🐙 Promptpack: Your ultimate end-of-year AI guide
What you need to know to make the most out of AI
Hi,
Throughout 2023, we shared a great deal of analysis and guidance about how to make the best use of today’s exciting AI tools. We decided to put them together in a single, widely-accessible note, to open up the variety of the resources our team at Exponential View have put together for you.
In this final Promptpack of 2023, we cover:
Feeling left behind? Quick getting-started tips.
How to talk to your AI — the art of prompting.
Our ten favourite prompts of 2023.
What could go wrong, and what to do about it?
Getting started
Choosing your AI
There are many AI tools entering the market every day, but the most common foundation models are the best starting point. Most of them have a free version or a free trial.
The fact that these are general-purpose does not mean that they necessarily perform less well than more specialised models (for example, one that is specifically marketed towards content creators.)
Proprietary AI such as OpenAI’s ChatGPT and Anthropic’s Claude-2 is owned and controlled by a specific entity with restricted access, while open-source AI (Meta’s Llama 2, Mistral’s Mixtral, Cohere) is publicly available with its source code accessible for modification and distribution by anyone.
Proprietary AI is often more user-friendly and better supported, making it easier for beginners to learn and use.
Therefore, I would recommend starting with either ChatGPT or Claude. If you are going to use AI specifically for research, you can directly start with Perplexity or Elicit. If you’ve been following EV for some time, you’ll know that Perplexity is
’s favourite AI tool — and he discussed the ins-and-outs of it with the co-founder and CEO Aravind Srinivas.
Figure out your first use case
The next step is to figure out, in your own workflow, where AI is most appropriate. Some questions to ask are…
What tasks take up more time than necessary?
What kind of work do you wish could be improved: is it more creative work, research, or strategic? It may be a blend!
Do you struggle with a blank page and need help getting started?
Fundamentally, AI tools complement you in two ways:
🧠 An extra brain: it helps you think better,
👋 And extra pair of hands: it helps you do things faster.
And finally, consider data privacy and risks — check the guidelines from your organisation.
As I already mentioned, I first got started with ChatGPT — and it still is the tool I personally use the most. Two specific features — alongside the chatbot — made by OpenAI are extremely useful.
GPTs are a great way to personalise your LLM and have it remember prompts: My colleague
wrote about how he used GPTs to create useful business personas.
Code Interpreter (now called Advanced Data Analysis) is a powerful tool for data analysis.
prepared two guides for you this past year:
How to talk to your AI — the latest in the art of prompting
Prompt engineering, at its core, is about mastering the art of communicating with AI. Imagine you’re giving instructions to a highly intelligent, yet quite literal-minded assistant.
That’s essentially what a prompt is in the world of AI — a way to tell these systems precisely what you’re looking for. The response you get hinges significantly on how you frame your questions. The science is coming out now, and we have some clues on how to best interact with these AIs. Here’s what we know thus far:
🎯 Be precise about what you want, and add context. The more context the model has about your request, the better it will be able to tailor its outputs to your circumstances. For example, ask it to take a certain persona — for instance, an expert ornithologist with knowledge of a specific geography if you want to learn about birds in your area; or tell it why what you’re doing is important, or who the audience will be for a certain text.
📚 Provide reference text. Giving the model reference text can help in two ways. First, if instructed to do so, it can mimic the style and format of a text of your choosing, which helps you get exactly what you need. Second, asking the AI to use data from the reference means that it can work from your specific data.
🪺 Ask for multiple possibilities. Generative AI models are probabilistic, which means that they will give different outputs if you query multiple times. A good way to make sure you have a high-quality output is to ask for multiple possibilities, and you can choose the best.
🧱 Ask it to break problems and tasks down into components. A lot of evidence is coming out in favour of the “doing things in a step-by-step manner” prompting strategy - it seems that getting the model to break down a task into its component parts improves the quality of its work.
🕯️Experiment! Some of the most effective prompting techniques can sometimes feel counterintuitive, and came out of some surprising experimentation. For example, some research has found that appeals to emotions (e.g., “This is really personally important to me”) produce better results. We don’t understand the mechanism, but it works.
💡 For more advanced prompting techniques, check out my guide on how to use AI to improve your prompts.
Our 10 favourite prompts of 2023
Prompt 1: Apply the concept of [Y] to [X]. Give me 7 options. Let’s take this step by step to make sure we have the correct answer.I would suggest experimenting with the following: Apply the concept of modularity to innovation in synthetic biology. Give me 7 options. Let’s take this step by step to make sure we have the correct answer.
Prompt 2: Is there a relationship between [X] and [Y]? If so, give me 7 examples.Try with the following: Is there a relationship between lasers and geopolitics? If so, give me 7 examples.