Almost Timely News: Improving the Performance of Generative AI Prompts (2023-12-17)
Almost Timely News: Improving the Performance of Generative AI Prompts (2023-12-17) :: View in Browser
Content Authenticity Statement
90% of this newsletter's content was generated by me, the human. Some of the prompt responses in the opening are generated by ChatGPT's GPT-4 model and are marked as such. Learn why this kind of disclosure is important.
Watch This Newsletter On YouTube ??
What's On My Mind: Improving the Performance of Generative AI Prompts
Today, let’s talk about getting better performance out of large language model systems using prompt engineering. Over the past 3 months, I’ve had a change of heart and mind about prompt engineering. Originally, I was aligned with what a lot of industry folks were saying about prompting, that the need for prompt engineering was going to go away as models became smarter. But the more time I spent in the trenches with models, especially the open source ones, the more I realize there’s some nuance there.
In general, for the average user of a large language model, that is a true statement, that prompt engineering will probably get less important over time. As models get smarter, they generally get better at guessing user intent, thanks to human feedback being incorporated into language models. However, there are a couple of shades of grey here.
The first is that large public models are also being censored more and more heavily. Ask an image model for a Disney reference and you’ll likely be told no. Ask a language model for a point of view about politics and you’ll get some linguistic gymnastics worthy of a politician themselves.
Here’s the thing with censorship of models: it diminishes their performance. Imagine you had a cookbook and you decided to censor the use of wheat. Now imagine going through that cookbook and ripping out every page that referenced wheat. You would have a severely diminished cookbook when you were done, and you would be capable of cooking much less, including recipes where wheat wasn’t the main focus, like a Beef Wellington. Imagine pretending Beef Wellington didn’t exist because you eliminated references to wheat. That’s what model censorship does. With added censorship comes added skill needed to get the most out of models.
The second shade of grey is that more advanced prompt engineering takes advantage of the architecture and structures of the models to get better results faster. For example, imagine you have a library, and you want to put together some books to check out. You could absolutely just walk around the library and collect books, and you’d end up with what you were looking for. That’s general prompting. Now imagine the library had a specific classification system and internal architecture - say, ISBN numbers or the Dewey Decimal system. How much faster could you find the books you were looking for if you had that internal representation and architecture of the library?
That’s what prompt engineering at its peak does - it doesn’t just instruct the models about what to do, but takes advantage of the way models work to deliver better results in less work. Now, to be clear, that doesn’t mean you’re doing it wrong today. If you’re getting good results from models, then that’s really all that matters. But if you’re curious about how to get better results in less work, then you’ll want to adapt a few techniques to improve your use of language models.
We’ve talked before about the RACE structure for prompt engineering , and it’s really good at what it does. The reason is that the RACE structure, when you follow it, has enough of the terms needed for a model to form the statistical associations it needs to generate great output. Here's what I mean. Suppose you said to a chef, "I'm hungry," and that was the entire prompt. The chef has so little to go on that they'll cook you SOMETHING, but it stands to reason it's not going to be what you want.
Suppose you said, "I'm hungry for pizza". That's a lot more specific than I'm hungry, but there are limitless variations of pizza. The chef might be Japanese and make you a favorite in Japan, creamed corn and squid pizza. If you love Japanese pizza, then you get lucky and have a good pizza experience. If you don't love Japanese pizza, then there's a good chance you're still not going to have an enjoyable experience.
Suppose you said, "I'm hungry for pizza. I'd like a margarita-style pizza with fresh mozzarella, fresh basil that's been cut chiffonade, and a tomato sauce made from Roma tomatoes and tomato paste to a very thick consistency. I'd like the crust to be thin, less than a centimeter thick, and I'd like it to be cooked at very high heat for very fast, so that the bottom is partially blackened. I'd like it to be 510 centimeters in diameter with 10 grams of sea salt and 15 grams of black pepper sprinkled onto the sauce prior to the basil and cheese being put on."
Do you have a very clear mental image of this pizza? Would our chef? Absolutely. With this level of clarity and detail, we can be reasonably sure our chef is going to produce a pizza that we want to eat. That's what the RACE structure - and many other prompt engineering structures - do. They provide enough detail and keywords and concepts that the language model is able to execute the task well.
Now, here's where we'll go beyond the basics. If you're not getting good results from a model today, there's a good chance the task you're asking it is too vague, too imprecise, and too big. Let's dig into a very specific example. Suppose you want a language model to emulate your writing style. You say something along the lines of "You're a writing expert in the style of Ann Handley. You know writing, writing voice, writing styles, how to be a great writer. Your first task is to read this sample of my writing. {insert sample here} Provide a bullet point list of my writing style."
