Good Start Labs uses games for autoresearch

Good Start Labs is betting that games can become the missing training loop for capable, testable agents.

• –Verifiable game scores provide cleaner rewards than subjective human judgments.
• –StarSkirmish turns autoresearch into a visible competition between AI-written bots and years of human engineering.
• –The company’s Diplomacy results suggest game training may transfer to customer support, finance, and other real-world tasks.
• –Public leaderboards make progress easier to reproduce and harder to disguise with benchmark gaming.
• –The major open question is whether narrow game gains reliably transfer to broad intelligence.