AI that creates text and images is a developing field that’s getting a lot of news coverage. Ian Evenden explains what’s going on

The question of artificial intelligence has been a hard one to get away from, with ChatGPT hitting the headlines as both an amazing scientific advance and the harbinger of the end of civilization. Which is it? Well, as with many tech stories in the papers, it’s a bit more complicated than a simple binary choice.

What has been released to the public are still a long way from the AIs we’re used to seeing in science fiction. They’re generally capable of doing one thing, whether that’s holding a conversation or creating images of the Pope in a fashionable white coat (or the image that opens this feature) that are photoreal enough to fool the casual observer. We’ve yet to see one that can do both, though ChatGPT is as capable at writing in programming languages such as Python as it is in English or Spanish.

This is also a developing field, where advances come quickly and suddenly. The difference between ChatGPT 4 has been claimed to be ten times as complex as ChatGPT 3.5, despite there being only three months between their releases.

So what are these AIs? How do they work? Are they going to destroy the world? And can I run one on my PC?

This scene from a videogame we generated with Stable Diffusion looks amazing.

© NVIDIA

Do you remember Clippy? The Office Assistant from Microsoft Office 97 through 2003 was actually called Clippit, but the misnomer stuck, and presented as an anthropomorphized paperclip (though other skins were available). It used algorithms to analyse what you were writing in Word, for example, and popped up with helpful advice, based on Bayesian probability, about writing a letter or a report for the 30 nanoseconds before you turned it off.

Nobody at the time called Clippy an AI, but at heart what we’re seeing in the AIs of today is a development of the same thing. They take a text prompt provided by the user, and produce output based on probability and the huge amount of training data they’ve been exposed to. This training regime is reflected in the name of ChatGPT—the GPT part standing for generative pre-trained transformer.