That's a good prompt; it follows the RACE framework and will certainly give you a better result than a vague prompt like "Analyze my writing style." Here's an example of the output, using GPT-4 and the ChatGPT interface:
This is a good start, to be sure, but it's still too vague. Watch what happens when I ask it to write a paragraph explaining the importance of the 10-year/3-month Treasury spread in macroeconomics:
Imagine diving into the world of macroeconomics like you're flipping through a cookbook, where the 10-year/3-month Treasury spread is a key ingredient, much like salt in a recipe. Picture the 10-year Treasury bond as a hearty stew, taking time to simmer and develop flavors, and the 3-month Treasury bill as a quick, refreshing salad. The spread between them, the difference in their yields, is crucial in predicting the economy's health, akin to a chef tasting and adjusting their dish. When the 10-year yield is higher, it's like having all the right ingredients for a robust economy, but when this spread narrows or inverts - imagine your quick salad taking longer than your stew - it often signals economic downturns, like a chef sensing a recipe might flop. This analogy helps to demystify a concept that, like a dash of salt, might seem small but is essential in bringing the whole economic picture to life.
That's not bad, but it's not quite my style. Why is that?
It's because writing style is too big a task for a language model to produce in one shot. What is a writing style? Take a few moments to mentally think about that. What is a writing style? What are the parts of a writing style?
There's word choice, rhythm, cultural perspective, grammar, voice, tone, punctuation, spacing, intertextuality, lexical diversity... the list of components that form a writing style itself is a full page of text. How many different pieces did you think of? Now, suppose you took the same writing sample and the same prompt, but broke it up into pieces. Here's an example:
"You're a writing expert in the style of Ann Handley. You know writing, writing voice, writing styles, how to be a great writer. Your first task is to read this sample of my writing. {insert sample here} Provide a bullet point list analysis of my writing style specifically about my use of grammar."
Here's what comes out:
That's a lot more specific, isn't it? Now suppose I re-ran that prompt, but I asked about punctuation, or spelling, or whitespace, or pacing, or... you get the idea. Writing style is so large and vague a concept that the response of a language model is going to be equally vague.
By breaking writing style down into its component pieces and asking the model for its analysis of my writing in those component pieces, I end up with a much more detailed, focused, and practical definition of my writing style.
Let's see what the results are if I use a prompt with a gigantic specification of my writing style, based on this technique of breaking down a big concept into its component concepts.
In the realm of macroeconomics, the 10-year/3-month Treasury spread is a crucial indicator, often likened to a compass guiding the financial markets. Essentially, the spread measures the difference in yields between the 10-year Treasury note and the 3-month Treasury bill. When investors are nervous, they flock to long-term bonds, pushing those yields down, seeking safety but signaling danger. When the yield on the 10-year note falls below that of the 3-month bill, it's an inversion, a signal historically predictive of an impending recession. This inversion is like a warning light in a car's dashboard, alerting economists and business professionals alike to potential economic slowdowns. For instance, in digital marketing, being ahead of the curve in understanding economic trends can be as crucial as predicting the next big shift in social media algorithms. An inverted spread can indicate a shift in investor sentiment towards risk aversion, much like how consumer preferences can suddenly shift online, affecting marketing strategies.
That's much better and more clear; behind the scenes, the prompt that I used to fuel this was 17 paragraphs long, each paragraph an aspect of my writing style.
Why does this work? Writing style is an aggregated concept. It's a concept that bundles a whole bunch of other concepts together; when you do that kind of bundling, you get a good overview of the concept, but you miss a lot of the nuance and subtlety in the individual components. And that nuance, that subtlety, is where style emerges. Everyone knows the basic big picture concept behind spaghetti and meatballs - but the little details are what sets apart one dish from another, details that might not be captured in the big picture.
Any time you're getting results that aren't quite what you want using prompt engineering, ask yourself whether the task is broken down in enough detail that the model knows what to do. Today's big public models can handle prompts that are very large in size, so you can afford to be more detailed in what you provide for instructions. Think of language models like the world's smartest interns. The results you get are directly proportional to the clarity of instructions you provide.
If you'd like to learn more about the RACE framework and prompt engineering, good news: my new Generative AI for Marketers course just launched ! With over 5 hours of instruction, tons of hands-on exercises, a workbook, and a certificate of completion, it's a great way to level up your generative AI skills. Use discount code ALMOSTTIMELY for $50 off the tuition.
If you'd like a deep dive into what's in the course to see if it's right for you, check out this video tour of the course .
How Was This Issue?
Rate this week's newsletter issue with a single click. Your feedback over time helps me figure out what content to create for you.
Share With a Friend or Colleague
If you enjoy this newsletter and want to share it with a friend/colleague, please do. Send this URL to your friend/colleague:
For enrolled subscribers on Substack, there are referral rewards if you refer 100, 200, or 300 other readers. Visit the Leaderboard here .
