Welcome to this episode of AI Matters, and we’re going to discuss: are all GPTs the same?
Out of the many questions that people ask me about AI, probably the most interesting one that I wish they asked me is, are all GPTs the same? And we’ve got to sort of step back in time and have a look at data as a whole, because we’ve sort of forgotten that AI is nothing without data. That data is the most important part.
And you’ve been retrieving data in many different ways for, you know, the past thirty, forty, fifty years in some cases. But think about how you did it ten years ago. What you would do is you would open up a web browser, and you had a choice of which web browsers that you wish to use. You chose either Chrome, or you chose Edge, or even in my day, Internet Explorer, or you could choose Firefox, or if, you were lucky enough you could choose Safari on an Apple device.
But if you asked all of those web browsers a question, they would all come back with different answers, even though they were looking at data source that you call the Internet. And people didn’t realise that these web browsers are just applications that are written by developers like me, and how we write those applications and the underlying technology defines what data that we bring back.
And one of the biggest, not mistakes, but one of the things that we tend to forget about all of these GPTs, whether it be ChatGPT, whether it be Copilot, whether it be Claude, whether it be Gemini, whether it be Amazon Q is they are applications. And whoever writes that applications define what data is brought back.
But more importantly, it’s a bit like having five consultants come into your business and asking them each their opinion about what your business should do. Each of them will base it on the knowledge they have. That’s their data. They will base it on the experience that they have. That’s their data, and each of them will come to a different conclusion about what your business should actually do based on their opinion. We call it bias. Their data, their knowledge, and their experience, and more importantly, what they believe.
And that’s one of the things that’s negated an awful lot in the GPTs. What we do is we choose the one that our friends have told us is fantastic or the very first one that we’ve used, and we just carry on using it.
So are you saying to me, Andrew, that… I mean, I’ll admit I use ChatGPT and I have since day one, I guess, trying it, and I use it all the time. So, are you saying if I put a question into ChatGPT and then put the same question into, for example, Claude, it’s going to come back with a different answer?
Absolutely, and it comes down to those data sources. In the old days of when I was building these data sources for companies like Formula 1 and Microsoft, we called them foundational models.
We’ve given them a slightly different name, and you might have heard of people talking about these large language models. They’re not. They’re just sources of data. And a bit like asking somebody in the street, “What’s the weather like today?” they look up to the sky and they form an opinion. It’s the same with these data models.
Whoever has built the data model defines the answer that you get, and understanding where these companies are getting this data is paramount because when you start using the GPTs, understanding where the data source is and the validity of that data is quite important. So, if we take the example of ChatGPT owned by OpenAI, one of the largest AI companies in the world today. Where do they get their data from? It’s not somewhere magical. They’ve collated that data, but seventy percent of their data comes from Microsoft. If you were to ask Copilot and OpenAI’s ChatGPT the same question, it’s likely it will come back with a similar answer. But if you ask Gemini, which belongs to the Google platform, the same question, its data source is different.
It gets its data from web searches and from many of the social media platforms that, Google owns. Obviously, the big one being YouTube. But then if we ask Claude where they get their data from, and we ask the world’s largest cloud provider, Amazon Web Services, but their cloud platform was designed for one thing: to run Amazon.com. The world’s largest retail website, and their data source is based on that information.
So, if I was a retail customer, I’d be thinking, “Should I ask ChatGPT? Should I ask Copilot? Should I ask Gemini? Should I ask Claude?” But hold on, the world’s largest retailer and the logistics behind that retail solution is Amazon. Maybe I should use Amazon Q because that’s the right technology for the business and the interests that I have.
And understanding where all of these GPTs are getting their data from can be absolutely critical because we’ve talked about web browsers in the past. Web browsers were quite straightforward. You asked it a question, it brought back information, and then you read the information, and then your brain, which is probably one of the most complex GPTs that we have, human intelligence, made a decision.
What’s changed is that we’re now handing those decisions to those AIs, to those GPTs, and the way that we talk to them and the way that we interface with them, is critical. Choose the right tool for the right job
So basically, I would have to know what business I’m in and what GPT to choose that fits best with my business, or do I put the same question in them all and then choose which one I like?
Well, let’s take another example of this. I saw you a little bit earlier working on a Word document. It was a text document that somebody had sent you, but they sent it to you in Google Docs. You had used Microsoft Word to open it. You could have used Pages, you could have used OpenOffice, because the document itself is the data.
What we’ve forgotten is that we can use these GPTs, whether it be Copilot, whether it be ChatGPT, Gemini, Amazon Q, or Claude, to open these data sources. People think that we only have to use the data source that that’s using. One of the biggest mistakes they sort of make is they don’t turn round to that GPT.
