Image: reddit.com · I taught an LLM to win the Cold War
An open-source Twilight Struggle engine for testing AI players
With no open-source version of the Cold War board game available, Claude implemented the engine; a GPT model then managed to beat the digital game's AI once.
Credits
Earliest dated source:
The full story
The maker long wanted a bot to play Twilight Struggle, the Cold War board game with hidden cards and multi-use events, but there was no open-source implementation. This summer they used Claude to implement the engine.
They then tested whether a language model, GPT-5.6, could play it. With lots of domain knowledge and trial and error it beat the AI in the digital version in at least one game, though it still makes mistakes such as starting a nuclear war after a failed coup. The code is public.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: I taught an LLM to win the Cold WarReddit poster: alekpinelPosted
The maker's post on r/ClaudeAI.
If it disappears: archived copy of Reddit: I taught an LLM to win the Cold War (opens in a new tab)
- The project: siestainsolaris.substack.com
Linked from the post.