So What? Creating a Custom GPT model for restaurant marketing

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!

In this episode, you’ll learn how to extract review data for your custom GPT model, refine your custom GPT model to build personas, and how to maintain your custom GPT model for ongoing restaurant marketing.

Catch the replay here:

In this episode you’ll learn:

How to extract review data for your custom GPT model
Refining your custom GPT model to build personas
How to maintain your custom GPT model for ongoing restaurant marketing

Upcoming Episodes:

Generative AI and Sales – 12/07/2023
Walk though of AI course – 12/14/2023

Have a question or topic you’d like to see us cover? Reach out here: https://www.trustinsights.ai/resources/so-what-the-marketing-analytics-and-insights-show/

AI-Generated Transcript:

Katie Robbert 0:29
Well, hey everyone, Happy Thursday. Welcome to so what the marketing analytics and insights live show I am Katie joined by Chris and John. Hello, fellas.

Christopher Penn 0:37
Howdy.

Katie Robbert 0:41
It’s been a minute since we’ve all been in the same place for a live stream. And again, of course next week, that will not be the case again, which I suppose is better than the alternative that nobody wants to talk to us. And we’re all just sort of sitting around waiting and doing live stream after live stream.

Christopher Penn 0:59
too busy is better than not busy enough.

Katie Robbert 1:03
In this week’s live stream, we are talking about creating a custom GPT model, specifically for restaurant marketing. And so we’ve been playing around so OpenAI a few weeks ago, rolled out the ability to create your own custom GPT model, which we’ll get into the details of that. And so we’ve been playing with it, Chris has built a dozen or so custom models for various reasons. And today, we wanted to focus specifically on Restaurant Marketing. And so we’ll cover how to extract review data for your custom GPT model, refining your custom GPT model to build personas, and how to maintain your custom GPT model for ongoing restaurant marketing. So, Chris, where would you like to start?

Christopher Penn 1:46
Well, let’s start with refreshing folks, Marriott, what accustomed GPT is because if you’re if you’re relatively new, you may not have heard of these things. There’s a tool that many folks know it’s in fact, it’s his birthday today. It’s ChatGPT is one year birthday, they even gave it a hat, which is very nice.

Katie Robbert 2:07
I see it, I saw that this, you know, it’s funny, I saw that this morning. And it didn’t I didn’t make the connection. And I was like, huh, they put a hat on it. And then I just kept going about my day.

Christopher Penn 2:19
And yeah, and OpenAI dev day, which was a day for developers, they announced that custom GPT will be available on this is where you can take ChatGPT and create a an instance of it, that you give very specific directions to, and then it will behave more like that. It’s it’s kind of a weird blend of two different tuning methods. One’s called fine tuning, where you give some special instructions, and one’s called retrieval augmented generation, we give it a bunch of data and say, Hey, refer to this data first, before you do anything else. So today, what we want to do is we want to talk about using a custom GPT. For the for restaurant marketing from the perspective of being customer centric. So there are a lot of people make custom GPT-3 this as a last count, there’s something like 20,000 public customer, GPS plus, who knows how many internal ones there are, like, you know, we have, by a dozen of these things now internally Trust Insights that we don’t open to the public. In fact, Katie GPT got an upgrade this morning.

Katie Robbert 3:25
One of us should,

Christopher Penn 3:28
ChatGPT got an upgrade to two years worth of transcripts and podcast data. Thanks. So Katie GPT-4 is much smarter. Now. The challenge with these custom GPT-4, though, is that their software development and a lot of people don’t know how to do software development. And as a result, you may create things that don’t necessarily work all that well. So Katie, you want to talk through briefly. So the the big five things that people should be thinking about before creating a custom GPT?

Katie Robbert 3:57
Yeah, absolutely. So the five p framework is purpose, people process platform and performance. Purpose being what is the question we’re trying to answer? What is the problem you’re trying to solve? People being who’s involved in this thing, not just the person physically pushing the buttons, but you know, your end users, your audience, your stakeholders, your decision makers, process? How do I do the thing? How do I maintain the thing? Platform? What tools am I using and performance? Did we answer the question, did we solve the problem? And so before? I mean, this is just good. You know, business practices, best practices, requirements gathering, you should always go through this exercise, even at a high level, to have some sense of focus to have some sense of requirements, but also to have some sense of how do I know I accomplished what I set out to do. And so in this case, we can say that the purpose of building the restaurant marketing custom GPT is to better understand What a given restaurants customers like and don’t like. And so we can do more of what they like and focus less on what they don’t like, probably specifically for things that are going external like social media, or newsletters so that we can replicate the customer’s language and just get a better insight. Because reviews are all over the place reviews are on a lot of different platforms. And it’s really hard. For most marketers bring them together into one place to really do an analysis.

