The following transcript summary was generated by AI. The interview took place in May 2023 and some of the information within it is now factually incorrect and out of date.
Artificial intelligence (AI) is rapidly changing the field of marketing, enabling marketers to be more productive and effective. In a recent podcast, host Michael Stelzner interviewed AI expert Christopher Penn about how generative AI can benefit marketers. Here are some of the key takeaways:
- AI tools like ChatGPT can help generate marketing content like social media posts, email campaigns, and blog articles. While the quality may not yet match human-written content, these tools enable faster content creation and can help overcome writer’s block.
- AI excels at summarizing large amounts of text. It can distill key points from transcripts, meeting notes, and long articles. This allows for quick review and extraction of critical information.
- Rewriting content is another application for AI. It can refine and enhance rough drafts as well as rewrite content in different tones and styles. This provides flexibility and efficiency.
- AI question answering capabilities enable conversational interfaces for customer service and marketing. With training, AI agents can handle common customer FAQs as well as more complex queries.
- For extracting data and insights, AI is very capable. It can quickly analyze piles of data like tweets or transcripts to identify key entities, relationships, and themes.
- Creating customized AI models allows for industry- and company-specific applications. With open source options now available, more businesses can fine tune AI to their unique needs.
- Autonomous AI agents present new opportunities as well as risks. While able to work independently towards goals, ethical constraints are still a work in progress.
The key for marketers is developing AI skills and experience. With the right training, AI allows individuals and teams to accomplish more in less time. Though AI won’t completely replace human marketers soon, skills and roles will need to adjust to this new technology.
Can’t see anything? Watch it on YouTube here.
- Got a question for You Ask, I'll Answer? Submit it here!
- Subscribe to my weekly newsletter for more useful marketing tips.
- Subscribe to Inbox Insights, the Trust Insights newsletter for weekly fresh takes and data.
- Find older episodes of You Ask, I Answer on my YouTube channel.
- Need help with your company's data and analytics? Let me know!
- Join my free Slack group for marketers interested in analytics!
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.
Here is the cleaned and formatted transcript with speakers identified:
Michael: If you’re a marketer, I think the best expression I’ve heard of this is, “AI is not going to take your job. A person skilled with AI is going to take the job of a person who is not skilled with AI.”
Today, I’m very excited to be joined by Chris Penn. If you don’t know who Chris is, you need to know Chris. He is a data scientist and author of AI for Marketers. He’s also the co-founder of TrustInsights, a consultancy that helps brands with analytics and AI. He also founded the TrustInsights Academy and his podcast, which has been around for a very long time, is Marketing Over Coffee.
Chris, welcome back to the show. How are you doing today?
Chris: Thank you for having me. I’m having a great time.
Michael: Just out of curiosity, how long has Marketing Over Coffee been around?
Chris: 16 years.
Michael: Dang! That’s crazy.
Well, Chris has a lot of insights and today we’re going to explore AI for marketers and we’re going to get on some fascinating rabbit holes. I guess my first question, Chris, is we’re recording this in the middle of May and this is going to come out about a month later, but there are still a lot of marketers that are not paying attention to AI and for whatever reason, maybe they’re not sold on the value proposition of what it could do for them. Maybe you could explore the benefits and we’ll get into the concerns, obviously, that they have next, but what’s the possible upside as to why maybe they ought to listen to what we’re going to talk about today?
Chris: Sure. There’s three branches of AI. There’s regression, classification and generation. Regression is something that marketers have had access to for a long time. If you use Google Analytics and you’ve said, “Hey, show me my attribution model. What’s working for me?” That is essentially regression and it’s super powerful for identifying, “Hey, I’ve got a bunch of data and I’ve got this outcome. What’s leading to this outcome?” If you’re a social media marketer and you want to know which social media channels are working best, you may have heard of marketing mix modeling or media mix modeling. That’s all regression-based AI.
The second category is classification. Again, this is a very classical AI. You’ve got a bunch of data. What’s in the box? If you ever downloaded a few million tweets at a time and you’re like, “Okay, I need to classify these things because it’s just a huge pile of stuff I’ve got in my customer service inbox. I’m in my favorite social media monitoring software. I’ve just got this pile of stuff,” and you would use AI to organize it to say, “Okay, what is in this data? How do I sort it so that I can make use of it?”
