| Adversarial Machine Learning |
|
The study of attacks on machine learning algorithms, and of the defences against such attacks. The context is usually jailbreaking. |
| AGI |
Artificial General Intelligence |
A hypothetical form of AI that would be much smarter than humans |
| AI |
Artificial Intelligence |
Field of computer research attempting to develop machines that think like humans |
| AI Ethics |
|
The study and practice of ensuring that artificial intelligence systems are developed and used in ways that align with ethical values and standards. |
| Algorithm |
|
A set of processes, instructions or rules used especially by computers to perform calculations, solve problems and process data |
| Alignment |
|
Field of AI research concerned with ensuring AI models stick to desired outcomes and stay as helpful, safe and reliable as possible - making sure that AI operates in accordance with human values, goals and preferences |
| Algorithmic Bias |
|
When an AI algorithm consistently produces unfair or biased results, often due to biases present in the training data. |
| ANN |
Artificial Neural Network |
A computational model inspired by the human brain, consisting of interconnected nodes (neurons) used for tasks like image recognition and language processing. |
| Autocomplete |
|
Also called predictive text, a feature of some (e.g., messaging or search) apps whereby they will predict the rest of a word, or the next word a user will type |
| Bias |
|
Unfair or inaccurate predictions or decisions made by AI due to imbalances or unfairness in the training data. |
| Chatbot |
|
A program that communicates with humans through text that simulates human language |
| ChatGPT |
|
A generative AI chatbot developed by the company OpenAI, that uses large language model pretraining |
| Deep Learning |
|
A subset of machine learning that involves artificial neural networks with many layers, enabling the model to learn complex patterns. |
| DeepFake |
|
Artificially generated content, often images or videos, that convincingly mimic real human faces or voices. |
| Diffusion |
|
A machine learning method involving adding random "noise" to data - usually images, and training the AI model to recover the original image or create an altered version of the original. |
| Emergence |
|
When an AI model shows new and unintended capabilities. Also called emergent behaviour |
| Fine-Tuning |
|
The phase following pretraining where a model is adapted to perform a specific task using a smaller, task-specific dataset |
| GAN |
Generative Adversarial Network |
An artificial neural network with two parts: a generator which generates content and a discriminator that tries to find flaws with the generated content. The generator uses the discriminator to improve the quality of its output and the discriminator uses the generator to improve its ability to spot errors. |
| Generative AI |
|
A type of artificial intelligence technology that can produce ("generate") various types of content, including text, imagery, audio and synthetic data from human prompting |
| GPT |
Generative PreTraining |
A type of large language model that pretrains with textual data to learn statistical patterns and structures of natural language before fine-tuning it on specific tasks (such as creating content) |
| Guardrails |
|
Policies and regulations built into AI models to prevent those models causing harm. For example, that the model doesn't create disturbing or unpleasant content and the model handles data responsibly |
| Hallucination |
|
A confidently asserted and plausible response from a generative AI model that is completely false |
| Jailbreaking |
|
Manipulating generative AI models through prompt engineering to bypass or defeat their programming, Thus making it possible, for example, to reveal information about how an AI model was instructed or induce it to respond in an anomalous or harmful way, counter to how it was programmed. |
| Labelling |
|
A stage in supervised machine learning that identifies content in raw data (such as images, video, audio, or text) and annotates it with descriptive labels to help provide the proper context for use of the data by an LLM |
| LLM |
Large Language Model/Language Learning Model |
A type of machine learning algorithm that uses a vast amount of textual data to process and learn the rules of human language or text |
| ML |
Machine Learning |
A branch of computer science mainly concerned with developing machines that can perform complex tasks without human intervention |
| Machine Learning Model |
|
A computational representation of a machine learning algorithm that can make predictions or decisions based on data. |
| Model Collapse |
|
A hypothetical outcome of later AI models being trained on predominately AI-generated output, leading to a deterioration in overall quality through the generation of repetitive, homogenous output. This may become a big problem as the web is filled with sloppy, low-quality AI-generated content that is used to train other AI models |
| Multimodal AI |
|
An AI model that can process and output different kinds of data such as text, image, video or voice |
| NLP |
Natural Language Processing |
A branch of AI research concerned with giving computers the ability to process and manipulate human language. ChatGPT and other generative AI models user NLP |
| PreTraining |
|
Feeding vast amounts of language data including books, web pages and other textual sources to identify patterns and establish connections between words |
| Prompt |
|
A user input to an AI model to receive a desired output or result, for example, "plan a birthday party for a three-year-old girl who likes elephants". The better and more detailed the prompt, the better the output. |
| Prompt Engineering |
|
The process of structuring an instruction that can be interpreted, understood and acted upon by a generative AI model |
| Reinforcement Learning |
|
A type of machine learning in which an AI model learns through trial and error using feedback from its actions |
| Stochastic Parrot |
|
(Disparaging) term to describe the capabilities of AI models; regardless of how plausible the output sounds, AI models have no more understanding of human language than parrots that mimic human speech. |
| Supervised Learning |
|
A type of machine learning where the model learns from labelled data, making predictions based on input-output pairs. |
| Training Data |
|
The dataset used to teach a machine learning model, consisting of input data and their corresponding correct answers or labels. |
| Transfer Learning |
|
A machine learning technique where a pretrained model's knowledge is used as a starting point for new, specialised tasks. |
| Transformer |
|
A type of artificial neural network architecture used in training generative Ai |
| Transparency |
|
The practice of making the processes and decisions of artificial intelligence systems open and understandable to users and other stakeholders. If they can see how processes work and decisions are made, users might have confidence in using AI |
| Turing Test |
|
A test of a computer's ability to display intelligent behaviour indistinguishable from that of a human. As originally suggested, Alice communicates with "Bob" and "Carol", but cannot see them. One of Bob and Carol is a computer, but Alice doesn't know which one. If, from Bob and Carol's responses, Alice cannot tell which one is a computer, then the Turing Test is passed. |
| Unsupervised Learning |
|
A machine learning approach where the model learns patterns and structures in data without explicit labels or guidance. |