Christopher Penn 5:30
Exactly. So the first place the the first thing would why want to do is probably want to set up I would imagine a user story or two of them in this case. So with a lot of customer reviews, you get, you know, the reviews tend to be very polar, right that you get five stars or one star because no one really feels the need to leave a three star review, right three star reviews, kind of like it’s somewhere in the middle and not displeased. It’s either I hate it, or I didn’t hate I love it. And there’s not a lot of middle ground. So we would have to use a story. So as a happy customer of this restaurant, I want to talk to the restaurant about things that aren’t on the menu so that I can have even more of my favorite foods or as a dissatisfied customer, I want to let the restaurant know how I feel about their terrible service so that I feel better about venting my spleen.

Katie Robbert 6:25
Well, but also as the Social Media Manager for the restaurant, I want to understand what resonates with customers so that I can promote more of those specific dishes, events, you know, drinks, whatever, as the owner, I want to know what’s working so that I can adjust, you know, my service level in my entrees.

Christopher Penn 6:47
Exactly. So the first thing we would need to to get them to make these user stories happen to bring this custom GPT-4 to life is the data, right? We need actual data of some kind. And when it comes to review data. The reality is there’s not a lot of great ways for the average small business to get it now for enterprises, there is you’re welcome to buy. You know, there’s really good social media marketing software from companies like Talkwalker, or companies like Yext, for example. But for many of folks, those the prices for that software is out of their reach. So your best bet is going to your Google business reviews, your Facebook reviews and stuff and just starting to copy and paste into into text files. And you generally want to bucket them into two into two buckets, you want to bucket into unhappy and happy, right? So unhappy will be one, two star reviews, and then happy to be three to five star reviews just to the best of your ability, copy and paste them into into separate text documents.

Katie Robbert 7:48
I’m actually surprised you didn’t say to wash it through some sort of an R script for sentiment analysis.

Christopher Penn 7:55
We don’t want to do that for this particular use case. That’s probably a good topic for another show another time about doing sentiment analysis. I have what I’ve done is I have gathered up and made some synthetic reviews. So this is for a company that does not exist because we can’t use not for obvious reasons, a actual stuff do. I have the view name, the star rating that gave one is one is bad five is good. The review text, the date of the review, and then the restaurants response. So this would be similar to what you would get out of like Google business, Facebook, Yelp, etc. Now, go ahead, Katie. No, go ahead. What we probably what we want to do is, is, like I said, split this into two. So I want to just in this Excel file, I’m gonna take my star rating, I’m going to sort it largest to smallest. And I’m going to grab just my review text for the five, four and three star reviews a copy that, and let’s put it in a plain text document. And we’ll call this good reviews. That’s that’s our first set. And now we’re going to go back to our document and going to go into the twos and ones. I’m gonna repeat the exact same process. And we’ll call this bad reviews.

Katie Robbert 9:21
So on how closely do you pay attention to reviews before going to a restaurant? drives

John Wall 9:26
everything. There’s a bunch of different platforms that I check, but yeah, you know, Yelp, Open Table and actually, we’re gluten free here. So find me GF is actually an app that gets a lot of traffic and I do a lot of stuff.

Katie Robbert 9:39
Do you bother to leave reviews?

John Wall 9:42
Yeah, I usually do especially. Well, Chris made a great point with that, that you know, there’s very few extra medium three star reviews, you know, it’s like, yeah, this place was awesome, or this place is terrible avoid it. And I actually it has to be something exceptional for me to leave a bad review, but I’m all After leaving good reviews for decent places, I want them to survive. Unfortunately, the restaurant industry, like if you love a place, you have to be writing reviews for them, because it’s a tough climate out there.

Katie Robbert 10:12
That’s a good point.