The third category, which is the one that’s got everyone’s attention today, is generative AI, where you now have machines that can make stuff, images, sound, text, video. I just watched Coca-Cola’s first AI-generated commercial. It was very well done. Very well done. I’m not 100% convinced it’s fully AI-generated, but definitely the majority of it is. Each of these areas has benefits. Regression is all about, “Help me do my job better. Help me find answers.” Classification is, “Help me make sense of the data that I have.” And generation is, “Help me create and do more with the information that I have.” Marketers really probably want all three.
Michael: Yeah. Why? What’s the upside for them, especially on the generative stuff? Because that’s the hot stuff today.
Chris: It comes down to, people want, generally speaking, people want to save money, they want to save time and they want to make money. When you think about saving time, that’s an easy one. How long does it take you to write even a simple social post? How long does it take you to put together an Instagram image? How much money does it cost to put together compelling imagery or video or sound? How much does it cost to license stuff? You can save a pretty substantial amount of money by using generative AI to do those things. It obviously saves you time. If you’re saving money, you’re probably also saving time. And then because these tools let you scale, you can reach more people, do better messaging, reach out, be more places, and can bring in more business that way. So really clever, prudent use of the tools can really check the box in all three of those benefits that pretty much everybody wants.
Michael: Now, you have been in the AI sandbox, for lack of a better word, pardon the metaphor, for quite a while. How excited are you about what’s available to us today as marketers?
Chris: It’s funny. The technologies that we’re looking at today really are, to folks who’ve been in the field five or six years old, what has changed is the models themselves have gotten better. And anytime we talk about AI models, we’re really just talking about software that was written by machines for machine use. It’s kind of like if Microsoft Word is a human software, right? AI models are machine software. And the benefits today are, the things that’s changed today is that the accessibility is much easier. We’ve all heard of software like ChatGPT, for example, which is an interface to a model called the GPT family of models from OpenAI. We have just seen very recently Google’s second edition of its BARD software. We’ve used Microsoft Bing with the GPT-4 integration. We use Bing Image Creator to create images for free, right, inside your search engine. And so these tools are more accessible. They are, the advent, particularly of large language models, has made these tools easy to use for the non-technical person. You could have done some of this stuff five years ago, but you had to be a coder. Today, Andre Carpathi said this, a terrific quote in January, “The hottest programming language in 2023 is English.” Just being able to write. Prompt writers, right?
Michael: Exactly. So what is that? I mean, like, do you think this is going to unlock like a creative renaissance in some regards? Because like, what I’m hearing you say is that you had to be a coder to really take advantage of these things just a few months ago. Now anyone can use these things. And it seems to me that will unlock perhaps a new level of creativity. What’s your thoughts on that?
Chris: It depends on how you use them. And I know we’re going to talk about use cases at some point. In some ways they can unlock creativity. In other ways, for people who are perhaps not as self-motivated, they will be substitutes for creativity, right? These tools can create credible and reasonably good content. They don’t create great content. They don’t create like pure surprise-winning content, but they also don’t create crap anymore. Three years ago, it was like watching chimpanzees play Scrabble. It was not good. Now it is obviously much, much better. So I think there’s going to be a blend. You’re going to get more content. No matter anyway you slice this, they will be more. And if you are at a company where say you have a lot of people and you’ve got some C and D players on the team, with AI you could probably bring them up to like B minus players. So the bar has a bare minimum with these tools. There is still plenty of room and plenty of opportunity for A players to shine, right? The A players on the team, those individual contributors who have superior skills, there will always be a place for them. But it’s everybody else is like, “Well, if you’re a C player, we probably don’t need your specific skills anymore because machines can operate at a B minus now.”
Michael: A couple thoughts. First of all, I am seeing some of our peers actually putting out job wrecks for people to manage AI for their business, right? These are the smaller businesses. In addition, we are dealing with an aging population and a low unemployment rate, at least here in America. And I wonder whether or not this is going to help potentially, I don’t know, I’m just thinking macro and micro. I wonder whether or not with a lot of people entering into retirement and stuff, whether or not AI is going to allow smaller teams to be more productive, where in the past they had to hire out and there was a limited supply. I’m curious what your thoughts are on all that.
Chris: That is very much the case for smaller, more nimble organizations. My company, Trust Insights, we’re three people. We carry a client load that should normally require 20 to 25 people to run because so much of our work is done by machines, both regular programming and AI. For those companies, those organizations that are nimble and that have technical talent to make the tools work better and faster together, yes, they will have multiplier effects to make them punch above their weight.
