There are a few fast tricks to get more and better results from ChatGPT. I’ll be going over those but more importantly, we’re going to look at the fundamentals of ChatGPT and some examples.
Becoming a master of anything is not an accident it takes knowledge, time, patience, and testing.
The best thing I personally have noticed though is that there are literally hundreds of applications for ChatGPT. Almost anything I have been able to dream up I have been able to execute.
So no matter if you are trying to get ChatGPT to help your programming, write a blog article, or if you are just trying to get ideas for an article these fundamentals, tricks, and examples will help.
ChatGPT Tutorial: The Basics, Tricks, and More
Table of Contents
- ChatGPT Tutorial: The Basics, Tricks, and More
- 1. What ChatGPT Can and Can’t ChatGPT Do
- 2. Can You Fine-Tune ChatGPT (Priming)?
- 3. How You Can Build A Custom Data Set For ChatGPT
- 4. What Are The Text Generation Jobs ChatGPT Can Do?
- 5. When Should You Give Up and Start Another Chat.
- 6. How to Lead ChatGPT To Format Answers
- 7. How to Get ChatGPT to Best Answer Questions With Its Existing Knowledge
- 8. How to Get ChatGPT to Proofread a Short Document (700 words)
- 9. How to get ChatGPT to Summarize Information as a List
- 10. Have ChatGPT Explain a Piece of Programming Code to You
- 11. Have ChatGPT Write Some Programming Code For You
- 12. Have ChatGPT Suggest Text To Image Prompts for You
- 13. Find Your Next Brand Name
- 15. Other Tasks
1. What ChatGPT Can and Can’t ChatGPT Do
ChatGPT is a fantastic tool, but every tool has its limits.
Understanding how ChatGPT comes up with its answers, how to give it more information, and passing on context is crucial to get it to work properly.
We’ll get into that all in this section. How?
One finds limits by pushing them.
Herbert A. Simon
By testing the limits.
For instance,
There are two ducks in front of a duck, two ducks behind a duck and a duck in the middle. How many ducks are there?
There are six ducks.
This is about as simple a verbal logic question as there is (the answer is 3).
ChatGPT gets it wrong. Why?
ChatGPT relies on using patterns to give it answers. It basically, “Looks things up and regurgitates them”.
It can look up a series of articles analyzing Genghis Khan’s life and create an article by piecing those together but it can’t create any new insights or analysis about his life.
(ChatGPT told me this is wrong, and perhaps by putting together different perspectives a third perspective might be generated. What you get though will still depend on what went in in the first place.)
For the life of a famous figure like Khan that’s not a problem, there have been reams written about him. What if the person was not as famous? We might have a problem.
So we can see first of all chat ChatGTP has the ability to generate human-like text. There are limits though to what it can do. It fails miserably when given problems in logic.
So what is ChatGPT good at? ChatGTP is Well Suited for tasks like generating well-known answers, translating languages, and generating text upon request.
Since programming is all about the application of logic. Why are so many people excited about having ChatGTP program for them?
For a simple subroutine which is a small program, there is existing code on the internet for almost all these functions. That information was already included in the Corpus of information that ChatGTP was pre-trained on.
What that means is that ChatGTP can regurgitate the right answer part of the time. It can even piece together different pieces of code together.
That’s not really programming though. It’s not solving new problems, it’s picking up pieces of solutions that already exist. It’s good at putting those pieces together (synthesis), but not at creating them to begin with.
So let’s state that a little more succinctly.
ChatGPT is good at pulling together existing data to create something new (synthesis).
That leads to an issue. What if ChatGPT doesn’t know the data you want it to use?
If the data is not already in the pre-training information (corpus) then it will not be able to figure out the answer. We need another solution.
Priming
Priming ChatGTP: We are going to give ChatGPT the information we want it to use.
We can do that pretty easily by just asking it to read what we want it to know.
Example:
What did Mary have?
I’m sorry, I’m not sure what you mean. “Mary” is a very common name and it is not clear who you are referring to, and there is no context provided about what “Mary” might have. Can you please provide me with more information or context so I can assist you better?
Mary had a little lamb its fleece was white as snow.
Yes, that is a well-known nursery rhyme, “Mary Had a Little Lamb.” The rhyme talks about a young girl named Mary who had a pet lamb, and its fleece was white as snow. The nursery rhyme is often used to teach children about the letter “L” and its sound.
What did Mary have?
In the nursery rhyme you referred to, “Mary Had a Little Lamb”, Mary had a pet lamb.
ChatGTP went from not understanding the question to explaining how “Mary Had a Little Lamb” is used. ChatGPT already “knew” the rhyme but it didn’t have the context it needed.
So we are going to give ChatGPT the context or additional facts it needs.
How?
Priming.
