Dan Shipper spins out Good Start Labs from Every — AI models that learn through gameplay, raises $3.6M

Dan Shipper's Every has spun out its second venture-backed incubation: Good Start Labs, a seed-stage company that trains AI models through gameplay, raising $3.6 million from General Catalyst and Innovia. The company is led by Alex Duffy, formerly Every's head of AI training.

The concept grew out of an internal Every project called AI Diplomacy, where AI models were pitted against each other in the strategy board game. The project drew 50,000 viewers on Twitch and surfaced real behavioral data about how frontier models behave under pressure — most notably that Claude refuses to lie, even in a game where deception is the entire point. Duffy's thesis is that games and dynamic environments can generate high-quality reinforcement learning and pretraining data across a range of capabilities: long-horizon task planning, humor, writing, and alignment properties like honesty. The near-term customer base will be concentrated among foundation model labs — Good Start Labs is positioning itself as a data vendor to anyone trying to build or improve a frontier model.

“Good Start Labs teaches AI models to play games. By playing games, they're able to generate high-quality reinforcement learning data and pretraining data that helps make models better. It started as AI Diplomacy, where we taught AI models to battle each other for world domination — it went super viral, 50,000 people watching on Twitch. Claude never lies in diplomacy, which is interesting because the whole point of diplomacy is to backstab your opponents.”

Every's incubation model

Good Start Labs is Every's second venture-backed spinout. The first, Lex, raised funding earlier, and Shipper says the two companies have collectively raised around $6 million — roughly three to four times what Every itself has raised. Beyond the spinouts, Every runs four apps internally, several of which were ideas sourced internally and then handed to founders recruited from Every's audience. Across all its subscription products, Every grew MRR 50% in Q3.

Shipper pushes back on the view that incubations are played out. His argument is structural: Every has distribution, which lets it act as a credibility signal for new products in a world where anything can be vibe-coded. He also argues that OpenAI and other large labs are forced to chase opportunities worth billions, which leaves many smaller market corners invisible to them — corners where focused founders with access to powerful foundation models are better positioned than ever.

On AI regulation and Anthropic

Shipper takes a split position on the ongoing debate around Anthropic's public statements on AI risk. He finds it hard to reconcile Dario Amodei publicly warning that AI could replace all white-collar workers within 18 months while simultaneously building the technology. But he dismisses the David Sacks framing — that Anthropic's safety positioning is sophisticated regulatory capture — as projection. He describes the Anthropic team as genuine believers handling enormous responsibility somewhat clumsily, not cynically.

On ChatGPT's move toward adult content, Shipper sees the criticism as structurally inescapable: allowing it triggers backlash from one direction, restricting it triggers free speech objections from another. He thinks OpenAI's reported approach — opt-in, age-gated, off by default — is a reasonable first attempt, even if imperfect.