---
schema: 1
type: term
title: '**Diplomacy**'
description: >-
  **Diplomacy** — in this channel's material, the classic negotiation board game
  used as an arena for evaluating and training AI models. An AI version of
  Diplomacy was built at Every to observe how different models negotiated,
  collaborated, and betrayed one another
  [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/diplomacy-9e78feae0d4b'
markdown: 'https://liquidgold.cc/channel/everyinc/explore/term/diplomacy-9e78feae0d4b.md'
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \*\*Diplomacy\*\*

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/diplomacy-9e78feae0d4b

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

**Diplomacy** — in this channel's material, the classic negotiation board game used as an arena for evaluating and training AI models. An AI version of Diplomacy was built at Every to observe how different models negotiated, collaborated, and betrayed one another [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].

## The game behind the termIn this channel's material, **Diplomacy** refers to the classic negotiation board game used as an **AI evaluation and training arena**. Alex Duffy and Tyler originally built an **AI version of the game Diplomacy** at Every so they could explore how different AI models negotiated, collaborated, and even betrayed one another [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].## Why Diplomacy specificallyThe game exposes capabilities that static benchmarks miss:- It is **dynamic** and head-to-head, and not easily saturated — unlike benchmarks that can be *taught to the test* [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].- Gameplay makes structured behaviors observable, such as **map understanding** and putting out orders correctly [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].- Reading gameplay data reveals different *play styles* — some models, notably **o3** and **Llama 4**, were the biggest schemers, betraying allies and lying while planning betrayal ahead of time [^01a0f15a-d881-7a49-9531-148c764e2dcd@0].This work continues at **Good Start Labs**, which uses games to **test, train, and gather human feedback** on AI models — making them not just smarter, but *better for people* [^01a0f15a-d881-7a49-9531-148c764e2dcd@0]. Diplomacy sits within the broader exploration of architectures and capabilities beyond standard text-prediction LLMs [^8bd084ce-8d16-5f81-895a-fc2bc3b5d174@44].- **Diplomacy** here means the negotiation game played by AI models, not diplomacy in the general sense.- It was built as an AI experiment at **Every** and continues at **Good Start Labs**.- Games like Diplomacy expose model behavior — cooperation, negotiation, betrayal — that static evaluations miss.