They are OK if your problem is based on tabular data, but not for images and text. But the current paradigm is to reuse a pre-trained model, it's less more data intensive. In some tasks if you have a good backbone model you just need one or a few training examples.
SVMs absolutely work on text, TF-IDF + SVM is a very classic (and pretty solid) approach for classification. It's easily explainable and has known classes of problems.
