A linear model can fit unstable coefficients when inputs overlap or the training sample is small. Regularization adds a penalty to the fitting objective. Ridge uses a squared coefficient penalty; lasso uses an absolute coefficient penalty and can set some coefficients to zero. The alpha setting controls the penalty strength.
The penalty acts on coefficient size, so input units matter. Put StandardScaler before the estimator and fit both steps on training rows. You will use an explicit alpha to make the experiment reproducible, not because one value is suitable for every dataset. Later lessons select settings with cross-validation. A regularized fit can have worse training error and better held-out error, but that tradeoff must be measured. Inspect the predictions instead of assuming that any penalty improves a model.