What qualifies as AI?

AI is becoming a popular buzzword these days, and you are likely to see it used in marketing everywhere pretty soon. Everything will be AI, just like everything was in the cloud a decade ago, or like everything is on the blockchain today.

Both in the case of the “cloud” craze as with the present “blockchain” craze many of the services offered had no reason to be either in the cloud or on the blockchain. Worse, some weren’t but still claimed so. But hey.

The “AI” buzz word is going to run into a similar problem. So perhaps it’s worth clarifying what qualifies as “AI”.

In many ways intelligence is in the eye of the beholder. You can say your calculator is intelligent. It can calculate 23523 times 37275 in a split second and it would take you 30 plus seconds to do it long hand. If you put two kids next to each other and one did a multiplication task in 10 seconds while the other took 2 minutes you’d probably call the first kid smart.

You’d also call someone smart if they could name all 195 countries in the world. Doctors learn names of all the bones in the human body in medical school, and I know I could never do that. Doctors are smart. Of course, you could also look it all up in Google these days. It’s just a database really.

Some of these databases have a bit of intelligence to them. The other day I tested out a phone app called Ada, which does a pretty good job at diagnosing your symptoms through a series of multiple choice questions. It’s what we used to call an expert system. It’s also just a database: apparently there are around 10,000 diseases and only about 200 symptoms. The 10,000 diseases are a combination of the 200 symptoms. But on top of the database is a little engine that goes down decision trees of symptoms, trying to guess what other related symptoms you might have trying to confirm an assumption. It sure beats you having to think of all the symptoms and Googling them and then trying to match them. In the end Ada gives you several possibilities for your diagnosis with probabilities that they match yours. Doctors are concerned that this app might make mistakes— of course, human doctors misdiagnose up to 20% of the time.

Finally, there is chess and Go, in both of which computers excel and now routinely beat human players.

So we are at the so called Singularity yet? No not yet. All of the above we’ve had for a while, it’s just gotten bigger and faster as computers got bigger (in terms of disk and RAM, not physical size) and faster. And for some reason people just aren’t very impressed and aren’t ready to call these applications particularly intelligent, at least not in the way humans are.

People really got warmed up to the concept of AI when we got speech and image/face and voice recognition along with autonomous robots, such as self-driving cars. It’s pretty interesting actually that humans consider intelligent only something that can be social: recognize the world around it, be able to move around and listen to spoken words. Math doesn’t count.

Be as it may, the question is why it was easier to create machines to do math than to recognize faces. It’s actually a misnomer because recognizing faces also involves solving mathematical equations, but the math for recognizing faces is actually a lot, lot harder than the math to do just about anything else taught in mathematics curriculum at universities. Modern mathematics does not provide the shorthand tools to do the sort of things required and the brute force method of trying out all combinations is too slow even with the latest, greatest microprocessors. So how do computers solve this math? We’ve actually built a special kind of computer on top of a regular computer. This special kind of computer basically imitates the way the human brain works. It’s a general purpose learning machine called a neural network. Rather then tell it how to do something, we teach it how to learn and then we release it on training data — almost the way a childlearns. Amazingly, by the way, it’s not a straight forward process, and there are many pitfalls. “Teaching” a neural network is an art in preparing the training set and setting a bunch of parameters (such as rate of learning). Prepare a bad training set, set the parameters wrong, and the neural network gets “stuck”. It doesn’t learn very well. Or, run the training set too many times and it over-learns, essentially memorizing the data, but is then not capable of generalizing to new data.

When such a machine learns something we do not know how to convert this knowledge into any other form, how to summarize it. We can only run the machine on input data and it will spit out the result (for example the name of the person from a picture, or text from voice). The machine itself has no insight why it came up with a certain answer.

This is the true artificial intelligence.

It may be that the true sign of intelligence is lack of shorthand. If you can break it down into rules, it is no longer intelligent, it’s just code.