I’ve been using a version of ChatGPT since 1999 while playing the video game Age of Empires. The game is made by Microsoft, which now owns 49% of ChatGPT and has played a hand in shaping the development of the model.
Playing with and against the AI in that game has given me a deep understanding of its strengths and weaknesses. And now it is all coming to fruition as Interact is very quickly implementing AI into our recommendation system builder to make it state-of-the-art and increase growth dramatically.
How ChatGPT works
ChatGPT is like a horse with blinders on that can only see 10 feet ahead.
A horse can run 55 miles per hour and is roughly 10 times stronger than a human. But with blinders on and the ability to only see 10 feet ahead, they are pretty useless without specific direction on what to do and which direction to run in.
(of course this could be cruel to the horses, this is just an example).
ChatGPT is a strong, fast machine that can’t see anything around it or ahead of it
That’s how it works. ChatGPT was trained by guessing what the next word was in a sentence, and then humans told ChatGPT whether it got the word correct or not.
ChatGPT is not looking at other contexts when making these guesses, and it cannot see past the next word. It is not a sentient being, and it cannot reference other content when it is making outputting text.
What ChatGPT is good and not good at
Some context on what the strengths and weaknesses of ChatGPT are.
ChatGPT excels at:
- Going faster than humans when given a set of steps
- Being reliably consistent
- Expanding on given points
ChatGPT cannot:
- Think multiple steps ahead
- Connect higher-order ideas together
- Automatically fact-check itself as it generates
How to use ChatGPT (with example)
Specific steps of how to use ChatGPT.
Given that ChatGPT cannot see ahead it will often fail when given a general prompt. To illustrate let’s bring it into the real world. If you were to tell ChatGPT to “build a house” that won’t work.
But if you give it a specific set of steps and run them one at a time, building on the previous step, ChatGPT will do a great job executing on each step. (again, it can’t see around or in front of it, so you need to break things down and feed it one chunk at a time).
So instead of:
“Build a house”
You say.
- Build a foundation for a two story house
- Build a frame for that foundation
- Add a roof to the house
- Add plumbing to the house.
Now sometimes you have these steps and ChatGPT will still make errors, not to worry that just means you have to get more specific with each step.
Expand on the ChatGPT prompt.
- Build a foundation for a two story house by first pouring concrete and then adding re-bar in a horizontal pattern with equal spacing before the concrete sets.
- Build a frame for that foundation by constructing 2×4 vertical beams all around the exterior of the house and for interior walls.
- Add a roof to the house by creating a frame for it that peaks in the middle and then laying roofing tiles on top of the frame.
- Add plumbing to the house by first adding sewer plumbing, then water, then HVAC
And you can continue getting more and more granular until the prompt gets to a point where ChatGPT is able to execute on each step properly. Depending on your use-case you may need to be very specific.
But again, since ChatGPT can just “go” once you set up these prompts it will save you immeasurable time once the prompt is ready.
Creating content with ChatGPT
Let’s say I want to write a plan for how to achieve FIRE (Financial independence, retire early) using ChatGPT.
If I prompt ChatGPT it does a pretty good job.
But the post is very generic and if I want it done in my specific format I need to be able to guide it just like the steps of building a house. So I have the following series of prompts.
- Write a plan for how to achieve FIRE
- Now expand on each step of the plan as if you were speaking to someone in their late 20’s who has $100,000 saved so far.
- Now re-format each step to include one paragraph intro followed by bullet points with specific steps to follow
- Now add a specific example to each step of the plan after the bullet points.
- Now add an intro that explains why someone would want to achieve FIRE and a conclusion encouraging people that they can achieve FIRE by following the plan.
- Now put it all together into one cohesive blog post.
Here’s what it give me now, formatted the way I want. I can go more specific with each step and ask ChatGPT to recommend images to include etc. you can see how it breaks down into steps.
Fine Tuning the Model (API Only)
Fine Tuning is where you teach the AI how to make foundations, roofs, etc. in your particular building style. If you just tell it to lay a foundation, it will use what it has learned from the internet and make a generic foundation.
But if you want it to make foundations in the way you make foundations you have to teach it how you make foundations. This is done by uploading a minimum of 100 examples showing how you make foundations and creating a fine tuning model for foundations.
Then every time you prompt the AI to make a foundation it will pull from what you’ve taught it about how to make a foundation and follow your style.
Self Learning AI models (API Only)
To make your Fine Tuning model a self-learning model, you will need a way for users (running with the analogy here, your home buyers) to indicate whether the houses built with the AI are satisfactory to them or not, and why.
Then you can add the satisfactory results back into the Fine Tuning model and update it. In this way your customers begin teaching the AI what they like, and as tastes change so does the AI, it’s like evolution, survival of the fittest.
This is a ChatGPT zoomed out guide
I’m not doing a deep dive into any specific area, but rather trying to paint a story of how ChatGPT works and how you can actually begin using it. The over-arching structure of the thing will grow and evolve but these principles have been the underlying framework for 20 years and you can bank on some version of how it works to exist for a while.