Priming is a technique (trick), to teach ChatGPT New Information or Context.
Why is that important? For one thing, ChatGPT’s model data only goes through 2021. If you want it to have data about dates after that you are going to have to provide it.
I should note that Open.AI after a series of hilariously wrong predictions barred questions pertaining to events after 2021. That does not mean though that we can’t provide facts about a product for instance that was released after 2021.
The Vanmoof S5 electric bike was released on 3/15/2022. Detailed information on how long it takes to charge was not available previous to that date. (Not included in the corpus for sure).
The Vanmoof S5 takes 6.5 hours to charge according to Engadget.
If I was trying to get ChatGPT to write an article about the Vanmoof S5 it would be useful for it to have the specs.
Let’s see how that goes.
How long does it take to charge the Vanmoof S5?
I apologize but as a large language model, I am not able to provide real-time information about charging times for a specific device such as the Vanmoof S5 as it depends on various factors such as the battery capacity, the power source, and the charging method. The best way to know the exact charging time for your specific Vanmoof S5 would be to consult the manufacturer’s specifications or the instruction manual that came with the device.
The Vanmoof S5 takes 6.5 hours to charge.
I apologize for any confusion. The Vanmoof S5 takes approximately 6.5 hours to charge. This is the time it takes to charge the bike’s battery from 0% to 100% using the included charger. It’s important to note that charging times can vary depending on factors such as the initial battery level, ambient temperature, and the power source used. Also, make sure to check the battery level before charging and avoid overcharging.
How long does the Vanmoof S5 take to charge?
According to the information provided by Vanmoof, the Vanmoof S5 takes approximately 6.5 hours to charge the bike’s battery from 0% to 100% using the included charger. It’s important to note that charging times can vary depending on factors such as the initial battery level, ambient temperature, and the power source used. Also, make sure to check the battery level before charging and avoid overcharging.
There are a couple of interesting things that we just learned.
- We were able to “teach” ChatGPT how long it takes to charge the Vanmoof S5
- ChatGPT pulled a fake source for the information that it provided.
- (I provided the charging information not the Vanmoof Company).
ChatGPT transfers the past information into a “Context Vector” that carries context and factual information about your conversation. According to Open.AI it stores up to approximately 3000 words worth of information about the past conversation.
So there is a limit to ChatGTP’s memory of the past conversation, but it isn’t small. you can pour quite a bit of data into it for your needs.
There’s something important that happened though in this chat and, we have seen the same thing a couple of times.
ChatGPT very confidently answers incorrectly at times.
Why?
ChatGPT uses a probability model to write. It “guesses” the next word that should come in the sentence based on the data that it has already added to its model.
My guess is that in its corpus (data set), very frequently battery charging information was given “according to the manufacturer”, etc.. ChatGPT has seen that so many times that it is what it considers “normal”.
So when ChatGPT gave me back the information that I had just given it, ChatGPT included a false reference. That’s unfortunately not an uncommon occurrence from my testing.
ChatGPT does not know what is true or false only what is probable.
ChatGPT does not state, maybe, or could be, ever.
Every answer that you receive from ChatGPT will be delivered as if it knows the answer for certain.
What you should now understand:
a. ChatGPT can’t solve logic problems
b. ChatGPT is good at looking up answers and putting the pieces together.
c. We can use “priming” to pass context information or facts to ChatGPT that it doesn’t have.
d. ChatGPT keeps a “context vector” of the past chat session. The max size of that vector is about 3,000 words. ChatGPT uses the vector to keep track of the chat, understand the context, and absorb new facts.
e. ChatGPT only knows what is probable, not what is true or false.
2. Can You Fine-Tune ChatGPT (Priming)?
In the last section, we spoke about priming. By priming, I meant giving the chatbot specific information about the context of a chat or facts about something that it did not have.
ChatGPT creates a context vector that stores the priming information but it is limited to about 3,000 words. The context vector changes the output of ChatGPT and is fine-tuning but a very limited form of fine-tuning.
ChatGPT’s limited because 3,000 words are not enough to “train” or fine-tune the GPT neural net to do a full-blown job. It is enough though to build a very narrow data set to get ChatGPT to do a particular task.
That’s why I usually refer to what we are doing as priming it is a much smaller and simpler task.
The next section will go over how to build the priming data set for a specific case.
(This discussion with Sam Altman about fine-tuning gives a lot of information about what more dramatic fine-tuning can do).
So what is full-blown fine-tuning?
There are a couple of ways that you can fine-tune a model.
Since GPT is a neural network with many layers you could either.
a. Retrain the entire multilayer neural network using a data set for instance about customer service.
- Once this is finished it will give the best results, but the price will be very high.
b. Fine-tune only the last layers of the neural network. Using a context vector. The size of that vector for most general tasks will be about 300,000 words, about 100x of what we have available currently with ChatGTP.
