Making friends with ChatGPT
ChatGPT, the large language model (LLM), was released on November 30th 2022. Just over a year later, while highly controversial, it has become almost a household name and most organisations have begun to explore how to use it and its AI cousins. We're all used to talking to humans, who will often pick up clues in body language or tone of voice and go beyond the superficial meaning of what we say; LLMs can't do that, so we need to learn how to talk to them.
When a new tool comes out, we tend to focus on doing the same things we were doing before, just in a different – often quicker or more effective – way. But ChatGPT doesn’t just speed things up; it lets us do things we couldn't do before. And that's where the fun starts.
It’s all about words
The first thing to understand is that ChatGPT works using text prompts. This means that the key to using it well is to learn to write good prompts. There are already people employed as “prompt engineers”, whose job is to structure text so that it can be interpreted and understood by generative AI models. Fortunately, AI prompts are written in ordinary language, so it’s possible for any of us to learn how to write good prompts and get the best out of these new tools.
The next thing to understand is that ChatGPT is a conversational tool. You talk to it in ordinary English and it will answer you the same way. But question and answer isn’t the same as conversation. With ChatGPT, you don’t need to tell it everything with your first prompt. ChatGPT 3.5 (the free version) has memory. Not indefinite memory, but, as long as you don’t log out, around three pages of text will be remembered, so you can refer back to previous prompts and answers and refine until you achieve what you need.
Reformulation
ChatGPT has been trained on a vast corpus of texts and, unsurprisingly, it has real skill with words. Its answers are almost always well written, with correct grammar and punctuation. If you give it an existing text, it can rewrite it, ironing out any written errors, adding in catchy titles etc.
Given this skill, it makes sense that one of the most powerful functions of ChatGPT is its ability to summarise. You can copy and paste a text that is difficult to understand (a scientific concept, for example) and ask it to summarise, simplify, or translate it into another language. You can also ask for the answer to be given as bullet points or in table form. And once you have a table of information, you may ask for changes to be made, or request it in a standard format compatible with other software such as Excel.
A collection of experts
If you wanted to know about botany, you’d probably go to a botanist. And if you wanted to know about quantum physics, you’d try and find a quantum physicist to talk to. ChatGPT is an expert in all subjects, so we can ask it about anything. The problem here is that, unless we specify otherwise, we end up talking to a “default expert”, which is essentially a combination of all these experts: instead of speaking to an expert in the relevant discipline, we have a mix of botanist, quantum physicist, veterinarian, systems analyst, historian, linguist... The result of this is that the answers end up being mediocre or generic. So the first thing we have to decide is: who do we want to talk to?
Picture ChatGPT’s knowledge base as a tree structure, with the information for each discipline growing from a different branch. To better exploit the depth of its knowledge, we want it to ignore almost all of the millions of branches. Then, rather than dabbling around through the entirety of information available, it will limit itself to the field we are interested in and travel farther along the specific branch and its relevant offshoots to provide us with an answer.
Assign it a role
It’s a good idea to start each prompt by specifying which expert we want ChatGPT to be. We might say, “I want you to answer as if you were the greatest expert in botany in the entire UK” or, “I’d like you to answer as if you were a civil engineer consultant to the British government.”
That phrase “as if you were” is key. If you say, for example, “I want you to be Lord Nelson”, ChatGPT is likely to respond, “I can't. Lord Nelson is dead and I’m an AI.” Our prompt shouldn’t ask the LLM to be something or someone, but to act “as if it were.” Using this magic phrase “as if you were”, you can use specific people in a prompt: “Answer as if you were Einstein” if you want to talk about relativity, for example, or “Answer as if you were Freud” if you want to talk about psychology.
You can also ask ChatGPT to speak from an area of knowledge – “a botanist specialising in fruit trees” – or a specific role or point of view – “I want the perspective of a Labor Party MP”. But bear in mind that the latter is asking for an opinion. The answer may reflect a valid viewpoint, but won’t actually correspond to the actual attitude of any real life person.