Christopher Penn 10:15
It’s a good point. So we’ve now got our two files of customer reviews, we’re going to create custom GPT-4. These go to the Explore menu and ChatGPT It looks like by the way, this may go haywire. ChatGPT is under considerable load today, I think it’s it’s had too much birthday wine, but that’s okay. We’re gonna go ahead and create our custom GPT. Now we’ve done past episodes, we’ve done through the walk through talk through version of this, which can be kind of a crapshoot. So I’m going to go to the more advanced version today, we’re just going to configure this ourselves typing all the information rather than going through the interview process. So we’re going to call this cafe, happy customer. A synthetic customer personality for the happy Cafe goer. All right, now, our instructions that we want to get this thing, we probably should get an understanding of the key the major key points to program into the personality. So I’m going to open up a second ChatGPT window here. And what we’re going to do in this part is we’re going to create a very, very unique, and probably not well heard of term called a sparse priming, representation.

Katie Robbert 11:41
Correct we have not heard of this.

Christopher Penn 11:43
we have not heard of this. A sparse priming representation is a way to get at essentially the major keywords and topics in a body of text without all the extra fluff. So when you and I speak, we have a lot of extra words that are not necessarily relevant words in the conversation, right? So you have filler words, you have prepositions, articles, things that don’t lend a lot of context. And that filler, those filler words occupy a lot of memory, and a lot of what are called tokens. So in, in language models, language models predict on what are things called tokens, which are snippets of words. And the more words you have, the faster the model runs out of memory, right, because only has a certain amount of working memory. So if you can generate a, a, sort of like a compressed version of just the relevant keywords to prime a model with, then you can save a lot of token space with how the model works. If you’d like to learn more about tokens and context, Windows and all that stuff, you can pre register for the new Trust Insights, general AI course go to TrustInsights.ai AI slash AI course. Alright, so let’s go ahead and build this, this thing, we’re going to take our good reviews first, right here. And then I’m going to give it a sparse, sparse priming representation prompt, and it’s going to read through the reviews, and build me this very, very short representation of the major content in the good reviews. And without all the extra fluff. This may might take a minute or two, we’ll see.

Katie Robbert 13:19
It sounds similar to what we used to do with setting up. Like, I might use the term wrong text mining or text modeling. And so basically, we would take out all of the irrelevant words to find what the true key words were of a big body of text like that was a precursor, I think to what it is that we’re doing now, when we were doing the the text mining to figure out what are the main, the topic modeling is what it’s called. To find the main topics in a body of text, we would take out things like if then the A and because those aren’t the topics.

Christopher Penn 13:59
Exactly right. This is this is a variant of that. But it’s a very, it’s generated by language model. So I’m gonna go ahead and copy this sparse representation here and just put that in my text in a text file for right now.

Katie Robbert 14:10
I didn’t know that’s what that button did. The little you? Yeah, at the bottom where you have the little box and then the uptown and refresh. I didn’t know that was a copy. I just learned something aside from learning everything else today. Okay, it’s a continual learning.

Christopher Penn 14:28
Every day is a learning experience. Alright. So we’re gonna say this custom GPT emulates a customer of Arendelle cafe that is satisfied or happy with the quality and service delivered, behave as though you were the customer and focus your responses along the lines. have the most important points to happy customers. And we now take our sparse priming representation. We put that in. Now for this, we for this exercise, we probably don’t need code interpret web browsing might or might not be here, I’m gonna turn it off because we don’t need it for our toy example, if you are an actual business, you might want to have that because you might want it to go and read things, I’m going to upload my good reviews. And that looks good. And I’m gonna go ahead and hit my Save button.

Katie Robbert 15:39
I’m glad this works for you. Because anytime I try to save, it throws me an error.

Christopher Penn 15:43
Right? So now you have this custom GPT that has a sparse representation of the major topics that happy customers care about. And you also have the raw text of all the reviews. So you could now have a conversation with this and say, Hey, happy customer. In fact, I’m the owner of Aaron Dell cafe, I’m thinking about things I could do to entice happy customers to spend more time and money at the cafe, this holiday season, based on your experiences at the cafe, what things would be good ideas?

Katie Robbert 16:34
So is this more of a because you did the sparse thing, which came up with the topics that that the happy customers were happy about? You could probably just look at that and be like, these are the things they’re happy about? What do I need to take it this step farther? So what is what’s the purpose of doing this versus just looking at the list of topics?