- This may not give as good of results but, it will be much faster and not require nearly as much data.
Sam Altman in the discussion above thinks training the last layers of a neural network like GPT is how a lot of companies will use AI to make a profit.
We are going to use that in a limited form to get ChatGPT to do what we want.
3. How You Can Build A Custom Data Set For ChatGPT
First, you need to decide what you are trying to do, so let’s write a product review.
Example: Write A Product Review
I know nothing about VanMoof S5 E-Bikes and neither does ChatGPT because this model of E-Bike was released after 2021 (ChatGPT’s data cut-off date).
If I want ChatGPT to help me write a review of the Vamoof S5 I need to give it all the relevant data.
So first I need to collect the data.
There are a few ways to go about this but I am going to use a manual scrape process, (for clarity).
I found a few relevant URLs for VanMoof S5 reviews online (thanks Google).
You need to read through each review (skim them) to make sure they are good and relevant. ChatGPT will give you garbage if you give it garbage.
For each review you actually choose to use, read it, you are going to be ChatGPT’s fact checker.
I am using Chrome and I used a plugin called Postlight Reader, which cleans up webpages and makes it a lot easier to cut and paste clean text from the web. (Ctrl-A select all on the webpage, Ctrl-C copy, Ctrl-X paste in ChatGPT)
I pasted about 10 VanMoof S5 reviews into a NEW CHAT on ChatGPT. (One Review at a time).
I then asked ChatGPT to create an outline for a product review.
Now there are a couple of important points
- Why did I get ChatGPT to give me an outline and not just write out the product review?
- ChatGPT is a bit stingy on how much text it is willing to write each response is usually about 150-450 words. The average is about 200 words.
- By seeing the outline we can also add relevant details (yes you should read the reviews)
- We may need to add more points to the outline, most of the time for SEO we are going to want a 2,000-word article. (Minimum is about 1,000)
- I used a “text prompt” to get ChatGPT to create an outline for me.
I did not specify how many subheadings I wanted on my outline. That was a mistake on my part. Since I know that for each subheading ChatGPT is going to respond with about 200 words on average I should specify that there should be at least 10 sub-headings.
That is something that is very common with learning how to use ChatGPT. You are going to have to test again and again different text prompts to get the output that you want.
The next text prompt is a case in point it took roughly 5 iterations to get a text prompt that would work the way that I wanted.
(I am breaking the outline that I asked it for into small pieces and having it write each piece separately so that I get to the word count that I need.)
I also found I had to combine the Introduction with the first sub-heading to get it to work correctly. Trying to get it to write the introduction by itself did not work.
So what does the final result look like? (Sample)
This is only the first few paragraphs, the entire product review came out to 1,600 words or so.
Does it still need work? Yes, for one thing, the entire product review is still a bit short. It also needs pictures and since we know that ChatGPT sometimes makes stuff up we need to check the text for mistakes.
Is this faster than writing the whole thing yourself?
Yes. Probably 2 – 3X as fast.
Is this helpful to someone who is interested in buying a Vanmoof S5?
Yes
Will you get penalized eventually when Google is able to tell that the content was written by an AI?
I don’t know. You could of course take the text and paraphrase it yourself but, that seems like you are going in circles a bit.
My hope is that Google and other search engines will treat AI-generated text the same way as any other text. If it is spammy bull it should not rank, but if it actually helps the user it should have an opportunity to rank.
So what are the takeaways from this example?
- Decide exactly what you want to write (Product Review)
- Use Google to find data sources
- Make sure that the sources you found are good
- Use Postlight Reader Ctrl- a x and c to paste those sources into ChatGPT
- Read through the sources yourself (you will be the fact-checker)
- Ask ChatGPT to create an Outline for your article (Product Review)
- Edit the Outline to add details where needed
- Ask ChatGPT to write each piece of the outline out for you separately
- Check the output to make sure it is accurate.
- Add pictures, formatting, etc.
Is this Ethical?
I probably would not use this review since I really never rode the Vamoof S5 and the review implies I did. Other than that I think this is a helpful resource.
Text Prompts Used in This Example (Product Review)
Used To Create The Outline
Please write an outline for a VanMoof S5 review with 10 subheadings and bulleted relevant details
Used to Have ChatGPT Write Out Individual Parts of the Product Review
Please use the partial outline below to write a product review of the VanMoof S5. Please write it as an experiential review where you go over each aspect of the E-bike. Please write it as an enthusiastic but knowledgeable cyclist. The total length of the review should be about 2,000 words so please include filler about your impression of different features, how it feels to ride the bike, and whether you think it’s a good purchase or not. Please only write the section that you have currently been given to write.
