Alex Duffy

Alex Duffy co-created AI Diplomacy with Tyler Marques to evaluate how language models negotiate, maintain relationships, and make strategic decisions. He is the cofounder and CEO of Good Start Labs, which was incubated at Every, and an Every contributing writer as of September 2026. His subsequent work has explored whether training models through Diplomacy improves their performance beyond the game.[4][3][7]

The project’s foundations took shape in 2025, when Duffy and Marques built AI Diplomacy both as a benchmark for complex model behavior and as a game that a public audience could follow.[6]

In his account of building the project, Duffy recalls that the initial approach converted the board, units, and possible moves into sprawling lists. Models made illegal moves, failed to form coherent strategies, and stalled games. He credits collaborator Sam Paech with asking whether a person could play from the same context—a question that prompted the team to replace exhaustive data listings with plain-text summaries of territorial control, map connections, and immediately relevant moves.[6]

The team also gave agents persistent relationships and private diaries. Each model could update its assessment of another player from −2 for an enemy to 2 for an ally. Diaries recorded goals and reflections on moves, negotiations, and outcomes; generated turn summaries helped agents decide what to do next. The repository documents how this memory combines with current board information and pathfinding that identifies nearby threats and opportunities to inform negotiations and orders.[6][4]

Duffy frames context engineering as communication: “At its core is the deeply human skill of communicating clearly.” His redesign emphasized giving models an understandable account of the situation they needed to act within.[6]

Duffy treated presentation as part of the benchmark’s purpose and chose a Twitch livestream to make the games approachable. He credits Marques with transforming his initial interface attempts into an interactive Three.js browser interface with streaming built in. Duffy supplied the accompanying audio, using Udio and mixing software to create the 20-minute loopable track “Diplomacy Toujours.” He launched AI Diplomacy onstage at the 2025 AI Engineer World’s Fair; Latent.Space’s recap named him the Tiny Teams best speaker and featured a lightning podcast conversation about the project.[6][1]

Publishing as Alexander Duffy, he coauthored Democratizing Diplomacy, submitted on August 10, 2025, with Samuel J Paech, Ishana Shastri, Elizabeth Karpinski, Baptiste Alloui-Cros, Tyler Marques, and Matthew Lyle Olson. The authors reported optimizing textual game-state representations so that a 24-billion-parameter model could reliably complete matches without fine-tuning. They also introduced Critical State Analysis, a protocol for rapidly iterating on and examining key game moments in depth, alongside case studies of persuasion and aggressive playstyles.[5]

Duffy’s later work used Diplomacy as a training environment, extending beyond the earlier harness for unmodified models. In an article published in February 2026, he reported that his team fine-tuned Qwen3-235B on thousands of Diplomacy rounds and observed over 10 percent improvement on other games such as Hanabi and Wordle. He also reported better performance on Tau2, a customer-support benchmark, and AssetOpsBench, an industrial-operations benchmark, without specifying improvement percentages for those two tasks.[7]

Duffy argues that games can teach transferable behaviors because they provide goals and feedback on decisions. He identifies context tracking, shifting priorities, and strategic communication as capabilities exercised by Diplomacy that also matter in customer support, where information may be incomplete and requests can change.[7]