I remain convinced that 90% of what vendors are calling machine learning or AI is three linear regressions in a trench coat.
Part of my skepticism is years of wading through obvious bullshit from lazy marketers slapping “Now With AI/ML” stickers on all kinds of products that are barely one linear regression. That shit gets old fast.
There is real machine learning, and it is quite useful in the narrow use-cases it’s designed for. Phone cameras are getting pretty good at figuring out if something looks like a face and not focusing on it, for example. Transcription software only writes “not for prophet” on every third attempt.
But artificial intelligence doesn’t exist. It’s just a silly label stuck onto the side of a computer that is too complex for its creators to explain. We already have plenty of systems that are unaccountably racist, so I don’t see why we need to create more.
Non-human intelligence does exist though: it’s called dogs, and we’ve been using them to help us with all manner of tasks for thousands of years. Computers are not even a tenth as smart as the stupidest dog. I have chickens that are better at self-driving than a Tesla, and not one of them has tried to kill me while crossing the street. Though they do get a bit pecky at dinner time if I’m too slow getting to the seed box.
But that’s not my main frustration with the hype around AI/ML.
My main frustration is that stepwise linear regression is extremely useful and would probably do everything that 95% of organisations actually need but because it lacks fashionable tech cool points people don’t use it. They are wasting huge amount of time and effort doing something overly complicated when a simple and easy way of doing the same thing exists.
In fact, we could go one better.
Stop Optimising Bullshit
The best optimisation you can make is to eliminate work that you don’t need to do. As management guru Peter Drucker said decades ago “There is nothing so useless as doing efficiently that which should not be done at all.” Yet that’s exactly what we are spending huge amounts of engineering time on.
Doing unnecessary things faster or cheaper is a waste of effort. The overheads of what machine learning algorithms need is massive overkill for what the majority of customers need. There are huge gains to be had from simple applications of linear regression and Little’s Law, mostly to identify things that we should simply stop doing. We could be making these simple, incremental improvements instead of wasting a lot of time and money on magic beans.
Too much of modern computing is using human blood and sweat to teach machines how to hit us in the face harder and with greater accuracy.
Because AI/ML has taken over from Big Data (remember that?) we have this obsession with putting surveillance devices everywhere to collect more data to feed into a machine that we don’t really need, and most of us don’t even want. Huge amounts of nonsense is getting fed into systems that are stupider than my chickens so they can make dodgy predictions that lazy managers follow unthinkingly. If we replaced them with an Excel spreadsheet and a stepwise linear regression, we could get better results, faster, and not waste a bunch of resources in the process.
But then a lot of mediocre white men would have to find something else to do, and we can’t have that, I suppose.
Pingback: Linear Regression Is Better Than Machine Learning - Gestalt IT