RL infrastructure

- RL training infra: exists, but mostly "RLHF"-style
- Want options besides "wait for big lab tuning APIs"
- How can we plan ahead?
- Unknowns:
- Cost to do RL for agentic tasks?
- How small can the models be?
- Generalization across tasks?
- How to design effective rewards?
- Opportunities:
- Open-source infra for DIY
- Services for agentic RL

PyTorch

unsloth

Hugging Face

Ai2

Create a fine-tuned model