t-SNE is a nonlinear dimensionality reduction method designed for 2D or 3D visualization: it preserves local neighborhoods, which makes groups appear. Distances between clusters and their sizes are not reliable, and the result depends on the perplexity hyperparameter and on randomness.
Example
TSNE(n_components=2, perplexity=30, random_state=42).fit_transform(X_s)