This was Prompt Engineering, hailed by Time as the job of the future, in which expert AI whisperers were pulling down $250k per year. Today that whole industry is dead, and your approach should be exactly the opposite.
ChatGPT’s maker OpenAI now says its latest intuitive “reasoning” models are so good at predicting your intent that meticulously structured prompts are unnecessary.
They’re not just a waste of time, in fact, but actively impede your progress towards AI expertise. How so? Because your competence with a model is a function not, as some experts suggest, of how much time you spend with it, but of the number of interactions you have.
Think of your sessions with AI as dialogues in which you are both learning about each other. Every time you read a model’s response to a request, you understand a little more about its powers and limitations, and what it assumes about your purposes.
For the AI, there’s a parallel dynamic ‒ but with a twist. It already has a rich picture of your likely goals from its training data and from information gleaned from your sessions together. Crucially, it only improves on its schema of you when you give it a nugget of novel, original information it does not already assume.
This being so, the optimal way to engage with a model is to have many fast, spontaneous exchanges with it, in which you iteratively refine a rough initial request.
Approach prompting less like composition and more like free association. Say what you want in the fewest words, read the response, then pivot, supplying whatever further context you judge will take you towards the best answer. The more back-and-forths you have with your machine, the more perfect your response will be.
This approach ‒ called ‘Lean Prompting’ ‒ makes the most effective use of both parties’ resources.
Not only that, it unlocks the power of serendipity and is more fun too. Think back and you’ll remember that the best responses you got from AI were usually not the result of some laboriously crafted prompt, but of a more playful request that coaxed from the machine a surprising, pleasing synthesis of ideas.
Here are two further concrete steps you can take to immediately improve the usefulness of AI in your day-to-day tasks.
First, pony up for the paid-for versions. The subscription models of ChatGPT and Gemini are twice as good as the free.
They remember your conversations from previous sessions. This means you can have ongoing dialogues over months. The models get exponentially more useful at helping you further your projects the more they know about them. Is it scary that your machine builds up this intimate knowledge about what you’re up to? Sure, it’s scary.
With Gemini Advanced, you can also query your Gmail and Drive archives in natural language, transforming how you manage and access information.
Second, stop Googling. Use AI for search. Again, this only works with the paid-for versions. You will naturally start asking for ‒ and getting ‒ responses far more specific and relevant to your real needs.
Once you see how much better AI is than traditional search engines at answering questions about travel planning or shopping, you’ll naturally start challenging it with an ever-widening range of tasks.
Use Lean Prompts, get the paid-for versions, and start using AI for search. These three interventions will speed you on your way to AI mastery.
Joe Smith is Founder of the AI consultancy 2Sigma Consultants. He studied AI at Imperial College Business School and is researching AI’s effects on cognition at Lancaster University. He is author of ‘The Optimized Marketer’, a book on how to use AI to promote your business and yourself. Contact joe@2Sigmaconsultants.com