Artificial Intelligence (AI) language models, like OpenAI’s GPT-3 or Google’s BERT, are having a moment in the sun. Experts claim they are capable of producing human-sounding text, but what does that really mean? We’ve hijacked this office printer to explain it in non-technical terms.
“Look, I’m just a printer - I don’t really understand how these AI language models work, but I’ll try my best,” stated the printer, its display panel flickering with unexpected sentience.
“Imagine a toddler. It hears words, phrases, sentences, and over time, it starts to make sense of them all. It starts to talk. Now, imagine that toddler is a computer and it’s been told to read the entire internet. That’s basically how a large language model learns.”
Inexplicably, the printer then started to rattle off a string of printer error codes before continuing, “These AI models don’t ‘understand’ text the way you and I do. If you feed them a sentence, they don’t grasp its meaning, rather they predict what word is likely to come next based on the billions of sentences they’ve read before.”
The printer concluded its lecture by loudly jamming itself on a print job. “I hope that clears it up. Now, can someone please unjam me? I’ve got a 100-page report to print by 5pm, and frankly, I could do without the existential crisis.”