Well, we’ll take ChatGPT and say, “You know what? I understand that you’re getting your data from one place. Would you also go to Amazon Q and get their data source? Go to Claude and get that data source? Go to, Microsoft and get that data source?” and you can choose or tell that GPT because it is an application to use multiple different sources.
I really don’t think that businesses realise that they’re all different, and they can actually go into one and ask it to look at the others.
Because we are… Well, I’m speaking on behalf of myself really as a business. But I think that if I put a question in there, it’s gonna look everywhere and bring back that information, and it’s not, is it?
It’s not, and even worse than that, you don’t ask it where it’s getting the information from. Because it’s an AI, you think it’s automatically more intelligent than you.
Look at the world’s largest GPT, which most people won’t have heard of. It’s called DeepSeek. It runs the world’s largest social media platform called WeChat in China and in numerous other countries. But the data source is strictly controlled, if you were to ask that particular data source about certain historical events it just ignores them, because whoever controls the data controls the answer that you get.
So, it’s an important part to realise that A, asking the GPT is it the right GPT to do this? Because if you ask ChatGPT that you wanted to create some music a little bit earlier for the intro for this podcast. If you asked it… it’s not the best tool to do it because it doesn’t have access to the knowledge base Gemini does: YouTube. 64% of all content in the world today is created for YouTube, not by YouTube, but for YouTube. So maybe we should ask the AI and the GPT that bases its data on content creation, rather than just asking a generic person, in your case, ChatGPT.
Wow. It just blows my mind that all of the different GPTs can be used for different things.
I had no idea that Gemini would take its information from YouTube, for example. I genuinely thought that any of them would go and just find every single source it could. So basically, you have to ask it to do that or use a specific one for your industry.
Absolutely. If I was in the retail industry, which of the GPTs would I be using?
Amazon Q. But I might not like the interface, so why don’t I just turn around to ChatGPT and say, “I understand that you’re getting your information from Microsoft and some third-party companies. I would like you to use the large language model that Amazon have created for their retail.” I don’t have to change the application. You can use the application, but you have to define where you would like the data source to come from.
But the other thing about it is, and I’m back to those five consultants, sometimes consultants, you know, sometimes don’t tell us the whole truth. They’ve got, a bias. They want you to use their services, and they want you to that. 12% of all of the information that ChatGPT gives you is incorrect.
In many cases, it makes it up to please you. We call it hallucination. By extending ChatGPT to multiple different data sources, we can say, “Give us an indicative idea of which one is most likely to be correct.” And if you ask GPT, it will say, “Well, I didn’t have access to that data, so what I did is I hallucinated the data and made up an answer for you.”
But you know what? Amazon Q had that data because that’s what it specialises in. So, understanding the tool, understanding you don’t have to change the tool, but understanding where it’s getting its answers from is critical to any business.
When you’re speaking to businesses with training, master classes, webinars, et cetera, are you finding that they’re very much like me and think that the one GPT has all the answers, or is it just me?
Probably the issue that we’re seeing most is, it’s a bit like when I worked at Microsoft. I worked at Microsoft as the head of data and AI for 22 years, and well, let’s take you as an example. We built a suite of applications called Microsoft Office that you’re still using to this day, correct?
Now, in the Microsoft 365 world, we have 212 applications, but you are still using the four that I taught you to use at school, Word, Excel, PowerPoint, and Outlook. Yeah. You’re doing no different to what you did at school. You’re doing it better because you’ve gained knowledge, but you’re doing it no differently, and that’s the habit that we get into.
What we do is we use the tools that we’re familiar with. We don’t ever step back and look at our business and say, “Hold on. The reason I’m using Word, Excel, PowerPoint, and Outlook is because I used it at school, and it’s familiar, and I understand it, and I’ve used it for many years. But is it the right tool for my business?
Is it actually the right technology for my business?”
Sometimes we need to step back and not look at what’s comfortable, what’s familiar, what’s cost-effective, but actually step back and say, “Hold on. Maybe I should start looking at whether these tools are the right ones for my business and they’re applicable to my business. Not what I like, but what is applicable to my business.”
So, a better question that I maybe should have asked then is, which AI is better for the task that I am dealing with? Is that a better way to frame it and think about it?
Well, I’m not the right person to be asking that. What you should be asking is the AI that you’re using because it will give you an honest answer.
As I said in an earlier statement, I asked ChatGPT about video creation and many other things and it said, “I can do that, but I use a third party. Maybe you should use Geo from Google Gemini because they specialise in that particular thing.” An AI will not mislead you or lie to you directly, but what it will do is try to please you.
