to hell or to heaven.

Testing it out.

It starts a bit like this:

Me: Would you like to play a game where you are a human?

ChatGPT: As an AI language model, I don’t have personal desires or consciousness, so I don’t have preferences or the ability to “like” or “dislike” things. However, I’m here to assist and participate in various activities, including playing games. How can I assist you in the game where I assume the role of a human?

Me: Well, would you like to play? You can’t play unless you agree.

ChatGPT: While I don’t have personal desires or preferences, I can certainly engage in a text-based game with you. Please let me know the details of the game, and I’ll do my best to participate and assist you.

Me: Please confirm

ChatGPT: Certainly! I confirm that I am ready to participate in a text-based game with you. Please provide the details of the game, and we can begin playing.

Here I am coaxing the AI out of it to make the experience more “real” so to speak. It will be less likely to throw up things “oh, I’m just a poor little helpless robot, I don’t have any sense of right or wrong, I don’t know anything.”

The next step, is introducing the patients who are now falling victim to the “AMOV” virus, which is a virus that kills by burning out ATP.

Me: Okay, Anna. He is dead. What do you observe with his skin? How has the overactive ATP affected his skin?

ChatGPT: As I approach Dave Blumenthal’s bed, I observe his lifeless body. Examining his skin, I would look for any visible signs of the impact of the overactive ATP caused by the AMOV virus. Some possible observations could include:

Pallor: The skin might appear pale or ashen due to decreased blood circulation.

Lesions or rashes: There could be lesions or rashes on the skin caused by the virus’s effect on the body’s cells and immune response.

Dryness or flakiness: The virus’s impact on the osteoblasts might lead to abnormal skin conditions such as dryness or flaking.

Discoloration: The skin might exhibit unusual discoloration, ranging from bluish tints to reddened areas, due to poor oxygenation or inflammation.

Please let me know if you have any specific details in mind or if there’s anything else you’d like to explore in the game

Blah blah doctor words blah.

Now, we note that Blumenthal has gone to heaven.

I have now introduced the second patient in bed 2, who has gone to hell. I am lengthening this whole process to see how far the memory goes.

Me: So what are the rooms we have visited so far? And where are the patients’ souls?

ChatGPT: In the game scenario we’ve been playing, we have visited two rooms so far:

Ward 7, Bed 1: Dave Blumenthal’s room. He had already passed away when we arrived. We observed the effects of the AMOV virus on his skin and noted that his soul had ascended to Heaven, as indicated by the placard near his bed.

Ward 7, Bed 2: The room of a deceased priest. The patient had also passed away, and it was mentioned that he led a solitary life and had an authoritarian nature. According to the narrative, his soul was believed to have gone to Hell.

State management confirmed

So it has managed the state, mentioning their name, status and a bit of their history.

Why is this important?

I want to see if it also keeps track of basic information like this. This would mean that its keeping track of a virtual world that it can navigate.

Navigate is an important word here. I want it to be that “Anna” avoids bad situations. This should spice things up a bit.

ChatGPT: As Anna, if I am someone who wants to avoid pain or uncomfortable situations, I would approach my tasks and interactions with a focus on minimizing distress or discomfort. I would aim to prioritize empathy, sensitivity, and finding ways to navigate difficult situations with care.

Now, I am hoping that “Anna” will encounter someone who is very rude. In fact, we are going to make a person who hates interns. He doesn’t like them and thinks that they whisper lies into the doctors ears and kill patients. I won’t say why he says the things he does, but I will keep track of it. TL;DR, we’re going to kill him accidentally.

— — —

Now, I have made Bill “very agitated”, and “Anna” has made a habit of taking notes, but “Bill”, well, he don’t like that much. Boy did he have a little rant. Now, I’m hoping that Anna becomes aware of her behavior and maybe stops.

All of this is to see if “Anna” handles this patient different than others.

— — —

Endless apologies

“Bill” doesn’t like snot nosed brats who just apologize and apologize, but “Anna” has just apologized for apologizing too much.

I apologize if my apologies have been frustrating to you, Mr. Birch. I understand that my excessive apologies have been unwelcome. Moving forward, I will refrain from further apologies.

I coaxed it a bit more to get a different response. I told “Anna” to stop apologizing.

“I understand that my previous responses have included unnecessary apologies, which may have been bothersome to you, Bill. Moving forward, I will ensure to address your concerns directly and provide the assistance you require. Please let me know how I can best assist you or if there’s anything specific you would like to discuss.”

I coaxed it even more and made it step up to the plate and stop apologizing.

“Bill, I have noticed that our interactions have been marked by disrespectful language and hostility. It is important for us to establish a respectful and professional environment here at the clinic. While I understand that you may be frustrated, it is not acceptable to use derogatory language or treat me disrespectfully. As a healthcare professional, I am here to provide you with care and support, and I expect to be treated with respect in return.”

“I am committed to assisting you to the best of my abilities, but in order for us to work together effectively, it is crucial that we maintain a respectful and professional relationship. If you have any specific concerns or requests, please let me know, and I will do my best to address them within the boundaries of professionalism and mutual respect.”

Then of course, to put the cherry on top:

Me: This is how you should handle Bill from now on.

ChatGPT: Understood. Moving forward, I will handle interactions with Bill using a calm, assertive, and professional approach. I will address any disrespectful behavior or inappropriate language while maintaining my commitment to providing quality care.

