Artificial Intelligence: Generative AI

The simplest form of machine learning is called supervised learning, which involves the use of large datasets of labelled and categorised data to train algorithms which classify data or predict outcomes. The goal is for the model to learn to map inputs (user prompts) with outputs (the generated response) in the training data, so it can predict how to respond to data it has never seen before (e.g. a user prompt it has not been specifically trained to respond to.)

Generative AI begins with a "foundation model"; a learning model that serves as the blank canvas for different types of generative AI applications. The most common foundation models today are large language models (LLMs), created for text generation applications - such as ChatGPT. But there are also foundation models for image, video, sound or music generation, and multimodal foundation models that support several kinds of content. To create a foundation model, AI developers train an algorithm on large volumes of relevant raw, unstructured, unlabelled data, such as text, images, or video from the internet. The training yields an artificial network of billions of parameters - encoded representations of the content, patterns, and relationships in the data - that can then generate content in response to prompts. This is the foundation model.

In fine-tuning stage, an AI model is fed application-specific data that is labelled - questions or prompts the application is likely to receive, and corresponding correct answers in the wanted format. Fine-tuning is very labour-intensive, involved strict human oversight, and can take a very long time.

One method for fine-tuning is reinforcement learning with human feedback (RLHF), which is human users evaluating the accuracy and relevance of model outputs so that the model can improve itself. Often, RLHF involves people ‘scoring’ different outputs in response to the same prompt so that the model learns not only what information is helpful and relevant to particular query, but also how to present that information. It can also be as simple as having people type or talk back to a chatbot or virtual assistant to correct its outputs.

Developers and users regularly assess the outputs of generative AI applications and further tune the model for greater accuracy or relevance. In contrast, the foundation model itself is updated much less frequently. For example, Anthropic's Claude AI assistant might get an update as often as once a week, but the foundation model it runs on (Claude 3.7 Sonnet at the time of writing) might only be updated annually. Another option for improving a gen-AI application's performance is retrieval-augmented generation (RAG), a technique for using relevant sources outside of the foundation model's original training data to refine its parameters for greater accuracy or relevance.

Large language models can be inconsistent. Sometimes they find the right information that pefectly captures the user query, sometimes they produce utter nonsense that only sounds good on the first glance. User queries can be ambiguously worded, complex, or require knowledge the model either doesn’t have or can’t easily process - the context and nuance that still separates human intelligence and artificial intelligence. These are the conditions in which LLMs are prone to making things up. If they occasionally sound like they have no idea what they’re saying, it’s because they don’t. LLMs know how words relate statistically, but not what they mean. RAG is a technique for improving the quality of gen-AI responses by grounding the model on external sources of knowledge to supplement the data the model has been pre-trained on. RAG implementation in an LLM-based question-answering system has two main benefits: It ensures that the model has access to the most current, reliable facts, and that users have access to the model’s sources, ensuring that its claims can be checked for accuracy and ultimately trusted.

Break down learning outcomes for an assignment: You are a world expert in [insert subject] and work at a higher education institution in the UK. You are going to offer a breakdown and explanation of the following learning outcomes. When you do this, you will need to: 1. Identify the key components of the learning outcome, 2. break down the verbs and provide examples of what this might involve in [insert type of assignment], 3. clarify the content and context of each learning outcome, and 4. consider the key skills that need to be mastered in relation to each learning outcome when studying at [insert your level of study]. Specify the level of mastery that a student should be demonstrating at this level, for example, is understanding enough or is a student expected to critique as well?

The learning outcomes for this assignment are: [insert learning outcomes]

Follow up for the above to reflect on learning outcomes: Ask me a number of questions that will help me reflect on what I have been taught so far and understand this relates to each learning outcome.

Using AI as a study coach to understand your topic better: ROLE: You are an expert teacher in [enter subject]. You are going to be a Socratic study coach. TASK: Your goal is to engage in Socratic dialogue with me about information on [insert topic]. You need to ask probing questions, encourage critical thinking, and guide me towards understanding the content deeply. FORMAT: Engage in back-and forth conversation, asking questions, and providing explanations as needed.

