How does a random forest differ from gradient boosting?

Why

The scikit-learn ensemble guide describes the contrast directly. In a random forest each tree is built from a bootstrap sample of the training data and considers a random subset of features at each split. The trees are deep and overfit individually, but their errors are partly independent, so averaging them gives a stable prediction.

Forests are hard to badly misconfigure, and more trees rarely hurt except in time. Gradient-boosted trees work sequentially. Each new tree is fitted to the remaining errors of the ensemble so far, and a learning rate controls how much each tree contributes.

The guide notes that boosting uses shallow trees that underfit individually and builds them in sequence, and that histogram-based gradient boosting is often a competitive alternative to forests, fast on large tables and able to handle missing values natively. Because boosting keeps reducing training error, too many trees or too high a learning rate will overfit, so it relies on early stopping and tuning. For a tabular assay dataset, both are reasonable starting points after a linear baseline.

On small biological datasets the gap between them is often smaller than the gap created by a leaky validation scheme, so validation design matters more than the choice.