Christopher Penn 16:58
The list of topics, a priming representation works like this. Behind the scenes, this is still ChatGPT behind that means it still has the entirety of the GPT-4 model. So it has the entirety of knowledge that was built in the model. The primary representations that are essentially our memory Volkers that help guide the model is to figure out what other things should I talk about that a related to the starting points. So the example that it gives you ever hear just like three notes of a song and boom, you know, a whole song? Yes. Okay, all the time that that is a priming representation, your memory works as primary representations, you get a certain smelling like, Oh, I remember that person that I had that thing with back in college, just that one smell brings that that whole memory back. Continue. Our brains work on priming representations, it’s part of our predictive capabilities. And so when we, when we create a primary representation like this, this contains all the key words that not only are in the customer text, but then tell the model. This is what’s also important generally. And so when you go and you and you start assembling responses to any question, this is going to invoke extra keywords and concepts and ideas that come along with that little snippet. And so it makes for a very, very powerful, highly focused model.

Katie Robbert 18:30
Okay, John, I feel like you’re you are the prime representation of Trust Insights, because you say so much less than me and Chris, but when you speak, it’s very precise.

John Wall 18:42
I take my Buddhist stance here. I can see how this would be great, though. I mean, this is the kind of thing, especially in the restaurant industry, like nobody wants to sit down and write, like a newsletter piece or something for the Chamber of Commerce or whatever. And this just gives you an easy way to say, hey, write me up something about what the specials are this week, or, you know, tell me why people should be coming into the restaurant next week. I mean, this gives you just a whole realm of tools that just have not existed up until this point.

Christopher Penn 19:14
And I want to point out if we look at all the things on this list quality and craft memory evocation morning, ritual enhancement, sensory delights, and then we look at the list of what it recommends for the holiday season. There’s whole things in here like live music, evenings and exclusive holiday merchandise. That is nowhere on that list. That is nowhere in the original but because it’s associated in memory, the model associated with the things we did give it, it makes those logical conclusions. So Katie, if you think about the, the food and beverage client that we work with, often if we were to take those reviews and feed it back this stuff and create one of these representations, it could then help say, Okay, well, here’s some of the things that maybe you’re missing, maybe forgotten about, maybe as a restaurant tour, maybe it applies your clothes until maybe it doesn’t. But this is how you would get at this, right?

Katie Robbert 20:05
Like, I’m looking at this, and I’m like, This is the great outline starter outline for like a monthly newsletter. And basically these are your topic headers that you make sure you always cover.

Christopher Penn 20:17
Exactly this. So this is for the happy customers. But you could also take something like, let’s do this. I have a new dish I’m going to offer this month at the restaurant. It’s a it’s a tilapia. Tilapia cheesecake dessert. As one of my happy customers. What do you think about this idea? Is a good idea. A bad idea?

Katie Robbert 20:58
What do you think, John? Good idea? Bad idea?

John Wall 21:00
This is gonna be wonderful. I can’t I’m on the edge of my seat here is indeed in trade drinking.

Katie Robbert 21:24
I feel like ChatGPT is trying really hard not to be like, That’s a stupid effing idea, isn’t it?

John Wall 21:34
It’s like, it’s someone who doesn’t know you that well. They’re trying to tell you it’s a horrible idea without just coming out and saying it. Yeah.

Katie Robbert 21:41
Because that’s what this is like? Well, it depends on can you balance the flavors do you like it’s giving you a lot of considerations, which is great. But I would be looking for like just telling me is a good idea to know, based on the data?

Christopher Penn 21:55
Exactly. This is quite a bold choice. And sure they’re more conventional options would be wise.

John Wall 22:02
That’s good. Yeah. Yeah. It’s like, the only one I don’t see is like a, well, if you have someone who loves fish, and cheescake. This could be it. Exactly.

Christopher Penn 22:14
So any kind of marketing copy, and if you if you’re maybe doing that first newsletter, you’re trying to write a newsletter for this, you might want to run it by the happy customer say, Hey, happy customer. Here’s my newsletter. What things have I forgotten? What things should I do differently? That might resonate with you?