Give it a personality
The role you assign to ChatGPT can also affect its voice: if you ask it to answer as if it were Sir David Attenborough, for example, whatever subject it is talking about, it will find metaphors and concepts from the world of ecology and natural history.
Asking it to write about critiquing poetry in the voice of Sir David resulted in a text that included the following:
“Consider, if you will, a poem as a vibrant ecosystem, each word a living organism contributing to the delicate balance of meaning. The choice of language, rhythm, and metaphor, akin to the intricate web of interactions in nature, creates a tapestry that unfolds before our senses. A skilled critic, much like a seasoned naturalist, recognizes the interplay of these elements and seeks to unravel the threads that bind them together.”
In contrast, this is an extract from the exact same prompt with the name changed to Boris Johnson:
“Now, poetry, you see, is a bit like politics – full of twists, turns, and the occasional eloquent flourish. So, let's dive in, shall we? First things first, poetry is like the English weather – it can be a tad unpredictable. One moment you're basking in the warmth of a well-crafted metaphor, and the next, you're caught in the rain of clichés. A proper critique, then, requires a keen eye for detail and a readiness to weather the storm of artistic expression.”
Define the audience
If you are used to providing your business information for different stakeholder groups, you probably already understand that the same information needs to be expressed differently depending on who the intended audience is.
You can ask ChatGPT to frame its answer for any audience, from a class of five-year-olds, to a PHD in Astrophysics, to a Board of Trustees for a mental health care charity. You can also ask it to write in a particular style – formally, casually, academically… In each case, the answers will be completely different and designed to engage with the specified audience.
Refine your questions
Remember that the answers you get from ChatGPT will depend greatly on how you phrase the question; small changes to the prompt can make a huge difference to the answer. So don’t settle for the first answer; tweak the prompt; then tweak it again! Model yourself on the stereotypical toddler who keeps asking, “Why?” and push ChatGPT to improve on its answer.
Remember, too, that the responses are greatly improved when you provide good ingredients for ChatGPT to work with. You can start by thinking of this in a literal sense: you might say, “Give me a recipe to cook for lunch”, or you could look in the fridge and say, “These are the ingredients I have available [list ingredients]. What can I cook for lunch?” Clearly the latter will give you a more useful and relevant answer. If you keep refining and adding restrictions or details – how long you have to do the cooking, what pans you have available, how many people are going to eat the meal, etc – the answer will become closer and closer to the best possible answer to the problem.
Now apply this technique to any question you might ask: keep refining and restricting, adding in parameters that encourage ChatGPT in the direction that interests you, and you will get better and better answers.
ChatGPT isn’t Google
Unsure of how to use this new tool, many people are asking simple factual questions such as, “What is the capital of Italy?” Yes, ChatGPT will (probably) give you the right answer, but it’s a question you could easily have asked Google. If you want to know what happened at the Battle of Jutland, you’d be as well going to Wikipedia. Again, ChatGPT will give you an answer – and it will provide a complete and unique essay on the subject – but it’s unlikely to include any information beyond what can easily be found via Wikipedia and the corresponding references and sources.
These simple factual questions go nowhere near exploiting the power of ChatGPT. What’s more, the answers it gives are answers that sound plausible, rather than answers that are correct, so it may include very convincing, but ultimately incorrect, information. When an LLM does this, it is said to be hallucinating.
It’s important to focus on the things that ChatGPT can do that Google and Wikipedia can’t. It can handle very complex questions and lots of information. And, as we said earlier, you can converse with it as if you were talking to another person, asking it to review and refine its answers. For all intents and purposes, the complexity of the prompt adds no additional cost. It’s like asking a printer to print a doodle, or to print the Mona Lisa: either way, it’s going to use a sheet of paper. Using ChatGPT, a simple factual question takes a few seconds to answer, and an absurdly complex question also takes a few seconds.
Leverage the strangeness
We saw earlier how we could ask ChatGPT to answer a question in the voice of an expert in an unrelated discipline. While the examples we gave were purely for fun, this juxtaposition of skill sets and subjects can be an absolute treasure trove of creative ideas.
