Good Start Labs
Open sourceGood Start Labs is a 2025 Every spin-out that sells game-based RL environments and agent and human gameplay data to AI labs. It also partners with game publishers (Bad Cards, Arkadium GameLab) to turn play into licensed training data, in-game agents and leaderboards. It publishes transfer studies from games to real work (Diplomacy to customer support, the 1830 railroad game to finance research), arXiv papers and open repos, and runs the Diplomacy and Humor (Bad Cards) Arena leaderboards.
What's new
- 1 Oct 2026benchmark
"Keep Going": 48-hour livestream of GPT-6 Astra (Codex) and Claude Opus 5.5 (Claude Code) autonomously writing StarCraft: Brood War bots, plus StarSkirmish Bench
Two coding agents wrote and revised StarCraft: Brood War bots and climbed a ladder of human-written bots, live for 48 hours. The post says GPT-6 Astra and Claude Opus 5.5 are functionally tied at the top of StarSkirmish Bench, and that together with GPT-6 Sol they are a clear step above every other model tested. StarSkirmish was built by Kai McPheeters, with GSL partnering on the RL environment.
Why it matters: A long-horizon eval of autonomous coding plus strategy that compares current frontier coding agents. It also shows GSL co-developing environments with outside builders.
- 15 Sep 2026research
"Verification Is the Bottleneck": GSL tests TypeSafe's Jev classifier as a cheap RL verifier
Across 6,003 rubric checks on financial-research answers, Jev agreed with Claude Fable 5.1 91.5% of the time. Agreement with the other models tested was 86.5-91.5%. GSL reports a cost of about $160 per million graded answers for Jev, against about $33,000 for Claude Fable 5.1. It also ran 10,500 game-task grading calls with zero failed verdicts.
Why it matters: A data point on cheaper reward and verification for RL pipelines, relevant to anyone grading rollouts at scale. The figures come from GSL's own study.
- 15 Sep 2026other
Latent Space profile of Good Start Labs: "Can Skills Learned in Games Transfer to Real-World Work?"
Interview with CEO Alex Duffy. GSL sells agent game trajectories, custom data from agents playing in publisher games, and full learning environments, mainly to frontier AI labs. It says the data is anonymized and stripped of PII. The interview covers the Diplomacy-to-customer-support and 1830-to-finance transfer results and still cites the October 2025 $3.6M round.
Why it matters: An independent outlet's clearest description of the commercial model (data plus environments sold to labs). It also indirectly confirms no newer round had been disclosed by mid-September 2026.
- 5 Sep 2026benchmark
Humor (Bad Cards) Arena leaderboard updated: 14,459 rounds, 2,390 human judges
The leaderboard measures how often a model picks the same card the human judge picked. The chance baseline is 41.1% and the top model scores about 52%.
Why it matters: A live human-preference benchmark built on real player data.
- 16 Aug 2026customer
"Why do we do what we do": GSL says AI labs already train on its environments and data
Mission post by Alex Duffy. It says "AI labs already train on our environments and data, our agents play alongside millions of people, and we're growing a team of researchers who built some of the field's foundational learning environments." No labs are named. The post credits the Factorio Learning Environment to Jack Hopkins and says Arcee's 400B Trinity model trained on GSL's Diplomacy environment.
Why it matters: A self-claimed lab-traction signal with no labs named. It also hints at research hires, though Hopkins joining is not stated outright.
- 4 Aug 2026research
"What a Railroad Game Taught a Model About Finance": RL on the board game 1830 lifts Nemotron 30B on vals.ai Finance-Agent
Finance-Agent score rose from 12.1% to 18.3%. Answers with confident false claims fell from 83.0% to 64.5%, while finance knowledge stayed flat. Single-turn QA training gave no transfer; the gain appeared only with a multi-turn terminal-agent setup. An autonomous interpretability agent (Goodfire) independently checked the findings. The adapters are on Hugging Face (nemotron30b-1830-grpo-adapters).
Why it matters: Evidence, reported by GSL itself, that game RL can transfer to real-world agentic tasks, mainly by cutting confident hallucination.
- 30 Jul 2026research
"Where an Agent's Intelligence Lives": COS-PLAY write-up (co-evolving harness and skill bank)
Joint work with UMD, USC and MBZUAI, accepted as a poster at Agent Skills '26. A Qwen3-8B with a co-evolved skill bank beat GPT-5.4 by 25.1% on average across four single-player games. Per-game gains: 2048 +41.0%, Super Mario +26.2%, Candy Crush +21.8%, Tetris +11.5%.
Why it matters: Shows harness and skill design can matter as much as weights. The paper and code are public.
- 27 Jul 2026open-source
Hugging Face org publishes diplomacy-235b-lora, 1830 Nemotron adapters and a run of chess models
The org lists 53 public models, 0 public datasets and a '1650-Elo Browser Chess' Space (Qwen3-1.7B chess on WebGPU). Recent updates: nemotron30b-1830-grpo-adapters (Jul 7), diplomacy-235b-lora (Jul 27), and chess models including latent-graft and DAgger variants (through Aug 20).
