Introduction
Welcome to the Goldsmiths AI LibGuide! Here you will find a bunch of resources to guide you through the minefield that is generative artificial intelligence. There are video resources; articles on responsible usage; links to College policies; tips and tricks to effectively engineer your prompts - and so much more!
Please note that we will try our best to keep this guide as up to date as possible, however given the rapid pace of technological development in the AI area we may fall behind. We'll be reviewing the content here termly - but if you see something that is critically out of date, or if there is something we have forgotten to include - just let us know via the feedback form (on the Feedback tab at the top of the page!).
What is generative AI?
Glossary
A sequence of rules or calculations used by a computer to complete a task.
The process of trying to ensure that AIs decision making is in line with human values and goals.
An area of computer science and AI development exploring the ability of computers to view and understand images and videos as humans do. Applications include object recognition, facial recognition, navigation (think self-driving cars), augmented reality and text extraction.
Subset of machine learning that uses neural networks to simulate the decision-making power of the human brain by identifying patterns in data and generating a specific output. AI deep learning powers many generative AI tools we use.
AI generated / manipulated audio and video, designed to look, sound or act like someone else - often politicians, celebrities or powerful people.
A method of AI content generation that uses diffusion to generate content. The model is trained by adding random noise to the training data, then reversing this process to recover the data. Once trained, the model can then generate data by starting with random noise and applying the learned noise removal process. This method is behind popular AI image generators like Dall-e and Adobe Firefly.
Process and frameworks used to ensure that AI outputs can be understood and trusted by humans.
The phenomenon of AI models generating incorrect or misleading responses - often (but not always) caused by insufficient or biased training data, or incorrect assumptions or decisions made by the AI model.
A deep learning architecture / framework used to generate new data based on a dataset. There are 2 components: a generator and a discriminator which compete against each other. The generator creates fake data which is tried against the discriminator (which has access to the training data). The discriminator feeds back to the generator any differences which the generator uses to generate more realistic data. This is the method used in the creation of deepfakes.
A classification of AI model that can generate text, code, audio, image, video or other media content from a user prompt.
An AI large language model that has been trained on a huge amount of unprocessed natural language in text form, using a specific architecture (or method) - in this case the Transformer architecture. The basis of OpenAI's Chat-GPT.
A computer hardware component traditionally used to process computer graphics (image and video), but more recently used to power and accelerate machine learning and AI processes.
The process of preparing data ready to train an AI model. For example, if you are training a model to identify a dog, you would need to label your training images and videos that contain a dog so the AI knows what to look for.
A type of AI model that has been trained on lots of text to carry out language related tasks. Part of the natural language processing field of AI research.
A field of AI involving algorithms that can learn by examining and finding patterns in sample data.
The field of AI that is dedicated to analysing and synthesizing human speech and text.
AI inspired by the brain. A system of artificial intelligence nodes that are all interconnected, with the ability to quickly and easily transfer data between themselves - like neurons in the brain.
The settings of an AI model that determine how an output is generated. During training, an AI model learns how to adjust these to get a result. GPT-4 has 10 trillion parameters.
The user input used by an AI model generate a response or result.
The point in time when rapid technological advancement becomes uncontrollable and irreversible i.e. artificial intelligence surpasses human intelligence and is able to perpetually improve itself without human intervention.
Information used to teach AI models to recognize patterns and make decisions. Sometimes training data can be made up if there is limited real-world data to draw from.
A test to establish whether an AI model can communicate by text indistinguishably from a human being. See Alex Garland's film Ex Machina (2014) for a cinematic exploration of this!