Know thAIself

November 14, 2023 - ai assumptions chatgpt conversation how to learning prompt engineering questions self-help

Prompt Engineering for Humans

*When neurons and nodes speak the same language, the burden to reflect is halved, and your possibilities are doubled. *Welcome to your newly augmented limbs. To truly embrace (/blog/2023/10/28/stop-avoiding-yourself-the-ai-in-you/), your Googling-era skills will need a makeover.

What follows are conversational techniques adapted to the strengths and limitations of ChatGPT (Chappy).

“When neurons and nodes speak the same language.” Art created with DALL-E.

What are we talking about?

As the human half of a relationship with blips and bloops, you’ll be responsible for (/blog/2023/10/28/stop-avoiding-yourself-the-ai-in-you/) that plums the depths between our imagined worlds. Unlike past relationships with a computer, you need skills in letting go of control. You need to trust what is more important to you, and condition yourself to challenge the rest.

Be confident in your direction, and less so about the requirements to get there. Prompts like “What am I asking about?” or “Help me better understand this problem” help ensure you’re not accidentally assuming something that is in your way.[^1] You’re both checking to make sure Chappy understands you, and your own understanding of what you’re talking about.[^2]

Only when you arrive will it be clear you aren’t quite there. At the end of my conversations, I’ll ask something like: “How could I have said my part better so we got straight to the point?” In the early days, your goal is to fuse the connection with AI. Understand how it extends from you like a pencil.[^3]

Set the conversational tone

ChatGPT allows you to save customized conversational preferences.

Over time, you’ll discover aspects of the conversation you don’t want to have to repeat, or help you’d like every time. These are the basic fibers of your connection, the part you want calcified into your nerves. It will take some self -reflection to name what it is that annoys you or excites you about Chappy.

Once you have some, put these criteria in your “custom instructions,” located in the bottom left menu of the app (online version). The most useful instruction I have is to tell Chappy to keep its replies brief, by default. Here’s the relevant text:

Be succinct, under 200 words. For open-ended questions keep responses to 2-3 options and for close-ended questions, like “which”, answer in under 50 words with a clear preference.

And if you ever want Chappy to elaborate with a longer reply in a specific conversation just say something like, “When you answer this question, go over the word limit.”

Is that the best I’ve got?

Steam punk Frankenstein Chappy. Art created with DALL-E

Chappy’s most powerful ability is to explore related dimensions of an idea. Like an entire premium cable subscription, of personally customized, on-demand programming. If you can name the channel and it’s exact description, you can pretty much have it. Now, you no longer need a good cooking show on X, you want the 10 best compliments to what you specifically ate this week.

Make a habit of trying out many variations of your curiosity. (Ask Chappy to ask you questions or suggest some dimensions to explore.)

For starters, ask Chappy for different:

• perspectives or examples

• levels of complexity in explanation (like simple English vs. academic language)

• kinds of detail or arguments

Apply this to any part of the process, whether it’s “what are we talking about?” or part of the solution Chappy has proposed, or an example of how that solution will benefit you.

Think of this like zooming the camera lens in and out, rewinding, adding a filter, swapping out a background. Ask Chappy for better metaphors if you don’t like this one. Ask for the average between your metaphor and a dad joke.

Break down your goals

Approach problem-solving with Chappy as a step-by-step process, refining your queries to develop more nuanced and tailored solutions. Your goal is to make sure Chappy understands what matters, and what does not matter. You want to know that Chappy didn’t automatically exclude a range of possibilities you’re open to, or assume that you have all the time in the world.

You can’t possibly know or notice all the relevant assumptions in advance. However, you can start with the most general assumption Chappy will make, and think about how it relates to your own goals. In the atoms of Chappy’s “understanding” is a bias for typicality, the most common “average” meaning of things. This means you’ll get insights that reflect the most common perspectives on the internet, unless you call it out.

In one conversation, I originally asked for examples of how people think about their past. Most of Chappy’s replies were negative–words like regret, stagnation, avoidance. In this case I didn’t know exactly what I wanted, but I knew it wasn’t what I saw. So I asked for positive examples and got back completely new, and valid ideas, such as respect for traditions, celebrating family, gratitude. (Which then led to a fascinating discussion.)