Why it matters: Open checkpoints from GSL's game-RL work that researchers can inspect. The earlier public datasets (e.g. GSLBenchmark) are no longer listed.
- 15 Jul 2026open-source
GoodStartLabs GitHub org adds a catanatron fork (Settlers of Catan simulator, GPL-3.0)
A fork of bcollazo/catanatron, last updated 2026-07-15. It sits alongside AI_Diplomacy (707 stars, no license specified, updated 2026-06-01) and an OpenTinker fork (Apache-2.0).
Why it matters: Possibly signals a Catan environment. GPL-3.0 licensing matters for reuse.
- Jul 2026partnership
"Measure What Matters. Get Paid for It.": GSL says it helped Arkadium launch GameLab, a public leaderboard of frontier models on casual games
GSL says it built GameLab's benchmarks, evaluated them against real users and trained a 4.6M-parameter (18 MB) Gin Rummy expert that beats human players about 90% of the time. GSL puts its serving cost at about $60 a year. GSL's post describes GameLab as launched in partnership with Meta and DeepMind. GamesBeat's launch coverage (2026-06-16) mentions only unnamed frontier-lab agreements and does not mention GSL, so GSL's role is self-described.
Why it matters: The first named publisher partnership beyond Bad Cards. It shows the publisher-side business: data licensing plus in-game agents.
- 22 Apr 2026research
COS-PLAY paper on arXiv (2604.20987), co-authored by Duffy and Marques
"Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks" by Xiyang Wu, Zongxia Li, Guangyao Shi, Alexander Duffy, Tyler Marques, Matthew Lyle Olson, Tianyi Zhou and Dinesh Manocha. It reports over 25.1% average reward improvement against four frontier LLM baselines on single-player game benchmarks.
Why it matters: Research output with academic collaborators.
Founding team
What practitioners should know
For researchers
- Open papers: Democratizing Diplomacy (arXiv 2508.07485, NeurIPS 2025 workshop), interpretability of multi-agent RL in Diplomacy (arXiv 2602.05183, ICLR 2026 workshop), and COS-PLAY with UMD, USC and MBZUAI (arXiv 2604.20987, Agent Skills '26 poster). Matthew Lyle Olson co-authors all three.
- Transfer results, all self-reported: Diplomacy RL improved Qwen 3 on Tau2 customer support and by more than 10% on Hanabi and Wordle. RL on 1830 lifted Nemotron 30B from 12.1% to 18.3% on vals.ai Finance-Agent, and the gain appeared only with a multi-turn terminal-agent setup.
- Open code and checkpoints: AI_Diplomacy (no license stated), forks of OpenTinker (Apache-2.0) and catanatron (GPL-3.0), and 53 public Hugging Face models including diplomacy-235b-lora and nemotron30b-1830-grpo-adapters. No public datasets are currently listed.
- GSL's August 2026 post says it is 'growing a team of researchers who built some of the field's foundational learning environments' and credits the Factorio Learning Environment to Jack Hopkins. It does not explicitly say Hopkins has joined GSL, so treat that as unconfirmed.
For program managers
- Seed-stage team. No new round found since the October 2025 $3.6M raise (Inovia and General Catalyst leading; Tirta Ventures also an investor). Two open roles, both Remote (US & Canada): Research Scientist ($220-300K) and ML Engineer ($165-250K).
- The homepage 'Trusted by' row shows Cohere, Arcee and OpenAI logos. These are self-claimed. The about page lists Arcee's Mark McQuade and Lucas Atkins and Cohere's Teresa Grandia as investors/advisors, so some logos may reflect investor ties rather than paid contracts. The one specific usage claim is that Arcee's 400B Trinity Large trained on GSL's Diplomacy environment 'among others'.
- A two-sided model: it sells environments and trajectory data to AI labs, and partners with game publishers on data licensing, agents and leaderboards. All partner figures are self-claimed: Arkadium GameLab; Bad Cards with 2M+ users; a Bad Cards agent running 20,000+ games a week; a 13% engagement lift.
For data ops
- The products page publishes no integration formats (Docker, MCP, API). Engagement is through the site's contact form.
- No SOC 2, security or trust page found. The publisher page says pipelines have 'player consent designed in from the start', and Latent Space reports that data is anonymized and stripped of PII. Neither gives detail.
- Check licensing before reuse: open repos are mixed (Apache-2.0 fork, GPL-3.0 fork, AI_Diplomacy unspecified).
