You manage an AI research lab balancing rapid capability development against mounting alignment risks. Push too hard and your revenue crashes from safety failures; move too cautiously and competitors pull ahead. This incremental game embeds AI safety as functional mechanics rather than narrative window dressing, forcing genuine strategic tradeoffs between progress and stability.

About this game

AI Lab is a browser idle game about training and selling AI models.

Phase 1 is free and complete: it takes about 40 minutes to reach your first research paper, and the game continues after that. Phases 2 to 5 will be a paid game on Steam.

Click to make compute. Buy GPUs. Train a model, deploy it, and earn money every second until it goes stale and you need a better one.

The choices come from real machine learning. Push Model Complexity up and your models score higher, but they go stale faster and lose more AI Safety. A Safety Check costs 20% more compute per run and cuts the AI Safety that run loses. Models earn more at high AI Safety, up to 1.5x.

Each paper starts the lab over with a speed bonus. Progress saves in your browser.

After you use a part of the game, a Lab Notes page appears. It explains the real idea behind that part, like overfitting or RLHF.

**Returning players**
- Phase 1 is longer, and publishing no longer ends the game.
- Controls appear one at a time, each with one line on what to do with it.
- Rent Compute earns at most 30% of what your models earn.
- Upgrades have their own tab.
- The full list is in the 2.0 update post.

**How this game was made**
Loss Curve was made by one developer working with an AI programming assistant. Parts of the code were written by the AI assistant under the developer's direction, including the interface icons, the sound effects, and the store art, which are all drawn or synthesized in code. Game text was drafted by the developer and polished with AI assistance. No image generators, AI voice, or AI audio tools were used, and nothing is generated by AI while you play.