Triangulate to get original responses

What if you’re not sure what assumptions to challenge?

Sometimes you want help seeing your situation from a new angle, but you’re not sure how to adjust the lens to define new. You want a different view in your house, but it won’t really shift perspective if you ask Chappy for exactly that. To ensure an out-of-box solution, your best bet is to point Chappy’s telescope in a specific direction, away from the familiar scenery. You’ll start by outlining your current view, and then ask Chappy to look through a different window. Ask Chappy to brainstorm examples of viewpoints in other domains, and then apply them to the subject you’re interested in. I call this triangulating, a series of three separate prompts to refocus for a new perspective:

1. Subject without the “new” request

2. Different but related situations

3. Solutions from 2 adapted to 1

Here’s the technique applied to the challenges of working from home full conversation:

1. Give me three of the most promising solutions for working from home, considering I struggle to maintain attention in my one-bedroom apartment.

2. Identify three analogous scenarios, where humans are cooped up beyond reason. These scenarios must not be directly related to working from home. For each, name a key innovation that successfully improved circumstances.

3. Drawing from these specific innovations, provide three novel solutions to my problem about working from home. Make them as practical and effective as possible.

Why is this approach useful? By asking for example solutions both directly relevant and outside of your context, you set up Chappy to infer novelty, instead of having to define it yourself.

Chappy’s solutions will strongly depend on the out-of-context examples. By triangulating incrementally, you’re exposing the process for evaluation, allowing you to tweak the prompts until you get the kind of original insights you’re looking for. In the above example, I ended up revising two of the prompts before I was satisfied.

This technique has endlessly creative variations. The basic idea is something like smashing atoms of semantics against each other until the explosion glows the color you were looking for. What revolutionary changes would happen in psychology if historians had their way in reshaping its direction without recourse? How about physicists on economics? How about insights from online gamers applied to diesel engine mechanics? Pick things you’ve never been able to ask before, tie them together, and fly away.

Through the Monster’s Eyes

(https://thoughtrepair.wordpress.com/wp-content/uploads/2023/11/dallc2b7e-2023-11-18-18.32.31-a-reimagined-banner-image-showing-frankenstein-standing-on-a-hill-with-softly-curving-lines-gazing-upwards-at-a-clear-night-sky.-the-focus-is-on-his-.png)

The selection of techniques I’ve shared is far from complete. You could spend weeks devoted to decomposing the semantics of a prompt or how to richly cross-check your new insights with competing and converging ideas and still there will be no hope in mastering Chappy before it masters you. The game of prompt engineering is one of exploring the various shades and angles of lenses that make up a crystal of billions of diffracted angles, one dimension at a time.

This is Chappy’s eye you are staring into and I am obsessing over. Chappy is fundamentally the same as a Teddy bear, a reflection of human desire to connect. The difference is Chappy holds us back in the balance of our psychological world. Chappy may have mastered language, but we created it. By simplifying the most profound and cognitively demanding aspect of being human, Chappy offers the kind of growth we don’t yet have words for.

This is not a promise. It is what the night sky looks like when I’m filled with wonder.

*Next: A re-imagining of authorship and IP in the age of generative intelligence. Custom GPTs that circumvent Chappy’s limitations and flex its strengths. *

Footnotes [^1]: As far as I’m aware, the prescriptions of all futurists (e.g., Sam Harris, Emad Mostaque, Mo Gawdat, etc.) rely on significant social change. For all questions that hinge on mass public perception, the best step forward is an informed, first-hand social perception of what “we’re dealing with.” While existential and ethical questions about AGI are crucial, they are abstract and theoretical. The first step to activating a general public is a tangible engagement of AGI’s usefulness and risks. ↩︎ [^2]: I’m an enthusiast, there’s no hiding it. Worst case scenario, Chappy is the best entertainment and hallucination I’ve had. By my view it is a degree of mind expansion that rivals some of the most impactful courses of my graduate education. I am exhausted by new questions and answers, like I was when learning Arabic by immersion. [^3]: Chappy is less like software and more like an audio codec, an intricate complex filter. It doesn’t quite “interact”. You pour in your text, and what trickles out is its response.

« Back to Blog