Training Data Building the GitHub for RL Environments: Prime Intellect's Will Brown & Johannes Hagemann
126 snips
Feb 10, 2026 
Johannes Hagemann, researcher and co-founder building environment design and RL tooling. Will Brown, AI researcher and co-founder focused on large-scale training infra and post-training workflows. They discuss Environments Hub, environments as evals and product surfaces, harnesses and agent interaction, democratizing frontier training, cybersecurity sims, recursive language models, synthetic data, and companies becoming AI research labs.
AI Snips
Chapters
Transcript
Episode notes
Harnesses Are Environment Components
- Harnesses are part of environments that define how models interact with tools, sims, or subagents.
- Use the environment abstraction because 'agent' or 'harness' alone is too narrow for the many interaction patterns to come.
Treat Environments As A Toolkit
- Use environments as a general toolbox: for SFT, distillation, RL, A/B tests, and prompt optimization.
- Reuse the same environment protocol to generate filtered training data and run systematic comparisons.
rCI Collaboration Drove Platform Work
- Prime Intellect collaborated with rCI on large-scale pre-training and post-training infrastructure.
- This partnership pushed them to build compute orchestration, training, and inference features needed by frontier labs.