- Because models are effectively "black boxes" the training manuals have to be produced by trial and error.
- Because models "hallucinate" it's critical to approach their outputs with reasonable skepticism.
- Because models train on truly massive datasets, they "know" a LOT of things.
- But, that also means that domain knowledge is incredibly useful in getting AI's to generate quality outputs.
- Because the models are not "intuitive" you should specify clearly what results you want.
- Most Generative AI systems are "pedantic" and will give you exactly what you ask for, not more, not less.
- Sometimes giving them more flexibility can be interesting, but usually it doesn't give you what you want.
- Prompting for things that are likely in the training data generally yields better results.
Consider the following progression of prompts and results made using DALL-E 2 and 3:
A corgi
a corgi with high depth of field
a corgi with high depth of field and Rembrandt short lighting
a full body shot of a happy corgi in a field of flowers with Rembrandt short lighting
a full body shot of a happy corgi in a field of flowers with Rembrandt short lighting and high depth of field
- Note that there are a large number of ongoing lawsuits related to GenAI and copyright. Be judicious when copying styles.
A stylish woman in the style of a renaissance painting
A stylish woman in the style of Picasso
A stylish woman in the style of a Pixar movie
ChatGPT:
Prompt: Write a short story, 300 words or less, about a man who won the lottery
In a small, unremarkable town, Thomas lived a life of routine as a grocery store clerk. His existence was a series of predictable days until a whimsical purchase changed everything. Thomas bought a lottery ticket on a day when his dreams felt particularly out of reach, tucking it away in his wallet as a tiny beacon of hope.Weeks passed, and the lottery draw slipped by unnoticed. It wasn't until a mundane apartment cleanup that Thomas rediscovered the ticket and, on a whim, checked the numbers. To his utter disbelief, he had won. The realization was a slow dawn, transforming his world from monochrome to technicolor. Thomas's win turned him into a local celebrity overnight, but the money didn't change who he was at his core.
Choosing wisdom over extravagance, Thomas donated to local charities, explored the world, and, most importantly, opened a small bookstore. This became a cherished community spot, a tangible reflection of his lifelong passion.
Years later, Thomas's life was rich with experiences, far removed from the predictability of his former existence. He found happiness not in the wealth itself but in the opportunities it brought to live authentically and give generously.
Thomas's story is a reminder that life's most profound changes often come from the most unexpected places. Winning the lottery didn't just alter his life; it allowed him to impact others positively, proving true wealth lies in giving and living fully.
Prompt: Rewrite the story in the style of an Axios article
**Title: Small Town Clerk Turns Lottery Win into Community Haven**Why it matters: In an unexpected twist of fate, a grocery store clerk from a nondescript town turned a lottery win into a life-changing opportunity not just for himself, but for his entire community, proving that true wealth lies in the impact one has on others.
The big picture: Thomas, a man whose life was defined by routine, found his world turned upside down when he discovered a forgotten lottery ticket in his apartment. The ticket, bought on a whim, ended up being worth millions, catapulting him from obscurity to local fame overnight.
By the numbers:
- 1 lottery ticket: The catalyst that transformed Thomas's life.
- Millions in winnings: Though the exact sum remains undisclosed, Thomas's win was substantial enough to reshape his future.
- 1 small bookstore: The dream Thomas turned into reality, using his winnings to open a community space that quickly became a local treasure.
Between the lines: Unlike many lottery winners who succumb to the temptations of sudden wealth, Thomas approached his windfall with a sense of responsibility and purpose. He recognized the opportunity to make a lasting difference in his town.
What they're saying: "The money didn't change who I was. It simply allowed me to do more for the people around me," Thomas shared. His sentiment reflects a deeper understanding of value and happiness, rooted in community and personal passion rather than material wealth.
The impact: Thomas's story is more than just a feel-good tale of luck and generosity. It's a reminder of the transformative power of giving back and pursuing one's passions. His bookstore has become a hub for connection and learning in the town, embodying the idea that the greatest impact comes from sharing one's fortune.
The bottom line: In a world where wealth often equates to personal gain, Thomas's journey underscores the profound satisfaction and joy found in using resources to enrich the lives of others. His legacy is a testament to the idea that the richest individuals are those who contribute to the fabric of their communities.
Prompt: Rewrite the story as a story for kindergartners
Once upon a time, in a cozy little town, there lived a man named Thomas. Thomas worked at a grocery store, helping people find their favorite snacks and fruits. Every day seemed the same for Thomas, like flipping the pages of a book with no pictures. But Thomas had a secret wish—he dreamed of adventures and a splash of color in his days.One sunny day, Thomas decided to buy a magical ticket. This wasn't just any ticket; it was a special lottery ticket that could make wishes come true. He tucked the ticket into his wallet, like a tiny treasure, and almost forgot about it as the days danced by.
Then, one day, while cleaning his small apartment, Thomas found the magical ticket again. With a fluttery heart, he checked the numbers and—surprise!—his ticket was the winner! Thomas's eyes sparkled like stars; his dream was about to come true.
With his newfound treasure, Thomas didn't buy a castle or a mountain of toys. Instead, he thought of his friends and neighbors in the town. "What can make everyone happy?" he wondered.
Thomas had an idea brighter than the sunniest day. He used his treasure to open a magical place—a bookstore! But this was no ordinary bookstore. It was a place where every book was an adventure, waiting to whisk you away to lands of dragons, talking animals, and secret gardens.
