Game-trained AI shows transferable skills to financial research via environment design

Good Start Labs trained a 30-billion-parameter model inside the 1830: The Game of Railroads and Robber Barons, a deterministic strategy game with a stock-market mechanic, and then evaluated the model on financial research tasks. The company reports that only the version trained with a multi-turn terminal-agent design—one that uses tools to explore its environment, plan, and adapt in real time—improved performance on the Finance-Agent benchmark, while a single-turn design improved only in-game objectives.

The experiment was designed to test whether habits learned in a game could transfer outside the game. In the game, players bid on railroad-company stock to build logistics networks, mirroring workflows in financial research. Good Start Labs set up tasks where models retrieved information from a database, structured it in a spreadsheet, reasoned over it, and produced calculations—tasks representative of finance workflows but executed within the game environment.

Good Start Labs was spun out of Every in October 2025 with $3.6 million in funding from General Catalyst, Inovia, Every, and angel investors. The company’s co-founder and CEO, Alex Duffy, previously observed frontier models playing Diplomacy on a 2025 Twitch stream and noted differences in strategic behavior, such as OpenAI’s o3 planning future betrayals while Anthropic’s Opus 4 refused to lie. Duffy concluded that training on games with verifiable outcomes could teach strategic thinking.

Duffy and co-founder Tyler Marques launched Good Start Labs to explore which games and training designs teach useful, verifiable skills. Duffy argues that the way a game is presented to an AI—whether via images, natural text, or code—changes what the model learns. He describes the company’s goal as figuring out how to design learning environments that teach specific capabilities, a question also explored in a co-authored paper titled COS-PLAY: Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks.

Duffy reports that newer, more capable base models tend to perform better in game environments, but differences persist along axes like betrayal, collaboration, and theory of mind. He adds that while more capable models may need less handholding to finish a task, the harness and environment design remain critical when the goal is to shape how the model works—treating “the environment as curriculum.”