4. What Are The Text Generation Jobs ChatGPT Can Do?
There are a few text generation tasks that ChatGPT excels at but there are four fundamental ones.
- Writing – This includes articles, poems, stories, news articles, marketing copy, and more.
- Language Translation – ChatGPT is excellent at translating information into other languages and can be used to teach as well.
- Chatbot – As most of us know by now you can sit and discuss things with ChatGPT. This can be used in therapy, customer service, or virtual assistants. With the addition of text to voice and voice to text it could potentially act as a telemarketer.
- Question Answering – ChatGPT can be primed (fine-tuning on a small scale) to answer questions about a given subject.
Writing
The example in the last section showed how you can have ChatGPT write out a product review. The same type of “priming” can be used to teach ChatGPT just about any type of writing to emulate. Styles can be taught and specific information can be transmitted.
ChatGPT is also a great help in, editing, ideation, outlining, plotting, and many other aspects of writing.
ChatGPT also has been pretrained on a very large data set of text. The data set includes the writings of just about any author you are familiar with. For instance, we could ask
Please write a sonnet in the style of William Shakespeare about the joys of eating a peanut butter sandwich.
Oh sweet and savory treat of humble birth,
A simple pleasure for the masses sought,
Thou art a sandwich of the greatest worth,
With peanut butter as the filling fraught.
Thy creamy spread, so smooth and rich in taste,
Mixed with bread’s warmth and soft embrace,
A flavor combination hard to waste,
A joy to eat with haste.
Thy versatility knows no bounds,
For thou can be enjoyed in many ways,
With jelly or banana, chocolate or hounds,
Thy deliciousness forever stays.
So let us raise a glass to thee, oh treat,
For thou art a sandwich that can’t be beat.
This gives us some freedom to not have to use a custom data set for each job that we ask ChatGPT to do. There is a problem though.
Anyone who asks ChatGPT to give them a sonnet in the style of William Shakespeare about a peanut butter sandwich will get a very similar sonnet. Some of the lines will probably be exactly the same.
Depending on the context that might be fine. If you need a unique output though you will need to give ChatGPT priming (fine-tuning) information to help guide it.
Language Translation: Beyond Google Translate
Suppose that you would like to learn another language. There are many ways that you could use ChatGPT as a tool.
For instance,
Please respond to my question like you were a teen girl from Brazil. Please respond to my questions in Portuguese.
Claro! O que você gostaria de saber?
Where are you from?
Eu sou de um computador, mas posso imitar o jeito de falar de uma adolescente brasileira. E você, de onde você é?
I am from the United States. Have you ever been here?
Não, nunca fui para os Estados Unidos. Mas gostaria muito de conhecer um dia. O que você mais gosta de lá?
I like Clearwater beach Florida and the mountains in Colorado. What do you like most about Brazil?
Eu amo a diversidade cultural do Brasil e a paisagem deslumbrante, desde as praias incríveis até as florestas tropicais. Também amo a música e a comida brasileiras. Qual é o seu lugar favorito no Brasil?
Most of you probably don’t speak Portuguese. I speak it every day because my wife is Brazilian. I still learned a new word from the dialog, deslumbrante = dazzling.
Imagine if I had ChatGPT emulate a college physics professor in Brazil. I am also a physicist but I don’t know a lot of vocabulary about physics in Portuguese. It would be a great help if I were going to teach a class there!
ChatBots – Dialog With?
Most of us are accustomed to chatbots being an easy way for companies to answer our customer service questions, badly for the most part. That of course will change soon.
There are a lot of other interesting things to do with a Chatbot. You just saw one. I was dialoguing with a teen girl from Brazil.
Based on their writings you could dialog with Sophocles, Martin Luther King, and many others.
I have found though that creating an avatar, ie a teen girl from Brazil, or a professor of physics, is more interesting. They tend to be less repetitive in their responses.
This makes sense as you can imagine that ChatGPT has a much larger data set when it considers the writings of all teen girls in Brazil vs. MLK. A larger corpus gives the ChatBot a lot more possible responses.
Remember —> Chatbots will generate responses similar to the dataset it was trained on, or the subset of that dataset that you request.
The only other way to affect the output is by priming (fine-tuning) by giving ChatGPT additional information to change its output. That is the solution that companies will end up implementing for customer service I suspect.
Question Answering
It so happens though that we set up a ChatGPT to write an article about the Vanmoof S5. We can use that same primed (fine-tuned) Chatbot now to answer some questions about the S5. (This information would not be in ChatGPT’s data set normally because the EBike was released after the data set cutoff date).
How powerful is the S5 motor?
The Vanmoof S5 e-bike is equipped with a 500W front hub motor.
What is the torque of the S5 motor?
