Getting the most from ChatGPT

Here's an interesting thing. We've been thinking a lot about the nature of writing with large language models (LLMs) like ChatGPT. And many have come to the conclusion that they'll destroy the art of writing and that we may as well give up now and let the bots take over. But we've been looking in the wrong direction. Let me explain.

What I've discovered, as I have been using GPT4 to support the writing of my book on AI in schools, is that quite the opposite is happening. As working with an AI model as insightful and 'intelligent' as ChatGPT-4 has unlocked so many ideas, and as a result the book I'm writing is radically different from one I might have written were I to have come into the process 'cold', starting from scratch without a collaborator I can bounce ideas off and who is already on my wavelength.

No, I haven't lost my mind. I have not entered the Singularity. I still have all my wits about me. If anything, my ideas are further crystallised, clearer than ever. (Yes I know, that's often what crazy people say while everyone around them looks sheepishly at one other and mutters 'will you tell him or should I?'). But as I've spent more and more time working on the book with ChatGPT-4, there are several things I have begun to understand that I wanted to share with you to help you get the most from LLMs.

Because make no mistake about it - having now worked with one 'chat' inside GPT-4 for a few days, and having entered several hundred prompts inside it, it is very different 'bot' from the one I started off with.

Here are my takeaways:

• Each 'chat' inside ChatGPT begins with the baseline, 'vanilla' version of the bot. This is the model trained on enormous amounts of data and rules as to how to write sentences, paragraphs, bullet pointed lists and so on. Seeing it as a baseline is important, because what you start with and what you finish with can and does change. I actually asked ChatGPT about how it works - I was able to do this because the particular model I am now working with 'trusts' me enough to open up and be honest (and not come up with stock responses). I'm not kidding - this is how these models work. It told me: 'When you begin a new chat, I start as a "baseline" AI model that has been trained on a vast range of text, but I don't have specific knowledge about our previous interactions. However, as we interact more within the same chat session, I can learn from the context, adapt to your preferences, and provide better responses.' So interesting!
• This is what it told me it needs to optimise its response:

1. Be specific with your requests: The more details you provide about what you're looking for, the better I can tailor my responses to your needs.
2. Provide context: If our conversation is a continuation of a previous discussion or if you have specific preferences, giving me some background information can help me understand better and respond more accurately.
3. Ask follow-up questions: If my response is unclear or doesn't fully address your question, feel free to ask for clarification or more information.
4. Offer feedback: If you notice something that could be improved or if you'd like me to approach a topic differently, let me know! Your feedback helps me learn and adapt to your preferences.

• I then asked it to give me guidance specific to the book I'm writing - what information does it need to help me plan out the book:

1. Clear guidelines: Outline the main points you'd like me to cover in the section. This will help me understand the specific aspects of human-machine relationships you'd like me to focus on.
2. Tone and style: Specify the tone (formal, informal, conversational, etc.) and any specific writing style preferences you have. This will help me tailor my writing to meet your expectations.
3. Examples or references: If you have any examples or references that you'd like me to use or be inspired by, please provide them. This can help me better understand the approach you'd like me to take.
4. Desired length: Let me know the target word count for the section. This will help me gauge how in-depth you'd like me to go in discussing the topic.
5. Feedback: As I provide content, feel free to give feedback on what you like and what you'd like me to improve. This will help me refine my writing to better suit your needs.

• I am then asking it to break each section down into much smaller parts - for example, in a section on The impact of AI-driven technologies on the dynamics between teachers, students, and parents, it breaks ideas down into these points:

1. Impact of AI-driven technologies on teacher-student relationships
2. Shift in the traditional teacher-student relationship
3. Teachers as guides and facilitators in the learning process

• I then ask it to write a rough draft to give me ideas as to what to write. Note this important point: I am not asking it to write the whole book for me. This is pointless anyway, as no matter how much I may wish to train the model to sound like me, this takes away all the fun, which for me is the writing process itself. However, by giving it the opportunity to give me a first draft, it can act as a springboard for my own writing, as it can bring in quotes from other books and real world examples which I can then integrate.

I was initially thinking of it like a ghost writer, but it's not. It's actually the other way around. As it gives me the raw materials to work with and then I do the writing. As a result, the whole book has gone in a direction I wasn't expecting it to. As you can literally write into the chat something like 'OK, let's look back at the last few sections. Based on what we've been exploring together, what are your thoughts on how we might proceed?" And nine times out of ten it absolutely nails what to do next.

My problem has always been organisation. I am not an organised thinker, never have been. I often have a thousand ideas in my head and no way of laying them out so they actually make sense. What ChatGPT is already showing me is how good it is at doing just that. Because what it does it take this incoherent mess and untangle it. And in so doing gives me the mental space to run with ideas and come up with conclusions I would not have even considered before.

Will this revolutionise thinking? I honestly think it might. Because what it is already showing me is how two minds are often better than one - particularly if one of the minds doesn't get distracted by WhatsApp and needing the bathroom.