How to get the AI to answer a particularly difficult question.

How to get the AI to answer a particularly difficult question. somebody Fri, 09/08/2023 - 09:02 pm

Remember the old TV series, "Star Trek"? Do you recall the episodes where Captain Kirk would outthink an android or computer by asking it a series of questions? Well, believe it or not, that is an effective strategy for dealing with today's Artificial Intelligence.

There are a few different processes that could be used to have a large language model answer a difficult question by leading up to it with a series of simpler questions:

Step-by-step questioning: This involves breaking down the difficult question into a logical series of simpler questions that build on each other and provide context, eventually leading up to the final difficult question. The model is essentially "walked through" the reasoning step-by-step via the questions.
Providing background information: You can provide the model with background information and facts relevant to the difficult question in a summarized format before asking the final question. This gives the model context to draw from when answering.
Query reformulation: With this approach, you start with a simplified version of the difficult question, get the answer, then iteratively modify and expand the question to get closer to the actual difficult question. The model leverages information obtained from previous iterations.
Explanatory questioning: Here you ask a series of questions that explain aspects of the concepts involved in the difficult question, elucidating the key factors and considerations needed to reason through and answer the final question.

So in summary, various tactics of questioning the model incrementally and providing contextual information can aid large language models in tackling difficult questions and reasoning challenges. The step-by-step approach leads the model through the logical reasoning chain to arrive at the answer.

Here is a real world example of that process. There is a legal document in the Real Estate Books AI library titled "Privacy Policy". Note how I was able to get the system to give me the information I was looking for from this policy:

Question: What is the explanation of your privacy policy?
Response: Apologies, but I cannot provide an explanation of the privacy policy based on the given document content. Please consider rephrasing your question or conducting a keyword search within the document itself.

Question: Can you summarize the privacy policy for me?
Response: I apologize, but the privacy policy document does not provide a concise summary within its content. I recommend reviewing the document carefully for a complete understanding or conducting a keyword search for specific information you are looking for.

Question: What information does the privacy policy document provide?
Response: The privacy policy document provides information on the collection of personal data, the use of cookies and similar technologies, the purposes for which personal data may be used, how personal data may be disclosed, the rights and choices of users regarding their personal data, information about third-party websites and services, data transfers, data retention, the age restriction for the services, and the contact information for any questions.

Question: What information does the privacy policy provide about the personal data collected?
Response: The privacy policy provides information about the personal data collected when interacting with the website, services, APIs, and other interfaces. The types of personal data collected include technical information (such as IP address, browser, and device information), information provided directly (such as email address and phone number), Prompts (text, documents, or materials submitted to the AI services), and Outputs (responses generated by the AI services based on the Prompts).

Now, this is the actual question I wanted answered:

Question: What happens to the personal data collected according to the privacy policy?
Response: The personal data collected according to the privacy policy is used for various purposes such as providing and improving services, creating user accounts, facilitating payments, communicating with users, enforcing acceptable use policies, and complying with legal obligations. The personal data is not shared or sold to third parties, except when legally mandated. Users also have rights to access, correct, delete, and port their personal data, but the prompts and outputs created by user interaction with the service may be retained for training AI models. The data is processed and stored in the United States, and the privacy policy may be updated as needed. For more information, users can contact info@RealEstateBooksAI.com.

Of course, the next question would be, "Why didn't you say that in the first place?"

But I doubt the AI would understand the question.