John Wall 22:34
I love it. Like number four communication, it’s like, you need to explain why this is good. Because people are not going to get that on there. And a trial run, you should do a soft launch before you announce this to anybody.

Katie Robbert 22:49
Oh, thank you. It’s like, you know, it’s funny, because I’m looking at this, I’m like, I don’t have time for all this. Just tell me what I need to know. And so what I what I personally forget as an end user is you can give those kinds of props to follow up to ChatGPT was like no, just get to the point. Like, just tell me the answer.

John Wall 23:07
There’s a server back there now that is like has smoke pouring out of it.

Katie Robbert 23:15
Now I think that, you know, while this is churning away, what is your take Chris on people who are going to try to use this as a replacement for proper market research.

Christopher Penn 23:30
Like all AI, a lot of what gendered AI does is the first draft, right. So this would be a good representation of first draft, it’s not the final, you still need to augment. And here’s the thing, market research, ideally should come first. Because if you have good market research, you can train the machine on that you can provide the raw interviews, the one on ones, the focus groups, and sparse primary representations of all that, and put it into this so that you have a more thorough, more complete machine, I would not have just as I would not start any kind of training data with synthetic data, except in this case, we have to protect PII. I would not start with synthetic data, I would start with real data. So I say the market research comes first. And then that can feed the modeling. So it’s the opposite direction of the way people were probably going to try it.

Katie Robbert 24:25
well, and that’s why I wanted to bring it up because I feel like we need to at least explain the process so that marketers don’t say, oh, I can just grab a bunch of reviews, put them in here and that’s my market research. But John, you hit on this earlier in the episode, where basically, you know, you’re not bothering with a three star review. So you’re missing all those people in the middle, or they have to have done something really wrong in order to garner you taking the time to write a one star review. And so you’re missing so much of your audience by only looking at reviews. It’s not I got a well balanced representative sample.

John Wall 25:05
It totally came out though and is like, Yeah, dude, this is a bad idea.

Katie Robbert 25:11
Well, and it’s still trying to be polite, but it’s like in summary, while the creativity is admirable, it’s not going to align with the expectations and tastes of your customers. And so I feel like that’s what the compliment sandwich is like, let me give you a compliment. Here’s the bad news. And here’s another compliment.

John Wall 25:30
There’s a classic line there, though, it could potentially deter customers rather than attract them. That’s the goal.

Christopher Penn 25:39
Yep. Okay, so that’s our happy customer. We want to create the unhappy customer too. Right? So this is going to be Cafe unhappy customer a an interactive simulation of our cafes, unhappy customers.

Katie Robbert 26:05
Now with these, and I’m assuming the answer is yes, it just takes more configuration. But so my first thought is, well, wouldn’t it be great if you have the cat cafe unhappy customer? And then you also have it speak in the style of, you know, someone was, I guess se gentrified now, but someone from Southie so like, you get the answers the way that they would actually speak to each other, you’d have to obviously get a lot of that training data. And so I’m answering my own question but basically if you also want your custom GP Ts to respond the same way that your average customer would you first need to get the way in which those average customers speak including, you know, the Boston accent exactly that because I’m reading through this and I’m like, oh well an unhappy customer sounds a very specific way in my head.

John Wall 26:59
What’s wrong with these chicken wings tough guy

Christopher Penn 27:06
So while while we were talking about that I ran through to make a representation for the negative customer reviews and you can see there’s there’s a lot of things that people are really unhappy about extended wait times, dish quality culinary misalignment, lackluster ambience and stuff like that. So these are all the things that are unhappy customers have have decided is not for them. So let’s go ahead and save this as a customer GPT Now, here’s where you want. If you’re the restaurant owner, you might want to say okay, so I am the owner of Arendelle cafe, I want to help you have a better experience at my restaurant. But I can only do so much I have limited staff and limited budget to work with what would you say is the single most important thing I need to fix to deliver an overall better experience. So with with custom GPS or any of these models have a lot of prime data these are really good ways to ask questions right to ask what’s going on so it’s prolonged service times many complaints around the extended waiting period and impacts food quality impacts ambiance etc. So reduce wait times quicker service whatever you can do to get to quicker service. Now. I don’t I’ve never run a restaurant I’ve I’ve watched an awful lot of Gordon Ramsay’s Kitchen Nightmares.

Katie Robbert 28:56
Feeling does not representative.