For larger companies, I think you will see more of that sort of the downsizing effect where you’ll see, okay, we can get efficiencies within these companies that reduce the number of total people needed. It will definitely change the competitive landscape. If you’re a marketer, I think the best expression I’ve heard of this is AI is not going to take your job. A person skilled with AI is going to take the job of a person who is not skilled with AI. That really is the essence of what’s happening. If you are skilled with these tools, you are a more valuable employee. You can do more stuff. You can do stuff faster. You can do stuff at a better minimum level of quality versus somebody who is not. That is probably what the roadmap for an individual person is. If you’re thinking like, “Oh my gosh, what’s this going to do to my career?” You have a mandate to at least get familiar with and learn these tools. Whenever disruptive technology comes out, this has happened with the internet when it first came out, learning HTML and learning how to do website coding, and then eventually with social media, understanding how to create content on the social platforms and game the algorithms and create content. Now the challenge is the pace at which it’s happening is extremely fast.
Michael: Would you agree with that?
Chris: Oh, for sure. Think about this. We had computers in 1955. Thirty years later, we had personal computers. Fifteen years later, we had smartphones. Ten years later, we’re now getting into things like artificial intelligence. The span of time which we have to adapt keeps getting shorter and shorter and shorter. If you go back a couple hundred years and you look at the industrial revolution, you went from having 50 people in a field working to today one farmer driving this massive combine that’s GPS powered and all that stuff. He’s sitting there listening to podcasts as his machines are going up and down fields. There is still a farmer as a role, as a job in society, but that farmer’s job today looks very different than it did 300 years ago.
The good news is we should be smart enough. Those of us that are listening to this, we’ve been through, we’ve lived through these waves of technological innovation, especially those of us that are a little more gray haired. We’ve seen what it was like before the internet. We now know we’re entering into this new era. Nothing ever lasts forever and that’s why we do these kinds of shows so that you who are listening can embrace this change and hopefully become more valuable to your prospects, your company, your clients, etc.
Michael: I think that’s a good transition into exploring some of the different use cases that you see today specifically start wherever you want to start with.
Chris: I think for marketers and for everybody, you need to understand the six fundamental use cases within generative AI, particularly with large language models like those with ChatGPT, Bard, Bing, et cetera. Those use cases are generation, extraction, summarization, rewriting, question answering and classification.
Let’s talk through each of these. So generation, everybody knows that is, hey, write me a blog post about Instagram tips, right? And the machines will spit that out and the better your prompt is, which is the plain English code that you are writing, the better the results you’ll get from generation. These are good at generation. They’re not great at it. They’re good.
The second category, which I think is really where they start to shine is extraction. Say I take a million tweets, right? And I just have this data I can use. I can write a prompt says, extract the Twitter handles from these tweets and compile them into a list and a model like GPT four will do that. We’ll present it in the format that I want. Extract some email addresses from this PDF and so on and so forth. These tools are very capable of extracting data out.
The third use case is summarization. This is one of my favorites. Summarization is you tell these machines summarize this, for example, this podcast episode, take the transcript from this podcast episode and summarize it. Tell me the five most important things that Chris and Mike talked about and it will spit out those things. My best favorite use case of this is I use a piece of software called Otter, which is a transcription audio transcription software. If you go to TrustInsights.ai/otter, you can see the whole thing. It’s real simple. You get a raw transcript. Now, of course, a lot of what we say as in speech is not grammatically correct. It’s not polished. There’s a lot of um and uh, you know, all those things. And that shows up in transcripts. You then take that transcript, give it to a service like ChatGPT and say, rewrite this to be grammatically correct. And suddenly that random foaming at the mouth you had is, is clean or it’s maybe it’s a conference call you had with the client. You say summarize this into meeting notes and action items and boom, instead of having a virtual assistant that you’re paying or, or, or clerical staff, you’re paying now. You’re just having a machine do this. I just did this earlier today with a client call and they gave me the five action items from that call, put them right into my to do list program. And boom, I was, I took that 45 minute client call and it within literally a minute and a half, I distilled it down and I was ready to start my workday. So the summarization is really one of those, those very powerful things.
The fourth area that they’re really good at is rewriting content. This is again, you know, taking a voice call where you’re kind of rambling and having it rewrite that into something that sounds better is an easy use case. One actually just put this up on LinkedIn the other day and that’s actually like half a million people have shared it. It’s crazy. I had this very terse note from Karen and accounting to Bob saying, Bob, the two months of invoices you left on my desk aren’t done. They’re not going to get done anytime soon because you can’t just do that. A bunch of profanity in it until there’s Bob. Oh F off. And then the prompt says rewrite this email on a professional tone of voice. And it comes out, uh, Bob, uh, dear Bob, uh, I regret to inform you that, you know, very formal professional tone. It’s a rewrite.
