Over the last few years, Artificial Intelligence (AI) systems in the form of chatbots have become widespread. Over the same time, they have been improving.
But what are they? And how do they work?
The answers to those questions can go a long way to explaining how much can you trust this remarkable and evolving technology. 3Blue1Brown's Grant Sanderson put together a fantastic video introduction into what these AI systems are and how they work.
How AI chatbots work is remarkable, but those workings come with some notable limitations. Although AI chatbots rely on some pretty sophisticated math to do what they do, they're not terribly good at actually doing math. At least not quite yet. This is one of the areas in which the capabilities of AI systems are being expanded and progress is being made.
Likewise, AI systems that use Large Language Models (LLM) designed to sequentially predict the next word in their responses to prompts aren't very good at inferring how those words might realistically relate in a scenario beyond how it encountered them in its training. You can see this by giving an AI system a simple mystery to solve. Like one from G.T. Karber's Murdle Jr.
Short of specifically training an LLM with the mysteries and solutions that it will be given to solve, it's not likely a generally-trained AI chatbot will be capable of reliably performing that task any better than a system that randomly picks out a suspect, murder weapon, and room from a game of Clue. While they might be able to assemble words together into something that looks and sounds like a solution to these kinds of puzzles, AI systems cannot infer how the words they really connect in scenarios they haven't been spoon fed.
On the other hand, if you know how the LLMs behind AI chatbots work, you can probably use them more effectively to perform the kinds of tasks they can be good at, such as summarizing information contained within the sources on which it has trained. Like any tool, the best results come from knowing how to use them.