The Current

Startup Tests Whether AI Trained on Games Can Improve at Real Work

Good Start Labs trained an AI on a railroad game, and one version improved at financial research. According to Latent Space, the training design made the difference.

useful research · for technical · September 16, 2026

Latent Space reported on Good Start Labs, a company turning games into training material for AI models. In an interview, co-founder and CEO Alex Duffy said, "Games have always been these underrated educational tools. They're super approachable. They're very human." The company was spun out of AI media and tools company Every in October, with $3.6 million in funding from General Catalyst, Inovia, Every, and angel investors, according to the report. Duffy said the idea came from a 2025 Twitch stream of frontier models playing the game Diplomacy, which he said normally takes "days or weeks to play." At the time he was head of AI training at Every. Watching the models play, Duffy observed that each frontier model behaved differently: OpenAI's o3 won games by planning a future betrayal, while Anthropic's Claude Opus 4 refused to lie and, in his words, "got destroyed." Duffy concluded that training AI models on games such as Diplomacy could teach skills like strategic thinking, particularly because such games have outcomes that can be verified. According to the article, the company trained an AI on a railroad game, and one version improved at financial research, with the difference attributed to how the training was designed.

  • Good Start Labs spun out of Every in October with $3.6 million in funding from General Catalyst, Inovia, Every, and angel investors
  • One AI version trained on a railroad game reportedly improved at financial research, per Latent Space
  • OpenAI's o3 won Diplomacy games by planning betrayals; Claude Opus 4 refused to lie and lost

What it means for you

This is an early-stage research idea: a startup is exploring whether teaching AI to play strategy games can make it better at unrelated business tasks like research. It's interesting, but it's one company's finding reported by one outlet, not a product you can use or a proven method. For now it tells you something about where AI training may be heading, nothing more.

Who should care

People who follow how AI models are trained, and anyone tracking research directions in AI capability. Not relevant to day-to-day tool users.

Skip this if

You're a small business or individual just trying to get useful work out of existing AI tools — this changes nothing about what you can do today.

Sources: Latent Spaceread the original

← All stories