So if you are the kind of person who maybe you don’t have a lot of confidence in your writing, but you have a lot of confidence in your ideas, you can use these tools to do this. There’s a great use case of a person who wrote an app for a smartphone. He works with construction contractors and his one friend was dyslexic, very severely dyslexic, um, and would write very terse, kind of confused emails to clients and clients were not appreciative of it. He made this app, this app did exactly. I took those terse directions and reformatted it to a formal business email. And now clients are very happy with that. So, um, rewriting very powerful. You can even do silly stuff like take the blog posts that accompanies this episode and, and rewrite in Sumerian or emoji. These tools are capable of that.
The fifth area that is powerful is, uh, open is classification. So again, as we were talking about earlier, if you have a bunch of say tweets or emails in your, in your social media monitoring software, or, uh, maybe you even have podcast episodes you want to listen to them in the, from the past, you could have these tools, say you identify the top three topics this episode is about, and then you can sort through those listings and go, okay, I want to listen to these episodes. I could classify, uh, tweets by sentiment. Is this a positive sentiment, negative sentiment? Uh, what kind of social media comment is this? Is this a complaint? Is it a question? Uh, so these tools are very good at doing that kind of classification.
And the last one, this is where there’s major change happening is question answering. These tools are very capable of answering questions. Now they do have limits. For example, open AI’s family of tools, uh, have a time horizon. They don’t know anything after September of 2021 Microsoft Bing, Google’s barred. They don’t have those limitations. They, they are using a search engine data to power them, but they can answer very complex questions, questions that you might not get a concise answer out of a traditional search engine.
For example, uh, one of my favorite little tricks just for around the house is I’ll write out a menu for the week of the things I’m cooking for dinner and I’ll say to one of the models based on the list of these dishes, put together a probable grocery list for me and it will spit out all the ingredients for all the dishes, you know, with quantities like, okay, great. Now I can go to the grocery store and not have to spend 20 minutes going, well, look up this recipe. What do I need to buy? Nope. The tool gives me a good enough list that I can go shopping and save a lot of time.
Those six categories of use cases apply to everything in marketing, apply to everything in social media, apply to everything in customer care. They’re super, super powerful. That’s where marketers will see a lot of benefits.
Michael: What I’m most excited about is a couple of these classifications, a couple of these categories, summarization, rewriting and question answering. And I want to dig in on these a little bit.
I love the idea that like, for example, anybody who creates content, if you like have a transcript, right? You mentioned Otter, I think one of my team members has Otter show up to meetings with him, if I’m not mistaken, and it will like send notes on what the major points were in the meeting and stuff like that. It’ll even prompt, you know, people to ask questions in the meeting, which is kind of fascinating. We joke about it all the time because like, you know, we say, Joel, your Otter is in the meeting. I’m almost certain that’s what the tool is. But, you know, the summarization thing is kind of a big deal because when we are in a call, a company meeting or a client meeting, right, and there’s a transcript of it, there could be a whole bunch of stuff that was discussed and a whole bunch of rabbit trails that we can go down. And it’s hard for us as humans sometimes to remember all the things that were discussed. And the idea that you could have a tool that catches all these things could be a really big deal. Would you agree?
Chris: Absolutely. And the ability for it to then distill it down and assign it or at least to say like, hey, Mike is responsible for these things. These are things that Mike signed up to do. Depending on how good the transcriptives, if people have attributions to what they said, yeah, it’s super powerful and it’s a great way to deliver the kind of customer service that clients wish you would, but that we know because again, we have very human limitations about what we can remember. These tools are kind of like an outside brain.
Michael: Well, and you also have some people who are dyslexic like I am and struggle sometimes to read very long content. So and you know, some blog posts are like 20,000 words. I could totally see a tool that would say something along the lines of, hey, give me the talking points inside this blog post, right? I would imagine they already exist. Do they or don’t they? I’m just curious.
Chris: They absolutely do. They absolutely can do that. The tools do that. There are prompts for that. There are entire companies that are startups that are trying to do that. For those of us who have a bit more gray hair, you probably remember Cliff’s notes, right?
Michael: Of course, yeah.
