In just the past year, advancements in artificial intelligence have introduced transformative new ways to automate complicated tasks. Your company won't be able to take full advantage of this new technology, however, if you don't understand how it works.

To help you make sense of all things A.I., we're building a living document to explain all the hard-to-understand terms around A.I. We'll continue to update this guide with everything you need to know about A.I. These definitions were written with the help of Tiago Cardoso, principal product manager at digital transformation firm Hyland.

Generative A.I :

Artificial intelligence programs capable of creating and generating "original" content. Recent advancements in A.I. have led to breakthroughs for image-generation models like Dall-E and large language models like ChatGPT, but the tech is also being used to create original music, video, and code.

Think of it this way: Generative A.I. is an extremely new technology, and the rules around its use are still being debated. As such, be careful about how you implement it in your business. The US Copyright Review Board recently determined that AI-generated art cannot be copyrighted, for example.

Training Data:

Sets of data that are processed by machine learning algorithms to improve their functionality.

Think of it this way: Datasets, which are often extremely large, are fed into machine learning algorithms to teach them how to respond to inputs. Once the data has been processed, it gets converted into a model. There are two main types of training for machine learning algorithms: supervised and unsupervised.

Supervised and unsupervised learning 

1. Supervised learning : Training in which each piece of data is paired with a label, which helps the machine learning algorithm understand the meaning of the data. An algorithm being trained to make a diagnosis based on X-ray scans, for example, would be trained on images labelled with the correct diagnosis. Think of it this way: An object detection model designed to identify fruits would be trained with many different pictures of those fruits, all paired with the correct labels. Through training, the algorithm would learn to identify the unique characteristics that define each fruit.

1. Unsupervised Learning: In unsupervised learning, the training data doesn't come paired with any descriptive labels. Rather, machine learning algorithms process large amounts of data, which are then grouped into "clusters" based on their similarities or differences. This style of learning is what allows ChatGPT to do all kinds of tasks, like holding conversations, writing stories, and answering questions. It wasn't trained to do any one thing specifically; it's been loaded up with a massive collection of text. Think of it this way: Alpha-Go, the A.I model that beat a world champion in the classic game Go, wasn't trained on any labelled information about gaming strategies; it just played the game enough times to master every possible winning pattern.

Neural networks/Deep learning:

One of the oldest, and for the last decade most dominant, designs for A.I. programs, loosely modelled on the organisation of neurons in the brain. A neural network consists of several layers of interconnected nodes, which act as the network's "neurons." Each node processes input data, performs calculations, and outputs the data to be reprocessed by the next layer of nodes. Deep learning is a class of especially large neural networks with hundreds of layers, which allows for even more connections.

Think of it this way: Most generative A.I. models are built with deep learning, with the largest neutral networks being large language models like ChatGPT, which have billions of "neurons."

Parameters:

In a neural network, parameters are the settings and weights that control how each "neuron" or node processes and transforms input data. You can imagine parameters as knobs on an old radio. Just like you'd adjust the knobs to improve the frequency, volume, treble, and bass of the radio, parameters are automatically fine-tuned during training to create an optimal output.

Think of it this way: Imagine an A.I. model built to analyse images of licence plates taken from a red light camera. Each "neuron"/node has a parameter responsible for turning the image's pixels into a sequence of text and numbers that the model can understand.

Natural language processing (NLP):

A specific type of A.I. designed to understand and interpret everyday language. NLP models are trained to break down a piece of language, either written or spoken, into machine-readable data.

Think of it this way: NLP models can be used to analyse documents, turn speech into text, translate between languages, and create advanced chatbots.

Transformer:

A highly advanced type of A.I. architecture that has hastened the revolution in generative A.I., and in particular the field of natural language processing, since being introduced by Google in 2017. Transformers use a process called "tokenization" to convert a string of symbols like this sentence into data, and then analyse that data to identify patterns.

Think of it this way: Nearly all modern natural language processing models, like OpenAI's GPT (Generative Pre-trained Transformer) family of models, are built using transformers.

Tokens:

Grammar elements that have been converted into data by a transformer. When you submit a query to ChatGPT, for example

Hallucinations:

Instances where an A.I., typically a large language model, generates content that appears plausible but is, in fact, untrue. The A.I. isn't intentionally deceptive, as it lacks awareness of the accuracy of its statements, leading to the term "hallucinations."

Consider this scenario: New York attorney Steven Schwartz employed ChatGPT to identify legal cases for citation in a legal briefing. Schwartz only realised that the cases generated by ChatGPT were hallucinated when he was required to provide copies of these cases.

API (Application Program Interface):

A software component facilitating the integration of someone else's program into one's own application without requiring an understanding of the underlying code. A.I. models are deployed and released through an API, enabling companies to monetize their technology by granting external parties access to the tech's services and capabilities.

Think of it this way: OpenAI has made APIs available for nearly all its A.I. models, and users incur charges based on the number of tokens utilised to process and output a query.