Memory, Mind Maps, and Meta-Prompts

This latest in my "How I Use Foundation Models" is definitely my most personal, and tw: cancer.

My memory is not what it was. My attention is not what it was. I recently had a touch of the Lymphoma, and I must say that I don't recommend it. Chemotherapy was a rough battle and it left me with some long-term side-effects, one of which is called "chemo-brain". In my case it manifests as reduced memory and an ADHD-like lack of focus. I've experimented with many different methods of coping and I still haven't found the perfect set of tools, but I have found LLMs to be incredibly helpful in two key roles: First, helping me visualize what I need to think about, and second, helping me "remember" aspects of a subject I've forgotten.

I've found mind-mapping to be somewhat useful in the past, but never really gravitated to it. Still, it's a handy way to visualize and brainstorm about all aspects on a given subject - and more importantly get those all down on "paper" in a form that's easy to "grok" quickly. I theorized this might be exactly what my "new brain" needed, and I have found it much more useful now than I ever did in the past, however I still wound up spending a lot of time returning to these maps as my brain circled back around with "just one more thing." The clip below is a good talk on the benefits of mind-mapping:

One good thing about being a geek in this era of LLMs is you quickly realize that these things are phenomenally good at producing text, and as a geek you realize how many inherently visual tools we've turned into common text-based formats - mind-maps are no different. So I asked GPT - can it help me produce mind-maps in markmap format (markdown for mind-mapping)? Of course it could, was the reply. Now I can simply ask it for help in brainstorming a subject and it can produce the initial mind-map, and we can work together to refine it. How about a simple contrived example, where I ask it for help building an LLM curriculum:

As an AI language model, your primary task is to aid in the generation 
of a comprehensive multi-layer mind map, using the Markmap format. 
This mind map should encapsulate various aspects of 
'designing an upskilling curriculum for building production solutions 
using large language models', each thoughtfully considered from 
multiple viewpoints. Your output should be both structured and detailed, 
showing clear connections between these perspectives.

Start by defining 'designing an upskilling curriculum for building 
production solutions using large language models,' then branch out 
to diverse subtopics, related challenges, and potential solutions. 
Furthermore, delve into different perspectives on this topic, 
including those held by various stakeholder groups.

Keep the audience of this mind map in mind: 'technical managers'. 
The level of detail and complexity should be appropriate for this demographic.

You will structure the output as follows:

The main node is the central topic.
The first-level nodes are subtopics.
The second-level nodes are detailed aspects or 
 perspectives about each subtopic.
Add more levels as needed for additional details or viewpoints."
Remember to keep the Markmap format in mind, i.e., 
 using Markdown syntax to structure the information..        

(I know what you're thinking - whoa, that prompt is massive, I don't want to write all that! - well, keep reading). This produces an initial mind-map that looks like:

This is a good starting point, but as I said you can then go deeper and refine the map with the help of GPT. I asked it to "Expand on the "feedback and adjustment" section with some ideas on how to gather that feedback, and an appropriate cadence for making those adjustments" and it produced a new mind-map with those refinements:

You can keep refining this map both with your own ideas and with those of the model, and have the model critique your ideas as well, making the overall outcome much stronger and more useful for you. I've found this to be very helpful, and it's pointed out areas that I hadn't even considered. It's not perfect, but I am definitely not perfect myself, and I believe with its help I produce more robust ideas and a better overall result. One note: please as always be wary of putting sensitive corporate or personal data into any online LLM system unless you have guarantees they won't use it for training.

The second area in which I've found GPT to be my ally in tackling my newly neurodiverse state is with "remembering" for me. I can simply give it a list of what I'm thinking about on a given subject and ask it something as simple as "what am I forgetting?" For someone who is now easily distracted, this can be a great way for me to come back to a task I'd started, priming my brain to get back into "the zone" for a given area I'm tackling. Let's take a similar scenario to the one above:

You can then keep the conversation going - ask it to go deeper on Interpretability, ask it for additional items on the list, talk to it about the items you already have on there and if they can be made stronger. I've found this a shockingly effective way to get my brain back in gear once it's gone on an ADHD misadventure. And the benefit of multiple chat histories is that since my ADHD misadventures are often on other work that I also need to do, I can task-switch more effectively by just coming back to the thread and asking it "Remind me, what were we talking about?"

I've also found it useful for travel planning - got a packing list? Give it the location, timeframe, and time of year and it can tell you what you're missing. In fact my travel planning prompt is a bit extreme as well (contents changed for mildly humorous effect):

Imagine that I'm preparing for a trip to 'Thailand' that will last for 
'two weeks' Based on the destination's climate, cultural norms, 
and potential activities, along with the duration of the trip, 
you need to suggest essential items or activities I may have overlooked. 
To aid you in this, I'm providing you with a preliminary packing list 
and a list of planned activities. Please consider each item and activity, 
identify potential gaps, and suggest additions to ensure a comfortable 
and enjoyable trip

Packing list:
- 2 swim trunks
- parka
- pith helmet

List of planned activities:
- Pak Chong market
- Muak Lek waterfall

Your task is to evaluate the given lists, cross-reference them with 
common needs and considerations for a trip to 'Thailand' for 'two weeks' 
and suggest items and activities that seem to be missing, 
but could significantly enhance the trip's comfort, safety, or enjoyment. 
Also consider contingency items for unexpected situations, 
such as changes in weather or health emergencies.        

I won't reproduce the full response, but I did love GPT's answer for my packing list:

Why do I have these long prompts for such simple tasks? Could I get by with smaller, easier prompts? Well, this is where I tease my next article 😉. Those prompts were all created with Meta-Prompt Engineering. This is just a fancy way of saying that I asked GPT to help me build a prompt for GPT, but it's such an incredible power-tool that I'm planning on doing a whole post just on that. To give you a taste, my first mind-map prompt started as:

Your job is to think about questions, topics, or concepts 
from multiple perspectives, and help me visualize these 
using a multi-layer mind-map in the markmap format.         

As you can see that prompt is a lot easier to write. Anyway, more on that soon! I hope you're enjoying reading this series as much as I'm enjoying writing it, and I hope you're finding some of the tips useful. As always, I love hearing from you and would love to hear how others are using these new tools in creative and exciting ways. I'd also be interested to hear from other neurodiverse both on whether and how they are finding these tools useful, and what other tools and tricks they use to make them the most effective version of themselves. I'm new to this, and love to learn.