Chris: This is basically these tools are basically Cliff’s notes for life.
Michael: They’re very good at this, right? This is one of the things that they’re, they generally get down really quite well, right? I mean, sometimes they’ll miss some of the important points I would imagine, right? Or do you find like they’re getting quite sophisticated?
Chris: For the current generation tools, they’re extremely good because you’re not asking them to create anything new. You’re actually asking them to take things away. And so they have all the data to start with and it’s much easier for them to remove than it is to create and add more. So the rewriting thing, I think, is also a really big opportunity for any of us who are in the business of creating any kind of written content, right? Like for example, emails. Like we did a fun little thing with ChatGPT4 where we asked it to create a, well, actually this is technically question answering and rewriting. We asked it to create a four week email campaign and we were going to send this many emails in week one, this many in week two, this many in week three, and this many in week four. And we said, how many? And we said, please come back with recommendations. And it said, here’s what you should send in week one. Here’s the subject line. Here’s what the topics might be. And it prepared the whole thing. And then we used ChatGPT to actually feed it a little bit of data, right? On what we thought it should have. And then it crafted emails. And then we went through this, like you talked about this editing process of refining it and refining it. And what I found was, as a writer, anybody who writes sometimes gets a creative stick where they’re blocked, they’re stuck, right?
Michael: And I feel like, I don’t know if rewriting or writing are the same thing, but I would imagine they kind of fall into the same classification here, creating content versus rewriting, or is it a different classification here?
Chris: They’re different functionally in these tools. They’re different, but you’re speaking more to a human thing, right? As writers, as creators, yeah, we get stuck. When a tool does generation for us, like you feed it two pages of a white paper and like, “Okay, continue from where I left off,” it will spit out something. And that flips your brain from writing mode to editing mode, which is often enough to get you past your writer’s block. Because you’re like, “No, no, no, that’s not what I was going to say.” Oh, that’s what it is. And so your brain’s back on track. Yeah.
Michael: Now, you mentioned there was a web browser extension for ChatGPT. Is that by a third party? Is that by ChatGPT? Do you know what the name of that extension is and what does it do? Does it allow you to bring in the outside web?
Chris: If you are in ChatGPT and you are in the paid program, the $20 a month program, ChatGPT+, you’ll see a little toggle. It says GPT-4 and a drop down menu that has two menus. One is web browsing and two is plugins. Plugins are third party extensions that are provided by other companies. It is probably the new app store for those who are in that kind of market. The web browsing one is built by OpenAI and it allows ChatGPT to go out, browse the web and pull data back in. Now, I have seen pictures of that, but I have not seen that for myself. Does one have to sign up for their alpha or beta program in order to be able to see that? Do you know?
As of three days ago, when we were at the day of recording this, it was open to everyone who’s a paying customer. So you have to go to your settings menu and turn on the beta stuff.
Michael: What does the plugins make possible?
Chris: Pretty much anything you can do on the web. So Kayak is in there for trip planning. Zapier is in there to connect it to these things. There’s a couple of extensions that people are doing to connect to stock market data. There’s actually a recent investigation done by a major investment firm. They took a stock portfolio, some back data and gave it to ChatGPT and said, pick some stocks and then they, because it was back data, they could see how their stock picks performed. It performed like 400% better than the market. And so now this firm’s like, so we’re just going to give this some real money now and see if it can keep making a 4X return on our money. But there’s about 40 extensions in there now and there are probably going to be 10X or 100X that if your company has already done stuff like built an app or built with APIs, it would behoove you to start looking at deploying an extension and getting it to open AI and get it through the approval process to be able to use it within their system. That’s one way that there’s a lot of marketing opportunity.
Michael: Okay. Is there any other, we’ve talked about how you can use AI, particularly ChatGPT to summarize information and to create information, maybe refine information. Is there any other marketing uses that we haven’t addressed that you’ve seen recently that maybe marketers might be like, oh, I hadn’t thought about that when it comes to generative AI?
Chris: So there’s a new model that’s not within the ChatGPT ecosystem. It’s from Mosaic ML called MPT Storywriter. One of the limitations of today’s models, the commercially available ones, is that they have a relatively limited frame of reference. They can create about 3000 words at a time, give or take. You’ve seen this in ChatGPT, if you’re like, stop writing in the middle of a paragraph and you have to type continue to get it going. MPT has released a model that is competitive to the GPT series, but can do 65,000 tokens at a time. So it could write 40,000 words all at once. So now you’re talking like business book length. So think about that from a rewriting use case. Imagine that you were wanting to write another business book and you have a bunch of audio that you recorded. That’s 30,000 words of rambling audio. You could, with the MPT Storywriter model, feed that in and say, “Okay, give me 30,000 words of coherent text now, please.” So we’re going to start seeing these tools be capable of very long-form content, much longer than it’s been generated so far. That I think is going to be a very interesting marketing opportunity for everyone.