For example, what might a marketing plan look like if it were designed by Einstein? Obviously the answer is going to be somewhat bizarre, but it is in this generation of strangeness, those disconnected and unexpected perspectives, that original and innovative ideas emerge.
Ask it to prompt you
We’ve said the trick to getting good answers lies in the creation of good prompts, so why not put the ball in ChatGPT's court and let it tell you what it needs to know?
Suppose your child has a problem at school and you need to write a letter to their teacher. Crafting a letter is a perfect task for ChatGPT to help you with. But immediately launching into a simple “write me a letter” prompt, is likely to result in a very generic response that will need several rounds of refining. Instead, you could start with, “I need to write to my child's school. What do I need to tell you so that you can write the letter?” ChatGPT will then ask you to provide all the information it needs: your child’s name, the name of the teacher, the reason you are writing, relevant details to include, desired outcome, etc. Providing all these things in your prompt will allow it to make the letter as personalised and relevant as possible.
Perhaps you need help writing a marketing plan. Again, a curt “I need a marketing plan for a business in [X] sector.” will generate a plausible, generic response, which is highly unlikely to be useful. So ask ChatGPT to help: “My business is in [X] sector. What information do you need in order to write a marketing plan for me?” It will provide you with a list of information that it needs to write a plan that is relevant to your actual business, not simply a stereotypical form-filling exercise.
Get the AIs talking
While ChatGPT is probably the best known LLM, there are others, such as Bing AI and Claude 2, as well as text-to-image models, such as Midjourney, Stable Diffusion or DALL-E, which take prompts and create original pictures. All of these come under the umbrella term "generative AI".
We’ve just seen that ChatGPT can help create prompts to help you ask it the right questions, so it’s reasonable to think it’s also going to be capable of creating prompts for other artificial intelligence software. So, if you use other AI platforms, you can ask ChatGPT to generate prompts for them. For example, “I want a campaign to launch a product named [name], to be aimed at [audience], with the following [characteristics], etc. Write a prompt for a text-to-image model to create a logo consistent with this brand positioning.”
Refine, refine, refine
While the initial responses from ChatGPT can be startlingly convincing, they can usually be improved on. So don’t simply settle for the first answer. You may see clear improvements that are needed: “The style is too formal. Make it more colloquial” or “It's too short. Elaborate a little more” etc. There may also be other specific corrections: “I see that you do not mention [topic] enough”, “Expand on point 2”, “Add more statistics”, “Restructure the text so there are five main points”, etc.
Even if you don’t spot any particular flaw, it’s worth pushing for more. You can simply use the “regenerate” button, or get conversational: “I was expecting more from you. Get your act together.” If you tell ChatGPT you expected more, you’ll get an apology: “I'm sorry to hear that you are disappointed, I'll do it better now.” And it will go ahead and produce another – usually better – answer.
Whatever the topic, corrections and some final polishing will always be needed and it can often take four or five attempts to reach a final answer that really meets your requirements.
The good, the bad, the hallucinations, and one final key
There are some things that ChatGPT is really good at. It’s brilliant at summarising, at simplifying complex concepts and at translating. It’s also excellent at generating creative ideas and brainstorming.
But there are also some things it’s very bad at. Don’t use it for the things Google already does – don’t ask it for factual data or to recommend website addresses. It’s not particularly good at programming, although it can be used to identify errors in code. According to many mathematicians, it’s currently not very good at maths.
Remember that anything written by ChatGPT needs to be fact-checked: it gives plausible answers, not necessarily true ones.
Finally, when ChatGPT has produced something that is perfect, that is just what you were looking for, you should ask it one last question: “How should I have written the first prompt for you to provide this answer immediately?” You can then use this information to help you write better and more efficient prompts in the future.
Mastering generative AI opens doors to a world of possibilities for your organisation. Making friends with ChatGPT is a relatively easy – and fun – first step on this fascinating journey.
ChatGPT and other generative AIs are amazing tools and can transform your business, but they need to be used judiciously. At Tantamount we have always used cutting-edge tech where appropriate. We also know when the human side is more appropriate. You won’t find yourself dealing with an AI if you get in touch to discuss your branding and marketing needs!