Thinking with ChatGPT

How do you think best? Given how important thinking is to us, we ought to have a ready answer. Do you? You might be tempted to answer: When I do x or am in place y. This is, however, not an answer to the question. You're describing an experience, not a process.

Many people can describe experiences, but it's the processes that matter. A chef understands the process and can make a great meal in any kitchen. The cook has a good experience but can't generalize or share.

All your best thinking is done in three circumstances: conversations, in writing, and under guidance. What do these categories share? An imposed structure.

Your thoughts are a mess. But you do know what you want. So you can co-shape that mess with another person. You can co-shape it with your past self in writing. You can have standard operating procedures shape it.

Empower your thinking

In the past, you could only do this haphazardly. Your thinking about goals and your thinking about the details got mixed up. You may have gotten fortunate by finding a good procedure. A good partner, a writing habit, a therapist, and good operating procedures.

But, all of those systems are flawed. They have their own goals and issues. They were not built for you. So you had to make do. However, we live in an age of AI. AI systems like ChatGPT can be given your goals. It can be shaped to your purpose.

Yet, most people do not understand this. A beginner approaches a system like this and pattern-matches it onto their preconceptions. A programmer? This must be a code generator. A student? A way to cheat on my homework. A non-technical person? “Something” that can talk to me. A curious person? This is advanced Google.

None of these perspectives enable them to get the full value out of the system. Perspective is vital. You must adopt the right perspective.

You don’t fail to get what you want because of the nature of the thing, but because you view the thing in the wrong light.

With the proper perspective, you gain more opportunities

So what is the right light to view such a system in? As a simulator. The Large Language Model has learned the patterns inherent in the space of human ideas and concepts. You interact with a specific part of that realm, a simulacra. It takes the potential inherent in the space and actualizes what is brought forth by your characterization. You characterize it with your input.

Think of it this way. Take the best actor in the world. They deeply understand the realm of acting. But how well will they perform if you wake them up in the middle of the night and ask them to do something without context?

No, a great actor needs a great playwright to actualize their potential. You must set the stage for them, you must look at the whole project to see how each part matters. Then you get a masterwork.

Today, you are mixing this high-level work with the low-level work. You get caught up in tedious details. You worry about low levels like spelling or formulation. You can hand all of this off to the AI. If you direct and guide it correctly. You can massively increase the value you get from the system. You will save time and frustration. You can have an assistant. A perfectly patient, knowledgeable, and quick assistant whom you must only direct.

The core idea of getting the system to do what you want is to tell it exactly what you want. Ask yourself what the actor awoken in the middle of the night would be able to do with your instructions. Why do you expect the AI to do any better?

Like any playwright, you need to give your actor instructions about their role, the context of their act, instructions for their actions, and your desired outcome. Then they can integrate all that, and you get a good result. Let's go over the points.

Role

Who or what is the actor? What sort of person could give you the results you want? This can be anything from “Explain painting to me like Leonardo DaVinci” over “Give me advice as an expert for interpersonal relationships” to “Act as a text correction software.”

Specifying who it is makes the system “zoom in” on a specific section of the space of human concepts, where certain ideas are more likely to emerge from it.

Context

What is your situation? Describe it. Describe the circumstances that made you pull the actor onto the stage. How does this matter in the play of your life? Do you need a document corrected? Provide it. Do you need advice on landscaping? Describe your garden. Do you want a sanity check on an idea of yours? Explain the idea to it.

You care about it doing useful work for you. So it needs to understand what sort of work you need to do so that it can do it better.

The biggest benefit is that it forces you to think deeply about your actions. Too often, we just drift into our work. Here you must take a step back and consider it. Extract the relevant expectations.

Describe it in a manner that would allow someone to provide useful advice. If you would ask an advisor and they would say, it depends, you haven't given it the things your solution depends on.

Instructions & Outcome

Any word that is an operator directed at the AI is an instruction. At the most basic level, write or tell me about x, evaluate and criticize x should be part of any prompt. Otherwise, the system is just guessing what you want. Intuitively you already do this in conversation.

