notes

What is AI and what are AI-related technologies?

Artificial Intelligence (AI) is computer systems that can make predictions, recognise patterns, organise and generate information based on existing data.
Examples include your email service filtering for spam, your phone providing predictive texts, streaming services providing suggestions or a revising tool helps you rewrite a paragraph.
A Large Language Model (LLM) is a type of AI technology is specifically designed to draft, summarise, explain and write human language.
ChatGPT is a product from OpenAI which allows you to interact with an LLM.

LLMs are trained on massive datasets of text, and use the probability of words to predict the next word in a sentence.
When asking a question, it doesn’t answer immediately. It instead builds a response by using its training to predict the following word or phrase.
An example is providing the prompt “The sky is…” and its training indicates that “blue” has the highest probability to come next, so the answer becomes “The sky is blue.”
Every word or phrase that the LLM adds to the response becomes context for the next prediction. LLMs do not retrieve words from a list stored in a database, its generating the words according to the patterns found in the data it has trained on.
More advanced models have been trained on other types of data, such as audio, images and video. Generating content such as that also follows the same prediction concepts as text does.

Pre-training: The model learns from a large dataset to understand how language works, and any common patterns.
Post-training: The model is refined using feedback, examples and evaluations to improve it’s ability to follow instructions.
Product Experience: ChatGPT provides a simple interface to interact with the model.
Ongoing Feedback: Researchers improve models over time.
One example is asking “Explain change management in one sentence.”. While it’s raw pre-training may provide a large explanation, post training ensures that the output is concise, focused answer which is most helpful to the user.

A prompt is the context, instructions and questions you ask an AI model.
Good prompts provide context, and have requirements for the expected output. There is no perfect method of prompting, and everyone has their own style of doing so.
Communication with a person should be clear and specific if you want them to follow instructions well. LLMs are trained on human language, hence they work the same way.
Rather than a simple request, providing information such as the audience, intended use, goals, attributes you want in the output, style or language, etc, will improve the output you receive.
Giving clear instructions with the task, context, output expectation and then reviewing the output using CLEAR is best practice.

Useful context answers three questions; What are you trying to do, What information should ChatGPT use, What would a good response look like.
If a person was doing this task, what context would they need to know to do this well, is a good question to ask yourself. ChatGPT will need that context too.

Task: What do you want ChatGPT to do, such as draft an email.
Context: Background information ChatGPT should consider, such as who the email is being sent to.
Expectation: How the response should look or sound, such as the email must be concise.

Input Context: Useful information ChatGPT can use to form a response, like a file.
Custom Instructions Context: How you want ChatGPT to behave, like asking it to write responses in a professional, friendly tone.
Memory Context: Information ChatGPT can remember for future conversations outside of the current prompt, such as your current job.

Files are useful when ChatGPT needs to work with specific information specific to your task it won’t be able to find on the internet, such as reports, spreadsheets, policies, meeting notes, templates, examples, drafts, etc.
For files, best practice is only giving it the files it needs, telling it what to look for in the file, review its output and only use information that isn’t confidential.
When you need generic information, asking ChatGPT to search for it is the right way to go. Its useful for finding recent information, public information, etc but this doesn’t replace your judgement and it is a good idea to ask ChatGPT for the sources it uses so you can go check them out yourself.
Giving ChatGPT specific links is also possible if you want it to use a specific website, such as government or organisational websites, to ensure it gets information from the correct place.
You can provide AI with primary sources to ensure it’s output remains accurate, such as ensuring an email response follows the HR communication guidelines.

AI is not a source of truth, always go back to the primary source for important use cases such as current information like world events, legal guidance, prices, citations, etc.
Hallucinations are AI generated responses which sound confident, but the information is inaccurate.
An LLM has its limits, it may not know the latest cutting edge information, company details or any resources not found publicly on the internet.
Adequate context is required for the LLM to use its patterns to generate a response accurate to what you want.
The same prompt can result in different responses, even from the same model. There are multiple valid ways to respond to a prompt, and the response can differ depending on which words and phrases are used to build out the full output.

ChatGPT is only a helper, and needs human effort before and after prompting to ensure it stay accurate. You must stay in control, including taking the responsibility for using it’s work and the decisions made based on it.
You must ask yourself is this a task that ChatGPT can help with, am I providing only the needed and allowed information, and how much human review will this need after ChatGPT has done its work.
Sharing information with ChatGPT must be done so carefully, and following your organisational policy surrounding this matter is best practice.
You can minimise how sensitive data is by redacting confidential parts with fake names, and place holders like “Client A”.