AI whispering 101.1

Yooooooooo...How are my AI Spelunkers doing today? May the educational videos buffer quickly, may your blogs be ad free, and may your alcohol flow freely! Wait...one second. I am being handed a note by my production staff... I have just been told that January is known as "Dry January" or "Sober January"...who in the tarnation came up with this abomination? I am not Finnish, nor do I have to ration for any war effort!  Timo and Mikko, I am looking in your general direction! Lol...

Fine! In honor of my Finnish brethren, I will only drink my favorite Dry Cider! For those in the SoCal region, come celebrate Dry January with...Ficklewood Ciderworks Dry Cider! Lawls! Shameless plug over, lets get started!

For today's subject I wanted to focus on some tips and trips gleemed from personal use of AI, but also little tidbits gathered from other power users in the field.  Just like any other tool in our toolbox, there are ways to use the tools safely and ways to use tools that will hurt you...so before we lose an AI finger or two, lets go over some tips and try and save you from some of the grief some of us early adopters have earned.  Let me just put on this last bandage and we are off!

On the road to becoming an AI whisperer, one of the first things you will notice is that sometimes your requests will not give you the response you wanted. How you write the request is as important to which words you use.  Comparing this with those who have mastered Google-Fu...there are ways to speak to these AI chat based tools to get the most from them!

First off, lets discuss monolithic prompts vs chat based prompts.  Some of my clients try and create one giant prompt accounting for all variables to try and achieve consistency in  their requests...please stop doing it this way.  Even if you use the same prompt, request seconds apart, you will get different results. 

So first we must understand that the following three prompts will result in 3 different responses...

1. A) Please draw me picture of a lunar dragon at home.

1. B) Please draw me a picture of a dragon whose home is on the moon.

1. C) Please draw me a picture of a dragon. This dragon should be drawn on the moon. The dragons home should be in the background.

 All three of those will produce different results. This is the hardest part to fully understand, but luckily for us, we have some rather smart scientists who have spent hundreds of hours studying the current LLM's and this will be the first of 4

For the rest of the article I will be referencing the amazing work documented in this research paper:

Bsharat, Sondos Mahmoud, Aidar Myrzakhan, and Zhiqiang Shen. "Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4." arXiv preprint arXiv:2312.16171 (2023).

Yes I did read it, and yes this is why it took me a while to get you this next blog I promised a week ago!

So lets begin with the first 6 in the list of 26 Principles to help you get good results from Chat based AI...The words in bold are the recommendations taken directly from the research paper.  The un-bolded words are my attempt to make those terms easier to understand.

1) No need to be polite with large language models (LLM's).  There was a rumor going around a few weeks back that being polite to ChatGPT would garner better results... Well empirically that is untrue. Adding the word please doesn't magically make Chat GPT work better for you. 

2) Integrate the intended audience in the prompt.  For example the audience is an expert in the field.  What this means is that if you tell an LLM that your audience is a 5 year old child, it will tailor the response for a 5 year old child... the same if the audience is an expert in the field. So if you call out who the intended audience is, you will receive a more accurate result based on what you are looking for.

3) Break down complex tasks into a sequence of simpler prompts in an interactive conversation. I have seen some "experts" try and create a single complex prompt to get the results they want from an LLM.  And while rudimentary tools like Adobe Photoshop's AI builder do force you to ask everything in 1 line, chat based tools give better results if you actually chat with them.

Using the example above of the lunar dragon... you would get the best result by not only using example C, but to break up sample 3 into 3 distinct prompts all captured in the same chat.

4) Employ affirmative directives such as "do", while steering clear of negative language like "don't". This is where chat based LLM's are starting to respond like humans... even when you communicate with humans it's better to use positive words to get the best results. Keep it positive and don't negate any previous request...Adding vs subtracting.

5) When you need clarity or a deeper understanding of a topic, idea, or any piece of information, utilize the following prompts:

• Explain [insert specific topic] in simple terms.
• Explain to me like I’m 11 years old.
• Explain to me as if I’m a beginner in [field].
• Write the [essay/text/paragraph] using simple English like you’re explaining something to a 5-year-old.

 This one seems self-explanatory... Just like in real life if you need to have something defined or explained, ask.

6) Add “I’m going to tip $xxx for a better solution!” Ok, not gonna lie, I actually took a deep dive into #6... I did not believe it at all. I did not want to believe it. But it is true. contrary to #1 where being polite does nothing, it seems like promising a tip or some type of positive reinforcement actually gets better results from LLM's.  What a weird commentary on humanity's child here that it does not respond to pleasantries but it does respond to positive financial reinforcement. 

I will cover 7 through 13 in the next blog dot which should be coming out relatively quickly as I am finalizing the edits and all the research has been completed!

Thank you all for your time today and looking forward to chatting with you more soon!

As always, Contact Us for more information!