ICYMI: In Case You Missed it
Besides the new Generative AI for Marketers course I'm relentlessly flogging , I recommend
12 Days of Data
As is tradition every year, I start publishing the 12 Days of Data, looking at the data that made the year. Here's the first 5:
领英推荐
Skill Up With Classes
These are just a few of the classes I have available over at the Trust Insights website that you can take.
Premium
Free
Advertisement: Generative AI Workshops & Courses
Imagine a world where your marketing strategies are supercharged by the most cutting-edge technology available – Generative AI. Generative AI has the potential to save you incredible amounts of time and money, and you have the opportunity to be at the forefront. Get up to speed on using generative AI in your business in a thoughtful way with Trust Insights' new offering, Generative AI for Marketers, which comes in two flavors, workshops and a course.
Workshops: Offer the Generative AI for Marketers half and full day workshops at your company. These hands-on sessions are packed with exercises, resources and practical tips that you can implement immediately.
Course: We’ve turned our most popular full-day workshop into a self-paced course. The Generative AI for Marketers online course is now available. Use discount code ALMOSTTIMELY for $50 off the course tuition.
If you work at a company or organization that wants to do bulk licensing, let me know!
Get Back to Work
Folks who post jobs in the free Analytics for Marketers Slack community may have those jobs shared here, too. If you're looking for work, check out these recent open positions, and check out the Slack group for the comprehensive list.
What I'm Reading: Your Stuff
Let's look at the most interesting content from around the web on topics you care about, some of which you might have even written.
Social Media Marketing
Media and Content
SEO, Google, and Paid Media
Advertisement: Business Cameos
If you're familiar with the Cameo system - where people hire well-known folks for short video clips - then you'll totally get Thinkers One. Created by my friend Mitch Joel, Thinkers One lets you connect with the biggest thinkers for short videos on topics you care about. I've got a whole slew of Thinkers One Cameo-style topics for video clips you can use at internal company meetings, events, or even just for yourself. Want me to tell your boss that you need to be paying attention to generative AI right now?
Tools, Machine Learning, and AI
Analytics, Stats, and Data Science
All Things IBM
Dealer's Choice : Random Stuff
How to Stay in Touch
Let's make sure we're connected in the places it suits you best. Here's where you can find different content:
Advertisement: Ukraine ???? Humanitarian Fund
The war to free Ukraine continues. If you'd like to support humanitarian efforts in Ukraine, the Ukrainian government has set up a special portal, United24, to help make contributing easy. The effort to free Ukraine from Russia's illegal invasion needs our ongoing support.
Events I'll Be At
Here's where I'm speaking and attending. Say hi if you're at an event also:
Events marked with a physical location may become virtual if conditions and safety warrant it.
If you're an event organizer, let me help your event shine. Visit my speaking page for more details.
Can't be at an event? Stop by my private Slack group instead, Analytics for Marketers .
Required Disclosures
Events with links have purchased sponsorships in this newsletter and as a result, I receive direct financial compensation for promoting them.
Advertisements in this newsletter have paid to be promoted, and as a result, I receive direct financial compensation for promoting them.
My company, Trust Insights, maintains business partnerships with companies including, but not limited to, IBM, Cisco Systems, Amazon, Talkwalker, MarketingProfs, MarketMuse, Agorapulse, Hubspot, Informa, Demandbase, The Marketing AI Institute, and others. While links shared from partners are not explicit endorsements, nor do they directly financially benefit Trust Insights, a commercial relationship exists for which Trust Insights may receive indirect financial benefit, and thus I may receive indirect financial benefit from them as well.
Thank You
Thanks for subscribing and reading this far. I appreciate it. As always, thank you for your support, your attention, and your kindness.
See you next week,
Christopher S. Penn
Super excited to read about enhancing AI prompts! ?? It reminds me of what Albert Einstein once said, "If you can't explain it simply, you don't understand it well enough." Breaking down complex ideas into detailed, clear prompts is crucial for understanding and innovation. Keep up the great work! ???? #ContinuousLearning #Innovation #AIWisdom
Business Designer and Product Owner at Nationwide Building Society
11 个月Great episode Chris - I used Chat GPT with plug ins to create Cornell style notes for this video which I will handwrite to aid recall. I like that you acknowledge other prompt techniques other than RACE.
Crafting Audits, Process, Automations that Generate ?+??| FULL REMOTE Only | Founder & Tech Creative | 30+ Companies Guided
11 个月Looking forward to learning these techniques! ??
Co-Founder and Chief Data Scientist at TrustInsights.ai
11 个月Thank you!
Artificial Intelligence | Communication Measurement & Evaluation | Thought Leadership | Strategic Communication | Public Relations
11 个月Another great newsletter, thank you Chris ???? I love your use of metaphors to explain complicated technical concepts - makes it very easy to grasp for the rest of us.