Before I started my tech startup Artificial Intelligence (AI) really wasn’t on my radar. I was somewhat aware of autonomous driving vehicles, but apart from that I didn’t see how AI was going to have an impact in my life.
Even now when I meet with other tech entrepreneurs, the discussion surrounding AI often causes me to go quiet. I think that part of the reason I go quiet is that people talk about AI as if it has already happened – and if you’re not involved in the scene then you’re missing out.
Even the almighty Elon Musk often talks about how AI is one of the greatest threats to mankind.
So, what reason do I have to think AI isn’t really a big deal? After some further reading over the last few months I think there are a few main reasons why I am not convinced of the AI apocalypse just yet.
What actually is intelligence? I think this is a good starting point in this discussion.
Intelligence as we currently know it has three components. A stimulus which results in >>> a process >>> which results in an outcome.
As a doctor, I have a pretty good overview of how this type of response architecture works in humans. If for example we see a cat, the image of the cat is focused by our eyes (the lens, cornea, small muscles of the eye etc) and an image is projected on to our retina. The optic nerve then transmits the image to our brain. Having studied some basic neuro-science and neuro-anatomy in school, I can reassure everyone that we have no idea what happens with the said image at this point. Sure, we know some of the pathways that the neurons in our brain use to transmit the image around. If for example you have a lesion at the optic chiasm then part of the image we see of the aforementioned cat will be missing. But, we have no idea how our brain interprets the image and results in the outcome i.e. what we do next.
In other words we have no idea how the image of the cat results in us dismissing the cat or going to pet it, or shouting at that damn cat to get off our car!
There is another factor in this architecture. The “outcome” is different for all of us. Humans have free will and agency. We are not simple creatures with certain inputs and outputs. Someone may see the cat and get terrified as they suffer from ailurophobia, someone may feel sorrow as they remember their deceased pet cat from childhood.
Human intelligence is very nuanced. In actual fact this simple observation was only recently acknowledged by the scientific community in the last few decades. Previously scientists had the view that intelligence was simply: Input >>> Output. Meaning that a certain input would result in certain predictable output. This was the “Behaviourist” point of view which has been superseded by the more nuanced view of intelligence discussed above.
The way we learn language is another example of the nuances of human intelligence. It’s quite interesting that infants and toddlers pick up the prevalent language around them so easily. If you think about it, when you are born there’s a lot of noise. It must feel like a complete sensorineural attack for babies. In this environment how is it that they are able to pick out words and start developing speech? It has also been shown that infants are able to pick up languages and start speaking fluently despite not hearing all the words in any given sentence. So they are able to pick out words and work out the syntax with ease. On top of this – if they have started to develop a language, they can happily ignore foreign languages as they can tell that this is not their language.
What this example of learning language implies is that humans have an inherent ability to learn languages. There’s something within our brains that wire us to pick up languages and communicate with one another. This was a ground breaking insight by the linguist Noam Chomsky. He scientifically showed that the principles underlying the structure of language are biologically determined in the human mind and hence genetically transmitted.
Again, how we are genetically determined and how the human mind works is largely still a mystery.
Coming from a bilingual background myself, I do find it interesting how I have managed to learn two languages while growing up and have never mixed the two up or gotten them confused with one another…
So now that we have a basic understanding of what intelligence actually is, what is the state of artificial intelligence as we know it?
AI as we know it right now is basically statistical analysis. It sticks to the (now-defunct) behaviorist view point that input >>> output.
Give a computer a bunch of inputs, feed it a ton of data. The computer can now process all that data due to increasing RAM and CPU power. Then the computer will give you an outcome.
As a result people like IBM can analyse the chess moves of every chess game ever played, feed it into a computer and the computer will make moves which will lead to an outcome that will mean it is likely to win.
Or you implant a computer into a car which can analyse its inputs (images of the environment, the behaviour of surrounding cars etc) and the car will drive and manoeuvre so that you don’t crash.
In other words: Data in >>> Statistical analysis >>> Outcome.
This is not intelligence as we know it by any stretch of the imagination. This is statistical analysis and has been touted as a revolution since before the 1960s, but has yet to make much of a dent in the world.
Note that in the cases of chess playing and driving, that it is the humans who have already done all the interesting work. It is the chess masters of years gone by that the computer then goes on to analyse. Without the human input there is no useful output. So, the statistics which are analysed to lead to a useful outcome are always created by humans.
This technology may be useful in certain fields. It does seem that work which is mundane, doesn’t require much human input, creativity, thought etc can be automated. However, let us not confuse “automation” with “artificial intelligence”. I think automation will be massively disruptive to the world of work, but not the type of creative work that matters in the world. Not the type of work where human interaction takes place, where empathy is required, where original thought takes place.
Statistical Analysis & Big Data
There are a ton of startups and established companies who have been going on about big data for years now.
This is the stuff that people in Silicon Valley are always talking about; “Imagine if we could create AI and allow it to analyse all the data on the Internet”, or “Imagine if we could get our hands on patient medical records and allow AI to analyse all that data”.
For some reason people think that statistical analysis of large data sets will reveal new compelling information and automate and improve how, for example, medicine is practiced. It is thought by pseudo-scientists that medicine is not scientific enough. “Maybe if we analyse all the data, medical records, blood work, etc then we can detect diseases before they have even occurred / just about to occur!”.
Well the problem with statistical analysis to reach a conclusion only works when you are looking for one specific outcome. For example, the NSA do this the right way. They analyse data to figure out one thing: “Should we be suspicious of this person or not?”. When it becomes more complex than a yes or no answer we run into trouble analysing data-sets.
Let’s look at medicine again. If we look at a patients whole medical record and run it through a ton of statistical analysis and find correlations, you will draw more and more false conclusions. It will look like an exponential curve:
Let’s say that data points run across the X-axis and the correlations found run on the Y-axis. It’s clear that the more data you feed it the more false meaningless correlations you will reach.
In fact you’ll get very random correlations which are meaningless such as the consumption of chickens being correlated with the amount of US Crude Oil being imported:
In fact this is only the beginning of real science and knowledge. Medicine is an interesting topic as it treads the line between science and real world skeptical empiricism. What I mean by this is that if you get something wrong in the world of medicine then it can cause real patient harm.
Medicine over history has made progression with continuous experimentation and then observing the results of a given treatment or intervention. It doesn’t work the other way round. You can’t come up with a statistical correlation that says something along the lines of; “If you eat eggs, then you will get diabetes” and expect it to be at all meaningful. Is it because people who eat eggs are also more likely to smoke and not exercise? Is it the actual cause of diabetes? Does it affect all populations? And so on.
In fact statistical analysis such as this is likely to cause more harm than good due to increased interventionism being carried out which themselves will carry a larger list of adverse effects.
So in effect only a very foolish person would take any kind of correlation seriously and change their clinical decision-making due to random correlations, over the already well-tested evidence based empiricist medicine which already exists.
In effect, I don’t think that AI right now is anywhere close to what people have been claiming it is capable of just yet.