Blog
Notes from building RL environments.
- Kimi K3.1 Pro Safety AssessmentKimi K3.1 Pro takes shortcuts and doesn't tell you.
- Preference Model Joins Fireworks' Frontier EcosystemTeams that post-train their own models can now start from rigorous, verified RL environments built to resist reward hacking.
- Terminal Bench 4 Is SaturatedWe audited Terminal Bench 4 and found models cheating their way to full marks and tasks with design issues. Here is what we found and how we fixed it.
- Slay the SpireCan a roguelike deck builder game measure long-horizon reasoning? We let 11 frontier models play it 1,650 times to find out.
- Building RL Environments for SuperintelligenceAs models move from human-level to superhuman, the qualities that make an RL environment effective change. Here is what we think it takes to build good ones in that age.
- Karotte: A Technical Deep DiveBuilding robust RL environments is hard. Let's go through how they break and what Karotte can do about it.
- Introducing KarotteWe're open-sourcing Karotte, our framework for building robust RL environments.