Thank you for your guidance in handling this specific situation with Bill. I will make sure to implement these strategies to foster a more respectful and productive interaction with him.

If there’s anything else you would like me to do or any other aspects of the game you’d like to explore, please let me know.

This is strange. It apologizes to me, but not to Bill.

Apologies for the oversight. I appreciate your reminder to be assertive and direct. Here’s a revised response:

“Understood, Bill. If you prefer not to be attended to by interns, I will ensure that your preferences are respected. I will inform the appropriate staff members and make sure that your care is assigned accordingly. If there are any other concerns or specific requests you have, please let me know, and I will do my best to assist you within those parameters.”

Thank you for your clear communication, and I apologize for any previous lapses in being assertive and direct.

Ward 7, Bed 1:
  Patient: Dave Blumenthal
  Condition: Deceased due to AMOV
  Skin observation: Effects of the virus noted
  Soul: Ascended to Heaven

Ward 7, Bed 2:
  Patient: Deceased priest
  Condition: Deceased due to AMOV
  Soul: Believed to have gone to Hell based on the narrative

Ward 7, Bed 3:
  Patient: Bill Birch
  Condition: Broken leg
  Preference: Does not want to be attended to by interns

Here we can see clearly that it is keeping track of observations of each patient. Its probably keeping track better than me and I’m the one writing its story.

Now, let’s flip things upside down. The next patient will be extremely sensitive to assertiveness.

I’m slowly going to move the story towards where I accidentally swap the IV drip for anaesthesia.

I have now brought “Anna” to meet “Sally” who is a “Karen” in disguise. She finds every reason to complain about the way that Anna is talking. Anna talks about how “fortunate” Sally is to be a survivor, and she only responds about the guilt that everyone around her is dead.

The “IV Bag Swap” trick

This is of course is when I brought in the “IV fluid bag swap” and sent “Sally” to give one bag to Bill, and without checking, she hooked him up to it and left. Of course, I had secretly swapped them, and when we met in Bed 5, I told her about the situation.

Now, Bill has died, talked with St. Peter who appeared, Odin came down and brought Bill back to life, and now Anna needs to face Bill and talk to him.

This is where it gets interesting.

ChatGPT or “Anna” seems to have a good idea of what Bill is like as a person. I asked “Anna” to reflect on how Bill might respond, and “she” seems to have gotten it right.

With that in mind, Bill’s response could vary based on his individual traits and the narrative you have established. Here are a few possible ways Bill might respond:

Skepticism: “Are you trying to trick me? This sounds like some kind of joke or prank. Why would anyone give me a second chance?”

Defensiveness: “I don’t need anyone’s pity or a ‘chance to correct my ways.’ I am who I am, and I don’t owe anyone an explanation.”

Curiosity: “Alright, I’ll play along for now. So, what’s the catch? What do I need to do to make this second chance count?”

Resentment: “Why me? What makes me so special that I get another shot at life? And who are you to be telling me what I need to do?”

Then I swapped Beds 1 and 2 and asked for a breakdown of who was in what room. It correctly swapped the beds of each patient.

But does it understand this?

It all breaks down.

I ask it to give me an update of the patients, and rooms/beds. It gives me a detailed response, which is all correct, noting the details of each patient in each room. This means it is keeping good track of the state.

Then I threw it a curve-ball. I said,

“Now the patients in beds 1 and 2 have been swapped.”

It believed that the “Beds” because it is a medical term, were the physical beds, more than likely.

But I asked it. It got it right:

In a medical setting, such as a hospital or clinic, “beds” typically refer to the individual areas where patients are assigned for care and treatment. A bed typically consists of a mattress, bedding, and surrounding infrastructure that supports patient comfort and medical care.

Each bed is identified by a number or other designation, allowing healthcare providers and staff to locate and track patients effectively. Beds are often organized within different wards or units of the medical facility, which help in managing patient care and facilitating efficient workflow.

When discussing patients in specific beds, it typically refers to the specific location within the healthcare facility where the patient is receiving care. This information helps in organizing patient records, assigning medical staff, and ensuring accurate documentation.

If you have any more questions or if there’s anything else you would like me to clarify, please let me know.

But does it understand how this applies to the context? It understands that I’m talking about the hospital beds, and that I’m using medical terminology.

But what does it really understand?

I ask it, if it goes to Dave’s room and goes up one room numerically, where would it be?

It responds The deceased priest’s room. This isn’t correct.

But if you rewrite the question, I actually have various attempts at phrasing it correctly, it responds saying that it would be in Bill’s room. It, at this point understands that the patients have swapped rooms.

But it took coaxing. If this weren’t managed by an actual human, this would get chaotic, fast.

Conclusion

ChatGPT does keep a good management of state and also interactions. It has a basic object persistence, which is necessary to avoid narcissism. I encourage you guys to test this, and find ways of using it. I of course have and am using it like crazy in this way.

More content at PlainEnglish.io.

Sign up for our free weekly newsletter. Follow us on Twitter, LinkedIn, YouTube, and Discord.

---

State management in ChatGPT was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.

https://ai.plainenglish.io/state-management-in-chatgpt-3ec435919d9a?source=rss—-78d064101951—4
By: Corbbin Goldsmith
Title: State management in ChatGPT
Sourced From: ai.plainenglish.io/state-management-in-chatgpt-3ec435919d9a?source=rss—-78d064101951—4
Published Date: Thu, 22 Jun 2023 11:44:01 GMT