![[redacted] learns AI](https://substackcdn.com/image/fetch/$s_!errE!,w_80,h_80,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6e3153aa-ec5e-4245-a303-fa25811e83ee_512x512.png)
![[redacted] learns AI](https://substackcdn.com/image/fetch/$s_!errE!,w_36,h_36,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6e3153aa-ec5e-4245-a303-fa25811e83ee_512x512.png)
The Prompt Engineering Grift
![[redacted]'s avatar](https://substackcdn.com/image/fetch/$s_!lWn2!,w_108,h_108,c_fill,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F17d07f92-f7cb-4c47-8487-c12ee130769c_3419x3021.jpeg)
Prompt engineering has become a pretty hot topic.
This is largely due to the amount of FOMO you can create by shouting about it. Here’s a new tool! It’s going to be super important! Now here’s a 30-page guide of all the prompts you can use to unlock its true potential! Oh and if you at just $19.99, a like, and a retweet, you’ll get the full 270-page document! If you enter these commands into your trusty little ChatGPT, you will unlock its true powers!
Ugh. What a grift. I guess it comes with the territory of having a shiny new toy.
To be clear - prompt writing is important, since the model works by predicting the next word using everything that came before it. But much of the “prompt engineering” stuff out there misses the mark by acting as a recipe book. Instead of teaching people how to think about prompting from the bottom up, it feeds on people’s innate desire for a secret shortcut to make them believe there are secret phrases. As if saying “open sesame” will make the model do everything you want it to do.
Instead, there is a simple process you can follow to actually learn to use these tools better:
Remember that ChatGPT is a probabilistic text prediction tool. It may display a level of intelligence, but it doesn’t know how to think, it doesn’t “know” anything, and it doesn’t understand anything. It is merely predicting an output, based on the input given, and the patterns it learned from its training data.
Treat ChatGPT as you would treat a text-to-image tool like Midjourney. Nobody is prompting Midjourney with “act like you are the world’s best photographer.” Nor do people complain that it “lies” when you get images back that don’t exactly match what’s in your mind. However, since ChatGPT imitates a conversation, it can be easy to slip into the habit of treating it like a human.
Apply some basic problem solving skills. Put the “engineering” in prompt engineering.
What does this look like, in practice? Below are some basic pointers:
Provide clear and specific prompts. While GPT models are great at contextual understanding, they are not magicians.1 Generic prompts will get you generic results. Even simple requests such as “summarize” can be confusing, because the model doesn’t know what you might consider to be important. Be clear with your intentions and expectations on what output you’re looking for.
Learn and respect the model’s structural limits. Since it’s a generative model, pushing it past its limits will return useless or incorrect results. Using a language model as a calculator, or a fact finder is a great way to get it to hallucinate answers. Pushing past its context window will lead to lower consistency.
Don’t expect perfect answers. Since GPTs are probabilistic generative, getting an exact output may take longer than editing an answer that’s 80% there.
Break down complex or multi-step tasks. Remember that the model can’t think or plan ahead. Asking it to think through multiple steps will increase probability of hallucination.
Treat every error as a testing / learning opportunity. I’ve seen many friends abandon using ChatGPT / Bing search after the model spat out wrong or insufficient answers. “I’ll come back in 6 months when it’s all better”, they say. Except that majority of the time, the problem had to do with the user. Think through why you are not getting the results you want. Iterate your prompts. This is the fastest way to learn how to use these tools.
Remember that the model doesn’t “think.” It predicts. The quickest way to waste time and frustrate yourself is to ask the model to “think through” an error. The initial output was likely wrong due to a combination of prompt quality or structural limitation. Continuing to iterate on wrong answers beyond a simple “correct any errors” will likely lead to errors compounding on themselves.
At the end of the day, none of this is rocket science. Don’t let grifters FOMO you into memorizing bunch of half useful recipes.
This is one of the major problems with many “role play” prompts. The model lacks any in depth understanding of what “best” means in the context of any jobs.