Machine Learning
Machine learning is the branch of computing in which a program improves at a task by adjusting itself against examples rather than by following rules a programmer wrote out. The result is a model: a set of learned parameters that maps an input to an output.
Three ways a model is trained
Supervised learning gives the model inputs paired with the right answers and adjusts it until its outputs match. Unsupervised learning gives it unlabeled data and lets it find structure, clusters or compressed representations. Reinforcement learning gives it a reward signal and lets it discover which actions earn more of it. Language models use a mixture. Predicting the next token is self-supervised: the text supplies its own answers, since the next word is already there. The tuning that makes a model follow instructions and decline some requests is done afterwards, partly with human feedback used as a reward.
Machine learning, artificial intelligence and language models
Artificial intelligence is the goal, machines doing things that needed a person. Machine learning is the method most of it now uses. A neural network is the kind of model most machine learning now trains. A language model is one such network, trained on text. The order matters for a business owner because it says what an assistant’s statement about you is: a prediction learned from examples, not a record looked up in a register. Change the examples it reads about you, on the pages it fetches, and the prediction changes with them.
A common question
If an assistant describes a business wrongly, can the description be corrected?
Not by correcting a register, because there is no register. What an assistant says is a prediction learned from examples, produced by a model whose parameters were fitted against text rather than looked up in a record. The lever a business has is the examples themselves: change what the pages an assistant fetches say about the business, and the prediction built from them changes too.
Related terms
Sources
This entry states mechanism and carries no figures. The site’s own measurements of what assistants say are described on the methodology page.