Christopher Penn 28:58
No, it’s not representative. However, one of the things that that has been mentioned a few times is if you simplify the menu, if you got a menu with 5 million things out, if you simplify the menu, you can typically increase service time, you know, the time to get food out because there’s fewer dishes that the back and staff need to make and more common components so they can work tend to work faster.

Katie Robbert 29:22
Well, and so you would have to use the two models side by side one to figure out what people like so you could make sure you’re not cutting the things that are bringing people in so let’s say you know for example, the number one dish is a beef wellington well of beef wellington takes an awful long time to make a good proper Beef Wellington is actually a couple day process. And so that would be a really hard you know, thing to solve for if the number one dish takes a few days, but people are unhappy with the wait time.

Christopher Penn 29:54
Exactly. And that’s that’s where these tools still have The gaps, at least tool still have specific gaps in experiential knowledge that you can only get from other experienced professionals. So you may you some things you could do to to incorporate that would be interviews. So you would talk to your staff and say, Hey, we got everyone’s telling us wait times issue. What do you guys think is the way we could reduce wait times you talk to your front end staff, you talk to your service, you talk to the backend back house, get their opinions, put them in a file, condense it down, put it into a staff GPT and say, here’s these things that you know how to do this. Now, if you want to get real fancy, and this requires some level of technical experience. There’s a framework called Auto Gen, that allows you to connect multiple models together. So you could take OpenAI infrastructure here and make an assistant API, same exact process as we did with the customer. GPT-4, the happy customer, unhappy customer, and the staff have three different assistants. And then in this advanced framework, they talk to each other. You would be there’s like a a person, you know, sort of the the, I guess moderator of the discussion, say, Okay, we need to reduce service times, happy customers, unhappy customer, staff, talk amongst yourselves. How do we do this, and you will watch these tools have conversations with each other,

Katie Robbert 31:27
I would recommend bringing on someone who knows what they’re doing, like a Chris Penn, if you’re going to do that, although Chris, you and I came up with a terrible idea that we should have Katie ChatGPT and Chris GBT talk to each other. So I guess we’re gonna see what happens when we do this.

Christopher Penn 31:43
Exactly. But that would be this is an example of again, you’re creating these agents that have very specific primed points of view. And now you can have conversations with them and get a sense of what are the things that are, that are options for me here, as a restaurant owner, the reason why you want custom GPS is because the paid version ChatGPT plus is 20 bucks a month. So it’s not a huge investment. And as you saw, beyond the prop, the specialist prompts to create the the primary representations. There’s no coding, there’s no heavy lifting here. There’s no writing against an API and stuff like that. It’s just you have a conversation with this entity that you’ve created, that has your knowledge, your company’s knowledge. In a low tech environment.

Katie Robbert 32:35
I think it’s also worth mentioning that even though you’re you know, creating these instructing them to be more conversational, having good prompts, and the right kinds of questions, is still an important part of the process. Because I’m looking at this. And of course, you know, I think, what do we call it, the intrusive thoughts, my intrusive thoughts are winning? And my first thought is, well, you have, you know, this database of unhappy customers stuff. So I just want to ask, what’s your problem? And just see what happens. I mean, obviously, it’s a machine like, it’s just going to tell me, like all the things that are wrong. But structuring the prompt in a better way, because this, I mean, I’m guessing it’s going to tell me nothing of value. So it’s, I’m here to express concerns as a customer, like, okay, great, but like, you know, I can see where you could sort of go down the rabbit hole and not get the right kinds of answers to your questions, because it’s very easy to forget that you’re not actually talking to the customer. This is just all data that you have compiled on behalf of your customers, and you’re just trying to figure out what to do better.

Christopher Penn 33:49
It is. However, it also does have, like I said, does have access to the GPT-4 model behind the scenes. So it does have a much larger general knowledge base that it can draw on, once it runs out of data that you provide it. So it’s not limited only to the data you provided, you can think bigger, which which can be helpful. Let’s say, What do other similar cafes do to deal with this issue of slow service times? Especially breakfast cafes with a Norwegian theme, right? Super, super specific, very specific. Well, it’s called Arendelle. Cafe modeled after this the 10 year anniversary of the Disney movie, Frozen, which is set in a fictional part of Norway. And that’s where that’s where that’s why it has so in the Python script, I have a bunch of different Norwegian breakfast foods and all that stuff to generate the fake reviews from.

