FDE Times: learn evals before fine-tuning, because most FDE job ads don't ask for deep post-training skills

FDE Times compared several FDE job descriptions and first-hand accounts to answer a common question from engineers moving into the role: should they invest in fine-tuning skills such as SFT, LoRA, DPO, GRPO and PPO? In the ads it reviewed, that demand came from only one employer, Together AI, whose FDE Post-Training and FDE Strategic Inference roles pair fine-tuning with inference engines like vLLM, TensorRT-LLM and SGLang, and with leading technical projects directly with customers.

An account from an FDE at inference platform Baseten points the same way: customers usually bring an already fine-tuned or open-source model, and the FDE's job is to serve it, for example cutting latency without raising costs. On the application side, the piece cites OpenAI's FDE work with Morgan Stanley, where the team tuned retrieval rather than the model and treated an eval set as mandatory, in line with OpenAI's guidance to start with prompts and an eval baseline before RAG or fine-tuning.

Why it matters for FDEs: the article suggests a learning order of evals and prompts first, then retrieval and data quality, then serving and inference, with fine-tuning last unless you're targeting infrastructure-layer roles. It also advises turning "fine-tuning experience" on a CV into a story with numbers: the baseline score, how the errors broke down, and why you did or didn't fine-tune.