The most recent successes in AI have come from large language models (LLMs) - AIs that do exactly that - model language. Their ability to tackle a wide range of problems shows how powerful language itself is. Everyone intuitively knows language is powerful, but we now have an effective model for it. And although all models are wrong, some are useful. LLMs are certainly useful.
It's surprisingly simple to build an LLM. To start, they took a simple idea from linguistics, "a word is defined by the company it keeps," and operationalized it. They take a lot of text examples, hide one of the words, and then have the AI try to "fill in the blank" with the correct word. By doing this over and over again, the AI is able to learn effective representations of words.
I'm struck by how similar this is to language learning. How many times in grade school did you complete fill-in-the-blank exercises? Or how apps like Duolingo are essentially one big fill-in-the-blank game. And by filling in the blanks, somehow you are able to intuit what the words mean.
Next, let's discuss the second phase of designing an LLM. After learning a vocabulary of words, the AI must learn to use them. For this part, they use a different learning technique. In a nutshell, they teach the AI to choose words so that the formed sentences are viewed favorably by humans. When it forms a good sentence, the AI receives a reward; when it forms a bad one, a penalty.
