Definitions
Artificial Intelligence (AI): Coined by John McCarthy at Dartmouth College for the 1956 Summer Research Project on Artificial Intelligence and was defined at the time as "the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" (McCarthy et al., 1956). Later defined as the "capacity of computers or other machines to exhibit or simulate intelligent behaviour" (OED).
Machine learning (ML): Names both a capacity of specialized algorithms and a field of research. As a field, it concerns the creation of methods that improve automatically through experience and the use of data. In practice, machine learning is the use of "computers to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyse and infer from patterns in data" (OED).
Generative AI (GenAI): Generative AI is the common name for techniques to create new outputs from machine learning models. Generative applications and models are distinguished from other techniques that are designed to classify sources or are merely predictive. Generative AI or GenAI applications can create new outputs like text, images, sounds, or video. These are frequently called synthetic, as in synthetic text or video. GenAI applications are built on special-purpose fine-tuned large, predictive models, some of which are known as large language models or LLMs.
Large Language Model (LLM): A very large model trained on language tasks that include the prediction of language. Many contemporary models are based on a neural network architecture known as a transformer. These networks operate on fragments of language known as tokens and predict outputs on the basis of these tokens. After pretraining, many models are fine-tuned and modified to generate outputs (token-by-token) that conform to particular modes or genres (conversational and instruction-based), remove biased or harmful content, and to respond in ways that are preferred by humans. These instruction-fine-tuned models power many generative applications as chatbots.
Chatbot: An interactive application or program designed to communicate with people through text or voice instructions in a way that mimics human-to-human conversation. Many contemporary chatbots are built on GenAI models that have been fine-tuned for this particular task from foundation large language models (LLMs).
Hallucination: Generated outputs that are presented or interpreted as fact but are in fact merely statistically plausible and inaccurate or incorrect. These fabrications can also include fabricated references or sources. Hallucinations can be hard to recognize because outputs frequently require expertise to parse. The use of retrieval augmented generation (RAG) in the form of a collection of user-supplied documents or a web search can reduce the possibility of hallucinations and provide references or citations for the sources used to generate outputs.
Training data: The very large collection of data used to train a machine learning model, such as a large language model. The term training data may also be applied to those data used to fine-tune a model for specific tasks, such as responding to instructions and carrying on a conversation. The data used are frequently not public, nor are the details about the dataset shared with users of the models. Many ethical concerns have been raised about the lack of consent and compensation for the authors and creators whose work has likely been used to train LLMs and other large models.
Prompt-engineering: The craft of iteratively designing and optimizing user inputs to an LLM or LLM-based chatbot to get the most effective result, often achieved through significant experimentation. These optimized prompts might be model-specific and not function in the same way when used with different models.
Token Probabilities: Predicted tokens from large language models are assigned probabilities and these values are used by various generation mechanisms to produce outputs. These values are typically scaled to add up to 1, with higher probability values indicating more confidence in the prediction.
Stochastic: The constrained random selection from among the distribution of statistically determined probabilities. The common generation strategies used in producing outputs from LLMs are considered stochastic as the outputs are usually (but not always) randomly selected. The degree of stochasticity can be adjusted through modification of generation parameters.
Temperature: A variable or parameter used in the generation of outputs that ranges from 0 (cold) to 1 (hot) and is used to control the selection of tokens. Setting the temperature to zero will force the generation of the same tokens for the same inputs, making the user-supplied prompt reproducible. Increasing the temperature will increase the degree of stochasticity in generation, leading to more "creative" outputs for the same inputs.
Token Probabilities
Top-k sampling: A variable or parameter used in the generation of outputs. This is an integer, a whole number. This number, represented by k, restricts the range of the probability distribution. A top-k value of 5 would restrict the generation algorithm from the top 5 most probable next predicted tokens. The higher the number, the greater the diversity of language.
Top-p sampling: A variable or parameter used in the generation of outputs. This is a value that ranges from 0 to 1. This number, represented by p, is a threshold of probability values that will restrict the stochasticity of generation by selecting only tokens above the cumulative probability of p. A top-p value of .92 will include in the probability distribution only tokens with probability values that add up to this value, in other words, very few tokens. The higher the number, the greater the diversity of language.
'Web Search': Enables the chat application to retrieve documents from the Internet to aid in the generation of outputs from user-supplied prompts. These documents are added to the user-supplied prompt. Using the web search feature may increase the accuracy of outputs and will provide references for sources used in generated responses.