Can Skills Learned in Games Transfer to Real-World Work?
Good Start Labs is turning games into training material for AI models, co-founder and CEO Alex Duffy told Latent Space. The company was spun out of AI media and tools company Every last October with $3.6 million in funding from General Catalyst, Inovia, Every, and angel investors.
Duffy said the idea came from a 2025 Twitch stream of frontier models playing Diplomacy, which he said normally takes days or weeks to play, during his time at Every as head of AI training. He observed OpenAI's o3 winning games by planning a future betrayal, while Claude Opus 4 refused to lie and was defeated. Duffy concluded that training models on verifiable-outcome games like Diplomacy could teach skills such as strategic thinking, and wrote in a later Every article that fine-tuning a model on Diplomacy improved performance on customer support and industrial operations benchmarks.
Good Start Labs recently trained a 30B model inside 1830: The Game of Railroads and Robber Barons and tested it on financial research tasks. The published results compared single-turn question answering with a multi-turn terminal agent that uses tools to explore, plan, and adapt. Both designs improved their in-game objectives, but only the terminal-agent design improved performance on the Finance-Agent benchmark.
Duffy said Good Start Labs compares every new model and that more capable models tend to be better at games but continue to diverge on personality axes including betrayal, collaboration, and theory of mind. He said a more capable model needs less handholding to finish a task, but for treating the environment as curriculum, the harness matters more, not less.
Duffy said the company mainly sells data and learning environments, with frontier labs as its main customers. The data includes agent trajectories from games and custom data for game publishers, anonymized and stripped of personally identifiable information before being sold to model developers. Learning environments are full games that models can play end to end, with tasks using the game engine as a verifiable source of rewards.
Good Start Labs is also training a general model from expert models built for specific games. Duffy said today's evidence supports that goal-directed execution matters and reasoning transfers, citing the 1830 finance task, Diplomacy training producing a better customer support agent, and improved tool use downstream from every environment they have built. He said how broadly and reliably game skills transfer to real-world work remains an open question.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Good Start Labs trained an AI on a railroad game — and one version improved at financial research. The difference was the training design.