But you can do much better. Being conscious about the interaction of your instructions and how they will lead to your goal is very valuable.

Given that the system only “thinks” forwards, it generates each word based on several pages of your previous conversation. It takes your prompts and its answers. It then adds a word. Then it repeats the process. This is why it struggles with telling jokes. A comedian sets up the punchline in the story. The model tells the story first and then tries to find a punchline. Understanding this, you know why you need to think ahead.

Take and adapt this prompt

Given that it can't have a plan, you must give it one. I'll give you my favorite blueprint for a prompt.

[context] - - > Repeat your instructions back to me. Evaluate and criticize [the idea]. Adopt the perspective of [3 different customers] and give me positive and negative feedback. Tell me how to improve [the content]. Self-criticize your response, was it correct?

Let's break it down.

• - - > is a separator from the context that marks out that this is an instruction section.

Repeat your instructions back to me. This makes sure you can see that it understood your instructions. Cancel the answer and adjust your prompt by telling it, “no, do x, not y,” “less wordy,” or “more formal”... You can save a lot of time by seeing if it fundamentally understood your instruction as you meant them.

• Evaluate and criticize [the idea]. This is its basic action with your goal. This is our last point, the desired outcome. Specify it clearly and shortly here so the model pays extra attention to your goal. This could also be written about, give me suggestions...

• Adopt the perspective of [3 different customers] and give me positive and negative feedback. This is an example of an additional instruction. Here you can and need to think out of the box. It is more creative than you. It can brainstorm over the sum of human knowledge and provide insights. Ask it to “zoom out” on the idea. Let the AI break your framing of the problem so you can see the whole scene and not just your current picture.

• Tell me how to improve [the content]. This is where you ask for advice more generically. Again, this depends on the last of your four points, your desired outcome. Ask it for help. Most of what you will get here is trivial and useless, but a gold nugget is often mixed into the dirt.

• Self-criticize your response, was it correct? You might be worried about how these models tend to sell you hallucinations. They can just make stuff up. This is a result of the model's limitations. It predicts text forwards. But it can also tell you if it just made stuff up. It is not attached to its opinion. It just tells you that its answer was wrong if you ask.

These self-corrections won’t give you the truth if it is not clearly recognizable in the world. Yet, it will help you if the model just made stuff up. This partially eliminates a certain category of errors, but not all.

Any one of these steps is valuable. But the magic happens with their integration. I can tell you all about the ingredients and recipes. Yet, one only becomes a chef through experience.

These steps can and must be combined to serve your purpose optimally. How exactly? This is a matter of experience and intuition. As a rough guideline, the more important your question, the more elaborate the prompt.

Finally, don’t think you have to do this all the time. Using AI as a search engine is fine if that is what you need now. Use a tool suited to the job, not the best one you could use.

Clear communication will now be valuable beyond human interactions

In conclusion, the most important lesson here is that you need to learn to clearly communicate the actions that will lead to the fulfillment of your goals. Note how expressing this clarity of what is required for a purpose is valuable beyond your interactions with AI systems.

You have seen why the best perspective on such systems is that they are universally capable actors. You now know how to direct them. Learning to master the interactions of role, context, instruction, and outcome in your thinking is one of the most valuable skills of our time.

CTA

Where does this leave you?

In a place with ideas.

To get actual value out of this, you need to execute.

Follow the steps below to 10x the value of the time you invested into reading this:

1. Tool: Go get ChatGPT.

2. Goal: Pick a moderately complex problem that bugs you in life.

3. Role: Consider the sort of person who could help you solve it. Tell the model to “Give me advice from the perspective of [description of person].”

4. Context: Describe the problem in detail. “I am facing… My situation is… I have tried…”

5. Instructions & Outcome: Tell it what you want it to do and what you want to achieve with that. “--> Evaluate and criticize, give me advice, brainstorm solutions, … [add operators that are suited to your purpose] to help me get [desired outcome] out of this.”

6. Respond: See what the AI tells you. Now construct a new prompt based on the principles above. Dig deeper. Find a solution.

I hope that this has helped you.