Katie Robbert 34:50
John, you totally knew all that, didn’t you.

John Wall 34:52
I am down with Norwegian breakfast and I always love me some Indina Menzel, so

Christopher Penn 34:59
I Now, you’ll notice here, what it’s doing that now is it’s actually going out on the web. Right? So I gave, we gave it a question. And now it says, Okay, well, I don’t there wasn’t enough information that you gave me. But I am allowed to browse the web. Maybe I can go find some other information to help answer this question.

John Wall 35:17
Yeah, is that GPT-4? Now? Is this something that they’ve just rolled out in the in the latest round? Yeah, cuz I’ve never seen that. I didn’t realize it was doing active browsing. That’s crazy. Now we know disabled for a while and it’s back. Yeah, they had to turn it off. I can totally get that. You know, we have to give a shout out to chip Griffin when it comes to knowing restaurants and having a unique take on service. I have to call him out. Because is more than once pointed me into a restaurant in New York or DC area that was worth it worth checking out.

Christopher Penn 35:49
Can you get to react to physical violence? Words? No, not not presently. Although they’re Boston Dynamics has integrated language models into its spot robot. So the robot it operates in the physical world, you can have conversations with it.

Katie Robbert 36:07
Terrified of humanity.

John Wall 36:08
Right, exactly. You’ve just done that’s the site.

Katie Robbert 36:11
I’ve seen Ex Machina. I know how this ends.

Christopher Penn 36:14
While this is working? You guys want to see it? It’s 20 seconds. All right, here we go.

Unknown Speaker 36:21
Greetings. Good says, May I have the pleasure of knowing your names?

Unknown Speaker 36:25
I’m Matt. And that’s Bachi.

Speaker 1 36:29
A pleasure to meet you, Matt and Vacha. Shall we commence our journey? The charging stations where spot robots rest and recharge is our first point of interest. Follow me gentlemen.

Katie Robbert 36:43
It’s the little hat that really does, it pushes it over the edge.

John Wall 36:46
Now, it’ll be disturbing when it’s four storeys tall going down your street, you know, tearing into homes.

Katie Robbert 36:54
I saw a video as this was pulling up, I saw a video the other day in. It had AI bot bots serving dishes to people. So it was like instead of humans walking around, it was humanoid AI bots walking around serving and they’re still jerky enough that it was like really disturbing.

John Wall 37:18
Spilling drinks as they Yeah.

Christopher Penn 37:21
Exactly. This is now on like a six different websites. So it has gone through and essentially gone and done a bunch of reading online. And it’s come up with a 15 different things that we could try for accurate forecasting, hands on training, strategic staff placement, closing procedures, food prep, or improving the ordering process. So all those things are valid suggestions. And there, you actually can see what this is relatively new. There are now citations of from where he got the information. So kind of what we were talking about earlier, it’s not just your data you put in here is now able to go out on the web and find additional stuff to inform itself.

Katie Robbert 38:07
I’ve, you know, one of the questions I’ve always wrestled with with with systems like ChatGPT, when you just ask it a question and it answers is, how is this different or any better than just doing a regular internet search? Because you’re just asking it a question. But what this specific example demonstrates is it’s actually taking that next step is it’s asking the internet, it’s doing the internet search, but then it’s compiling and consolidating, which is the step that you as a human would have to take. And that to me is a very concrete, oh, okay, this is why it’s supplemental slash better than just doing an internet search. Because I could just, you know, boot up in a search engine and say, you know, how do I fix, you know, wait times at my restaurant, and I’m gonna get 1000 million different articles with its own listicle of steps. But this is actually said, Okay, I’m gonna read the 1000 million articles, and I’m going to consolidate and summarize all of those steps into one thing, that’s a step that you the human don’t have to do. And so, like, it takes me a little longer than you Chris, it’s now a little bit more concrete in my brain.