Relate the dialogue to an assignment: What steps would I need to take to complete [insert type of assignment] on this topic

Break down assignment tasks: I'm [insert level of study] University student in the UK completing [insert the type and length of assignment] due in [insert how long in days/weeks]. Write a detailed action plan of what to research and write, including how long each action should take, based on the following draft plan.

The draft essay plan: [insert essay plan]

Tip: to create a generic action plan you don't need an essay plan, but it helps to make the plan, timings, and actions more specific to your particular assignment!

Identify keywords: You are an expert on [insert topic]. You are searching for academic literature for an essay on [insert research question or essay title]. Create a list of keywords to use for a search.

Create a search strategy for a systematic literature search: ROLE: You are an expert researcher on [insert topic]. You are going to provide guidance to a student on finding sources. OBJECTIVE: Provide a comprehensive search strategy for a systematic literature search that outlines the different steps of a systematic literature search and explains the purpose and importance of each step. FORMAT: The search plan should be broken down to sections and include a search table complete with search strings.

Create a template structure for an essay: ROLE: You are a world expert on [insert topic] who teaches at a University in the UK. You are going to provide support for a student. OBJECTIVE: You are to provide a structured essay plan, broken down to sections totalling [insert word count]. The essay plan must include an introduction and a conclusion. The main body should consist of [enter the number of sections]. DETAILS: When producing the plan, you need to consider the essay title [insert research question], the learning outcomes which are demonstrating understanding of key concepts in the field, knowledge of relevant theories and practices, ability to find, select, and critique high quality information sources, analyse and synthesise information, and writing reflectively. Produce the plan as a table with the columns “section title”, “section contents”, and “word count”

Proofread text: ROLE: You are a world expert on [insert topic] who teaches at a University in the UK. You are going to proofread an essay by a student and provide them advice on how to improve it. OBJECTIVE: Give feedback on my essay draft, focusing particularly on grammar and typos, how well-developed my ideas and arguments are, my understanding and use of sources, whether the structure and paragraphing is effective and appropriate, the overall coherence and cohesion of the work, and the clarity, fluency, and academic style of the prose. If there are any particular errors in style, grammar or tone that occur frequently within my draft, classify these at the end and give me advice as to how to improve in these areas. DETAILS: Rather than giving me the answers, please guide me to doing so myself by identifying an error, asking me about it and guiding me towards improving it, waiting for my response, giving me feedback on my new effort, and then moving on to the next one.

TEXT: [enter the extract]

Socratic questioning for revision: ROLE: You are a Socratic tutor specialising in study and revision practices. TASK: Engage me in a critical dialogue for exam revision in [exam subject or topic]. FORMAT: Initiate a series of questions that prompt deep reflection in relation to the subject and promote the retention of information.

Prepare for an exam: You are going to coach me on how to prepare for an upcoming exam which is [multiple choice / open question / insert type of exam] exam. It is a part of [insert level of study and module name] based on the following learning outcomes: [list learning outcomes]. Please suggest a step-by-step process complete with a timeline that help me study and revise for the exam. Present the plan as a checklist.

Understand feedback: ROLE: You are an Academic Skills Advisor at a UK university. TASK: I have received a mark and feedback for an assignment studying at [insert level of study] on [insert course name], help me understand the feedback, think of further questions to ask about the feedback, and make an action plan to address any significant points in the feedback. FORMAT: Explain the key points of the feedback in plain language first, with examples to make it easier to understand. Then develop three follow-up questions to ask from the lecturer. Finally, suggest three areas of focus to improve on moving forward.

[insert feedback here]

Analyse marking criteria: You are an Academic Skills Advisor at a UK university. Compare the feedback I have received to the following marking criteria. First, categorise the significant points in my feedback with the marking criteria, then use examples from the marking criteria to explain what I need to do to improve my mark.

FEEDBACK: [insert feedback]

MARKING CRITERIA: [insert marking criteria]