Admit it, you don’t understand A.I.
I’m sure a larger percentage of people in my circles actually do understand, but unless you personally have built something called A.I. yourself, you’re just guessing. #BAYESIAN
You need to grok that most people don’t understand the difference between grokking and understanding.
And of the people who have built some A.I.-related code or tool, you’re just hoping.
Those of you following my rants… er, posts… know there are a few themes I keep bringing up. On one hand - a cold metallic and plastic robot hand - A.I. will definitely be taking jobs. A lot of jobs. No really, almost all jobs for the common people, and soon. I’m sticking with this 100%. It will hit like a tsunami too. We need a plan.
On the other servo whirling x100 strength clamp-claw of a hand, most A.I. - yes, most - are just parlour tricks, although some are snake oil, and just a few are outright scams.
In Carl Sagan’s The Demon-Haunted World: Science as a Candle in the Dark, he brilliantly presented a Baloney Detection Kit. This was a very general method and best practices list to commit the act of SCIENCE.
Remember, science is a verb, and should always be thought of like a verb, as in ‘doing.’ It’s not a thing, like a pair of scissors, it’s an action, like cutting, together with rules, such as ‘don’t run with them.’ And, while I’m on the topic, did you read The Demon-Haunted World? If not, do so.
Ok, back to A.I.
We need a baloney detection kit for A.I. But what would this look like?
I’m reminded of one of my favourite business models, ISO, which stands for the International Organization for Standardization, which of course should be IOS, so we’re off to a rough start even agreeing on what word order we should standardize on. Yeah, I can’t make this crap up. To their credit, they backronymed it (or is it hackronymed) and claim it is based on the Greek word ‘isos’, which means ‘equal.’
Why are they my favourite business model?
Well, here’s how it works, you pay them to come over to your business and tell them what you do for a living. If you’re not sure what you do for living, which they will tell you don’t, you can pay them to help you realise this. They also charge you more money to tell other people you hired them. They will charge yet again even more money if you deviate from what you told them. And they charge you to keep them around charging you. #BRILLIANT, #MAFIA , #CATCH-22 #DUEDILIGENCE
Before we go too far here, do keep in mind, most of what is being called A.I. today is actually just giant data sets mined for patterns, called Machine Learning (M.L.). This can produce some fun and intriguing results, but unto itself, this is not really A.I. It is a tool. One needs to apply the tool then. Not all A.I. uses or needs M.L. (this is a heavy statement a lot of data scientists and people claiming they are building A.I. don’t seem clear on).
A simple example, you want to identify if there is a hot dog in a photo. First, you need to find as many photos of hot dogs as you can, and right here is the most important part; you must confirm first these are actually photos of hot dogs. The ‘learning’ part of machine learning is about humans teaching the machine (don’t go anthropomorphising the machine here just yet). This takes a lot of work. You can often improve the results if you also have a lot of photos of things that are similar to hot dogs, but are not hot dogs, and again, humans need to confirm each of these as well. Let’s just call all of that ‘vetted data’ (this is where the real value is).
If you are lucky, you might find a stash of pre-vetted data. This is much of what data scientists seek out. Or more accurately, the VCs and investors hoping to leverage data scientists to seek out. The money people don’t care or understand any of it, they just want an advantage.
If you feed the machine enough hot dogs, and enough not hot dogs, the machine will do a pretty good job of identifying a new photo it has never seen before as having a hot dog, or a not hot dog.
Did I mention how many human-vetted photos of hot dogs (and not hot dogs) you need to get useful results in the future? There is no exact answer here, but it is reasonable to assume a good starting number of 10,000 unique hot dog photos, but even 100K or 1M might be needed. Again, this is a lot of work for a hot dog.
But, hey, once you have this, now you can use this to identify hot dogs pretty well, probably better than most humans, and definitely faster, and it will do it 24/7. If you build a machine to do this to replace a human that was doing this, then you can call that A.I.
It should be noted, I have a pretty low bar to pass under to be called A.I. For me, a simple spring is technically A.I. in my book. Read this in my scratch pad, to understand why.
Now, if you need to identify a pickle, you’ll need a source of thousands or hundreds of thousands of vetted photos of pickles. But, one thing that happens as you do this, is if you did indeed identify the hot dogs first, and you trust the results, you might be able to use this to tag and remove the hot dogs from new photos that include pickles, which makes identifying the pickles a little easier.
I use a simpler analogy often, which is if you want a system that recognises brown M&Ms (you know, to be Van Halen Compliant), then we also as a by-product will have a system that can identify the green ones, red ones, in fact, all of them. You might then be able to I.D. peanut vs. other types of M&Ms.
Where things get a little scary is when you ask a Machine Learning system to do the opposite, and present you with what it thinks is a hot dog. Ever see what A.I. does to hands?
If I had to sum up the poignant point it would be:
Real A.I. applications should DEMONSTRATE tangible and consistent improvements over traditional approaches or human performance in specific tasks.
Rather what we see is investments (of money, mind-share, belief) in things that sound good on paper, but can’t be demonstrated. #StringTheory
The parlour trick here is showing the marks, er, uh, investors how the machine can seem to easily identify a hot dog, but then claim it knows how to make a hot dog.
It’s a disgusting job, but to do one’s due diligence, you need to see for yourself how the hot dogs are made.
Give a thumbs up if this was insightful.
