Artificial intelligence software is nothing without data.
The tools develop intelligence through machine learning, a process that allows computers to “learn” on their own, without requiring a programmer to tell them each step. Feed a computer massive amounts of data, and it eventually can recognize patterns and predict outcomes.
Key to this process are neural networks, mathematical systems that act like a computerized brain, helping the technology find connections in data. They’re modeled after the human brain, with layers of artificial “neurons” that communicate information to one another. Even experts don’t necessarily understand all the intricacies of how neural networks work.
Large language models, or LLMs, are a type of neural network that learns to write and converse with users; they back all of the chatbots that have swooped onto the scene in recent months. They learn to “speak” by hoovering up massive amounts of text, often websites scraped from the internet, and finding statistical relationships between words. When these systems pattern-match, it can lead to feats of creativity: A chatbot can create song lyrics closely matching Jay-Z’s style because it’s absorbed the patterns of his entire discography. But LLMs don’t have awareness of the meanings behind words.
Parameters, which are numerical points across a large language model’s training data, dictate how proficient it is at its tasks, such as predicting the next word in a sentence.
In the future, some researchers say, the technology will approach artificial general intelligence, or AGI, a point at which it matches or exceeds the intelligence of humans. The idea is core to the mission of some artificial intelligence labs, like OpenAI, which lists achieving AGI as its goal in its founding documents. Other experts contest that AI is anywhere close to achieving that kind of sophistication, with some critics contending that it’s a marketing term.
