Why does no one know how AI works?

Why does no one know how AI works?

AI is in the news, and not just news written by AI. OpenAI’s ChatGPT and other, similar options are opening up artificial intelligence to use by individuals and not just corporations or tech startups. It’s definitely tempting to use a tool that can – for example – write your marketing content for you. But how does it do that? What is actually going on inside the ‘black box’? Can you trust the resulting content if you don’t know how it’s produced?

I’ve been working with computers for over 25 years, and I have a good idea of how they function. When people say they don’t know how computers work, or a website for that matter, I’m not surprised as most people aren’t as much of a geek as I am. Also, most people simply don’t need to know.

So when I read that “we” – meaning fellow geeks or tech journalists – say we don’t really know how AI works, I was a bit confused. Is it really that technical? or is it just sufficiently complex to put off even the geekiest of geeks?

I found an explanation (and insights) in Mo Gawdat’s book Scary Smart. It’s a must-read for anyone using AI. or feeling like it might take their job one day. (#spoiler: it will)

Mo explains why we don’t really know how AI works, and the answer is actually very simple.

So, in case my (usually less geeky) followers also don’t know how AI works, I’d like to share Mo’s concept, with my own spin on top.

Note: This is meant as a high-level simplification of how building AI works; the actual process is far more complex, but that’s why you read my blog.

Isn’t AI just a really BIG computer?

Let’s start with the difference between the computers we use and the AI that we are being told about every day. To illustrate, here’s a problem: how do you tell the difference between a number 3 and a number 8?

For us, we learn as children that the shapes of these numbers are different and the different shapes have different meanings.

For a computer, the difference is not shape but code. Developers write code that enables the computer to process numbers (which includes telling them apart!) The code is so efficient that a computer does this faster and more accurately than humans. This is why we all use computers.

A computer does not understand numbers the way we do, it is simply able to process numbers in a way that matches our instructions, and gives us the correct mathematical result. If we really want to know how a computer solves a calculation, we can look at the code and follow the steps.

AI is different.  The code in an AI was not written by a human. The code is created by the AI itself…

How does and AI make an AI?

A developer building an AI doesn’t write the code. They start with an idea of what they want to achieve, provide known input, and have an idea of the output they are expecting.

If we want to train an AI to tell the difference between a 3 and an 8, we could give it code that we already know works. That’s easy, but what if we wanted it to tell the difference the same way a human can, by “seeing” the numbers and recognising their different shapes? That would need some more advanced code that can analyse and then process patterns.

At this point, the developer could start coding, or they could use a software robot (a ‘bot’) to do it for them.

The developer builds two bots: a build and a tester.

The builder bot assembles the code, varying it slightly to produce different iterations. In this way, the builder bot can create millions of variations of the same code at the same time. Something humans would take months or even years to do.

The tester bot tests each iteration to see how well it performs against the desired task. It reports back to the builder bot what code worked and what code did not. Code that doesn’t work is discarded, and code that works is kept and then built upon further.

This process means that millions of code variations can be built and tested, often in seconds instead of months or years. If the code isn’t perfect the first time, no problem. Like the proverbial infinite number of monkeys with typewriters eventually writing Shakespeare, builder bots will eventually build code that works perfectly.

So, in our example, the bots build and test code until it can recognise a 3 and an 8 from the images it is given. This sounds like it shouldn’t work but it does, remarkably well.

We now have code that works but exactly what that code is, the developer doesn’t know. They could take a look and work through the code, and for our 3 and 8 example that wouldn’t take too long, but imagine the AI we have created has read every book and 800 billion web pages…. working through that could take lifetimes.

When we say we don’t know how AI works, it’s because we can’t know.

This sounds complicated

AI is so incredibly complicated no human can ever understand how it works in detail. The companies building AI systems celebrate and take the credit when they work, and pass the blame to the ‘black box’ when they don’t.

As Mo Gawdat says, “We are heading into an era of civilization where these machines could be completely in charge of our lives … aspects of our lives are already run by AI, social media apps for example are run by machines that no one, not even their creators, truly understand. We are handing our fate over to the complete unknown.”

This might sound worrying when you think about using AI – to create marketing content, for example – but like any tool, you just need to know how to use it and understand its limitations. After all, between generating the content and posting the content, is you.

It’s an exciting time, and a time of great change. I find it fascinating and enjoy reading all about it, from doomsday scenarios to fantastical ways that it will solve all of humanity’s problems. Personally, I sit somewhere in between these two extremes and I’ll be making more content about AI as it develops (fast!).

If you want to keep up to date with the content I create then sign up for the email below or follow me on LinkedIn and YouTube.