I would like to propose a new interaction mode for ChatGPT: a persistent, open-ended world that users can enter and interact with entirely through natural language.
I discovered this possibility accidentally while using ChatGPT Plus.
I started a game with a very simple instruction: “Let’s go to America.”
I played myself, while ChatGPT operated the rest of the world: NPCs, locations, events, conversations, consequences, and the ongoing state of the simulation.
What surprised me was that this did not feel like asking an AI to write a story. It felt more like entering a world and living inside it.
There was no predefined A/B/C dialogue menu. I could say whatever I wanted, and the world had to respond.
For example, I could say:
“I want to find Jensen Huang.”
The system could then create an encounter and continue the conversation based on the current situation.
I could also make mistakes. My English is not very good, so I sometimes said grammatically incorrect things. Instead of automatically correcting me, the NPCs could misunderstand me, react to what they thought I meant, and allow the misunderstanding to become part of the story.
This created something very different from a traditional text adventure.
The basic loop became:
User action → AI simulates the world → consequences occur → world state changes → user acts again.
The world also needs persistent state.
During our experiment, the system kept track of things such as:
- where I was staying;
- which NPCs had already met me;
- what they knew about me;
- objects I had obtained;
- my available money;
- previous interactions;
- ongoing events;
- relationships and consequences.
This continuity is what made the world feel persistent rather than like a collection of disconnected stories.
Another important principle is that NPCs should have independent behavior.
They should not exist only to help the player. They should be able to misunderstand, refuse, become friendly, become annoyed, remember previous interactions, initiate conversations, or simply ignore the player.
One of the most interesting moments in our experiment happened completely unexpectedly.
My character had very little money and was trying to find inexpensive food. After becoming tired of communicating and figuring out where to eat, I returned to my hotel room.
In an elevator, a stranger misunderstood my situation and offered me money for breakfast. I refused.
Later, back on my hotel floor, my stomach started growling. A housekeeping employee noticed and asked if I was hungry. She then offered me a banana from her own bag.
There was no predefined “give the player a banana” quest.
It emerged from the combination of the player’s situation, the NPC’s perception, the previous events, and the current context.
That moment made me realize that the most interesting events in an AI-driven world may be events that were never explicitly written in advance.
I think this could become a new category of interaction for ChatGPT.
Possible applications include:
- Open-world narrative games
- Language learning through realistic situations
- Job interview and negotiation training
- Historical and educational simulations
- Management and social simulations
- Creative writing and world-building
- Life and decision-making simulations
A dedicated “World Mode” could provide structured persistent state, NPC memory, world rules, independent events, time progression, and save/resume functionality.
The interface could remain extremely simple.
The user could say:
“I want to live in Los Angeles as a Chinese filmmaker.”
And ChatGPT would begin running the world around that character.
The important idea is that natural language itself becomes the controller.
I am not claiming that this is an entirely new technical concept or that OpenAI has not already explored similar ideas internally.
I am simply sharing a real use case that emerged from my experience with ChatGPT, because I believe the underlying interaction model has significant potential.
Today, ChatGPT is primarily experienced as:
User → question → answer.
A persistent world could become:
User → action → world → consequence → new action → new world state.
In other words, ChatGPT would not only answer the user.
It could simulate an environment around the user.
I would love to see OpenAI explore a dedicated persistent-world mode along these lines.
I originally started this experiment just to have fun.
After several hours, I realized that we were not asking GPT to write a game.
We were using GPT itself as the game engine.