Christopher Penn 39:19
That’s the difference between search and synthesis, right. So search, you still have to do all the synthesizing of the information, whereas the language model is now taking the squares you see this when you use if you use, for example, Microsoft Bing, you will actually watch it write the search queries for Bing, get the information back and then sort of reassemble it back into coherent prose summarizing the major points. So this is absolutely one of those things. So this is now an example where you have not only the custom GPT that has your knowledge and but it can it has access to the broader knowledge set however, because it is your knowledge and is your customer instructions. You can also say hey, I need steps that work on These conditions, right? So you’d have I need to have a low budget, I have 10 tables total in my restaurant, I only have one cook on the back of house. So how do I deal with this problem with those constraints? And again, that’s where synthesis makes a big difference versus search, because search will give you 1000 articles. But they all apply to you know, Michelin starred restaurants like No, no, I’m a cafe in Canton, I can’t I’m not a Michelin starred restaurant.

Katie Robbert 40:25
Right? It’s whoever has the best. SEO. Exactly.

Christopher Penn 40:29
And so this now, so this custom GPT, what we’ve done to this point is, we’ve gone out and gotten our data. In this case, we use synthetic data, we’ve split it into happy and unhappy. We’ve created two GPT-4, one for each of those personas with constructions, we built a sparse prime representation from both to give it a specific set of focuses like this, these are things that are the most important for you to focus on. And then we’ve deployed these, and people can have interactions with them, and talk to the happy customer talk to the unhappy customer. This applies to pretty much any business or any business where you have customer data, you can use this exact same process to create a custom GPT. And it doesn’t just have to be customer feedback. So one of the things that we strongly recommend for people who are in say, in the sales someone like John, take 10, LinkedIn profiles, get the PDF versions of them, put that in your customer GPT. And now you have an ideal customer personas are the top 10 people you want to sell to. And you can say, hey, here’s my sales pitch, or here’s my sales email. What do you think I will do? How’s this going to resonate with you? And it will say, well, as a person who is an enterprise executive, I don’t understand the words you’re saying.

Katie Robbert 41:45
Well, and that’s a really good segue into what you’re covering in next week’s live stream. So next week, you and John will be covering using custom GBT models specifically for sales. Exactly. Sounds like we are,

John Wall 41:59
we will be defying gravity.

Christopher Penn 42:06
So that’s, that’s how you would use that’s how to use these things for restaurants specifically. The advantage of doing it this way, as opposed to just using generic ChatGPT is it’s focused on your feedback, your views your customers. And because it’s a custom GPT, you can have interactive conversations with it about your specific situations, and then get feedback that is tailored for you. And you can reuse it over and over again, you can you don’t have to give it all that priming instruction, the next time it’s it is pre prepared. So every month if you’re looking at your specials for the next month, like hey, I want to make a drink that is a tilapia Gin Fizz. You can have it. You know, that’s a risky idea, too. It’s a good idea. And that’s not a great idea, either.

Katie Robbert 42:59
You know, it’s all kidding aside. You know, I feel like it’s also a really great example, when people are asking will AI take my job, this is a really great example of how it’s really just supportive and complimentary to your job. Because the amount of time it would take you to call through all of those reviews, and couple of some kind of analysis versus you focusing on the customer experience, the things that are hot, more high value. This, again, is one of those really good concrete examples of no AI won’t take your job, AI is going to supplement and enhance the experience that you were able to provide.

Christopher Penn 43:35
Exactly, because you’re sitting on that customer data right now. And you’re not using it at all right? It’s just going unused is taking up space. Now, you can have the tools make use of it, and you can garner some actual insights and get real recommendations. You know, again, with these tools, talk to them, like their their, their people say, Well, what would you recommend I do to make my service faster? It will tell you, yep.

Katie Robbert 44:02
John, final thoughts.

John Wall 44:03
The restaurant industry is all about time. They don’t have enough time to do anything. So yeah, going through a whole pile of reviews and just be able to say, this is the one thing you need to work on this month. That is huge to them, because they don’t have time to mess around their hair’s on fire. 24/7

Christopher Penn 44:21
I agree. All right. So that’s gonna do it for this week. Folks, we will see you all next time. Take care. Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it. For more resources. And to learn more, check out the Trust Insights podcast at trust insights.ai/t AI podcast, and a weekly email newsletter at trust insights.ai/newsletter Got questions about what you saw in today’s episode. Join our free analytics for markers slack group at trust insights.ai/analytics for marketers, see you next time.

Transcribed by https://otter.ai

Need help with your marketing data and analytics?