Merve Noyan Stopped Writing Training Scripts — Her Agent Just Fine-Tuned 18 Models Solo for $11.40

I watched Merve Noyan’s “Your Agent Can Now Train Models” talk three times this week. It went up on the AI Engineer channel three days ago, hit 17,300 views in 72 hours, and now sits as the second-most-watched talk on the entire @aiDotEngineer feed — beaten only by the “CI/CD Is Dead” pitch from Hugo Santos two slots above it. Both are screaming the same thing in different keys: the loop where a human writes a training script, picks a GPU, watches loss curves, and pushes a checkpoint is about to look as quaint as configuring Tomcat by hand.

So I stopped watching and started testing. Over the last two days I sat in front of Claude Code with the hf-llm-trainer skill installed, typed plain-English instructions like "fine-tune Qwen3-0.6B on this dataset," and let the agent pick the hardware, write the TRL script, submit the job to Hugging Face Jobs, monitor Trackio, and push the finished weights to the Hub. I did this 18 times across SFT, DPO, GRPO, and one vision-language model run. Total bill: $11.40. Total time I spent writing Python: zero lines. I stopped writing training scripts at task 4, when I realized the agent was making better hardware choices than I would.

This is the part of MLOps that just got automated, and it’s not a toy demo. The skill ships from…