How to Get the Best Out of an ML Chat System

Some recommendations to improve your interactions with an ML based system, such as a ML chat like ChatGPT https://chat.openai.com/chat:

1. Structure your requests
• Structure your requests in a way that is explicit and large in structure. This will help the ML algorithm "unroll" your request and provide a more structured and in-depth response. Avoid using too much implicit structure, as this may confuse the ML. 
2. Rephrase your request with previous reply content
•  Rephrase your request using words and themes from the previous replies, as the ML may provide additional content to explain its inability to respond. By using these words and themes, you can bring your request into a form that the ML can provide response content for.
3. Change the content order within your request
• Try changing the order of the structure of your requests, as the ML's approach to the response may be overly constrained by the input order. 
4. Use analogies to structure your request
• Use analogies to structure your request can also help strengthen the structure of the response. You can do this by asking the ML to find an analogy between the subject of your request and a reference that has a lot of structure already assimilated by the ML. The request might be phrased as: 
• How is unstructured-request related to structured-knowledge?
5. Anchor the request structure in known complementary relations
• A powerful way to structure a request is to provide complementary input. The idea is that you provide a well know implied relation, or you implicitly provide a well know relations such as that A is different to B, and you guide the ML's response by telling it to use the provided relation as structural base for the response. Additional complementary attributes can be added to shape the response.  The request might be phrased as: 
• How is structured-input related to unstructured-request?
• How is complementary-input-1 different from  complementary-input-2 in the context of unstructured-request?
• If structured-input is true, how is unstructured-request false?
• How does structured-input  imply unstructured-request?
• Debate unstructured-request as a dialog between reference-profession and complementary-profession?
6. Generalize, refine, strengthen, weaken your request
• Add extra words and context to generalize, refine, strengthen, or weaken your request to shape the response. This adds implied structure and helps guide the ML in its response. 
7. Use negations
• Use negations in your request to help broaden the scope of the ML's replies.
8. Change the point of reference of the request
• Change the point of reference of the request to change the basis on which the response is built. For example, request 'how does A become B' might be rewritten as 'how does B derive from A'. 
9. Limit implied structure within requests
• Limit implied structure within your requests, as the ML may struggle to infer this structure on its own and may make mistakes.
10. Reset the chat thread
• In a chat mode, previous requests may also over-constrain the ML's replies, so it may be best to reset the chat session before continuing. 
11. Request alternate reply forms
• Request the answer in 'reverse order', 'from other point of view' (see change of point of reference above), 'as a semantic graph', 'as code' (e.g. Python). Request reply as a 'rap song',  a 'fairy tale', a 'debate/dialog between ...'.
12. Request alternate views without key elements
• Rephrase the request it include a removal of a key element of the 'normal' request response. This will allow the system to present how that key element relates to the rest of the response. The request might be phrased as:
• What is input without key-element-from-previous-reply-to-input?
13. Limit lower bounds of request details
• Request 'over X reply-category'. For example, a semantic graph with over ten nodes and one hundred edges.
14. Use your imagination!
• Finally, remember to use your imagination and think creatively about how to qualify your requests to shape the response of the ML system. There are many ways to do this, and using a variety of these techniques can help improve your interactions with the ML system.

All original content copyright James Litsios, 2022.