How AIs (Like ChatGPT) Learn

I first wrote about this video a few years ago, when artificial intelligence first captured the media's attention. Back then, we were impressed that algorithms could help you pick out your Christmas gifts on Amazon, suggest new music for you on Spotify, and do their best to capture your attention on websites. 

AI is seemingly everywhere now. And what is surprising to me isn't its prevalence or impact ... it is the speed and breadth of its adoption. Last night, at an early holiday dinner, we talked extensively about how easy to adopt and how accessible tools like ChatGPT are (and almost none of us were Nerds). 

With all that being said, I think it's wise to have a basic understanding of the things that most impact our lives. Even if you're not a big fan of AI, understanding its growing powers will benefit you. 

The video below is a bit simple in its explanations, but it describes some fundamental concepts worth understanding.

 

CGP Grey via Youtube

The video is engaging and easy to understand. It focuses on genetic algorithms (which is one type of machine learning) and ignores some of the other more complicated techniques and approaches. For example

In my experience, it is more useful for an executive to understand what tools like this can do rather than how they work. Likewise, it is better to understand when to use tools like this rather than knowing precisely how to use them. But, a cursory understanding like this video still adds value. 

As machine learning gets more complicated and evolves, it gets harder for humans to assess their output or process accurately. Here is something to consider:

How do you know that the answer it gives is the answer you seek?

Just because an algorithm responds quickly and confidently doesn't mean it's right.

While bots can deliver impressive results, their decision-making processes can be opaque. This presents both a risk and an opportunity. Artificial intelligence seems cool, but artificial stupidity is scary ... and making mistakes at light speed rarely results in good outcomes.

It's human nature to feel safer when we understand something. It's human nature to envision machines making human-like decisions, just faster. But we are quickly going beyond that ... way beyond that!

In the past, algorithms were static while data changed. But now we're in a different world - one where the algorithms themselves evolve and dramatically adapt to handle different types of data. While this might sound like a subtle distinction, it represents a fundamental shift in how AI systems learn and operate.

One of the challenges of understanding exponential technologies is that their progress isn't linear. This makes it difficult for humans to accurately gauge how rapidly these tools will advance in capability. As algorithms grow more complex, they will increasingly operate in ways that may not be fully understood by their developers (and even less so by their users). As a result, we'll likely find ourselves using AI to solve problems and accomplish tasks that aren't even on our radar today. 

It might sound strange, but it doesn't matter why a bot makes a decision or what inputs it uses to make the decision. What matters is whether it accomplishes its goal and how its performance and level of decision-making rank in relation to its prior performance and other options (and, perhaps, whether the bot is biased).

Part of what makes artificial intelligence exciting is that it can do a lot of things well that humans are really bad at. And, even when you're using an AI in your domain expertise, it can be a great first step to save you time and effort. 

It's a brave new world, and not only is Big Brother watching, but algorithms are, too.

Live long and prosper!