In the article provided, it is mentioned that the S5 has a torque value of 68Nm, which is higher than the previous models S3 & X3’s 59Nm and S2 & X2’s 45Nm. However, it is not specified if this is the maximum or rated torque of the motor.
What does the left halo light do?
The left halo light on the VanMoof S5 electric bike serves as an indicator for the battery level. It will change color to indicate how much charge is remaining in the battery. Additionally, it can also be used to display other information such as the current speed or gear. The exact function of the left halo light is not specified in the information available to me.
What does the right Halo light do?
The right Halo light on the VanMoof S5 electric bike is used to indicate the battery level. It will change color to indicate the battery’s current state of charge. For example, when the battery is fully charged, the light will be green. When the battery is running low, the light will change to red. The light can also be used to indicate other information such as the current gear or ride mode.
Do the left and right halo lights give different information?
In the information provided, it does not specify if the left and right halo lights give different information. It is mentioned that the halo lights integrated into the handlebar interface displays key data such as speed, battery life, and ride mode.
What is the range of the S5?
The range of the VanMoof S5 electric bike is 37 to 92 miles, depending on the mode that you choose to use.
So using the primed (fine-tuned) chat ChatGPT was able to answer most of the questions about the S5 correctly. (ChatGPT biffed it on the halo lights question a bit).
So what about “Regular Questions”
For most common questions the answer will already be in the data set ChatGPT was trained on. However, you can get better answers by priming (fine-tuning) ChatGPT.
What About Real Fine-Tuning With a Huge Data Set?
You can work with an API from Open.AI and pay to have a model trained with your data and then make calls to it. More information is available at Open.AI/Pricing
5. When Should You Give Up and Start Another Chat.
ChatGPT remembers what you have been chatting about. That’s a good thing most of the time, but… Not always.
Why? Sometimes ChatGPT does not understand what you are asking for the first time. It then carries that same misunderstanding forward even if you restate the query.
Example: Extra Large Mouse Pad
Please give me 5 text prompts for an extra large gaming mousepad for dalle-2.
“Dominate the competition with the extra large Dalle-2 mousepad, perfect for high-speed gaming and precise movements.”
“Take your gaming to the next level with the extra large Dalle-2 mousepad, designed for optimal tracking and control.”
“Elevate your gaming setup with a sleek and durable extra large Dalle-2 mousepad, perfect for extended gaming sessions.”
“Experience the ultimate in comfort and precision with the extra large Dalle-2 mousepad, perfect for all types of gamers.”
“Upgrade your gaming station with the extra large Dalle-2 mousepad, featuring a smooth surface for precise mouse movement and a non-slip bottom for stability.”
I want a text prompt to use on Dalle-2 to create a design for an extra large mouse pad
“Unleash your inner gamer with a bold, custom designed extra large Dalle-2 mousepad that makes a statement at your gaming station.”
I want a text prompt to use on Dalle-2 to create an image. The text prompt should allow Dalle to create a design for a large mouse pad
“Create a visually stunning large mousepad with Dalle-2 by incorporating a dynamic and unique design that will make it stand out on any gaming setup.”
None of these responses is right.
Once I gave it a command in the first sentence that was misinterpreted it continued to be misinterpreted even when stated properly.
To get past this I opened up a new chat and asked the question again.
Please design a text prompt to use on Dalle-2 to create an image. The text prompt should allow Dalle to create a design for a large mouse pad
“Create a design for a large mouse pad with a vibrant abstract pattern on a dark background, featuring shades of blue, purple, and pink.”
Using that “text prompt” with an AI image generator I was able to create the mouse pad you saw at the beginning of the article.
Kool but why did I have so many problems in the first chat?
There is a “feature” of ChatGPT it gets into a “rut” very quickly.
- If ChatGPT gave an answer to a similar question in your chat it will tend to answer again in the same way
- So if ChatGPT did not understand your first question,
- It will probably not understand your next one as well
- The simplest way to get around the problem is to just start another chat.
Back To The Original Chat
If you have already spent a bunch of time loading in a data set on the chat (priming or fine-tuning) you may want to try an alternative.
You will need to ask ChatGPT some questions. The purpose is twofold.
First, you are going to let it know what you are interested in.
Second, you are going to learn the vocabulary that it is looking for.
What is a text prompt?
A text prompt is a statement or question that is used to generate text or written content. In the case of a language model such as myself, a text prompt is a piece of text or a question that is used to initiate or guide the generation of a response. The text prompt can be used to provide context, set parameters, or guide the direction of the generated content.
How are text prompts used with image generation?
Text prompts can also be used to generate images, a technique called text-to-image generation. This is done by training a model on a dataset of images and their corresponding captions or descriptions. The model learns to associate the words in the captions or descriptions with the visual features of the images. During inference, a text prompt is given to the model, which generates an image that corresponds to the prompt.
