I looked it up a while back, and some site I found said the definition of AI changes often, which frankly made a lot of sense. For AI today, the “AI” part of it is LLMs — Large Language Models. And, like, if that is “intelligence” then I’m the king of France.
What is an LLM?
So let’s start with how an LLM works. Unlike _some_ charlatans 🙄, I actually think you can get better results if you know how the thing works. In this case, that thing is an LLM. An LLM essentially finds needles in haystacks. That is to say an LLM predicts the next “token” based on the training data and the preceding “token.”
Let’s make up a simple example. Let’s say we have “training data” that is like 100 sentences. Every group of words forms a sentence. For illustrative purposes, these sentences are all simple sentences that generally have a subject, a verb, and a direct object. Are you with me so far? Ok so if a new sentence input by the user follows this same pattern, it is considered valid because it uses the same pattern as the training data. If it does not follow this same pattern, it is considered invalid.
With the big LLM chat bots, they subdivide the sentences and even the words to compare those to the training data. For instance, I have seen that with some of the chat bots, you can significantly misspell words, and the chat bot is able to respond sensibly. That is because it is able to determine what you input, as if you hadn’t misspelled the words simply by determining which “tokens” (words) were likely correct given the circumstances. If I were to compare it to people, I’d say it’s like context clue maxxing.
The reason LLM chat bots use copious amounts of energy is because most people are bad spellers — haha jk. Mostly. Graphics processors (GPUs) only got efficient enough fairly recently to be able to provide useful responses quickly. Whereas a human brain uses relatively little electricity (approximately 12-20 watts of electricity — roughly as much as a dim lightbulb), a single GPU typically uses far more power, with the top-of-the-line GPUs requiring 700 watts.