Prompt Engineering is a bug, not a feature

And it’s not engineering

AI will become more intelligent

When talking to children, it’s important to remember that their brains and intelligence are still developing. You need to choose your words carefully and use more examples and stories. Once children grow up and become educated, you no longer need to speak to them in a "childish" manner.

The so-called “Prompt Engineering” is similar to how we talk to children. We use various prompt techniques not because we should to, but because Large Language Models (LLMs) are not yet intelligent enough. Given the current pace of AI development, LLMs will soon "grow up," and we will be able to write prompts as if writing to an educated adult.

Trial and error, not engineering

LLMs are artificial neural networks with millions to trillions of parameters. While we may know which prompts yield good results, even top AI experts don’t know which specific parameters are responsible because they’re too many.

We’ve developed prompt techniques through trial and error rather than fully understanding how LLMs work and reason. Somehow, we know these techniques are effective and reproducible. This is similar to how herbs were used before modern medicine: we knew they worked but not why.

If trial and error is the method, computers can do it better and faster. Computers can generate and test many prompts quickly. We should let them handle it.

The real and old challenge

If you are not working on a project with a specific LLM and only want to learn about AI, don’t spend too much time on “Prompt Engineering”. AI will get more intelligent, and there’ll be many tools to combine all proven tactics and generate prompts for you.

However, the real and old challenge remains: even if we have a perfect prompting tool, there’s no good prompt if you don’t know what you want. Knowing what we want and communicating it clearly has always been a challenge, long before AI emerged.

Perhaps we should focus more on knowing ourselves and learning how to write well. Then we’ll not only get better results from LLMs but also from other humans.

Originally posted on LinkedIn: https://www.linkedin.com/posts/zhuoqun_prompt-engineering-is-a-bug-not-a-feature-activity-7220428288438009856-x7eX