Interesting post. Did the difference in wall clock training time take the reduction in cold start time into account? Seems like that could be a significant factor for small jobs and negligible for large ones.

higher abstraction than Tinker, more flexible than OpenAI RFT. i like integration to production inference, so i can switch between training and inference for continuous learning.

Will the rate limits go higher? How about other models? Qwen 2.5 is nice but 3 is nicer