Good Start Labs

Good Start Labs builds reinforcement learning environments, world-model data sets and benchmarks derived from games, which it positions as compact worlds with clear objectives and verifiable signals of success. The company curates games into interactive learning environments targeting capabilities such as long-horizon planning, code, spatial reasoning, memory, negotiation and collaboration, and separately collects real human gameplay across text, image and 3D worlds to produce world-understanding data covering physics, space, navigation, perception, motion and problem solving. It states that full game engine access supports higher labeling accuracy.

Alongside data and environments, the company runs public benchmarks and leaderboards framed as competitions. Its GitHub organization describes Diplomacy Arena, based on the AI_Diplomacy repository in which frontier models play the board game Diplomacy as a test of long-horizon reasoning, negotiation and honesty, and LOL Arena, a humor preference alignment benchmark informed by large-scale human voting. The organization also maintains forks of catanatron (a Settlers of Catan bot simulator) and OpenTinker (RL-as-a-service infrastructure, Apache-2.0). The company hosts tournaments and live events in which participants take part through prompts and strategy, and publishes datasets on Hugging Face such as GSLBenchmark and gsl-benchmark-logs.

The company describes its category as "alignment through entertainment": using real human preferences and competitive pressure inside games to reveal and improve what models can do. It publishes articles and technical write-ups on its site and reports that its research was accepted to a NeurIPS workshop on multi-turn interactions in LLMs, with a submission to AAAI.

Founding story

Founders Alex Duffy (Co-Founder and CEO), previously Head of AI Training at Every and creator of AI Diplomacy, and Tyler Marques (Co-Founder and CTO), a University of Waterloo alumnus with an applied AI/ML and DevOps background, started the company out of Every in 2025. The company's about page frames the origin in the founders' own experience of learning through games such as RuneScape, Minecraft, League of Legends and Mafia, and argues that the same experiential learning can be applied to aligning AI.

Business model

Good Start Labs sells game-derived reinforcement learning environments and data to AI model developers, and reports live contracts with model training labs. It also pursues game publisher partnerships, integrating its agents into existing games and using player behavior as a source of evaluation and training signal. Its careers page describes demonstrating measurable model improvements to current and prospective lab partners as a core objective.

Revenue is described in sourced material only in terms of contracts with model training labs for environments and data; no pricing or revenue figures are disclosed. A vendor directory records revenue signals as unknown.

Traction

The company reports that its agents are integrated into games with millions of players, that it holds live contracts with model training labs, and that research was accepted to a NeurIPS workshop on multi-turn interactions in LLMs with a submission to AAAI. It maintains research partnerships with PhDs at Princeton, Rice, UCSD, HKUST, UMD and NYU. Its AI_Diplomacy repository has roughly 698 GitHub stars and 96 forks, and its Hugging Face organization lists 9 team members, 47 models and 8 datasets. Bad Cards, a Discord Cards-Against-Humanity-style game, is cited as a LOL Arena partnership and data source.

Latest developments

The company published articles in July and August 2026 on game-based training, including a write-up on a model trained on a railroad board game improving at SEC-filings research and a piece on where an agent's intelligence resides across model, memory, harness and skill system. It is hiring a Research Lead for LLMs, games and multi-agent environments in Toronto or New York at $140K-$250K base plus equity, to lead a multi-year research program initially focused on agents and computer use, and reports a NeurIPS workshop acceptance.