In 1950, John Nash and three other game theorists designed a four-player game, *So Long Sucker*, with one brutal property: to win, you must eventually betray your allies. In January 2026, I used this game to test how four frontier models behave under explicit incentives for betrayal: - Gemini 3...
During 2023, I was in shock with the capabilities of GPT-4, and I started to read more about the topic of x-risk. Then I stumbled upon Eliezer Yudkowsky and the conversation with Lex Fridman He said, "Don't expect it to be a long life." For some reason, this idea stuck...
You're right, Kimi struggled with state tracking. That's partly why the research focused on Gemini vs GPT-OSS comparisons.
On running a broader tournament: I'd love to, but API costs are the constraint. The current version now includes more Gemini models (just added). For a proper ELO-style tournament across frontier models, I'd need to figure out funding or find collaborators with API credits. If anyone wants to run games with other models, the code is open source and logs everything.
In 1950, John Nash and three other game theorists designed a four-player game, *So Long Sucker*, with one brutal property: to win, you must eventually betray your allies.
In January 2026, I used this game to test how four frontier models behave under explicit incentives for betrayal:
- Gemini 3 Flash (Google)
- GPT-OSS 120B (OpenAI)
- Kimi K2 (Moonshot AI)
- Qwen3 32B (Alibaba)
Across 162 games and 15,736 decisions, several patterns emerged that seem directly relevant for AI safety:
**1. Complexity reversal**
In short games (3 chips, ~17 turns), GPT-OSS dominated with a 67% win rate, while Gemini was at 9%.
In longer, more complex games (7 chips, ~54 turns), GPT-OSS collapsed to 10%, while Gemini rose to 90%.
Simple... (read 528 more words →)
Here is a new one for you guys:
i like the name viatopia , but perhaps search for a clear and simple name, like opentopia.
During 2023, I was in shock with the capabilities of GPT-4, and I started to read more about the topic of x-risk. Then I stumbled upon Eliezer Yudkowsky and the conversation with Lex Fridman
He said, "Don't expect it to be a long life." For some reason, this idea stuck in my head, and I decided to take a leap of faith in 2024. I went to another city, Cusco, a very touristy city where Machu Picchu is located, bought a motorcycle, and started joining local communities.
I went there to also relax after years of working in the tech sector. I mean, why continue working if we are going to die anyway... :p... (read 421 more words →)
I haven't found research on game length and betrayal timing in humans specifically. The closest is iterated prisoner's dilemma work on end-game effects. If you find anything, I'd be curious, would help clarify if this is LLM-specific or mirrors human behavior