After reading through papers and the code, I'm still not clear if AlphaFold and the likes do actually calculate the 3d structure. My current feeling they are more like information retrieval engines. They find statistical 'similarity' features between the primary structures of the query protein and a database protein (or a family of those) with a known 3d structure, and that known fold is the 'calculated' 3d structure. This would mean that if there is no 'similar' protein with a known 3d fold then the program cannot 'predict' anything. I realize this could be a stupid layman question but here it goes.
Alphafold calculates the "folds" throughout a protein, thus its protein structural prediction. It uses convolution-based calculation, which is a modification of co-variance. The reason it doesn't use co-variance directly is complicated, but can result in artefacts.
Thus it looks for "simultaneous" amino acid mutations in protein alignments and believes those represent close spatial proximity - based on previous supervised learning of proteins, comprising an alignment and a known structure.
It has won two CASP competitions (world competition for protein folding prediction). The training is likely deeper both in data sets and ANN/CNN/RNN "layers" (deep learning jargon) than previous protein structural prediction models and hence its success. "Deepmind" the company behind it will be Tensorflow experts, rather than structural biologists.
I don't use it but have seen its structural accuracy first-hand. The accuracy statistic RMSD was impressive.
It is important to note it is built on the back of a long term effort by this community as to calculating protein structural prediction. So it's not "radical technology", they've generally increased the accuracy of established methods. Again they've likely done this with more extensive training.