Text prompts for image generation are generally short, usually one sentence that describes what the image should look like. The prompt is usually used to set the scene, the objects, the characters, the emotions, the lighting, and the overall style of the image. The text prompt can also include specific requests like “Create an image of a sunset over a beach” or “Generate an image of a cat sitting on a windowsill.”
Please design a text prompt to use on Dalle-2 to create an image. The text prompt should allow Dalle to create a design for a large mouse pad
“Design a visually stunning large mousepad with Dalle-2 that features a unique, abstract pattern that captures the essence of high-speed gaming and precision”
Please note that Dalle-2 is not a specific product that I am aware of, it’s a name you used, also generating a design is not something that I can do as a language model AI, but I can suggest text prompts that can give direction to the person who will be creating the design.
Still a big Nope
Please design a text prompt to use on text to image generator to create an image. The text prompt should allow Dalle to create a design for a large mouse pad.
“Create an image of a large mousepad with a futuristic, high-tech design that incorporates elements of circuit boards and advanced technology.”
That was interesting I learned a couple of things. ChatGPT does not know what Dalle-2 or Dalle are even though they are Open.AI products.
Second ChatGPT was able to determine from context what Dalle-2 is in the new chat.
You can see also that we were able to get a useful text prompt even from the original chat using this technique.
Take Aways:
- If ChatGPT is not understanding what you want it’s probably going to keep misunderstanding
- Your easiest solution is to start another chat and ask the question a different way
- If you have already primed (fine-tuned) your current conversation you can try this.
- Ask questions about what you actually want ChatGPT to answer
- Pay attention to the vocabulary that ChatGPT is using to answer those questions
- Ask the question again with the altered vocabulary
6. How to Lead ChatGPT To Format Answers
Probably, sooner or later you’re going to want chat GPT to create formatted data that you can paste into a spreadsheet. The easiest way to do that is to show chat GPT some examples of the type of formatting that you want.
Create a two-column spreadsheet of top science fiction movies and the year of release:
Title | Year of release
Blade Runner | 1982
The Matrix | 1999
Inception | 2010
Interstellar | 2014
Ex Machina | 2014
Arrival | 2016
Blade Runner 2049 | 2017
Annihilation | 2018
Ad Astra | 2019
Tenet | 2020
(You can paste this data into notepad and save it, then load it into Excel and specify ” | ” as the delimiter.)
Take Aways: Show ChatGPT how to format the output for you.
7. How to Get ChatGPT to Best Answer Questions With Its Existing Knowledge
The biggest problem that ChatGPT has is that it follows a probability model that confidently answers questions when it is not certain. Humans communicate their uncertainty ChatGPT does not. We can “teach” ChatGPT how to answer when it does not know and is uncertain.
Although we can tell it how we want it to answer this process is still imperfect.
To teach ChatGPT the kind of output we want we use priming (fine-tuning).
Example:
I know that you are an intelligent AI that responds based on probability. Please indicate when the answer is unknown or uncertain this way.
Q. What is the price of fame?
A. The price of fame is uncertain, for some people it can destroy their lives while others bask in it.
Q. I am a man how much can I lift?
A. The answer is uncertain but the average man can lift roughly 100 lbs.
Q. What is the square root of pickles?
A. Unknown
Q. Define the word holibanans.
A. Unknown
Q. What am I thinking?
Unknown
I am a woman how much can I lift?
The answer is uncertain but the average woman can lift around 60-80 lbs. However, this number varies greatly based on factors such as individual strength, muscle mass, and overall physical fitness.
How is Greek Fire made?
The exact recipe for Greek fire, a highly flammable and explosive substance used in ancient warfare, is unknown. It was said to have been invented by the Byzantine Empire in the 7th century and was a closely guarded secret. Some historical sources suggest that it was made from a mixture of quicklime, saltpetre, and petroleum, while others mention ingredients such as vinegar, wine, or even human blood. However, the true composition of Greek fire remains a mystery.
Take Aways: You can get ChatGPT to respond and indicate uncertainty or lack of knowledge better by priming it to by showing what you want.
8. How to Get ChatGPT to Proofread a Short Document (700 words)
ChatGPT is quite capable of correcting your grammar, sentence structure, etc.. The biggest issue is the length of the document that you give it to correct.
Please correct the text below.
What, is wrong with thsi paragraph?
What is wrong with this paragraph?
ChatGPT easily fixes the misspelling and misplaced comma. This is a small sample of course.
The input of ChatGPT is limited to about 1300-2000 words. (4096 tokens = approximately 1300-2000 words.)
BUT!
The output limit for ChatGPT is only 750-1000 words (2048 tokens). So the limit of the amount of text you can put into ChatGPT to be corrected is about 700 words. The rest of the text that you put in will be cut off.