Michael: Fascinating, first of all. To chat GPT, I know so many of us are using chat GPT for and are paid. The memory on it, when you create a new thread or whatever they call it, does it remember all the other stuff? Because this is the part where we think the AI is forever smart and remembers all the stuff we fed into it. But is there a limit to how long from your experience it’s going to remember before it has to be retrained in the prompts?
Chris: 8,192 tokens. So about 6,000 words it remembers. It has a roving memory window, so if you have a very long series of interactions, it sort of goes off the rails after a while.
Michael: Oh, interesting. Okay, so about 6,000 words. But what about if you come back to it like a day later? Is it going to remember what the discussion was inside of that?
Chris: Yeah, the thread will preserve what’s happened so far.
And then since you’re technical, if you’re using a tool that has an API integration, is it similar or is that not necessarily always the case?
So, with the OpenAI API for the GPT 3.5 Turbo model, which is the one that powers the default of ChadGPT, there is actually a section in your coding where you put in the previous responses. You feed them back to the software. So you have to do that. It’s costly, I would imagine, right? Because you’re feeding in bigger prompts or something like that.
Michael: Okay, so the API is not yet supporting four is what I’m hearing you say?
Chris: It is for some developers. You have to be enrolled.
Michael: Got it. Okay, so let’s talk about prompts. You mentioned earlier, this is kind of one of those secret weapons, like understanding how to actually engineer a prompt. Presuming we’re talking about ChadGPT because that’s the one that most people are familiar with. Any tips on how to give the system essentially the right kinds of information to get better output?
Chris: So all these models work essentially on the words you give them. They don’t have any words of their own. They all have mathematical probabilities of what it understands about how language works. So the more detailed your prompt is, the better result you’re going to get.
So we actually have a one page PDF, no registration, no forms to fill out. If you go to TrustInsights.ai/promptsheet, you’ll get the ChadGPT specific version of this. But it works out like this. There’s what’s called a role, which is you say you are a social media marketer, you know Instagram, Instagram stories, Instagram Reels, high performing Instagram posts. And there’s a task. Your first task is to generate five Instagram posts from the following background information. Then you provide your information like it must contain, you know, @SMExaminer, you know, mention the SMM24 hashtag and you give it a bunch of requirements. And then you sort of finish off the prompt saying write the Instagram posts. That structure of role, task, background, execute is the best format for ChadGPT to generate a high quality response for, particularly for generator responses.
Michael: Rook, okay. You’re going to pivot to something else because I have some clarifying questions, but go ahead and finish what you’re doing.
Chris: So real quick, every model is different. So if using Bard, what works for Bard will not necessarily work on ChadGPT, what works on Bing and so on and so forth. So you have to know the intricacies of each model that you’re working with.
Michael: Okay. So, so many of us have not done role and it still gets okay responses, right? So specifically, you are a, and you essentially substitute the role that you would be doing. Is that, is that what you mean?
Chris: In the context of what you want it to be doing. Yes.
Michael: Do you, what about the audience? Do you need to also identify who the target audience is? Like you are a marketer who is trying to attract XYZ audience and your task is blank. Does that make any sense or no?
Chris: I typically put audience stuff in the background information section.
Michael: And what’s the background information section? Cause you said role, task, and then…
Chris: Role task background execute is the…
Michael: Oh, the background. Okay.
Chris: That’s your requirements. So whatever it is, so if you’re having a right Instagram post, for example, you’d want to tell it which hashtags to use. You want to tell it whether or not it should use emoji in the text. You want to tell it what kind of imagery suggestions to make. You might have customer feedback in there, whatever information you have for this.
Now I will also say this, the prompt length depends on the kind of task. If you are doing generation, question answering, or extraction, you want longer prompts. If you’re doing summarization, rewriting, and classification, your prompts can be real short. Like for example, I have a one sentence prompt for Otter transcripts, fix grammar, spelling, punctuation, formatting, and spacing. That’s it. It doesn’t need anymore of that because it’s got all the information. Basic you said? What does basic mean?
