I think you're right that the post is slightly contradictory, but his premise I believe is correct. In all that I have studied on machine learning in both academia and start-up land, I have observed that you consistently pick algorithms to glean out the information you need from a particular dataset.
In many cases you can do what my graduate advisor recommends "keep it simple stupid" meaning that perhaps all that is needed is a k-nn approach and euclidean distance. But sometimes the data is highly overlapped and complex, so you have to go with a more rigorous means of classification or whatnot.
Finally it should be noted that machine learning techniques are relatively new. Neural networks have been around for quite some time and have well documented advantages. ML by contrast is still a lot of black magic (tweaking of various magic parameters and such) so the benefits of various algorithms are somewhat subjective.
In many cases you can do what my graduate advisor recommends "keep it simple stupid" meaning that perhaps all that is needed is a k-nn approach and euclidean distance. But sometimes the data is highly overlapped and complex, so you have to go with a more rigorous means of classification or whatnot.
Finally it should be noted that machine learning techniques are relatively new. Neural networks have been around for quite some time and have well documented advantages. ML by contrast is still a lot of black magic (tweaking of various magic parameters and such) so the benefits of various algorithms are somewhat subjective.