AI-Powered DIY Coaching: Getting Started (Part 1)
Introduction — A quick analogy of LLM AI
Have you ever read one of those “Create your own Adventure” books?
Goes something like this:
You have just time-travelled to the Jurassic era. You and your buddy, Bud, slip on your back packs and go exploring. You come across an abandoned dino nest on the ground. There are left over eggshells lying around.
Big eggshells. There is one set of massive footprints leading off in the direction of the distant river. There are also a bunch of similar but much smaller footprints going in the same direction. You are about to follow them to take some photos for your Instagram account, when you hear a noise from the nest.
Bud grabs a handy stick and pokes around moving eggshells. At the bottom of the nest you discover an unhatched egg. The egg moves. You hear tapping sounds. You hit ‘record’ on your camera. You capture the little guy’s hatching. He is so cute, despite the sharp teeth.
If you decide to adopt him and go home — turn to page 36.
If you decide that nature must take its course, and you will leave him there — turn to page 27.
If you decide to take him with you and see if you can reunite him with his Mom — Turn to page 43.
Now depending on which page you jump to, you get to decide your own adventure. Personally I’m rooting for page 43 — but hey — it’s your adventure…
Using a AI Large Language model (LLM) like ChatGPT, Gemini, Llama or Claude as your Coach has some parallels to one of these “Create your Own Adventure” (CYOA) storybooks.
The limitation the CYOA books is that there is a limit to the number of options at the end of each section (2 or 3 at most), and the ‘cheat’ is that the author creates a limited number of storylines and tries to create options that reuse sections of the story interchangeable no matter what story arc the reader picks. If they did not do this, the scale of the writing would become massive and unmanageable.
CYOA books are typically written for children who tend to be an unsophisticated audience and will accept the illusion. The complexity and scale to successfully create this for an adult work of fiction would be massive and unmanageable.
A very simplistic metaphor for LLMs is that they have a similar structure to a CYOA story. A CYOA story will have approximately 100 sections. Depending on the choice offered at the end of each section these will be stitched together in different ways to create different stories. We can regard these as the parameters that can be used to create each story.
LLMs end up with slightly more parameters. GPT-1 has 117 million. GPT-2 has 1.5 billion and GPT-3 has 175 billion. GPT-4 onwards is not disclosed.
A LLM with 100 parameters would have a very limited set of answers. If you can imagine a CYOA story with 175 billion sections you would have an adventure book you could read for the rest of your life, and then some.
LLMs are not dissimilar to this analogy. The other thing is that no-one believes that CYOA books are truly intelligent or self-aware. The only intelligence at play in the book is that of the author in creating a series of story sections that can be weaved into a bunch of different stories and create the illusion that the reader is creating their own adventure.
LLMs are also not intelligent and not self-aware, but the models and architecture that underlies these models convincingly create this illusion. Sometimes too well. You will see disclaimers on all LLMs about the fact that they make mistakes. Popularly this is referred to as Hallucinations.
This brings up an image of a LLM taking the day off to smoke some Magic Mushrooms. This is not accurate. All that is happening is that the LLM combined a bunch of parameters unsuccessfully and the output is factually incorrect. An example of this is where a lawyer used an LLM to research case law. The LLM created very believable but ultimately fictitious cases. You can read about it here: Lawyer Used ChatGPT In Court — And Cited Fake Cases.
Get Gio Pera’s stories in your inbox
Join Medium for free to get updates from this writer.
To assign human emotions or intent to the LLM would be incorrect. Going back to the CYOA example is that instead of going to page 43, you accidentally went to page 34. You end up confused. It’s sort of the same story, but does not quite fit. There are probably enough context clues to make you work out you made a mistake, but since we are still talking about our buddy Bud and we’re in the Jurassic era it might take reading a few pages before you realise your error.
Same problem for the lawyer. The case law it made up seemed correct — it fitted the context. That is the real problem with LLM hallucinations — they tend to fit the context and appear highly plausible.
Using an LLM as a Coach
Just because it’s called an AI (Artificial Intelligence), does not mean it’s intelligent. It’s an LLM (Large Language model). An incredibly sophisticated model that creates believable responses in your language.
Remember, that it can and will occasionally make mistakes. Always trust your gut. If the advice or suggestions the LLM makes don’t gel with you or make you feel uncomfortable, then you should listen to that.
Can I use an LLM as a Coach? Yes. Just be alert and vigilant.
Can you drive your car to work? Yes — but be responsible. Wear your seatbelt, use your indicators, keep your eyes on the road. Follow the rules of the road. Does that make driving completely safe? No — but there is a difference between reckless disregard and a cautious risk.
You need to apply the same principles when using an LLM as your Coach.
I would consider using an LLM as a coach to be similar to getting a self help book that offers advice, strategies and guidance on how to excel at something, grow in some way or try a different approach to specific challenges. The difference being that an LLM probably has access to numerous (millions or more) books, articles and text that cover the specific topics you want coaching on, and depending on the quality of the information, will present guidance and advice based on this. This is both a pro and a con.
In computing and data there is a well-known acronym known as GIGO (garbage in, garbage out). If you consider this as a cooking analogy — using rotten and poor ingredients to make a meal will result in a sub-standard meal no matter how good the recipe or chef.
If you present a chef with a pantry or larder full of ingredients, they will prod, sniff and tap all the ingredients and only pick the best quality ones to go into the pot. The ones that smell iffy, will get tossed.
LLMs don’t have this level of discernment. They ‘learn’ whatever is fed to them. Sometimes this can be quite humorous. Earlier this year, the X (Formerly Twitter) LLM, Grok created some humorous fake news about the sun’s odd behaviour that had experts baffled. What actually happened is that Grok has access to X feeds, and ingested a bunch of tweets by comedians and people who were making jokes about the eclipse. Since Grok has no concept of what is true or false, it happily input this data, and when asked about the eclipse created some odd fake news (Grok Fake News).
If you use a self-help book you will be getting the thoughts and concepts of a single individual who is qualified or an expert in their field or discipline. If you see a coach or specialist in any other discipline such as doctor, you will trust they are suitably qualified and able to dispense medication. With an LLM — the LLM will have absorbed ludicrous amounts of information and have a hodgepodge of information to draw upon. It has millions or billions of answers to draw on. Which is why you need to be discerning with the advice it dispenses.
If this is daunting or scary, think about the car analogy. Just like driving a car, you need to employ some strategies to ensure your safety. When the car was first invented and developed there were plenty of naysayers who spoke against this newfangled and unnatural mode of transportation. There were folk who told people to go back to using horses.
There were probably ‘experts’ who told everyone to go back to walking. For all we know if you go back in time far enough, there were probably experts who spoke out and discouraged walking on two legs, and recommended crawling on all fours.
As car technology improved, they have become safer and safer. Not perfect, but better and better. The other thing is that we have put measures in place to ensure that people are trained in using cars. Then age and license restrictions were put in place. This was not the case in the beginning.
We are at the beginning of the AI LLM revolution. It’s a bit like the Wild West. There are new models coming out all the time. Companies who produce them are still creating safety features — mostly after the fact. There are nascent legislative movements to try and regulate and create safety nets around LLMs, but it’s still early days.
I believe that LLMs can have many benefits and a tremendous positive impact on our lives. We just need to remember that this is a very sophisticated machine and needs to be used cautiously. I really like the convenience of being able to get into my car and drive wherever I want. Crawling….not so much
In the next post I’ll go through some prompts and use cases on how to use an LLM as a Coach.
If you want to experience coaching with a human coach — get in touch. www.demystify.co.za.
Images created with AI. Words my own.
