Choosing a First Model: Start with the Task, Then Run a Baseline

The article argues that model selection should aim to find a reasonable first candidate worth testing, not the single best model. It recommends defining the task, users, quality bar, and deployment constraints before looking at any model options. Teams should narrow the field to two or three candidates and build a small, representative test set that includes edge and bad inputs. The first run should be a simple, unoptimized baseline whose outputs are inspected closely, with failures treated as guidance for what to change next. Optimizations such as prompt tuning, retrieval, or fine-tuning come only after this baseline is established.