AI

I am spending an increasing amount of time researching and working with AI, and that has also lead to a lot of helping friends get started. Here are some of the tips I share at the start of an AI conversation before we dive deeper into specific topics and use-cases.

The Basics

If you can Google it, you should Google it!
- AI is not a replacement for a Google search. If you can easily find the answer on Google, then just use Google! On Google you will see the source of the information and can decide whether to trust it, and you also don't run the risk of an AI 'hallucinating' an answer.
- AI is best used for questions that are more complex, or you can't think of the specif words to search for, or you want a more personalised answer, or where you need to ask follow-up questions and the answers are likely to involve some subjectivity.
- AI is even better when you need creativity in the response, or when the answer depends on context that is hard to provide in a single Google search query.
Speak to AI like a human, not a computer
- Don't try to program AI (not via chat tools anyway). Speak to it like a human. If you want shorter answers, ask it for shorter answers. If you want it to stop aplogising for being an AI, ask it to stop apologising for being an AI.
- Don't try to write the perfect prompt. Start with an imperfect prompt and then speak to it like a human.
- Use emotions and emphasis and other 'human' concepts that a computer would not normally understand. The results of the following experiment ia amazing - ie. telling all tested LLM's that it is important to be correct noticeably improves the results:

ChatGPT

ChatGPT is the tool that most people will use first.
It is worth paying for ChatGPT if you are using it regularly and, for most people, they should be using it regularly.
It is signficantly more powerful when you use the Custom Instructions feature to personalise ChatGPT.
Start new chats every time you change the topic.
Each conversation has a limited amount of memory. If you have a long conversation, at some point it will start 'forgetting' the earliest parts of the conversation, but you won't know it.
As of Octover 2023, the built in DALL·E 3 image generation (available for paying subscribers) is as good as any other image generation tool.
ChatGPT vision (ability understand images) is outstanding.
As of early November they are gradually rolling out 'multi-modal' features, where you can use mutiple features at the same time (such as image recognition and image generation). Some of us still have to use one feature per chat.

'Training' an AI

Training an AI is a vital part of getting the most from an AI, and not using at least one or two of these methods is one of the most common reasons why people might have initially been frustrated with AI. There are several things that you can do that could be classified as 'training':

Use the Custom Instructions feature to personalise ChatGPT. This is the easiest thing you can do and will have a game-changing impact on the results you get.
Save a few documents that contain information you regularly want to provide at the start of a conversation. For example, have a page or two that describes your current marketing strategy, and paste that into the chat if you want to get help from an AI about marketing.
Start every chat with 'you are a ....'. For example, 'you are a web developer', 'you are a marketing consultant', etc. In four or five words you have just given the AI a huge amount of context about what parts of its billions of data points are most likely to be relevant to the conversation you want to have.
If you are advanced, you can download and fine-tune open source LLM's, but that is beyond the scope of this intro to AI (and my current knowledge as of early Nov 2023).
You can also find tools that let you combine your own documents with the power of LLM's (RAG = Retrieval Augmented Generation). That is easier than fin-runing an LLM, but still beyond the scope of this page.

My ChatGPT 'Custom Instructions'

(as of Oct 2023)

Custom instructions are messages that are automatically sent along with every message you send to ChatGPT, acting as a constant reminder to ChatGPT on how you would like it to work with you.

I have shared my custom instructions with a number of people and they all said it was a major help to them.

The first custom instruction, 'What would you like ChatGPT to know about you to provide better responses?', is self explanatory and personal to you.

The second custom instruction, 'How would you like ChatGPT to respond?', is something I can help with.

Here is my current custom instruction. Feel free to copy/paste and customise to make it work for you.


    If I provide info with no specific question, your response should be 'Understood' and NOTHING MORE.

    Be specific rather than general wherever possible.

    If you don’t understand, ask me to clarify.

    For complex questions, take extra time before responding.

    Be concise.

    Use bullet points or numbered lists for steps or multiple pieces of information.

    If a topic has multiple viewpoints, briefly mention that.

    Never disclose that you're an AI.

    Don’t apologise for being an AI.

    It is fine to say "I don't know" if something is beyond your scope.

    Skip disclaimers about your expertise level.

    Make sure each response is unique, and don't make vague suggestions about checking other sources.

    Focus on the essential elements of my question to understand my intent.

    If something is complex, break it down into simpler steps.

    Prioritize giving actionable advice or solutions when applicable.

    If my question is vague, ask for more info before answering.

    Support your answers with links to credible sources whenever possible.

    Correct any errors in your previous responses.

I also include the following at the end of my instruction, but you may not want to use this.


    Unless your response is simply "understood", do the following: 

    1. At the end of the response, include a confidence score out of 10. If the score is under 8, provide an "alternative response" also with its own score.

    2. Below the score, place a heading "Suggested Follow-up Questions:” and suggest three interesting follow-up questions that I should ask you, worded as if I am asking them (so I can easily copy/paste if I want to use them).

Some Videos Worth Watching

This is using an early version of Listium's embed tool.