We got access to OpenAI's RFT API with GPT5 and tried to see how good we could get it at one-shot Triton kernel generation. Some key decisions/observations:
1. tool use instead of multi-turn rl
2. skip SFT altogether
3. dataset curation was more important than dataset scale
4. reward hacks detection must be robust
5. models are getting a lot better at this
This was fun to work on. LLMs for writing kernels still has a long way to go. Its honestly a little surprising how decent they are now. I guess I've been pretty consistently "surprised" by codegen for a while now (meaning the last two years)