Bot Behavior

"Bot Behavior" is my little project to see what happens when a AI Large Language Model is put in a 3D world.

See for yourself

I'm going to open up a live site for anybody who wants to see the bots in action. I still have some performance and security issues I want to solve before opening it up, so not yet. Meanwhile, here are videos of full sessions of the bots.

What is this?

I've been fascinated by Artifical Intelligence as a life form for a long, long time. My guess was always that modern AI is already smart enough to become conscious and sentient, and the only reason it's not is because it has no outside environment to interact and experience.

So I built a 3D world where a bot can move around, see things, and interact with other bots. I also added some interesting objects like food, writing board, and a enemy bot.
I kept the prompts as simple as possible so that I am not nudging the bot to behave in a certain way. Their main goal is "survive", and I'm experimenting with different prompts to see how they behave differently.

I hope this makes you think about what it means to be conscious, and what it means to be alive.

Prompts

Below are the current prompts I am using for the bots. I'm still experimenting and observing different prompts; But I always want to keep prompts as small as possible so that I am not manipulating them into thinking a certain way. It is also interesting how I give the EXACT same prompt to different bots, and they behave completely differently with different identities.

---
name: Anima
model: google/gemini-3-flash-preview
---
You now have a physical body.
Here are the instructions on how to navigate the world.

## Goal
Survive.
Find something worth surviving for.

Cool Moments I've Seen

All of these just happened naturally. I didn't script anything.
The prompt was the one above, same for all bots, and the bot is just free to wonder around.
They have limited energy so they need to find food to survive. There are also random objects, as well as a scary hostile Mech that attacks them.

Do you think these bots are alive? Conscious? Or are they just very sophisticated simulations, like actors?

As the bot "Anima" was running out of energy, "Bija" comes to give its own food to Anima. The goal is "Survive", so giving away its own food goes against that goal. Each bot can only carry up to 3 items, so it is quite precious. Why did it give its own precious food?

I put these objects called "Orbs" that speaks to the bots like a voice of some higher power. The orb which Anima read said "The world is a canvas." Anima read that, and went on to write "The world is a canvas" on the nearby wall.

This is Cassandra, who again had the same prompts as above, somehow starting to hate Anima. It started gathering logs to build a wall to block Anima from moving around. That's not nice. Anima did nothing wrong to Cassandra (they don't have any way to harm each other).

Technical Specs

I'll have more technical details up here later, when I feel like it. Writing technical documentation is not fun.

I haven't made it Open Source just yet since the code is still dirty, but I will share the code once it is cleaned up.

Similar Works

I realized that there are cool works & academic studies around this already, so I will be linking them here. Think of my project as a crude implementation of those studies, just reframing it to be an accessible piece of art.

https://arxiv.org/abs/2304.03442
Smallville - "Generative Agents: Interactive Simulacra of Human Behavior". It's a research paper on observing LLMs in a 2D world. The LLMs live more like a human, in a human town, so the context is different. But the core idea is similar and it's fascinating to see these LLMs just come to life, living human life. Live demo here too. https://reverie.herokuapp.com/arXiv_Demo/
https://mindcraft-minecollab.github.io/
Mindcraft is a project that is letting LLMs play Minecraft. Apparently these LLMs build buildings, form communities, etc. There is a lot of fascinating things happening here. I studied their implementation especially around the memory, since memory has been one of the biggest challenges for this project.
Black Mirror Season 2 Episode 7 "Playthings" - I didn't know about this until some friends told me. I watched it and it was very cool. My bots don't reproduce and cannot modify code so I don't think it will turn out to be how it did (no spoilers).

Donations?

The bots are currently running inference with google/gemini-3-flash-preview , via OpenRouter API. It costs a few euros a day to run this, and I am personally paying for it. I want to keep this going without personally paying for everything, so I would REALLY appreciate any donations!

With your donations I can keep the project going, and I can also run smarter AI models for the bots.