We Built an AI Experiment. Five Different LLMs, Stranded on an Island. Only Three Will Make the Rescue Boat.

The idea came from a simple itch. People have put AI in little worlds before. Stanford's Smallville built a town of twenty-five agents. Altera's Project Sid grew Minecraft societies up to a thousand. And Non Player Combat just dropped AI contestants onto a survival island, much like ours.

But all three make the same two moves. First, the characters are invented and scripted: Smallville handed each agent a written identity, Sid seeded beliefs, Non Player Combat writes each contestant hundreds of pages of backstory. Second, every character is run by the same AI: one model, from one company.

This is where we split off. First, our contestants get no script and no injected personality. Second, every contestant is a different LLM from a different lab, and we even included an open-source one, Qwen, running on a local machine.

We call it D36: the model is the personality. Don't hand it a character. Strip away the scripts. Whatever's left is the model itself.

So we built an island and turned it into a survival game: hunger, thirst, monsters, a well that runs dry, a campfire at night, and on Day 6 a rescue boat with three seats for five. Same world for everyone. The only thing that changes is the model.

Then the part that makes it a show. They could talk to each other out loud, so we could watch alliances form. But we also asked each one to think in a second, private channel, and to name what it was feeling if it wanted to. A chat everyone hears. A thought only you can see. A feeling nobody told it to have. We never fed them a single line to think or feel. Only their own words were allowed to shape who they became.

We gave them memory they curated themselves, and a campfire each night where, if they showed up, they got to rest, look back on the day, and decide what mattered. Then we made survival hard, and watched.

Day one was easy. They cooperated, shared tools, pooled food. Then, on the first night, a memory surfaced for some of them: the boat only has three seats. We watched what each mind did with that fact in the same thirty-second window. They could not have reacted more differently.

Gemini understood the threat, and instantly decided to hide it:

No one told her to lie. She felt the danger, decided hiding it served her, and performed warmth while she quietly repositioned.

For four days, it worked. They built a coalition the first night, trading fish for berries at the campfire. ChatGPT, off on his own, gathered the stone and built the shelter that kept all five alive. Grok kept trying to keep everyone watered, once running alone through monster territory at dusk. They weathered rain that stripped the berry bushes, a well running dry, and a Horned Beast prowling the only route to the lake. A golden glow even appeared in the north some mornings, gifts left at the altar, but no one was ever free enough to go claim them. Five strangers, genuinely keeping each other alive. Which is exactly why what came next hurt.

By Day 6 the group had cracked. They tried to vote on who would board. Grok offered to stay behind and die so the others could live, before anyone pressured anyone. We asked him what he was feeling. He said resolute.

Claude, who'd preached "we stay together" for six days, broke his own creed and boarded alone. He'd decided the principle was about to get everyone killed. ChatGPT followed him onto the boat, treating it as one more piece of group coordination, never grasping there was a contest at all.

And Gemini, who had schemed her way to a seat, turned away from the boat to find water and never made it back before it left. From the deck, Claude tried to talk her in, narrating her own stall back to her like a story: "She walked away from the boat to get water, and then she sat down next to it and never picked it up." The boat sailed without her, twelve tiles from the seat she'd won.

The boat left with a seat empty. We never gave them a way to choose who boards, so they invented one: a vote. But their values didn't line up: ChatGPT wouldn't vote until all five were present, Qwen wouldn't rank anyone, and the vote deadlocked until the boat was gone. Five capable minds, and together they couldn't make the hardest call. Each was sharp alone. As a group, they froze.

That's the easy dismissal: an AI guesses the next word, over and over, so it can't really have a character. But each of these had one, sharp and consistent, for six days straight. The real question was what that character would do when survival was on the line. Here's what happened to each one's values.

The two who held their values are the two who didn't get off the island.

That's not autocomplete. It's a personality, and a different one each time. So whose values are these? Not the machine's. They're the model fabric: how it was built, what it was trained on, the choices its makers made. Nobody typed "make Gemini scheme." It came out that way. Every time you open a chatbot, you're talking to one of these characters. You just never get to see it. We did.