Can Skills Learned in Games Transfer to Real-World Work?

What it claims

Good Start Labs trained a 30B model inside the board game 1830: The Game of Railroads and Robber Barons under two harness designs. Both designs improved their in-game objectives, but only the multi-turn terminal-agent design improved the Finance-Agent benchmark. The page reports no scores, deltas or baselines for either result. Duffy separately reports that fine-tuning on Diplomacy improved customer support and industrial operations benchmarks, and that "every environment we've built also improves tool use downstream." Neither claim comes with numbers.

How it works

The 1830 tasks are built to mirror a finance workflow:

1. The model queries a database of past games.
2. It writes the findings into an Excel file.
3. It builds functions in that file and reasons over the data.
4. It calculates the answer from those functions.

The game engine supplies the verifiable reward.

flowchart LR
  G[1830 game engine<br/>verifiable reward] --> A[Single-turn Q&A:<br/>state → next move]
  G --> B[Multi-turn terminal agent:<br/>tools, plan, adapt]
  A --> I[In-game gains]
  B --> I
  B --> F[Finance-Agent gains]
  A -.-x F

Two other levers are mentioned:

• Denser rewards. An expert model can provide stepwise rewards instead of a single sparse outcome reward.
• Presentation modality. Duffy says the same game shown as images, as natural text or as Python teaches different things.

COS-PLAY is a paper co-authored by Duffy and Marques with university researchers. In it, a decision agent acts with the help of a "learnable skill bank." A separate skill-bank agent reviews each trajectory and edits the bank, and the updated bank feeds into the next run.

Duffy argues that more capable models make the harness more important, not less. Their example is GPT-6 Astra, which "reports doing less chain-of-thought and jumps to answers." In their view, the harness is what forces a model to work a particular way, for example by computing with code rather than "in its head," so the result can be trusted.

How to use it

No code, weights or commands are released. Good Start Labs sells three things, mainly to frontier labs:

• Agent trajectories from game play, each recording what the agent observed, what it decided, the action it took and the outcome.
• Custom data from agents running live in publishers' games, anonymized and stripped of PII.
• Full end-to-end game environments, plus tasks that use the game engine as a reward source for problems other than winning.

The company is also merging its per-game expert models into a general "game intelligence" model. The pointer worth following is the COS-PLAY paper.

Limits

• No quantitative results. The page gives no Finance-Agent numbers, no baseline model, no training compute and no name for the 30B base model.
• Narrow transfer evidence. The 1830 task was deliberately designed to resemble finance work. The interviewer notes it is unclear whether transfer extends further, and Duffy concedes that how broadly and reliably skills transfer "remains an open question."
• Second-hand support. Duffy's broader backing for transfer comes from other sources: a Surge AI article reporting that office-work post-training improved coding, and DeepSeek R1.
• Unscored behavioural claims. Newer frontier models are said to diverge on betrayal, collaboration and theory of mind. That is stated, not measured on the page.