Can Skills Learned in Games Transfer to Real-World Work?
Good Start Labs says the transfer showed up only when the game training made the model act like an agent, not just answer single turns.
The company trained a 30B model in *1830*, a railroad strategy game with stock-market mechanics, then tested it on financial research tasks. Both training setups improved in-game performance, but only the multi-turn terminal-agent design improved the Finance-Agent benchmark. Alex Duffy argues the harness is the point: it can force models to use tools, plan, and work in a verifiable environment. He calls the evidence a qualified yes, while saying broader transfer remains unresolved. Latent Space's note
The company trained a 30B model in *1830*, a railroad strategy game with stock-market mechanics, then tested it on financial research tasks. Both training setups improved in-game performance, but only the multi-turn terminal-agent design improved the Finance-Agent benchmark. Alex Duffy argues the harness is the point: it can force models to use tools, plan, and work in a verifiable environment. He calls the evidence a qualified yes, while saying broader transfer remains unresolved. Latent Space's note
score 5