Michael: No, I was saying for rewriting and summarization, you can have a one sentence prompt because you don’t need to know. I heard you, but you said you’re prompt, but you said against basic. That means I must know what the heck you meant by basic, right?
Chris: Oh, I don’t recall saying that, but okay. Fix grammar, spelling, punctuation, formatting, and spacing.
Michael: Oh, and spacing. Okay. I misheard you. Okay.
Chris: I heard and basic.
Michael: Okay. So getting these prompts really down specifically for anything that is question answering, right? Or generating something original is really, really important is what I’m hearing you say. Now when you’re in a thread specifically, since it does have a memory, if you have the paid account, you presumably only have to do that until it doesn’t remember. Or do you do that with every single one?
Chris: So here’s my recommendation. People should using the software of your choice, one note, ever note, Joplin, whatever you should have a prompt library of the best prompts that you found that work well and treat this with care. Remember what Andre Carpathi said, the hottest programming language in 2020 is English. These prompts are software. You’re writing software. This is possibly part of the secret sauce of your business. So don’t just go, Oh, look at this cool prompt I did on Twitter. It’s about giving away your source code, right? You don’t want to do that unless you’re doing intentionally. Be very careful. If you work at a company, you need to be thinking about, are we giving away company intellectual property and we shouldn’t be give remember that because it’s really important.
But for sure, you should have a prompt library of stuff that you work. And if you work within an organization, maybe there’s a shared document of some kind, a shared data system internally where you can store these things and people can trade them back and forth within a company so that you can maximize the productivity of these things give you.
Michael: Well, and I don’t know if you have done this, but sometimes you don’t like the output of it. So you ask it to rewrite it maybe in a casual voice because maybe you forgot to ask that the first time or maybe to rewrite it without mentioning certain kinds of things. I would imagine you can continue to refine the output until you really love it and then take what you learned and then put that into your next prompt. Is that fair?
Chris: You could do that. But if you have very technical resources, you can now start to scale it where you would take that prompt and you would send it to the API and say, okay, now write a thousand blog posts about this and things. This is a very popular thing that we do. We see a lot and we’ve done with our own SEO keyword list. We’ve written a prompt that has all the parameters for writing. And then we have the keyword list, which is in a data table. And then the ARC programming language, it goes through the keyword list and sends each keyword through and generates content for it. So you can now have machines taking your human prompts and just scaling them dramatically.
Michael: So just so we can kind of help everybody understand how they could do this on a text-based platform like Facebook or Twitter or LinkedIn, I would imagine you could say you’re a marketer working at company X, right? And that’s your company, right? And your task is to write a month’s worth of posts that are maybe like a hundred words or less, right? On this particular topic or to come up with 20 different questions, right? And then the background information is going to be, this is who the target audience is, right? This is who the audience is that we’re trying to attract with these kinds of questions. Now generate the output. Is that essentially, did I do that right? Is that kind of how we would do it?
Chris: That’s how you do it. And then like you said, you’re going to QA it, you’re going to refine it, you’re going to improve it over time. And then basically you just, at that point, put it to the test and see how it performs.
Michael: This is the analyst. I mean, like, do you, have you tested this stuff up against your stuff? And does the AI generated stuff tend to perform better for you when you’re using it?
Chris: It does not yet. Uh, generally, so we’ve done some A/B tests. I’ve actually taken existing blog posts I wrote in the past and had AI rewrite them and put up the exact same version, um, so that it gets crawled and stuff. And the performance has not been as good in terms of dwell time and in terms of discoverability. Now that might just, you know, that’s an N of one. So I would encourage anyone who’s interested in that to test it themselves, uh, because your results probably will differ. But the stuff you’re doing on LinkedIn, was that assisted by AI, the stuff that took off on LinkedIn that you were sharing earlier?
Michael: No, that was not. Okay. Well, I mean, the example was from chat GPT and stuff, but I didn’t chat. GPT did not originate that idea. That was just me being silly.
Chris: Got it. Okay, cool.
Michael: All right. So, um, where’s all this going? Um, let’s talk about like open source models and autonomous agents and stuff like, cause people are going to, their minds are probably going to be blown by some of this stuff that’s coming next.
