Comparisons
Pairs that get conflated in real metric reviews and real design docs — batch and online, precision and recall, data drift and concept drift, validation and test. Neither column wins; what decides is the problem. Each record leads with the confusion, because the confusion is the reason the record exists.
Fine-tuning vs retrieval
Fine-tuning is reached for to "teach the model our documents", which it does badly: weights are a poor database, the facts are blurred into the parameters, they cannot be updated without another run, and the model cannot cite what it does not retrieve. Retrieval is reached for to "make the model behave differently", which it does badly too: no amount of context changes a model's style, its output format or its decision boundary the way training on examples does. The strongest form of the fine-tuning side is that a small model fine-tuned on a few thousand good examples routinely beats a large model prompted with the same examples, at a fraction of the inference cost, for a narrow task. The strongest form of the retrieval side is freshness and provenance: the answer can point at the document it came from, and yesterday's change is live today. They are not rivals for most systems — a fine-tuned model reading retrieved context is the common shape. This domain teaches the training half: transfer learning, full and parameter-efficient fine-tuning, and embedding training; retrieval-augmented generation as a system lives in Agentic Engineering, at RAG Overview.
When the behaviour needs to change — a style, a format, a classification boundary, a domain vocabulary — and you have labelled examples of the behaviour you want. The knowledge is in the weights and updates only when you retrain.
When the facts need to change and change often — documents, records, prices — and the model's job is to read them. The knowledge is in the store and updates when the store does, with no training run.
| Dimension | Fine-tuning — change the model's weights on your data | Retrieval — keep the weights, put your data in front of the model at request time |
|---|---|---|
| What changes | The weights | The context the model sees |
| Good at changing | Behaviour, format, boundary, vocabulary | Facts, documents, freshness |
| Update cost | A training run and a redeploy | A write to the store |
| Provenance | None — it is in the weights | The retrieved source |
| Data needed | Labelled examples of the target behaviour | The documents, embedded and indexed |
| Inference cost | Lower — a smaller specialised model | Higher — longer inputs, a retrieval step |
| Characteristic failure | Forgetting; stale facts; overfitting the examples | Retrieving the wrong passage; ignoring it |
| Where it is taught | Here — foundation models and fine-tuning | Agentic Engineering |
Model families compared
Linear, trees, boosting, neural and k-NN — compared without naming a winner, and with the block that says where the comparison stops being true.
No column is a winner. Each family is a set of assumptions about the data — linear separability, axis-aligned interactions, additive residuals, a representation that can be learned, locality in feature space — and the right one is the one whose assumptions your data happens to satisfy at the size you have. The reflex answer “XGBoost” is the red flag this table exists to catch: it is often right on tabular data and it is never right as a reflex, because it skips the baseline that would have told you whether anything more than a linear model was needed. The where this comparison misleads block on every row is the part worth reading.
Count independent entities, not rows — a million events from ten thousand users is ten-thousand-sized data for generalising to new users. And the neural column flips completely with transfer learning: a pretrained model fine-tuned on two thousand images beats every other column on that task.
The "boosting wins on tabular" claim is true for a tuned model on tens of thousands of clean rows with a validation set. It is not true for three hundred rows, for a problem whose signal is linear, for a regulator who wants coefficients, or for a team with no time to tune — and the gap to the linear model is often inside the error bar.
The columns are not competing on the same input. Once a pretrained network has produced an embedding, a linear model or k-NN on top of it is often within a few points of full fine-tuning — so "neural for images" usually means "a neural representation, then whichever head is cheapest".
Interpretability is not one property. A linear model with two hundred correlated features and L1 selection is harder to explain honestly than a depth-three tree; and every column's "explanation" is undermined equally by a leaked or proxy feature, which the explanation will present with confidence.
Training cost is dominated by the number of runs, not the run: a boosted model tuned over two hundred configurations costs more than a network fine-tuned once. And the k-NN column's zero is a loan repaid on every query.
The model is rarely the slow part. Feature retrieval, a network hop and JSON serialisation usually dwarf any of these numbers, so a latency budget is a question about the serving path before it is a question about the family.
A low burden on the modelling side does not remove the burden on the serving side: every column's features must be reproduced at prediction time with the same code, the same freshness and the same point-in-time semantics. Trees remove the need to engineer features, not the need to serve them.
Native handling is convenient and dangerous in equal measure: it lets a tree model learn that "missing" predicts the target, which is fine until serving produces missingness for a different reason — a timeout, a new form — and the model reads the outage as a signal.
Calibration only matters if a downstream decision multiplies the score by a cost or compares it to a probability threshold; a pure ranker does not need it. And calibration measured on the validation set drifts with the base rate in production, for every column alike.
Neither behaviour is "correct". A tree that predicts last year's maximum for a record-breaking day is wrong; a linear model that predicts a negative price is wrong differently. The honest answer is a monitor on inputs outside the training range and a fallback for them, whichever family serves.
Irrelevant is not the danger — leaky is. Every column will seize a feature that carries the answer, and the more capable the model the more efficiently it does so; robustness to noise says nothing about robustness to leakage, which only a point-in-time audit provides.
Every column is wrong somewhere, and the question that finds where is the same for all of them: how much data, of what modality, at what latency, with what interpretability requirement, judged by which metric, at what cost to train and serve. A model family chosen before those are answered is a guess with a library name.