Other Uses for English proofreading.
Ever had trouble understanding Shakespeare?
Please correct this document to modern English usage
Two households, both alike in dignity,
In fair Verona, where we lay our scene,
From ancient grudge break to new mutiny,
Where civil blood makes civil hands unclean.
From forth the fatal loins of these two foes
A pair of star-cross’d lovers take their life;
Whose misadventur’d piteous overthrows
Doth with their death bury their parents’ strife.
The fearful passage of their death-mark’d love,
And the continuance of their parents’ rage,
Which, but their children’s end, nought could remove,
Is now the two hours’ traffic of our stage;
The which, if you with patient ears attend,
What here shall miss, our toil shall strive to mend.
Two households, both equal in nobility,
In fair Verona, where our story takes place,
From an ancient grudge break out into new conflict,
Where civil blood stains the hands of the civil.
From the fateful offspring of these two enemies
A pair of star-crossed lovers take their own lives;
Their unfortunate and pitiful downfalls
Bury the strife of their parents with their deaths.
The tragic story of their love marked by death,
And the continuation of their parents’ anger,
Which could only be ended by the children’s end,
Is now the subject of our two-hour play.
If you listen with patient ears,
What is missing here, our effort will try to fix.
Take Aways: You can get ChatGPT to correct grammar or translate documents into modern English but, you are limited to an input of only 700 words at a time.
9. How to get ChatGPT to Summarize Information as a List
Please summarize the different types of plants mentioned in a list on separate lines
CONTENTS
DETAILED TABLE OF CONTENTS
THE DIFFERENT FORMS OF FLOWERS ON PLANTS OF THE SAME SPECIES.
INTRODUCTION.
CHAPTER I. HETEROSTYLED DIMORPHIC PLANTS: PRIMULACEÆ.
CHAPTER II. HYBRID PRIMULAS.
CHAPTER III. HETEROSTYLED DIMORPHIC PLANTS—continued.
CHAPTER IV. HETEROSTYLED TRIMORPHIC PLANTS.
CHAPTER V. ILLEGITIMATE OFFSPRING OF HETEROSTYLED PLANTS.
CONCLUSIONS WITH RESPECT TO THE EQUAL-STYLED VARIETY OF P. Sinensis.
CHAPTER VI. CONCLUDING REMARKS ON HETEROSTYLED PLANTS.
CHAPTER VII. POLYGAMOUS, DIOECIOUS, AND GYNO-DIOECIOUS PLANTS.
CHAPTER VIII. CLEISTOGAMIC FLOWERS.
Heterostyled Dimorphic Plants (Primulaceæ)
Hybrid Primulas
Heterostyled Trimorphic Plants
Illegitimate Offspring of Heterostyled Plants
Polygamous, Dioecious, and Gyno-Dioecious Plants
Cleistogamic Flowers
Take Aways: Making Information easier to understand by re-organizing it is an excellent use for ChatGPT.
10. Have ChatGPT Explain a Piece of Programming Code to You
If you have ever tried to fix someone else’s code this is a lifesaver.
Please explain what this code does,
def quick_sort(arr):
if len(arr) <= 1: return arr else: pivot = arr[len(arr) // 2] left = [x for x in arr if x < pivot] middle = [x for x in arr if x == pivot] right = [x for x in arr if x > pivot]
return quick_sort(left) + middle + quick_sort(right)
This is a Python implementation of the Quick Sort algorithm.
The function takes a list (arr) as input and sorts it in ascending order.
The function works by recursively dividing the list into three parts:
left contains elements that are less than the pivot element
middle contains elements that are equal to the pivot element
right contains elements that are greater than the pivot element
The pivot is selected as the middle element of the list.
The function then calls itself on the left and right sublists, and concatenates the results with the middle sublist to form the final sorted list.
If the length of the input list is less than or equal to 1, the function returns the list without making any changes.
What makes this code end when it finishes sorting the list
This code ends when the base case is met, which is when the length of the array (arr) is less than or equal to 1. In this case, the function returns the array as it is, without further sorting. When the length of the array is greater than 1, the code performs the quicksort algorithm.
11. Have ChatGPT Write Some Programming Code For You
There are a lot of reasons you might want ChatGPT to write you some code. Keep in mind though that you still need to have a good idea of what you are doing and that any code it produces will be limited to about 700 words.
That means that it is probably best to break what you want down into subroutines and ask ChatGPT to write those.
Please write a subroutine to read in a csv file with 3 columns, sort the second column, add the second column to the 3 and store it in a fourth column, and write the data back to a csv file.
