Prompting – how I do it

Mastering AI tools can be a challenging task. At the onset, I was under the impression that simply feeding the AI straightforward commands would yield desired results. However, what seems ‘simple’ to me may not necessarily translate as such for the AI, and vice versa.

Take, for instance, when I requested an AI tool to provide a rough draft for a blog post about vector databases. I posed this command: “Write a comprehensive article about vector databases and their use in NLM tools.” The output was not what I anticipated. The question then is, why?

Here’s the crux: Language Learning Models (LLMs), like ChatGPT, aren’t privy to my personal expectations, intentions, background, or assumptions. They operate by generating the most likely response based on the input or prompt they receive. But, after several trials, I discovered ways to make these LLMs work for me, often exceeding my expectations. Here’s my approach:

Understanding the Nature of LLMs

LLMs are not repositories of knowledge. Instead, they function more like reasoning engines. They’re designed to provide the most probable outcome, which may not always be the most factual one, considering potential inaccuracies in their training data.

Implementing a Bottom-Up Approach

To effectively use LLMs, I prefer a bottom-up approach. For example, when creating an article about vector databases, I first shared the latest documentation of a few databases with ChatGPT. I then posed several questions to define a viable approach for my blog post. Once I had a rough structure in my mind, I conveyed it to ChatGPT, establishing a context for the LLM.

Here’s an example of such approach: https://chat.openai.com/share/3db07491-95b0-4ba5-a4f4-8e9f1ee2044c

Seeking Specific Improvements

The resulting blog post from ChatGPT was closer to what I envisioned, but it still had deficiencies. Rather than asking ChatGPT to rewrite the entire article, I specifically asked it to revise certain paragraphs or create examples to include specific information.

Refining and Reviewing

Finally, I compiled all the paragraphs and examples into a singular blog post and requested a review from ChatGPT. The feedback included some remarks about clarity and a lack of definitions. By asking ChatGPT for improved versions of these problematic paragraphs, I ended up with a well-crafted blog post.

My Learnings

After several months of working with LLMs, I’ve gleaned a few insights:

• Creating context is crucial since LLMs, such as ChatGPT, can’t read your mind (yet).
• Implementing a bottom-up approach can be beneficial.
• Engage in active question-and-answer sessions with the AI.
• Request the AI to ask you questions, helping it draw out necessary context.
• LLMs excel at combining information from a variety of provided data sources.
• Be vigilant for hallucinations – answers may include made-up sources or facts.

Looking Ahead

AI tools are poised to revolutionize our internet usage, with more tasks likely to be accomplished through conversations. However, just like how the advent of the internet didn’t render libraries obsolete, LLMs won’t replace search engines.

We’re in the early stages of AI adoption, right after the ‘AI winter’. Therefore, any AI tool should be considered as being in beta release.

A paradigm shift is on the horizon, ushering in an era of personal AI assistants that can handle tasks for us, provide brainstorming support, and engage in creative reasoning. The possibilities are vast and exciting. So, let’s embrace the journey of AI evolution together!