The children of the town loved Thomas's bookstore. It became their favorite place to explore, dream, and discover new stories. Thomas didn't just keep his treasure; he shared it, making his town a happier, more magical place.
And Thomas? He found that the greatest joy didn't come from what he bought, but from the smiles and laughter of his friends enjoying the stories. He realized that sharing your treasure makes your heart fuller and your days brighter.
So, Thomas and the children of the town lived happily ever after, with their hearts and shelves full of stories, proving that the best adventures come from sharing your dreams with others.
- Evaluate the outputs.
- Be especially wary of using GenAI in domains where you are not well versed.
- Check for common forms of social bias.
- Set explicit limits.
- 500 words or less
- Aspect ratios
- 6th grade reading level
- ...
- Ask for specific formatting
- HTML
- Bullet Lists
- MLA Style Citations
- ...
- Ask for specific styles
- Consider defining things like race, gender, ethnicity explicitly to avoid GenAI's inherent social biases.
- Long prompts work well, don't be afraid of using them.
- Utilize domain knowledge:
- GenAI's often know esoteric jargon.
- They respond well to domain specific prompting.
- Many models allow you to generate variations, or ask for modifications.
- Working in steps can be easier than defining everything in one prompt.
- Sometimes small changes to a prompt can result in large changes in output.
- Don't be afraid to do some things manually.
- If you're a talented photoshop user, let AI start and add your own final touches.
- Let ChatGPT produce a rough draft or an outline, and you write the rest.
- e.g. upload samples of your own writing or your companies style and ask it to imitate that style.
- Sample prompt: Analyze the text below for style, voice, and tone. Then write a new paragraph in 800 words in the same style, voice, and tone on Topic -[Topic]. [Paragraph]
- LLMs are good mimics, use phrases like...
- "in the style of cubism"
- "like an Axios article"
- "inspired by Pixar"
-
(This isn't supported by all models)
-
ChatGPT: use "system:" to make a system prompt, which tells the machine to use that for the whole conversation
- System: Write like an axios article
- System: Take the opposite view of claims I make, as if this were a debate
- System: Write using language that is accessible at a 6th grade reading level
-
Other systems have similar capabilities, try to find documentation for the systems you use.
Note: Many LLMs have incorporated this technique into their training data and via RLHF, so you may not get much additional benefit on such systems.
-
Key idea: ask the LLM to show its work.
- Some research has demonstrated LLMs succeed better when it has to do things step by step, and that more errors are introduced when it does things "all at once".
-
This can be done in a "few shot" or a "zero shot" way.
Here's a "few shot" example from a 2022 paper:
Instead of asking:
True or False: The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1.
They give the bot some examples of answering in a step by step style, then leave the final question for the LLM to fill in:
The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1.
A: Adding all the odd numbers (9, 15, 1) gives 25. The answer is False.
The odd numbers in this group add up to an even number: 17, 10, 19, 4, 8, 12, 24.
A: Adding all the odd numbers (17, 19) gives 36. The answer is True.
The odd numbers in this group add up to an even number: 16, 11, 14, 4, 8, 13, 24.
A: Adding all the odd numbers (11, 13) gives 24. The answer is True.
The odd numbers in this group add up to an even number: 17, 9, 10, 12, 13, 4, 2.
A: Adding all the odd numbers (17, 9, 13) gives 39. The answer is False.
The odd numbers in this group add up to an even number: 15, 32, 5, 13, 82, 7, 1.
A:
Here's a "zero shot" example. Instead of this word problem:
I went to the market and bought 10 apples. I gave 2 apples to the neighbor and 2 to the repairman. I then went and bought 5 more apples and ate 1. How many apples did I remain with?
Give the LLM the same prompt but add a tip to "think step by step:
I went to the market and bought 10 apples. I gave 2 apples to the neighbor and 2 to the repairman. I then went and bought 5 more apples and ate 1. How many apples did I remain with?
Let's think step by step.
Note: Many LLMs have incorporated this technique into their training data and via RLHF, so you may not get much additional benefit on such systems.
- Meta prompting is when you ask the LLM to make a prompt for you to use.
- I find this is especially useful in image generation contexts, for example:
Write a prompt to generate an image about the following:
Content: an elephant
Setting: the savanah
Style: digital illustration, HD
Yields:
Create a high-resolution, hyperrealistic digital illustration showcasing a majestic elephant standing in the middle of an expansive savannah. The image should capture the elephant flaring its wide, large ears against the sweltering heat. The detailed skin texture should embody the earthy grayish hue and exhibit the numerous wrinkles representative of its battle survival and age. Its long, curved ivory tusks should appear striking white, creating a significant contrast against its surroundings. The savannah should be portrayed as an infinite blanket of dusty, cracked earth with sporadic acacia trees. The sky should be colored a vivid, intense blue, with possible effects of heat waves rising from the ground. The golden, waving grasses and the merging of the horizon with the sky should be captured to represent the beauty and harshness of the Savannah. Manipulate shapes and forms, and features like layering effects to achieve a lifelike, realistic effect. The end result should be a vibrant, detailed representation of wildlife in its natural habitat, full of crisp, clear imagery.
Which yields:
- We're all learning more about how to better use these tools every day.
- Find other people who are using GenAI and ask them what works
- Or look online for similar resources.
- Ask for edits or proofreading on your writing.
- Ask it to find bugs in code, or write tests for a function
- Ask for a summary of input text.
- Generating custom images for articles, presentations, etc.
- Generating slides for a slideshow