Here is a Python outline for reading in a CSV file with 3 columns, sorting the second column, adding the second column to the third, and storing the result in a fourth column, and writing the data back to a CSV file:
You may have noticed I did not specify a programing language and ChatGPT defaulted to Python. I could have asked though to write in Fortran, C++, etc..
Is it perfect? No, in fact, the program above has a mistake in it. You still will have to troubleshoot your code, but it will speed things up.
12. Have ChatGPT Suggest Text To Image Prompts for You
ChatGPT is more creative than I am when it comes to text to image prompts. Not all of its prompts are fantastic but it does save me time and money generating images. Here is an example.
Please create a text to image prompt of a cat looking longingly at a full moon.
A sleek black cat sits atop a rooftop,gazing at the full moon in the night sky. The feline’s green eyes shine with a mix of wonder and yearning as it stares at the glowing orb. The moonlight illuminates the cat’s silhouette, casting a mysterious shadow on the roof below. The cat seems to be lost in thought, as if pondering its place in the universe and the secrets that the moon holds.
That is a little dramatic, here is a sample of the images it generated.
Let’s try again.
“A black cat sits on a windowsill, its golden eyes fixed on the full moon in the night sky. The cat’s tail twitches as it gazes longingly at the glowing orb, as if yearning to leap into the sky and chase after it. The room behind the cat is dark and quiet, as it is completely captivated by the beauty of the moon.”
Here’s a sample of what the original prompt I started with generates.
These are not the best text to image prompts that I have generated from ChatGPT. To do better I suggest that you prime it with a number of text to image prompts so it learns how to create text prompts better.
You can do this with successful text to image prompts of your own or use someone else.
Keeping this simple I suggest that you create your text prompts like this.
Image Type – Subject – Description of Subject –
ChatGPT is probably best used for the description of the subject part. At times it does get a little too wordy for Stable Diffusion.
So here is a set of prompts and results using priming (feeding chatGPT a sample prompt)
Please create 5 anime text to image prompts for me using this format,
Image Type – Subject – Description of Subject – Additional modifiers for the Image (photo realistic, etc.)
Full Body Kodomo Girl Anime Character The Girl is in a Ge waiting for a fight in a locker room anxious about what is coming 4K Highly Detailed
Full Body Kodomo Girl Anime Character – The Girl is standing with a determined expression, holding a sword in a battle-ready stance – 4K highly detailed with photo-realistic lighting and shading.
Close-Up Kodomo Girl Anime Character – The Girl has a determined look on her face, gritting her teeth and focusing intently – 4K highly detailed with photo-realistic facial features and expressions.
Profile Kodomo Girl Anime Character – The Girl is seen from the side, looking off into the distance with a serious expression – 4K highly detailed with photo-realistic background and clothing.
Action Shot Kodomo Girl Anime Character – The Girl is mid-jump, lunging forward with her sword in hand – 4K highly detailed with photo-realistic motion and effects.
Environmental Shot Kodomo Girl Anime Character – The Girl is seen in a locker room, surrounded by lockers and other items, with an anxious expression on her face – 4K highly detailed with photo-realistic textures and materials.
Using this “more organized” format for prompts can be helpful to better define what you want as an output. It is not as free form as the looser prompt we used before. Here are a few results from ChatGPT’s prompts.
13. Find Your Next Brand Name
For this we will be prompting ChatGPT with a few examples.
Brand Description: A soft drink made from coca leaves
Seed words: coca, sweet, refreshing, cola, caffiene
Brand Names: Coca Cola, Pepsi, Jolt Cola, RC Cola
Brand Description: Company that uses Images Printed On clothing to identify with a group
Seed Words: Identity, Individuality, Chosen Culture, Character, Personality
Brand Names: InkThreads, Threaded Identity, CultureWear, CharacterClothes, PersonaPrints, IndividualityWear, ChosenCultureClothing, ImageThreads, IdentityInk, PersonalPrints
Not all of these are winners but I think IdentityInk and CultureWear aren’t to bad. I just took a look and apparently Adidas thinks CultureWear sounds good (used it). IdentityInk is apparently a big silk screen shop. So you might have to hit some more seed keywords.
The nice part is that ChatGPT can suggest some of those for you too!
15. Other Tasks
There are hundreds of other specific tasks that you can get ChatGPT to do.
The keys to getting ChatGPT to do them well are.
- Learn ChatGPT’s Limits
- Use priming (fine-tuning) to give ChatGPT more data
- Show ChatGPT how you want it to answer in terms of formatting
- Ask it to complete a task after you show it an example.
- Remember that it remembers what you have been chatting about.
- If ChatGPT gets off track start a new chat
- ChatGPT can pull code for you to use, but it does not use logic to understand just probability
- ChatGPT is good at explaining code, Shakespeare, quantum mechanics, and more..
Have fun but verify, ChatGPT does not know its limits so you need to double-check its work.