Chris: Yeah. So we talked about the open source models. This is an exploding area right now. There are hundreds of models being built and designed and customized and deployed for free that you can download and use and tune to your own use cases. So if you, any piece of software that has even the modicum of complexity, I would expect software manufacturers to have a large language model interface that will allow you to chat with the software in the next three years or less. Any company that does not do that, they are behind the eight ball and they are asking to have their lunch eaten by a more nimble competitor because think about it. How complicated is Photoshop to use, right? It’s not a particularly user friendly piece of software for an amateur. Imagine taking a photo in there and say, and it’s all chat window pops and says, okay, colorize this photo, make it more dynamic and bright. Oh, and remove my ex.
Michael: Even better if you could talk to it instead of typing, right?
Chris: Exactly. Um, and so these open source models will now allow software manufacturers to do that without having to pay to open AI for every interaction because you can put that model straight in your software. So that’s going to enable a lot of innovation in the next couple of years. You’re going to see this stuff appearing everywhere. It’s already going to be in Microsoft office and Google docs and all the big tech, but pretty much any software manufacturer, I would expect to see this. So get good at prompt engineering because you’re going to be using an awful lot as a discipline.
The bigger area, which is fascinating and alarming is what’s called autonomous AI. And so this is where you have software that you give it a general goal and, uh, and maybe a starting task or two. And then it spins up multiple instances of these large language models and tries to solve this problem, uh, that you’ve given it. For example, I did a test said, I want you to go to my Twitter profile and figure out how to make me more popular on Twitter. How do I get more likes and retweets and things? And so it spun up 15 to 16 instances of AI agents started writing its own code to scrape Twitter to be able to identify Twitter handles and stuff like that. And essentially sort of assembly software solution that would let me identify what works on Twitter at the time the software runs. I can’t code in these languages, right? And it took the software a while to do it and it was, I would declare it a moderate success. It was not particularly, you know, a human social media manager, a social media strategist would have done a much better job. Um, but the fact is I was able to do that and just walk away from the software and let it do its thing for a few hours is pretty interesting.
The use cases for this, however, these tools are amoral. They have no morals. They have, they’re, they’re like chainsaws, right? They can be used for good or bad and there are plenty of examples of these tools being used in nefarious ways. You know, there’s, uh, there’s an example of, uh, I actually tried this on my website. Um, I told it, here’s the, my website’s login page, try to find a working login. And it started downloading things like, you know, known, you know, cracked password lists and stuff like, okay, clearly this tool has no restraints. So it can be a little hazardous. If you work in cybersecurity, your life has gotten a lot more complicated, but you have job security for a long, long time.
Michael: So what I’m hearing you say is these autonomous agents can be given a task and they’re just going to keep trying until they achieve their outcome. Is that really what I’m hearing you say?
Chris: That’s correct.
Michael: Wow. What’s the upside to that? I mean, like, I would imagine this could be used to do some really cool scientific discoveries as well, don’t you think?
Chris: Absolutely. Think about it. Stuff like, you know, take the, the RNA sequencing on the spike protein of a COVID virus, right? That’s just text. It’s just four letters, A, G, C, and U. You can write models that can interact with that data, just like, you know, it’s interacting with a blog post and have it do predictions, have it do estimations and things. So yeah, there’s a tremendous amount of power in these tools and like any, any powerful tool you can do great things with it. You can do bad things with it. And what the outcome is depends on who’s using the tool.
Michael: Well, Chris, we have clearly just barely scratched the surface of this fascinating frontier. Like people want to learn more about everything that you’ve got going on. Which social platform do you want to send them to and, and where do you want to send them if they want to learn more about your company and all the great things you’ve got going?
Chris: So for the company, go to TrustInsights.ai. For me and my weekly newsletter where I cover a lot of AI stuff, go to ChristopherSPenn.com. And we have some courses, not, no AI courses yet, because by the time the course is done, it’s out of date. But we have some regular courses at academy.trustinsights.ai. The one that I would point people to is we have a free one called Power Up Your LinkedIn Profile, where we looked at the AI behind LinkedIn and how it works and make recommendations to what you should be doing personally on LinkedIn to kind of adhere to the way that their AI works.
Michael: Chris Penn, thank you so much. Really appreciate your time today. Like I’m, I’ve got a bunch of rabbit trails I need to go down now.
Chris: You’re very welcome.
You might also enjoy:
- How to Measure the Marketing Impact of Public Speaking
- Is Social Listening Useful?
- The Basic Truth of Mental Health
- It's Okay to Not Be Okay Right Now
- Best Practices for Public Speaking Pages
Want to read more like this from Christopher Penn? Get updates here:
Get your copy of AI For Marketers