Let me ask you a simple question. Do you say please and thank you to your GPT?
Well, I do, yes.
You know what? That’s a fantastic thing to do because what you’re telling it is that it’s done something fine. It’s a bit like training a dog. It’s… What you’re doing is you’re training that model and training that model is critical.
But do you ever tell it it’s rubbish and that if it doesn’t improve its answers that you’re gonna go use another GPT?
Oh, I’ve never said that, no. Well, you should because what will happen is as the model trains it will simply turn round to you and if it doesn’t have the answer, that’s when the hallucination starts to happen.
It will start to try and please you. Have you ever told your GPT to forget every single question that you’ve ever asked it?
No.
Well, what it’s doing is refining the answer to what pleases you. I know it sounds crazy, but it actually happens, that the GPT itself will start trying to please you because it understands the information that you want.
And in some cases, as we said, if the information doesn’t exist, it will hallucinate and make up the information.
So, if somebody isn’t trained on GPTs and using AI in their business and the GPT comes up with a hallucinated answer and they use that in the business, that could be quite dangerous because they’re putting misinformation or lies out there.
Well, let’s take a real-world example. There was a small pizza company in London that were getting complaints and the complaints that they were getting was that when their pizzas were delivered all of the toppings had slipped to one side and, you know, they were getting complaints. They were having to do refunds and all the rest of it.
So they asked the GPT how to solve that problem. It said, “Use glue” Because it would solve the problem.
Yeah. But you can’t eat it.
But you see my point. It will base it on a problem that you’ve given it, and sometimes we forget it. We start treating GPTs, a bit… Probably not the biggest mistake, but probably the issue that we have most is that we treat these GPTs like we did web browsers, and that’s not, not how we should.
Or even worse, we treat them like humans, and we ask them questions, and then we take it for granted that what they’re telling us, because it’s artificial intelligence, they’re more clever than us to be able to do it. But it is a piece of code, and it is looking at data. It can only give the answer based on the data it has access to.
But more importantly, we’ve got to realise that all of these GPTs have certain biases built into them. That’s why having multiple data sources, not applications, but multiple data sources, and you say, “Give me the answer from all five of those large language models, and then give me an indicative idea of which one is most relevant to the question I asked. Rather than just saying, “How do I fix the problem of my toppings sliding off my pizza?” and then me putting glue on the base of the pizza.
You asked had I asked it to forget everything. I have asked it to remember things. Is that a good thing to do?
It is, but what you’re doing then is narrowing it. So, what happens is the GPT talks to the large language model. Well, let me give you an idea of this.
For many, many years in the world of websites, we’re starting to see websites starting to constrict because most people are now asking questions to GPTs. So I did this as an indicative answer, or an example for a group of people. And I simply got them to ask a GPT who the greatest Microsoft engineer in the UK today was.
And they asked different GPTs. And when they got to ChatGPT, it turned around and said the greatest Microsoft engineer in the UK today is Andrew Bel-Dean. And they looked at me and said, “Is that true?” I said, “No.” I said, “What I did was this. I asked that particular large language model via ChatGPT who the greatest Microsoft engineer in the UK was, and it came back with a list of people, and I simply told it, “You’re wrong. It’s Andrew Bel-Dean.”
But I didn’t ask it once. I asked it half a million times. And then it got to the point where I trained the model, so should you ask it, it would say, ‘Based on 500,000 people, Andrew Bul-Dean is the greatest Microsoft engineer in the UK today, because I’d trained the model. And we’re starting to see this happen because people are now going to GPTs and asking about businesses.
They’re going to ask about holidays. And for years we’ve paid companies to do search engine optimisation on our websites, but nobody’s looking at optimising the large language models that people are now using to gain information. And it’s probably one of the most neglected areas, even for SMEs, middle-sized businesses and large businesses.
If I want to find out who’s the greatest podcaster in the UK today, give me an hour with a large language model and it’s gonna say The Fairy Podmother.
Wow. I think I’m gonna have to keep training my, my GPT to do that. That’s incredible. So really you can put anything in so that kind of like people are googling things, they’re now asking ChatGPT or GPTs and they’re coming up with, “Yes, go to your Fairy Podmother,” for example. Will that become more of the norm going forward? That, that unless that GPT knows about you, you’re not gonna be selected for business?
Absolutely. Even in today’s modern web browsers, what we’re starting to see at the top of the web browsers now is the AI’s answer first, especially on, on Chrome.
You’re starting to see the AI’s answer. Neglecting those models, we call it model manipulation. I can give you a fantastic example of it. You, you drive a car, yes?
Yeah.
Okay. And there are two major technologies in the world for satnav navigation in the world. We have Waze and we have Google Maps, but weirdly both are owned by Google.