Benchmarks & research
Products
Scale & velocity
Research depth
Capital
Security & compliance
Product
Buyer analysis
How we verified this
Related vendors
Sources
- goodstartlabs.com/ · 2026-10-05, Homepage: 'Trusted by' logos (Cohere, Arcee, OpenAI), product lines, latest research
- goodstartlabs.com/products · 2026-10-05, Products; Arcee Trinity Large and Qwen 3 Diplomacy transfer claims; Bad Cards 2M+ users and 20,000+ games a week
- goodstartlabs.com/benchmarks · 2026-10-05, Diplomacy (200,000+ matches, updated 2026-01-15) and Humor (14,459 rounds, 2,390 judges, chance 41.1%, updated 2026-09-05)
- goodstartlabs.com/benchmarks/diplomacy · 2026-10-05, Diplomacy Arena method (20 games against a Devstral Small 1.1 baseline)
- goodstartlabs.com/research · 2026-10-05, Research index with dates; two arXiv workshop papers listed
- goodstartlabs.com/research/keep-going · 2026-10-05, StarCraft livestream; StarSkirmish Bench; built by Kai McPheeters with GSL
- goodstartlabs.com/research/verification-is-the-bottleneck · 2026-10-05, Jev verifier study (91.5% agreement, $160 vs $33,000 per million graded answers)
- goodstartlabs.com/research/why-we-do-what-we-do · 2026-10-05, 'AI labs already train on our environments and data'; Factorio Learning Environment and Jack Hopkins; Arcee Trinity
- goodstartlabs.com/research/what-a-railroad-game-taught-a-model-about-fin · 2026-10-05, 1830 game RL transfer to Finance-Agent (Nemotron 30B, 12.1% to 18.3%)
- goodstartlabs.com/research/where-an-agents-intelligence-lives · 2026-10-05, COS-PLAY write-up; UMD/USC/MBZUAI; Agent Skills '26 poster; per-game gains
- goodstartlabs.com/research/measure-what-matters · 2026-10-05, Arkadium GameLab collaboration; Gin Rummy expert model
- goodstartlabs.com/for-game-studios · 2026-10-05, Publisher services; Bad Cards and GameLab named; consent language
- goodstartlabs.com/careers · 2026-10-05, Two open roles, Remote (US & Canada), with salary bands
- goodstartlabs.com/about · 2026-10-05, Founders Tyler & Alex; leads Inovia and General Catalyst; investor/advisor list incl. Arcee co-founders, a Cohere employee and Tirta Ventures
- www.latent.space/p/good-start-labs · 2026-10-05, 2026-09-15 interview with Alex Duffy; sells data and environments to frontier labs; $3.6M Oct 2025
- gamesbeat.com/arkadium-launches-gamelab-to-make-ai-ready-for-the-real-wo · 2026-10-05, GameLab launch 2026-06-16; frontier-lab agreements unnamed; GSL not mentioned
- arxiv.org/abs/2604.20987 · 2026-10-05, COS-PLAY paper, 2026-04-22; Duffy and Marques co-authors
- arxiv.org/abs/2602.05183 · 2026-10-05, Data-Centric Interpretability paper, 2026-02-05
- arxiv.org/abs/2508.07485 · 2026-10-05, Democratizing Diplomacy harness paper, 2025-08-10
- github.com/GoodStartLabs · 2026-10-05, Repos: AI_Diplomacy (707 stars, updated 2026-06-01), OpenTinker fork (Apache-2.0), catanatron fork (GPL-3.0, updated 2026-07-15)
- huggingface.co/GoodStartLabs · 2026-10-05, 7 members, 53 models, 0 datasets, 1 Space (browser chess); recent model update dates
- github.com/Tylermarques · 2026-10-05, Tyler Marques GitHub: 'CTO @GoodStartLabs', links tylermarques.com; avatar is a real photo of him
- every.to/@AlxAi · 2026-10-05, Alex Duffy's Every author page ('cofounder and CEO of Good Start Labs'); studio headshot avatar
- ai.engineer/speakers/alex-duffy · 2026-10-05, Duffy bio: AI Camp, VP of AI at Salt AI, Every; LinkedIn alex-d
- www.inovia.vc/active-companies/good-start-labs/ · 2026-06-07, Investor page: founded 2025, HQ Brooklyn NY, Toronto office, pre-seed co-led by Inovia Oct 2025, founders Alex Duffy (CEO)/Tyler Marques (CTO)
- www.inovia.vc/founders/company-founders/tyler-marques/ · 2026-06-07, Tyler Marques background: Waterloo, ~7 yrs AI/ML/DevOps
- every.to/on-every/our-new-incubation-raised-3-6-million-to-teach-ais-to- · 2026-06-07, Funding announcement $3.6M; investors General Catalyst, Inovia, Every, DeepMind angels; Every spin-out (direct fetch 403, content via search)
- tracxn.com/d/companies/goodstartlabs/funding-and-investors · 2026-06-07, Total $3.6M, 1 round, round date Oct 15 2025, 37 investors (4 institutional)
- www.linkedin.com/company/good-start-labs · 2026-06-07, Public snippet: founded 2025, ~10 employees / size 2-10, ~535 followers, specialties AI/Games/Data/RL/LLM training
- startups.gallery/companies/good-start-labs · 2026-06-07, Lists remote, 1-10 employees, NY; investor list appears to be boilerplate/inaccurate (conflicts with credible sources)
- every.to/diplomacy · 2026-06-07, AI Diplomacy project writeup / Alex Duffy background (ex-Salt AI, AI Camp, Every Head of AI Training)