AI Whispering 101.3

How are all my Cybersecurity Mavens doing out there? Hopefully all of your passwords are encrypted, and your encryption anything but MD5! Or SHA-1...lol

Today we continue with our AI prompt tips #14-#19 as we are still referencing the study that can be found here: (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). and while I am sure that we will have modifications to these rules as time goes by, these rules seem to be the best path forward to getting better results from the ANI LLMs out there! Also, please note that the first part of the recommendation in bold is copied directly from the referenced paper.

#14 Allow the model to elicit precise details and requirements from you by asking you questions until he has enough information to provide the needed output (for example, "from now on, I would like you to ask me a question to...") ...what? This one took me a while to get the gist of and is honestly one of my least favorite of these recommendations...its essentially asking the LLM to ask you questions to build the prompt.

Prompt example #1: I would like you to ask me a series of questions to help architect a house.

Prompt example #2: I would like you to ask me questions to help me design the perfect shirt.

#15 To inquire about a specific topic or idea or any information and you want to test your understanding, you can use the following phrase: “Teach me the [Any theorem/topic/rule name] and include a test at the end, but don’t give me the answers and then tell me if I got the answer right when I respond.”

This is rather straight forward...it essentially allows the LLM to be your study buddy and act like a well versed Flashcard partner.  I would just copy and paste the above prompt and fill in the data between the carats.

Prompt Example #1: Teach me the how to tune a starship to be able make the Kessel run in 12 parsec and include a test at the end, but don’t give me the answers and then tell me if I got the answer right when I respond.

#16 Assign a role to the large language models. This one intrigues me on a philosophical level in that...if you ask the LLM to be a teacher, it will add context thinking its a teacher and you are the student.  It will also change its language To address the intended audience better…

Prompt tip #1: Please respond as if you’re a fifth grade teacher.

This will respond in a language that’s more suitable for fifth grade students.

Prompt tip #2: Please respond as if you are a drill sergeant.

This will respond in a language that my father used my entire childhood… which is effective, but sometimes leads to having it psychiatrist on speed dial.

#17 Use Delimiters. Delimiters like “++” or “***” can be used to separate sections of prompts or to indicate the end of a response. This is another one that I like from a philosophical level… It seems just like with humans, Proper formatting and organizing your data can help with interpretation.

Having tested this and played with it a bit, you can use whatever type of delimiter you want really. if it makes better sense for you, then it will make better sense for the LLM.

#18 Repeat a specific word or phrase multiple times within a prompt. It seems that repeating words gives it more weight… so if you want the LLM to focus on a specific area or specific data set… Keep mentioning that area or data set.

Prompt tip #1: Please write a marketing blurb about our keto friendly cider, citing the low net carbs…keto friendly being that it is 2.7 net carbs per serving. it’s also organic and taste amazing, but I really want to focus on the keto, friendly aspect.

This prompt should write out a very keto friendly blurb for this product.

#19 Combine Chain-of-thought (CoT) with few-shot prompts. Not gonna lie, I had to look up some of these phrases myself. Let’s look at the definition for Chain-of-thought.

“Within the AI domain, chain of thought is characterized as the iterative and sequential process of cognitive ideation and reasoning, mimicking the human thought process.”

Lol… yeah, that didn’t help either…

So I spent a considerable amount of time on this particular one. And only after using AI for a long time do I think I understand it… it relies on the knowledge that AI every once in a while will give a bad answer… Correction, it’s not uncommon for AI to give incorrect answers. The difference in those two sentences is important.

Contrary to popular belief, the current gen of AI is far from infallible.  In fact, it fails much more than people realize...its just that those who don’t already know the data they are requesting don’t have the knowledge to know if it's true or not and accepts it as truth.

What this prompt tip is saying...if you teach the LLM your train of thought on how to get the answer then it can more easily answer with the more appropriate response.

For instance. There is a difference between:

The dog has 3 teeth cleaning bones.  His owner bought 2 more packages of teeth cleaning bones, each package containing 12 more teeth cleaning bones.  The dog was given 2 teeth cleaning bones during the weekend...how many bones do they have now?

Vs

The dog has 3 teeth cleaning bones.  His owner bought 2 more packages of teeth cleaning bones, each package containing 12 more teeth cleaning bones.  The dog was given 2 teeth cleaning bones during the weekend...given that the original amount of teeth cleaning bones was 3, and each bag holds 12 teeth cleaning bones. 2 new bags of teeth cleaning bones being 2x12 bones. 3 bones originally, 2x12 bones recently bought, and 2 eaten over the weekend...

how many bones do they have now?

Showing the logic, as simple as it may seem to you, will help the LLM provide the correct answer and not a hallucination or a false positive answer.

That’s all for today everyone, looking forward to the 4th and final installment of this guide soon...but until then, please make sure your servers have no red flashing red lights, your coffee be piping hot, and your patch rings report no errors! Until next time!

Interested in learning more? Contact Us!