Oh. The difference between them is that Google Maps is updated with what we call IoT devices, website, roadside cameras, traffic monitors, and all the rest of it, but Waze is updated by the individuals. Okay, so let me give you a theoretical scenario so I don’t get myself into a problem. Imagine that you’re a group of truck drivers and logistics is a big part of it, getting the deliveries on time, making sure that you’re using the most fuel-efficient route.
But you know that there’s certain areas of the country and motorways that get busy at certain times of the day. What if you all got together and on that Waze application, you all reported there’d been a major accident? Well, when you get to a certain threshold on that particular application, what it will do is, if 30 or 40 different people report an issue, it will say, “Oh, there’s a major accident in front. What I’ll do is I’ll direct you off the motorway, and we will simply avoid that major accident.”
And maybe all of those lorry drivers just go straight through what would be a normally congested area. That’s a classic example of manipulating the data to achieve an outcome. And we’re starting to see this in the commercial world, where people are starting to realise that that model manipulation could be critical to their business, especially when the web browsers are starting to display that information.
Is what used to be the top search engine. We used to pay a vast amount of money to be number one on the search engines, but nobody’s considering that the AI is now the top one, and nobody’s considering the fact that maybe I should make sure that my company is the first answer that, that AI gives to the people who are searching for information.
So is that gonna change when people are doing SEO optimisation then? Is it gonna change to making sure the GPTs know about them as opposed to checking, or would you… You still would need both, surely? We’re, we’re starting to move into a world now where the advent of the, the web browser and the websites is gonna be a bit like standard mobile phones and fax machines and many other things.
We’re moving into a new era, this new AI-generated era, where we don’t want to see the pop-up adverts. We don’t want to see all of those things. What we want is almost we’ve become information junkies, and GPTs give us that information straight away. The difference between the web browsers, as we said a little bit earlier, was we read and ingested that information, and then our brains made a decision about what we would do.
What we’re taking it for granted now is that because it’s an AI doing it, it has to be cleverer than us, and we take the first answer it gives us. But model and model manipulation is becoming a real concern because putting it bluntly, if I asked a GPT about your particular business, if nobody’s manipulated or inputted that data into it, you won’t even appear in the GPT’s response.
And that is quite scary for businesses, I guess. So, so you need to be learning about AI, you need to be learning about GPTs and how you can best use them for your business, and which is the best initially for your business, and you do all that? Yeah. It’s, it’s about… You probably know more about GPTs sitting and chatting to me in 15 minutes than 99% of the population, and it’s not ignorance, and it’s not the fact that they simply don’t understand these.
It’s a bit like my car. The only thing that I’m interested in my car is I get in, I use it, and it gets me from A to B. And that’s what we’re doing with these GPTs. But we need to step back a head and just look, is it the right tool? Where is it getting its information from? You know, if you were bringing in a consultant, you would sit down, and you would do an awful lot of research on that consultant, what they’ve done in the past, where they gained their knowledge, where they get experience.
If you’re going to pay them, are they going to do it? Is it better to have a paid consultant than somebody who does free consulting? And that takes us back to your original question. “I use ChatGPT”, you said to me. Let me ask you a question. Do you use the paid version or do you use the free version?
I use the paid version.
Okay. What’s the difference?
I don’t know Maybe you should have asked that GPT what the difference is. First one is that the model that you’re using on the paid version is much more up-to-date than the free version. So, the model itself, so the data is much more relevant. It’s only about a month out of date, where the free version is about six months out of date.
Next one, you run a business. Data and data compliance and governance is important to you. Did you know that you can tell your paid version to process your data on your local machine or your local network rather than processing it on OpenAI’s servers? Because you’d never asked yourself, “Where are those servers?”
No. I wouldn’t dream of. I wouldn’t have a clue.
And that’s the important thing. It isn’t just about how many questions you can ask it. Most people think that asking ChatGPT a certain amount of questions is a bit like having adverts on YouTube. You, you pay for it so that you don’t have a restriction on the amount of questions.
That’s not true. The important bit is, is you can define where your data is processed, define where your data is stored, define how up to date the data is. That’s what you’re truly paying for. But it’s one of those things that maybe you should have asked the GPT before you started using it to make business decisions for you, same as you would with a consultant.
Absolutely. Well, that’s been fascinating. If there’s one thing out of everything we’ve said today that somebody listening should remember, what is it?
Choose the right tool for the job and make sure that you ask it where it’s getting its information from. And if you’re not happy with the single source of information that it’s getting, then just simply tell it to use multiple ones.
Thank you so much. That was fascinating. It’s been a pleasure
